1 /* SPDX-License-Identifier: GPL-2.0+ */
3 * RCU expedited grace periods
5 * Copyright IBM Corporation, 2016
7 * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
10 #include <linux/lockdep.h>
12 static void rcu_exp_handler(void *unused);
13 static int rcu_print_task_exp_stall(struct rcu_node *rnp);
16 * Record the start of an expedited grace period.
18 static void rcu_exp_gp_seq_start(void)
20 rcu_seq_start(&rcu_state.expedited_sequence);
24 * Return then value that expedited-grace-period counter will have
25 * at the end of the current grace period.
27 static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
29 return rcu_seq_endval(&rcu_state.expedited_sequence);
33 * Record the end of an expedited grace period.
35 static void rcu_exp_gp_seq_end(void)
37 rcu_seq_end(&rcu_state.expedited_sequence);
38 smp_mb(); /* Ensure that consecutive grace periods serialize. */
42 * Take a snapshot of the expedited-grace-period counter.
44 static unsigned long rcu_exp_gp_seq_snap(void)
48 smp_mb(); /* Caller's modifications seen first by other CPUs. */
49 s = rcu_seq_snap(&rcu_state.expedited_sequence);
50 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
55 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
56 * if a full expedited grace period has elapsed since that snapshot
59 static bool rcu_exp_gp_seq_done(unsigned long s)
61 return rcu_seq_done(&rcu_state.expedited_sequence, s);
65 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
66 * recent CPU-online activity. Note that these masks are not cleared
67 * when CPUs go offline, so they reflect the union of all CPUs that have
68 * ever been online. This means that this function normally takes its
69 * no-work-to-do fastpath.
71 static void sync_exp_reset_tree_hotplug(void)
76 unsigned long oldmask;
77 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
79 struct rcu_node *rnp_up;
81 /* If no new CPUs onlined since last time, nothing to do. */
82 if (likely(ncpus == rcu_state.ncpus_snap))
84 rcu_state.ncpus_snap = ncpus;
87 * Each pass through the following loop propagates newly onlined
88 * CPUs for the current rcu_node structure up the rcu_node tree.
90 rcu_for_each_leaf_node(rnp) {
91 raw_spin_lock_irqsave_rcu_node(rnp, flags);
92 if (rnp->expmaskinit == rnp->expmaskinitnext) {
93 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
94 continue; /* No new CPUs, nothing to do. */
97 /* Update this node's mask, track old value for propagation. */
98 oldmask = rnp->expmaskinit;
99 rnp->expmaskinit = rnp->expmaskinitnext;
100 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
102 /* If was already nonzero, nothing to propagate. */
106 /* Propagate the new CPU up the tree. */
108 rnp_up = rnp->parent;
111 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
112 if (rnp_up->expmaskinit)
114 rnp_up->expmaskinit |= mask;
115 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
118 mask = rnp_up->grpmask;
119 rnp_up = rnp_up->parent;
125 * Reset the ->expmask values in the rcu_node tree in preparation for
126 * a new expedited grace period.
128 static void __maybe_unused sync_exp_reset_tree(void)
131 struct rcu_node *rnp;
133 sync_exp_reset_tree_hotplug();
134 rcu_for_each_node_breadth_first(rnp) {
135 raw_spin_lock_irqsave_rcu_node(rnp, flags);
136 WARN_ON_ONCE(rnp->expmask);
137 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
138 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
143 * Return non-zero if there is no RCU expedited grace period in progress
144 * for the specified rcu_node structure, in other words, if all CPUs and
145 * tasks covered by the specified rcu_node structure have done their bit
146 * for the current expedited grace period. Works only for preemptible
147 * RCU -- other RCU implementation use other means.
149 * Caller must hold the specificed rcu_node structure's ->lock
151 static bool sync_rcu_preempt_exp_done(struct rcu_node *rnp)
153 raw_lockdep_assert_held_rcu_node(rnp);
155 return rnp->exp_tasks == NULL &&
156 READ_ONCE(rnp->expmask) == 0;
160 * Like sync_rcu_preempt_exp_done(), but this function assumes the caller
161 * doesn't hold the rcu_node's ->lock, and will acquire and release the lock
164 static bool sync_rcu_preempt_exp_done_unlocked(struct rcu_node *rnp)
169 raw_spin_lock_irqsave_rcu_node(rnp, flags);
170 ret = sync_rcu_preempt_exp_done(rnp);
171 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
178 * Report the exit from RCU read-side critical section for the last task
179 * that queued itself during or before the current expedited preemptible-RCU
180 * grace period. This event is reported either to the rcu_node structure on
181 * which the task was queued or to one of that rcu_node structure's ancestors,
182 * recursively up the tree. (Calm down, calm down, we do the recursion
185 * Caller must hold the specified rcu_node structure's ->lock.
187 static void __rcu_report_exp_rnp(struct rcu_node *rnp,
188 bool wake, unsigned long flags)
189 __releases(rnp->lock)
194 if (!sync_rcu_preempt_exp_done(rnp)) {
196 rcu_initiate_boost(rnp, flags);
198 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
201 if (rnp->parent == NULL) {
202 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
204 smp_mb(); /* EGP done before wake_up(). */
205 swake_up_one(&rcu_state.expedited_wq);
210 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
212 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
213 WARN_ON_ONCE(!(rnp->expmask & mask));
214 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
219 * Report expedited quiescent state for specified node. This is a
220 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
222 static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
226 raw_spin_lock_irqsave_rcu_node(rnp, flags);
227 __rcu_report_exp_rnp(rnp, wake, flags);
231 * Report expedited quiescent state for multiple CPUs, all covered by the
232 * specified leaf rcu_node structure.
234 static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
235 unsigned long mask, bool wake)
239 raw_spin_lock_irqsave_rcu_node(rnp, flags);
240 if (!(rnp->expmask & mask)) {
241 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
244 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
245 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
249 * Report expedited quiescent state for specified rcu_data (CPU).
251 static void rcu_report_exp_rdp(struct rcu_data *rdp)
253 WRITE_ONCE(rdp->exp_deferred_qs, false);
254 rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
257 /* Common code for work-done checking. */
258 static bool sync_exp_work_done(unsigned long s)
260 if (rcu_exp_gp_seq_done(s)) {
261 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
262 smp_mb(); /* Ensure test happens before caller kfree(). */
269 * Funnel-lock acquisition for expedited grace periods. Returns true
270 * if some other task completed an expedited grace period that this task
271 * can piggy-back on, and with no mutex held. Otherwise, returns false
272 * with the mutex held, indicating that the caller must actually do the
273 * expedited grace period.
275 static bool exp_funnel_lock(unsigned long s)
277 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
278 struct rcu_node *rnp = rdp->mynode;
279 struct rcu_node *rnp_root = rcu_get_root();
281 /* Low-contention fastpath. */
282 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
284 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
285 mutex_trylock(&rcu_state.exp_mutex))
289 * Each pass through the following loop works its way up
290 * the rcu_node tree, returning if others have done the work or
291 * otherwise falls through to acquire ->exp_mutex. The mapping
292 * from CPU to rcu_node structure can be inexact, as it is just
293 * promoting locality and is not strictly needed for correctness.
295 for (; rnp != NULL; rnp = rnp->parent) {
296 if (sync_exp_work_done(s))
299 /* Work not done, either wait here or go up. */
300 spin_lock(&rnp->exp_lock);
301 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
303 /* Someone else doing GP, so wait for them. */
304 spin_unlock(&rnp->exp_lock);
305 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
306 rnp->grplo, rnp->grphi,
308 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
309 sync_exp_work_done(s));
312 rnp->exp_seq_rq = s; /* Followers can wait on us. */
313 spin_unlock(&rnp->exp_lock);
314 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
315 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
317 mutex_lock(&rcu_state.exp_mutex);
319 if (sync_exp_work_done(s)) {
320 mutex_unlock(&rcu_state.exp_mutex);
323 rcu_exp_gp_seq_start();
324 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
329 * Select the CPUs within the specified rcu_node that the upcoming
330 * expedited grace period needs to wait for.
332 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
336 unsigned long mask_ofl_test;
337 unsigned long mask_ofl_ipi;
339 struct rcu_exp_work *rewp =
340 container_of(wp, struct rcu_exp_work, rew_work);
341 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
343 raw_spin_lock_irqsave_rcu_node(rnp, flags);
345 /* Each pass checks a CPU for identity, offline, and idle. */
347 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
348 unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
349 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
352 if (raw_smp_processor_id() == cpu ||
353 !(rnp->qsmaskinitnext & mask)) {
354 mask_ofl_test |= mask;
356 snap = rcu_dynticks_snap(rdp);
357 if (rcu_dynticks_in_eqs(snap))
358 mask_ofl_test |= mask;
360 rdp->exp_dynticks_snap = snap;
363 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
366 * Need to wait for any blocked tasks as well. Note that
367 * additional blocking tasks will also block the expedited GP
368 * until such time as the ->expmask bits are cleared.
370 if (rcu_preempt_has_tasks(rnp))
371 rnp->exp_tasks = rnp->blkd_tasks.next;
372 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
374 /* IPI the remaining CPUs for expedited quiescent state. */
375 for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
376 unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
377 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
380 if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
381 mask_ofl_test |= mask;
384 if (get_cpu() == cpu) {
385 mask_ofl_test |= mask;
389 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
392 mask_ofl_ipi &= ~mask;
395 /* Failed, raced with CPU hotplug operation. */
396 raw_spin_lock_irqsave_rcu_node(rnp, flags);
397 if ((rnp->qsmaskinitnext & mask) &&
398 (rnp->expmask & mask)) {
399 /* Online, so delay for a bit and try again. */
400 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
401 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
402 schedule_timeout_uninterruptible(1);
405 /* CPU really is offline, so we can ignore it. */
406 if (!(rnp->expmask & mask))
407 mask_ofl_ipi &= ~mask;
408 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
410 /* Report quiescent states for those that went offline. */
411 mask_ofl_test |= mask_ofl_ipi;
413 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
417 * Select the nodes that the upcoming expedited grace period needs
420 static void sync_rcu_exp_select_cpus(void)
423 struct rcu_node *rnp;
425 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
426 sync_exp_reset_tree();
427 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
429 /* Schedule work for each leaf rcu_node structure. */
430 rcu_for_each_leaf_node(rnp) {
431 rnp->exp_need_flush = false;
432 if (!READ_ONCE(rnp->expmask))
433 continue; /* Avoid early boot non-existent wq. */
434 if (!READ_ONCE(rcu_par_gp_wq) ||
435 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
436 rcu_is_last_leaf_node(rnp)) {
437 /* No workqueues yet or last leaf, do direct call. */
438 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
441 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
442 cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
443 /* If all offline, queue the work on an unbound CPU. */
444 if (unlikely(cpu > rnp->grphi - rnp->grplo))
445 cpu = WORK_CPU_UNBOUND;
448 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
449 rnp->exp_need_flush = true;
452 /* Wait for workqueue jobs (if any) to complete. */
453 rcu_for_each_leaf_node(rnp)
454 if (rnp->exp_need_flush)
455 flush_work(&rnp->rew.rew_work);
458 static void synchronize_sched_expedited_wait(void)
461 unsigned long jiffies_stall;
462 unsigned long jiffies_start;
465 struct rcu_node *rnp;
466 struct rcu_node *rnp_root = rcu_get_root();
469 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
470 jiffies_stall = rcu_jiffies_till_stall_check();
471 jiffies_start = jiffies;
474 ret = swait_event_timeout_exclusive(
475 rcu_state.expedited_wq,
476 sync_rcu_preempt_exp_done_unlocked(rnp_root),
478 if (ret > 0 || sync_rcu_preempt_exp_done_unlocked(rnp_root))
480 WARN_ON(ret < 0); /* workqueues should not be signaled. */
481 if (rcu_cpu_stall_suppress)
483 panic_on_rcu_stall();
484 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
487 rcu_for_each_leaf_node(rnp) {
488 ndetected += rcu_print_task_exp_stall(rnp);
489 for_each_leaf_node_possible_cpu(rnp, cpu) {
490 struct rcu_data *rdp;
492 mask = leaf_node_cpu_bit(rnp, cpu);
493 if (!(READ_ONCE(rnp->expmask) & mask))
496 rdp = per_cpu_ptr(&rcu_data, cpu);
497 pr_cont(" %d-%c%c%c", cpu,
498 "O."[!!cpu_online(cpu)],
499 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
500 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
503 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
504 jiffies - jiffies_start, rcu_state.expedited_sequence,
505 READ_ONCE(rnp_root->expmask),
506 ".T"[!!rnp_root->exp_tasks]);
508 pr_err("blocking rcu_node structures:");
509 rcu_for_each_node_breadth_first(rnp) {
511 continue; /* printed unconditionally */
512 if (sync_rcu_preempt_exp_done_unlocked(rnp))
514 pr_cont(" l=%u:%d-%d:%#lx/%c",
515 rnp->level, rnp->grplo, rnp->grphi,
516 READ_ONCE(rnp->expmask),
517 ".T"[!!rnp->exp_tasks]);
521 rcu_for_each_leaf_node(rnp) {
522 for_each_leaf_node_possible_cpu(rnp, cpu) {
523 mask = leaf_node_cpu_bit(rnp, cpu);
524 if (!(READ_ONCE(rnp->expmask) & mask))
529 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
534 * Wait for the current expedited grace period to complete, and then
535 * wake up everyone who piggybacked on the just-completed expedited
536 * grace period. Also update all the ->exp_seq_rq counters as needed
537 * in order to avoid counter-wrap problems.
539 static void rcu_exp_wait_wake(unsigned long s)
541 struct rcu_node *rnp;
543 synchronize_sched_expedited_wait();
545 // Switch over to wakeup mode, allowing the next GP to proceed.
546 // End the previous grace period only after acquiring the mutex
547 // to ensure that only one GP runs concurrently with wakeups.
548 mutex_lock(&rcu_state.exp_wake_mutex);
549 rcu_exp_gp_seq_end();
550 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
552 rcu_for_each_node_breadth_first(rnp) {
553 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
554 spin_lock(&rnp->exp_lock);
555 /* Recheck, avoid hang in case someone just arrived. */
556 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
558 spin_unlock(&rnp->exp_lock);
560 smp_mb(); /* All above changes before wakeup. */
561 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
563 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
564 mutex_unlock(&rcu_state.exp_wake_mutex);
568 * Common code to drive an expedited grace period forward, used by
569 * workqueues and mid-boot-time tasks.
571 static void rcu_exp_sel_wait_wake(unsigned long s)
573 /* Initialize the rcu_node tree in preparation for the wait. */
574 sync_rcu_exp_select_cpus();
576 /* Wait and clean up, including waking everyone. */
577 rcu_exp_wait_wake(s);
581 * Work-queue handler to drive an expedited grace period forward.
583 static void wait_rcu_exp_gp(struct work_struct *wp)
585 struct rcu_exp_work *rewp;
587 rewp = container_of(wp, struct rcu_exp_work, rew_work);
588 rcu_exp_sel_wait_wake(rewp->rew_s);
591 #ifdef CONFIG_PREEMPT_RCU
594 * Remote handler for smp_call_function_single(). If there is an
595 * RCU read-side critical section in effect, request that the
596 * next rcu_read_unlock() record the quiescent state up the
597 * ->expmask fields in the rcu_node tree. Otherwise, immediately
598 * report the quiescent state.
600 static void rcu_exp_handler(void *unused)
603 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
604 struct rcu_node *rnp = rdp->mynode;
605 struct task_struct *t = current;
608 * First, the common case of not being in an RCU read-side
609 * critical section. If also enabled or idle, immediately
610 * report the quiescent state, otherwise defer.
612 if (!t->rcu_read_lock_nesting) {
613 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
614 rcu_dynticks_curr_cpu_in_eqs()) {
615 rcu_report_exp_rdp(rdp);
617 rdp->exp_deferred_qs = true;
618 set_tsk_need_resched(t);
619 set_preempt_need_resched();
625 * Second, the less-common case of being in an RCU read-side
626 * critical section. In this case we can count on a future
627 * rcu_read_unlock(). However, this rcu_read_unlock() might
628 * execute on some other CPU, but in that case there will be
629 * a future context switch. Either way, if the expedited
630 * grace period is still waiting on this CPU, set ->deferred_qs
631 * so that the eventual quiescent state will be reported.
632 * Note that there is a large group of race conditions that
633 * can have caused this quiescent state to already have been
634 * reported, so we really do need to check ->expmask.
636 if (t->rcu_read_lock_nesting > 0) {
637 raw_spin_lock_irqsave_rcu_node(rnp, flags);
638 if (rnp->expmask & rdp->grpmask) {
639 rdp->exp_deferred_qs = true;
640 t->rcu_read_unlock_special.b.exp_hint = true;
642 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
647 * The final and least likely case is where the interrupted
648 * code was just about to or just finished exiting the RCU-preempt
649 * read-side critical section, and no, we can't tell which.
650 * So either way, set ->deferred_qs to flag later code that
651 * a quiescent state is required.
653 * If the CPU is fully enabled (or if some buggy RCU-preempt
654 * read-side critical section is being used from idle), just
655 * invoke rcu_preempt_deferred_qs() to immediately report the
656 * quiescent state. We cannot use rcu_read_unlock_special()
657 * because we are in an interrupt handler, which will cause that
658 * function to take an early exit without doing anything.
660 * Otherwise, force a context switch after the CPU enables everything.
662 rdp->exp_deferred_qs = true;
663 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
664 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs())) {
665 rcu_preempt_deferred_qs(t);
667 set_tsk_need_resched(t);
668 set_preempt_need_resched();
672 /* PREEMPT=y, so no PREEMPT=n expedited grace period to clean up after. */
673 static void sync_sched_exp_online_cleanup(int cpu)
678 * Scan the current list of tasks blocked within RCU read-side critical
679 * sections, printing out the tid of each that is blocking the current
680 * expedited grace period.
682 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
684 struct task_struct *t;
689 t = list_entry(rnp->exp_tasks->prev,
690 struct task_struct, rcu_node_entry);
691 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
692 pr_cont(" P%d", t->pid);
698 #else /* #ifdef CONFIG_PREEMPT_RCU */
700 /* Request an expedited quiescent state. */
701 static void rcu_exp_need_qs(void)
703 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
704 /* Store .exp before .rcu_urgent_qs. */
705 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
706 set_tsk_need_resched(current);
707 set_preempt_need_resched();
710 /* Invoked on each online non-idle CPU for expedited quiescent state. */
711 static void rcu_exp_handler(void *unused)
713 struct rcu_data *rdp;
714 struct rcu_node *rnp;
716 rdp = this_cpu_ptr(&rcu_data);
718 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
719 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
721 if (rcu_is_cpu_rrupt_from_idle()) {
722 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
728 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
729 static void sync_sched_exp_online_cleanup(int cpu)
733 struct rcu_data *rdp;
735 struct rcu_node *rnp;
737 rdp = per_cpu_ptr(&rcu_data, cpu);
740 /* Quiescent state either not needed or already requested, leave. */
741 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
742 rdp->cpu_no_qs.b.exp) {
746 /* Quiescent state needed on current CPU, so set it up locally. */
748 local_irq_save(flags);
750 local_irq_restore(flags);
754 /* Quiescent state needed on some other CPU, send IPI. */
755 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
761 * Because preemptible RCU does not exist, we never have to check for
762 * tasks blocked within RCU read-side critical sections that are
763 * blocking the current expedited grace period.
765 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
770 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
773 * synchronize_rcu_expedited - Brute-force RCU grace period
775 * Wait for an RCU grace period, but expedite it. The basic idea is to
776 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
777 * the CPU is in an RCU critical section, and if so, it sets a flag that
778 * causes the outermost rcu_read_unlock() to report the quiescent state
779 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
780 * other hand, if the CPU is not in an RCU read-side critical section,
781 * the IPI handler reports the quiescent state immediately.
783 * Although this is a great improvement over previous expedited
784 * implementations, it is still unfriendly to real-time workloads, so is
785 * thus not recommended for any sort of common-case code. In fact, if
786 * you are using synchronize_rcu_expedited() in a loop, please restructure
787 * your code to batch your updates, and then Use a single synchronize_rcu()
790 * This has the same semantics as (but is more brutal than) synchronize_rcu().
792 void synchronize_rcu_expedited(void)
794 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
795 struct rcu_exp_work rew;
796 struct rcu_node *rnp;
799 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
800 lock_is_held(&rcu_lock_map) ||
801 lock_is_held(&rcu_sched_lock_map),
802 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
804 /* Is the state is such that the call is a grace period? */
805 if (rcu_blocking_is_gp())
808 /* If expedited grace periods are prohibited, fall back to normal. */
809 if (rcu_gp_is_normal()) {
810 wait_rcu_gp(call_rcu);
814 /* Take a snapshot of the sequence number. */
815 s = rcu_exp_gp_seq_snap();
816 if (exp_funnel_lock(s))
817 return; /* Someone else did our work for us. */
819 /* Ensure that load happens before action based on it. */
820 if (unlikely(boottime)) {
821 /* Direct call during scheduler init and early_initcalls(). */
822 rcu_exp_sel_wait_wake(s);
824 /* Marshall arguments & schedule the expedited grace period. */
826 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
827 queue_work(rcu_gp_wq, &rew.rew_work);
830 /* Wait for expedited grace period to complete. */
831 rnp = rcu_get_root();
832 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
833 sync_exp_work_done(s));
834 smp_mb(); /* Workqueue actions happen before return. */
836 /* Let the next expedited grace period start. */
837 mutex_unlock(&rcu_state.exp_mutex);
839 if (likely(!boottime))
840 destroy_work_on_stack(&rew.rew_work);
842 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);