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 the value that the 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, which is the
43 * earliest value that will indicate that a full grace period has
44 * elapsed since the current time.
46 static unsigned long rcu_exp_gp_seq_snap(void)
50 smp_mb(); /* Caller's modifications seen first by other CPUs. */
51 s = rcu_seq_snap(&rcu_state.expedited_sequence);
52 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
57 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
58 * if a full expedited grace period has elapsed since that snapshot
61 static bool rcu_exp_gp_seq_done(unsigned long s)
63 return rcu_seq_done(&rcu_state.expedited_sequence, s);
67 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
68 * recent CPU-online activity. Note that these masks are not cleared
69 * when CPUs go offline, so they reflect the union of all CPUs that have
70 * ever been online. This means that this function normally takes its
71 * no-work-to-do fastpath.
73 static void sync_exp_reset_tree_hotplug(void)
78 unsigned long oldmask;
79 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
81 struct rcu_node *rnp_up;
83 /* If no new CPUs onlined since last time, nothing to do. */
84 if (likely(ncpus == rcu_state.ncpus_snap))
86 rcu_state.ncpus_snap = ncpus;
89 * Each pass through the following loop propagates newly onlined
90 * CPUs for the current rcu_node structure up the rcu_node tree.
92 rcu_for_each_leaf_node(rnp) {
93 raw_spin_lock_irqsave_rcu_node(rnp, flags);
94 if (rnp->expmaskinit == rnp->expmaskinitnext) {
95 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
96 continue; /* No new CPUs, nothing to do. */
99 /* Update this node's mask, track old value for propagation. */
100 oldmask = rnp->expmaskinit;
101 rnp->expmaskinit = rnp->expmaskinitnext;
102 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
104 /* If was already nonzero, nothing to propagate. */
108 /* Propagate the new CPU up the tree. */
110 rnp_up = rnp->parent;
113 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
114 if (rnp_up->expmaskinit)
116 rnp_up->expmaskinit |= mask;
117 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
120 mask = rnp_up->grpmask;
121 rnp_up = rnp_up->parent;
127 * Reset the ->expmask values in the rcu_node tree in preparation for
128 * a new expedited grace period.
130 static void __maybe_unused sync_exp_reset_tree(void)
133 struct rcu_node *rnp;
135 sync_exp_reset_tree_hotplug();
136 rcu_for_each_node_breadth_first(rnp) {
137 raw_spin_lock_irqsave_rcu_node(rnp, flags);
138 WARN_ON_ONCE(rnp->expmask);
139 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
140 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
145 * Return non-zero if there is no RCU expedited grace period in progress
146 * for the specified rcu_node structure, in other words, if all CPUs and
147 * tasks covered by the specified rcu_node structure have done their bit
148 * for the current expedited grace period.
150 static bool sync_rcu_exp_done(struct rcu_node *rnp)
152 raw_lockdep_assert_held_rcu_node(rnp);
153 return READ_ONCE(rnp->exp_tasks) == NULL &&
154 READ_ONCE(rnp->expmask) == 0;
158 * Like sync_rcu_exp_done(), but where the caller does not hold the
161 static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
166 raw_spin_lock_irqsave_rcu_node(rnp, flags);
167 ret = sync_rcu_exp_done(rnp);
168 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
175 * Report the exit from RCU read-side critical section for the last task
176 * that queued itself during or before the current expedited preemptible-RCU
177 * grace period. This event is reported either to the rcu_node structure on
178 * which the task was queued or to one of that rcu_node structure's ancestors,
179 * recursively up the tree. (Calm down, calm down, we do the recursion
182 static void __rcu_report_exp_rnp(struct rcu_node *rnp,
183 bool wake, unsigned long flags)
184 __releases(rnp->lock)
188 raw_lockdep_assert_held_rcu_node(rnp);
190 if (!sync_rcu_exp_done(rnp)) {
192 rcu_initiate_boost(rnp, flags);
194 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
197 if (rnp->parent == NULL) {
198 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
200 smp_mb(); /* EGP done before wake_up(). */
201 swake_up_one(&rcu_state.expedited_wq);
206 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
208 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
209 WARN_ON_ONCE(!(rnp->expmask & mask));
210 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
215 * Report expedited quiescent state for specified node. This is a
216 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
218 static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
222 raw_spin_lock_irqsave_rcu_node(rnp, flags);
223 __rcu_report_exp_rnp(rnp, wake, flags);
227 * Report expedited quiescent state for multiple CPUs, all covered by the
228 * specified leaf rcu_node structure.
230 static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
231 unsigned long mask, bool wake)
235 struct rcu_data *rdp;
237 raw_spin_lock_irqsave_rcu_node(rnp, flags);
238 if (!(rnp->expmask & mask)) {
239 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
242 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
243 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
244 rdp = per_cpu_ptr(&rcu_data, cpu);
245 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
247 rdp->rcu_forced_tick_exp = false;
248 tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
250 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
254 * Report expedited quiescent state for specified rcu_data (CPU).
256 static void rcu_report_exp_rdp(struct rcu_data *rdp)
258 WRITE_ONCE(rdp->cpu_no_qs.b.exp, false);
259 rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
262 /* Common code for work-done checking. */
263 static bool sync_exp_work_done(unsigned long s)
265 if (rcu_exp_gp_seq_done(s)) {
266 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
267 smp_mb(); /* Ensure test happens before caller kfree(). */
274 * Funnel-lock acquisition for expedited grace periods. Returns true
275 * if some other task completed an expedited grace period that this task
276 * can piggy-back on, and with no mutex held. Otherwise, returns false
277 * with the mutex held, indicating that the caller must actually do the
278 * expedited grace period.
280 static bool exp_funnel_lock(unsigned long s)
282 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
283 struct rcu_node *rnp = rdp->mynode;
284 struct rcu_node *rnp_root = rcu_get_root();
286 /* Low-contention fastpath. */
287 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
289 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
290 mutex_trylock(&rcu_state.exp_mutex))
294 * Each pass through the following loop works its way up
295 * the rcu_node tree, returning if others have done the work or
296 * otherwise falls through to acquire ->exp_mutex. The mapping
297 * from CPU to rcu_node structure can be inexact, as it is just
298 * promoting locality and is not strictly needed for correctness.
300 for (; rnp != NULL; rnp = rnp->parent) {
301 if (sync_exp_work_done(s))
304 /* Work not done, either wait here or go up. */
305 spin_lock(&rnp->exp_lock);
306 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
308 /* Someone else doing GP, so wait for them. */
309 spin_unlock(&rnp->exp_lock);
310 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
311 rnp->grplo, rnp->grphi,
313 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
314 sync_exp_work_done(s));
317 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
318 spin_unlock(&rnp->exp_lock);
319 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
320 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
322 mutex_lock(&rcu_state.exp_mutex);
324 if (sync_exp_work_done(s)) {
325 mutex_unlock(&rcu_state.exp_mutex);
328 rcu_exp_gp_seq_start();
329 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
334 * Select the CPUs within the specified rcu_node that the upcoming
335 * expedited grace period needs to wait for.
337 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
341 unsigned long mask_ofl_test;
342 unsigned long mask_ofl_ipi;
344 struct rcu_exp_work *rewp =
345 container_of(wp, struct rcu_exp_work, rew_work);
346 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
348 raw_spin_lock_irqsave_rcu_node(rnp, flags);
350 /* Each pass checks a CPU for identity, offline, and idle. */
352 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
353 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
354 unsigned long mask = rdp->grpmask;
357 if (raw_smp_processor_id() == cpu ||
358 !(rnp->qsmaskinitnext & mask)) {
359 mask_ofl_test |= mask;
361 snap = rcu_dynticks_snap(rdp);
362 if (rcu_dynticks_in_eqs(snap))
363 mask_ofl_test |= mask;
365 rdp->exp_dynticks_snap = snap;
368 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
371 * Need to wait for any blocked tasks as well. Note that
372 * additional blocking tasks will also block the expedited GP
373 * until such time as the ->expmask bits are cleared.
375 if (rcu_preempt_has_tasks(rnp))
376 WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
377 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
379 /* IPI the remaining CPUs for expedited quiescent state. */
380 for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
381 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
382 unsigned long mask = rdp->grpmask;
385 if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
386 mask_ofl_test |= mask;
389 if (get_cpu() == cpu) {
390 mask_ofl_test |= mask;
394 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
396 /* The CPU will report the QS in response to the IPI. */
400 /* Failed, raced with CPU hotplug operation. */
401 raw_spin_lock_irqsave_rcu_node(rnp, flags);
402 if ((rnp->qsmaskinitnext & mask) &&
403 (rnp->expmask & mask)) {
404 /* Online, so delay for a bit and try again. */
405 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
406 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
407 schedule_timeout_idle(1);
410 /* CPU really is offline, so we must report its QS. */
411 if (rnp->expmask & mask)
412 mask_ofl_test |= mask;
413 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
415 /* Report quiescent states for those that went offline. */
417 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
421 * Select the nodes that the upcoming expedited grace period needs
424 static void sync_rcu_exp_select_cpus(void)
427 struct rcu_node *rnp;
429 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
430 sync_exp_reset_tree();
431 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
433 /* Schedule work for each leaf rcu_node structure. */
434 rcu_for_each_leaf_node(rnp) {
435 rnp->exp_need_flush = false;
436 if (!READ_ONCE(rnp->expmask))
437 continue; /* Avoid early boot non-existent wq. */
438 if (!READ_ONCE(rcu_par_gp_wq) ||
439 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
440 rcu_is_last_leaf_node(rnp)) {
441 /* No workqueues yet or last leaf, do direct call. */
442 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
445 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
446 cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
447 /* If all offline, queue the work on an unbound CPU. */
448 if (unlikely(cpu > rnp->grphi - rnp->grplo))
449 cpu = WORK_CPU_UNBOUND;
452 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
453 rnp->exp_need_flush = true;
456 /* Wait for workqueue jobs (if any) to complete. */
457 rcu_for_each_leaf_node(rnp)
458 if (rnp->exp_need_flush)
459 flush_work(&rnp->rew.rew_work);
463 * Wait for the expedited grace period to elapse, within time limit.
464 * If the time limit is exceeded without the grace period elapsing,
465 * return false, otherwise return true.
467 static bool synchronize_rcu_expedited_wait_once(long tlimit)
470 struct rcu_node *rnp_root = rcu_get_root();
472 t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
473 sync_rcu_exp_done_unlocked(rnp_root),
475 // Workqueues should not be signaled.
476 if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
478 WARN_ON(t < 0); /* workqueues should not be signaled. */
483 * Wait for the expedited grace period to elapse, issuing any needed
484 * RCU CPU stall warnings along the way.
486 static void synchronize_rcu_expedited_wait(void)
490 unsigned long jiffies_stall;
491 unsigned long jiffies_start;
494 struct rcu_data *rdp;
495 struct rcu_node *rnp;
496 struct rcu_node *rnp_root = rcu_get_root();
498 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
499 jiffies_stall = rcu_jiffies_till_stall_check();
500 jiffies_start = jiffies;
501 if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
502 if (synchronize_rcu_expedited_wait_once(1))
504 rcu_for_each_leaf_node(rnp) {
505 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
506 rdp = per_cpu_ptr(&rcu_data, cpu);
507 if (rdp->rcu_forced_tick_exp)
509 rdp->rcu_forced_tick_exp = true;
512 tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
516 j = READ_ONCE(jiffies_till_first_fqs);
517 if (synchronize_rcu_expedited_wait_once(j + HZ))
522 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
524 if (rcu_stall_is_suppressed())
526 panic_on_rcu_stall();
527 trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall"));
528 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
531 rcu_for_each_leaf_node(rnp) {
532 ndetected += rcu_print_task_exp_stall(rnp);
533 for_each_leaf_node_possible_cpu(rnp, cpu) {
534 struct rcu_data *rdp;
536 mask = leaf_node_cpu_bit(rnp, cpu);
537 if (!(READ_ONCE(rnp->expmask) & mask))
540 rdp = per_cpu_ptr(&rcu_data, cpu);
541 pr_cont(" %d-%c%c%c", cpu,
542 "O."[!!cpu_online(cpu)],
543 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
544 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
547 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
548 jiffies - jiffies_start, rcu_state.expedited_sequence,
549 data_race(rnp_root->expmask),
550 ".T"[!!data_race(rnp_root->exp_tasks)]);
552 pr_err("blocking rcu_node structures (internal RCU debug):");
553 rcu_for_each_node_breadth_first(rnp) {
555 continue; /* printed unconditionally */
556 if (sync_rcu_exp_done_unlocked(rnp))
558 pr_cont(" l=%u:%d-%d:%#lx/%c",
559 rnp->level, rnp->grplo, rnp->grphi,
560 data_race(rnp->expmask),
561 ".T"[!!data_race(rnp->exp_tasks)]);
565 rcu_for_each_leaf_node(rnp) {
566 for_each_leaf_node_possible_cpu(rnp, cpu) {
567 mask = leaf_node_cpu_bit(rnp, cpu);
568 if (!(READ_ONCE(rnp->expmask) & mask))
573 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
578 * Wait for the current expedited grace period to complete, and then
579 * wake up everyone who piggybacked on the just-completed expedited
580 * grace period. Also update all the ->exp_seq_rq counters as needed
581 * in order to avoid counter-wrap problems.
583 static void rcu_exp_wait_wake(unsigned long s)
585 struct rcu_node *rnp;
587 synchronize_rcu_expedited_wait();
589 // Switch over to wakeup mode, allowing the next GP to proceed.
590 // End the previous grace period only after acquiring the mutex
591 // to ensure that only one GP runs concurrently with wakeups.
592 mutex_lock(&rcu_state.exp_wake_mutex);
593 rcu_exp_gp_seq_end();
594 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
596 rcu_for_each_node_breadth_first(rnp) {
597 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
598 spin_lock(&rnp->exp_lock);
599 /* Recheck, avoid hang in case someone just arrived. */
600 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
601 WRITE_ONCE(rnp->exp_seq_rq, s);
602 spin_unlock(&rnp->exp_lock);
604 smp_mb(); /* All above changes before wakeup. */
605 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
607 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
608 mutex_unlock(&rcu_state.exp_wake_mutex);
612 * Common code to drive an expedited grace period forward, used by
613 * workqueues and mid-boot-time tasks.
615 static void rcu_exp_sel_wait_wake(unsigned long s)
617 /* Initialize the rcu_node tree in preparation for the wait. */
618 sync_rcu_exp_select_cpus();
620 /* Wait and clean up, including waking everyone. */
621 rcu_exp_wait_wake(s);
625 * Work-queue handler to drive an expedited grace period forward.
627 static void wait_rcu_exp_gp(struct work_struct *wp)
629 struct rcu_exp_work *rewp;
631 rewp = container_of(wp, struct rcu_exp_work, rew_work);
632 rcu_exp_sel_wait_wake(rewp->rew_s);
635 #ifdef CONFIG_PREEMPT_RCU
638 * Remote handler for smp_call_function_single(). If there is an
639 * RCU read-side critical section in effect, request that the
640 * next rcu_read_unlock() record the quiescent state up the
641 * ->expmask fields in the rcu_node tree. Otherwise, immediately
642 * report the quiescent state.
644 static void rcu_exp_handler(void *unused)
646 int depth = rcu_preempt_depth();
648 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
649 struct rcu_node *rnp = rdp->mynode;
650 struct task_struct *t = current;
653 * First, the common case of not being in an RCU read-side
654 * critical section. If also enabled or idle, immediately
655 * report the quiescent state, otherwise defer.
658 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
659 rcu_dynticks_curr_cpu_in_eqs()) {
660 rcu_report_exp_rdp(rdp);
662 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
663 set_tsk_need_resched(t);
664 set_preempt_need_resched();
670 * Second, the less-common case of being in an RCU read-side
671 * critical section. In this case we can count on a future
672 * rcu_read_unlock(). However, this rcu_read_unlock() might
673 * execute on some other CPU, but in that case there will be
674 * a future context switch. Either way, if the expedited
675 * grace period is still waiting on this CPU, set ->deferred_qs
676 * so that the eventual quiescent state will be reported.
677 * Note that there is a large group of race conditions that
678 * can have caused this quiescent state to already have been
679 * reported, so we really do need to check ->expmask.
682 raw_spin_lock_irqsave_rcu_node(rnp, flags);
683 if (rnp->expmask & rdp->grpmask) {
684 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
685 t->rcu_read_unlock_special.b.exp_hint = true;
687 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
691 // Finally, negative nesting depth should not happen.
695 /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
696 static void sync_sched_exp_online_cleanup(int cpu)
701 * Scan the current list of tasks blocked within RCU read-side critical
702 * sections, printing out the tid of each that is blocking the current
703 * expedited grace period.
705 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
709 struct task_struct *t;
711 if (!READ_ONCE(rnp->exp_tasks))
713 raw_spin_lock_irqsave_rcu_node(rnp, flags);
714 t = list_entry(rnp->exp_tasks->prev,
715 struct task_struct, rcu_node_entry);
716 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
717 pr_cont(" P%d", t->pid);
720 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
724 #else /* #ifdef CONFIG_PREEMPT_RCU */
726 /* Request an expedited quiescent state. */
727 static void rcu_exp_need_qs(void)
729 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
730 /* Store .exp before .rcu_urgent_qs. */
731 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
732 set_tsk_need_resched(current);
733 set_preempt_need_resched();
736 /* Invoked on each online non-idle CPU for expedited quiescent state. */
737 static void rcu_exp_handler(void *unused)
739 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
740 struct rcu_node *rnp = rdp->mynode;
742 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
743 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
745 if (rcu_is_cpu_rrupt_from_idle()) {
746 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
752 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
753 static void sync_sched_exp_online_cleanup(int cpu)
757 struct rcu_data *rdp;
759 struct rcu_node *rnp;
761 rdp = per_cpu_ptr(&rcu_data, cpu);
764 /* Quiescent state either not needed or already requested, leave. */
765 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
766 READ_ONCE(rdp->cpu_no_qs.b.exp)) {
770 /* Quiescent state needed on current CPU, so set it up locally. */
772 local_irq_save(flags);
774 local_irq_restore(flags);
778 /* Quiescent state needed on some other CPU, send IPI. */
779 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
785 * Because preemptible RCU does not exist, we never have to check for
786 * tasks blocked within RCU read-side critical sections that are
787 * blocking the current expedited grace period.
789 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
794 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
797 * synchronize_rcu_expedited - Brute-force RCU grace period
799 * Wait for an RCU grace period, but expedite it. The basic idea is to
800 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
801 * the CPU is in an RCU critical section, and if so, it sets a flag that
802 * causes the outermost rcu_read_unlock() to report the quiescent state
803 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
804 * other hand, if the CPU is not in an RCU read-side critical section,
805 * the IPI handler reports the quiescent state immediately.
807 * Although this is a great improvement over previous expedited
808 * implementations, it is still unfriendly to real-time workloads, so is
809 * thus not recommended for any sort of common-case code. In fact, if
810 * you are using synchronize_rcu_expedited() in a loop, please restructure
811 * your code to batch your updates, and then use a single synchronize_rcu()
814 * This has the same semantics as (but is more brutal than) synchronize_rcu().
816 void synchronize_rcu_expedited(void)
818 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
819 struct rcu_exp_work rew;
820 struct rcu_node *rnp;
823 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
824 lock_is_held(&rcu_lock_map) ||
825 lock_is_held(&rcu_sched_lock_map),
826 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
828 /* Is the state is such that the call is a grace period? */
829 if (rcu_blocking_is_gp())
832 /* If expedited grace periods are prohibited, fall back to normal. */
833 if (rcu_gp_is_normal()) {
834 wait_rcu_gp(call_rcu);
838 /* Take a snapshot of the sequence number. */
839 s = rcu_exp_gp_seq_snap();
840 if (exp_funnel_lock(s))
841 return; /* Someone else did our work for us. */
843 /* Ensure that load happens before action based on it. */
844 if (unlikely(boottime)) {
845 /* Direct call during scheduler init and early_initcalls(). */
846 rcu_exp_sel_wait_wake(s);
848 /* Marshall arguments & schedule the expedited grace period. */
850 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
851 queue_work(rcu_gp_wq, &rew.rew_work);
854 /* Wait for expedited grace period to complete. */
855 rnp = rcu_get_root();
856 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
857 sync_exp_work_done(s));
858 smp_mb(); /* Workqueue actions happen before return. */
860 /* Let the next expedited grace period start. */
861 mutex_unlock(&rcu_state.exp_mutex);
863 if (likely(!boottime))
864 destroy_work_on_stack(&rew.rew_work);
866 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);