2 * RCU expedited grace periods
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2016
20 * Authors: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
23 #include <linux/lockdep.h>
26 * Record the start of an expedited grace period.
28 static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
30 rcu_seq_start(&rsp->expedited_sequence);
34 * Return then value that expedited-grace-period counter will have
35 * at the end of the current grace period.
37 static __maybe_unused unsigned long rcu_exp_gp_seq_endval(struct rcu_state *rsp)
39 return rcu_seq_endval(&rsp->expedited_sequence);
43 * Record the end of an expedited grace period.
45 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
47 rcu_seq_end(&rsp->expedited_sequence);
48 smp_mb(); /* Ensure that consecutive grace periods serialize. */
52 * Take a snapshot of the expedited-grace-period counter.
54 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
58 smp_mb(); /* Caller's modifications seen first by other CPUs. */
59 s = rcu_seq_snap(&rsp->expedited_sequence);
60 trace_rcu_exp_grace_period(rsp->name, s, TPS("snap"));
65 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
66 * if a full expedited grace period has elapsed since that snapshot
69 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
71 return rcu_seq_done(&rsp->expedited_sequence, s);
75 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
76 * recent CPU-online activity. Note that these masks are not cleared
77 * when CPUs go offline, so they reflect the union of all CPUs that have
78 * ever been online. This means that this function normally takes its
79 * no-work-to-do fastpath.
81 static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
86 unsigned long oldmask;
87 int ncpus = smp_load_acquire(&rsp->ncpus); /* Order against locking. */
89 struct rcu_node *rnp_up;
91 /* If no new CPUs onlined since last time, nothing to do. */
92 if (likely(ncpus == rsp->ncpus_snap))
94 rsp->ncpus_snap = ncpus;
97 * Each pass through the following loop propagates newly onlined
98 * CPUs for the current rcu_node structure up the rcu_node tree.
100 rcu_for_each_leaf_node(rsp, rnp) {
101 raw_spin_lock_irqsave_rcu_node(rnp, flags);
102 if (rnp->expmaskinit == rnp->expmaskinitnext) {
103 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
104 continue; /* No new CPUs, nothing to do. */
107 /* Update this node's mask, track old value for propagation. */
108 oldmask = rnp->expmaskinit;
109 rnp->expmaskinit = rnp->expmaskinitnext;
110 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
112 /* If was already nonzero, nothing to propagate. */
116 /* Propagate the new CPU up the tree. */
118 rnp_up = rnp->parent;
121 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
122 if (rnp_up->expmaskinit)
124 rnp_up->expmaskinit |= mask;
125 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
128 mask = rnp_up->grpmask;
129 rnp_up = rnp_up->parent;
135 * Reset the ->expmask values in the rcu_node tree in preparation for
136 * a new expedited grace period.
138 static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
141 struct rcu_node *rnp;
143 sync_exp_reset_tree_hotplug(rsp);
144 rcu_for_each_node_breadth_first(rsp, rnp) {
145 raw_spin_lock_irqsave_rcu_node(rnp, flags);
146 WARN_ON_ONCE(rnp->expmask);
147 rnp->expmask = rnp->expmaskinit;
148 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
153 * Return non-zero if there is no RCU expedited grace period in progress
154 * for the specified rcu_node structure, in other words, if all CPUs and
155 * tasks covered by the specified rcu_node structure have done their bit
156 * for the current expedited grace period. Works only for preemptible
157 * RCU -- other RCU implementation use other means.
159 * Caller must hold the specificed rcu_node structure's ->lock
161 static bool sync_rcu_preempt_exp_done(struct rcu_node *rnp)
163 raw_lockdep_assert_held_rcu_node(rnp);
165 return rnp->exp_tasks == NULL &&
166 READ_ONCE(rnp->expmask) == 0;
170 * Like sync_rcu_preempt_exp_done(), but this function assumes the caller
171 * doesn't hold the rcu_node's ->lock, and will acquire and release the lock
174 static bool sync_rcu_preempt_exp_done_unlocked(struct rcu_node *rnp)
179 raw_spin_lock_irqsave_rcu_node(rnp, flags);
180 ret = sync_rcu_preempt_exp_done(rnp);
181 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
188 * Report the exit from RCU read-side critical section for the last task
189 * that queued itself during or before the current expedited preemptible-RCU
190 * grace period. This event is reported either to the rcu_node structure on
191 * which the task was queued or to one of that rcu_node structure's ancestors,
192 * recursively up the tree. (Calm down, calm down, we do the recursion
195 * Caller must hold the specified rcu_node structure's ->lock.
197 static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
198 bool wake, unsigned long flags)
199 __releases(rnp->lock)
204 if (!sync_rcu_preempt_exp_done(rnp)) {
206 rcu_initiate_boost(rnp, flags);
208 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
211 if (rnp->parent == NULL) {
212 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
214 smp_mb(); /* EGP done before wake_up(). */
215 swake_up_one(&rsp->expedited_wq);
220 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
222 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
223 WARN_ON_ONCE(!(rnp->expmask & mask));
224 rnp->expmask &= ~mask;
229 * Report expedited quiescent state for specified node. This is a
230 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
232 static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
233 struct rcu_node *rnp, bool wake)
237 raw_spin_lock_irqsave_rcu_node(rnp, flags);
238 __rcu_report_exp_rnp(rsp, rnp, wake, flags);
242 * Report expedited quiescent state for multiple CPUs, all covered by the
243 * specified leaf rcu_node structure.
245 static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
246 unsigned long mask, bool wake)
250 raw_spin_lock_irqsave_rcu_node(rnp, flags);
251 if (!(rnp->expmask & mask)) {
252 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
255 rnp->expmask &= ~mask;
256 __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
260 * Report expedited quiescent state for specified rcu_data (CPU).
262 static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
265 rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
268 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
269 static bool sync_exp_work_done(struct rcu_state *rsp, unsigned long s)
271 if (rcu_exp_gp_seq_done(rsp, s)) {
272 trace_rcu_exp_grace_period(rsp->name, s, TPS("done"));
273 /* Ensure test happens before caller kfree(). */
274 smp_mb__before_atomic(); /* ^^^ */
281 * Funnel-lock acquisition for expedited grace periods. Returns true
282 * if some other task completed an expedited grace period that this task
283 * can piggy-back on, and with no mutex held. Otherwise, returns false
284 * with the mutex held, indicating that the caller must actually do the
285 * expedited grace period.
287 static bool exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
289 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
290 struct rcu_node *rnp = rdp->mynode;
291 struct rcu_node *rnp_root = rcu_get_root(rsp);
293 /* Low-contention fastpath. */
294 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
296 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
297 mutex_trylock(&rsp->exp_mutex))
301 * Each pass through the following loop works its way up
302 * the rcu_node tree, returning if others have done the work or
303 * otherwise falls through to acquire rsp->exp_mutex. The mapping
304 * from CPU to rcu_node structure can be inexact, as it is just
305 * promoting locality and is not strictly needed for correctness.
307 for (; rnp != NULL; rnp = rnp->parent) {
308 if (sync_exp_work_done(rsp, s))
311 /* Work not done, either wait here or go up. */
312 spin_lock(&rnp->exp_lock);
313 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
315 /* Someone else doing GP, so wait for them. */
316 spin_unlock(&rnp->exp_lock);
317 trace_rcu_exp_funnel_lock(rsp->name, rnp->level,
318 rnp->grplo, rnp->grphi,
320 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
321 sync_exp_work_done(rsp, s));
324 rnp->exp_seq_rq = s; /* Followers can wait on us. */
325 spin_unlock(&rnp->exp_lock);
326 trace_rcu_exp_funnel_lock(rsp->name, rnp->level, rnp->grplo,
327 rnp->grphi, TPS("nxtlvl"));
329 mutex_lock(&rsp->exp_mutex);
331 if (sync_exp_work_done(rsp, s)) {
332 mutex_unlock(&rsp->exp_mutex);
335 rcu_exp_gp_seq_start(rsp);
336 trace_rcu_exp_grace_period(rsp->name, s, TPS("start"));
340 /* Invoked on each online non-idle CPU for expedited quiescent state. */
341 static void sync_sched_exp_handler(void *data)
343 struct rcu_data *rdp;
344 struct rcu_node *rnp;
345 struct rcu_state *rsp = data;
347 rdp = this_cpu_ptr(rsp->rda);
349 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
350 __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
352 if (rcu_is_cpu_rrupt_from_idle()) {
353 rcu_report_exp_rdp(&rcu_sched_state,
354 this_cpu_ptr(&rcu_sched_data), true);
357 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
358 /* Store .exp before .rcu_urgent_qs. */
359 smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true);
360 resched_cpu(smp_processor_id());
363 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
364 static void sync_sched_exp_online_cleanup(int cpu)
366 struct rcu_data *rdp;
368 struct rcu_node *rnp;
369 struct rcu_state *rsp = &rcu_sched_state;
371 rdp = per_cpu_ptr(rsp->rda, cpu);
373 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
375 ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
380 * Select the CPUs within the specified rcu_node that the upcoming
381 * expedited grace period needs to wait for.
383 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
387 smp_call_func_t func;
388 unsigned long mask_ofl_test;
389 unsigned long mask_ofl_ipi;
391 struct rcu_exp_work *rewp =
392 container_of(wp, struct rcu_exp_work, rew_work);
393 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
394 struct rcu_state *rsp = rewp->rew_rsp;
396 func = rewp->rew_func;
397 raw_spin_lock_irqsave_rcu_node(rnp, flags);
399 /* Each pass checks a CPU for identity, offline, and idle. */
401 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
402 unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
403 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
404 struct rcu_dynticks *rdtp = per_cpu_ptr(&rcu_dynticks, cpu);
407 if (raw_smp_processor_id() == cpu ||
408 !(rnp->qsmaskinitnext & mask)) {
409 mask_ofl_test |= mask;
411 snap = rcu_dynticks_snap(rdtp);
412 if (rcu_dynticks_in_eqs(snap))
413 mask_ofl_test |= mask;
415 rdp->exp_dynticks_snap = snap;
418 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
421 * Need to wait for any blocked tasks as well. Note that
422 * additional blocking tasks will also block the expedited GP
423 * until such time as the ->expmask bits are cleared.
425 if (rcu_preempt_has_tasks(rnp))
426 rnp->exp_tasks = rnp->blkd_tasks.next;
427 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
429 /* IPI the remaining CPUs for expedited quiescent state. */
430 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
431 unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
432 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
434 if (!(mask_ofl_ipi & mask))
437 if (rcu_dynticks_in_eqs_since(rdp->dynticks,
438 rdp->exp_dynticks_snap)) {
439 mask_ofl_test |= mask;
442 ret = smp_call_function_single(cpu, func, rsp, 0);
444 mask_ofl_ipi &= ~mask;
447 /* Failed, raced with CPU hotplug operation. */
448 raw_spin_lock_irqsave_rcu_node(rnp, flags);
449 if ((rnp->qsmaskinitnext & mask) &&
450 (rnp->expmask & mask)) {
451 /* Online, so delay for a bit and try again. */
452 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
453 trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("selectofl"));
454 schedule_timeout_uninterruptible(1);
457 /* CPU really is offline, so we can ignore it. */
458 if (!(rnp->expmask & mask))
459 mask_ofl_ipi &= ~mask;
460 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
462 /* Report quiescent states for those that went offline. */
463 mask_ofl_test |= mask_ofl_ipi;
465 rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
469 * Select the nodes that the upcoming expedited grace period needs
472 static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
473 smp_call_func_t func)
475 struct rcu_node *rnp;
477 trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("reset"));
478 sync_exp_reset_tree(rsp);
479 trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("select"));
481 /* Schedule work for each leaf rcu_node structure. */
482 rcu_for_each_leaf_node(rsp, rnp) {
483 rnp->exp_need_flush = false;
484 if (!READ_ONCE(rnp->expmask))
485 continue; /* Avoid early boot non-existent wq. */
486 rnp->rew.rew_func = func;
487 rnp->rew.rew_rsp = rsp;
488 if (!READ_ONCE(rcu_par_gp_wq) ||
489 rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
490 /* No workqueues yet. */
491 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
494 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
495 queue_work_on(rnp->grplo, rcu_par_gp_wq, &rnp->rew.rew_work);
496 rnp->exp_need_flush = true;
499 /* Wait for workqueue jobs (if any) to complete. */
500 rcu_for_each_leaf_node(rsp, rnp)
501 if (rnp->exp_need_flush)
502 flush_work(&rnp->rew.rew_work);
505 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
508 unsigned long jiffies_stall;
509 unsigned long jiffies_start;
512 struct rcu_node *rnp;
513 struct rcu_node *rnp_root = rcu_get_root(rsp);
516 trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("startwait"));
517 jiffies_stall = rcu_jiffies_till_stall_check();
518 jiffies_start = jiffies;
521 ret = swait_event_timeout_exclusive(
523 sync_rcu_preempt_exp_done_unlocked(rnp_root),
525 if (ret > 0 || sync_rcu_preempt_exp_done_unlocked(rnp_root))
527 WARN_ON(ret < 0); /* workqueues should not be signaled. */
528 if (rcu_cpu_stall_suppress)
530 panic_on_rcu_stall();
531 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
534 rcu_for_each_leaf_node(rsp, rnp) {
535 ndetected += rcu_print_task_exp_stall(rnp);
536 for_each_leaf_node_possible_cpu(rnp, cpu) {
537 struct rcu_data *rdp;
539 mask = leaf_node_cpu_bit(rnp, cpu);
540 if (!(rnp->expmask & mask))
543 rdp = per_cpu_ptr(rsp->rda, cpu);
544 pr_cont(" %d-%c%c%c", cpu,
545 "O."[!!cpu_online(cpu)],
546 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
547 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
550 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
551 jiffies - jiffies_start, rsp->expedited_sequence,
552 rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]);
554 pr_err("blocking rcu_node structures:");
555 rcu_for_each_node_breadth_first(rsp, rnp) {
557 continue; /* printed unconditionally */
558 if (sync_rcu_preempt_exp_done_unlocked(rnp))
560 pr_cont(" l=%u:%d-%d:%#lx/%c",
561 rnp->level, rnp->grplo, rnp->grphi,
563 ".T"[!!rnp->exp_tasks]);
567 rcu_for_each_leaf_node(rsp, rnp) {
568 for_each_leaf_node_possible_cpu(rnp, cpu) {
569 mask = leaf_node_cpu_bit(rnp, cpu);
570 if (!(rnp->expmask & mask))
575 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
580 * Wait for the current expedited grace period to complete, and then
581 * wake up everyone who piggybacked on the just-completed expedited
582 * grace period. Also update all the ->exp_seq_rq counters as needed
583 * in order to avoid counter-wrap problems.
585 static void rcu_exp_wait_wake(struct rcu_state *rsp, unsigned long s)
587 struct rcu_node *rnp;
589 synchronize_sched_expedited_wait(rsp);
590 rcu_exp_gp_seq_end(rsp);
591 trace_rcu_exp_grace_period(rsp->name, s, TPS("end"));
594 * Switch over to wakeup mode, allowing the next GP, but -only- the
595 * next GP, to proceed.
597 mutex_lock(&rsp->exp_wake_mutex);
599 rcu_for_each_node_breadth_first(rsp, rnp) {
600 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
601 spin_lock(&rnp->exp_lock);
602 /* Recheck, avoid hang in case someone just arrived. */
603 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
605 spin_unlock(&rnp->exp_lock);
607 smp_mb(); /* All above changes before wakeup. */
608 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(rsp->expedited_sequence) & 0x3]);
610 trace_rcu_exp_grace_period(rsp->name, s, TPS("endwake"));
611 mutex_unlock(&rsp->exp_wake_mutex);
615 * Common code to drive an expedited grace period forward, used by
616 * workqueues and mid-boot-time tasks.
618 static void rcu_exp_sel_wait_wake(struct rcu_state *rsp,
619 smp_call_func_t func, unsigned long s)
621 /* Initialize the rcu_node tree in preparation for the wait. */
622 sync_rcu_exp_select_cpus(rsp, func);
624 /* Wait and clean up, including waking everyone. */
625 rcu_exp_wait_wake(rsp, s);
629 * Work-queue handler to drive an expedited grace period forward.
631 static void wait_rcu_exp_gp(struct work_struct *wp)
633 struct rcu_exp_work *rewp;
635 rewp = container_of(wp, struct rcu_exp_work, rew_work);
636 rcu_exp_sel_wait_wake(rewp->rew_rsp, rewp->rew_func, rewp->rew_s);
640 * Given an rcu_state pointer and a smp_call_function() handler, kick
641 * off the specified flavor of expedited grace period.
643 static void _synchronize_rcu_expedited(struct rcu_state *rsp,
644 smp_call_func_t func)
646 struct rcu_data *rdp;
647 struct rcu_exp_work rew;
648 struct rcu_node *rnp;
651 /* If expedited grace periods are prohibited, fall back to normal. */
652 if (rcu_gp_is_normal()) {
653 wait_rcu_gp(rsp->call);
657 /* Take a snapshot of the sequence number. */
658 s = rcu_exp_gp_seq_snap(rsp);
659 if (exp_funnel_lock(rsp, s))
660 return; /* Someone else did our work for us. */
662 /* Ensure that load happens before action based on it. */
663 if (unlikely(rcu_scheduler_active == RCU_SCHEDULER_INIT)) {
664 /* Direct call during scheduler init and early_initcalls(). */
665 rcu_exp_sel_wait_wake(rsp, func, s);
667 /* Marshall arguments & schedule the expedited grace period. */
671 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
672 queue_work(rcu_gp_wq, &rew.rew_work);
675 /* Wait for expedited grace period to complete. */
676 rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
677 rnp = rcu_get_root(rsp);
678 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
679 sync_exp_work_done(rsp, s));
680 smp_mb(); /* Workqueue actions happen before return. */
682 /* Let the next expedited grace period start. */
683 mutex_unlock(&rsp->exp_mutex);
687 * synchronize_sched_expedited - Brute-force RCU-sched grace period
689 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
690 * approach to force the grace period to end quickly. This consumes
691 * significant time on all CPUs and is unfriendly to real-time workloads,
692 * so is thus not recommended for any sort of common-case code. In fact,
693 * if you are using synchronize_sched_expedited() in a loop, please
694 * restructure your code to batch your updates, and then use a single
695 * synchronize_sched() instead.
697 * This implementation can be thought of as an application of sequence
698 * locking to expedited grace periods, but using the sequence counter to
699 * determine when someone else has already done the work instead of for
702 void synchronize_sched_expedited(void)
704 struct rcu_state *rsp = &rcu_sched_state;
706 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
707 lock_is_held(&rcu_lock_map) ||
708 lock_is_held(&rcu_sched_lock_map),
709 "Illegal synchronize_sched_expedited() in RCU read-side critical section");
711 /* If only one CPU, this is automatically a grace period. */
712 if (rcu_blocking_is_gp())
715 _synchronize_rcu_expedited(rsp, sync_sched_exp_handler);
717 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
719 #ifdef CONFIG_PREEMPT_RCU
722 * Remote handler for smp_call_function_single(). If there is an
723 * RCU read-side critical section in effect, request that the
724 * next rcu_read_unlock() record the quiescent state up the
725 * ->expmask fields in the rcu_node tree. Otherwise, immediately
726 * report the quiescent state.
728 static void sync_rcu_exp_handler(void *info)
730 struct rcu_data *rdp;
731 struct rcu_state *rsp = info;
732 struct task_struct *t = current;
735 * Within an RCU read-side critical section, request that the next
736 * rcu_read_unlock() report. Unless this RCU read-side critical
737 * section has already blocked, in which case it is already set
738 * up for the expedited grace period to wait on it.
740 if (t->rcu_read_lock_nesting > 0 &&
741 !t->rcu_read_unlock_special.b.blocked) {
742 t->rcu_read_unlock_special.b.exp_need_qs = true;
747 * We are either exiting an RCU read-side critical section (negative
748 * values of t->rcu_read_lock_nesting) or are not in one at all
749 * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU
750 * read-side critical section that blocked before this expedited
751 * grace period started. Either way, we can immediately report
752 * the quiescent state.
754 rdp = this_cpu_ptr(rsp->rda);
755 rcu_report_exp_rdp(rsp, rdp, true);
759 * synchronize_rcu_expedited - Brute-force RCU grace period
761 * Wait for an RCU-preempt grace period, but expedite it. The basic
762 * idea is to IPI all non-idle non-nohz online CPUs. The IPI handler
763 * checks whether the CPU is in an RCU-preempt critical section, and
764 * if so, it sets a flag that causes the outermost rcu_read_unlock()
765 * to report the quiescent state. On the other hand, if the CPU is
766 * not in an RCU read-side critical section, the IPI handler reports
767 * the quiescent state immediately.
769 * Although this is a greate improvement over previous expedited
770 * implementations, it is still unfriendly to real-time workloads, so is
771 * thus not recommended for any sort of common-case code. In fact, if
772 * you are using synchronize_rcu_expedited() in a loop, please restructure
773 * your code to batch your updates, and then Use a single synchronize_rcu()
776 void synchronize_rcu_expedited(void)
778 struct rcu_state *rsp = rcu_state_p;
780 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
781 lock_is_held(&rcu_lock_map) ||
782 lock_is_held(&rcu_sched_lock_map),
783 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
785 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
787 _synchronize_rcu_expedited(rsp, sync_rcu_exp_handler);
789 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
791 #else /* #ifdef CONFIG_PREEMPT_RCU */
794 * Wait for an rcu-preempt grace period, but make it happen quickly.
795 * But because preemptible RCU does not exist, map to rcu-sched.
797 void synchronize_rcu_expedited(void)
799 synchronize_sched_expedited();
801 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
803 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */