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);
14 static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp);
17 * Record the start of an expedited grace period.
19 static void rcu_exp_gp_seq_start(void)
21 rcu_seq_start(&rcu_state.expedited_sequence);
22 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
26 * Return the value that the expedited-grace-period counter will have
27 * at the end of the current grace period.
29 static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
31 return rcu_seq_endval(&rcu_state.expedited_sequence);
35 * Record the end of an expedited grace period.
37 static void rcu_exp_gp_seq_end(void)
39 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
40 rcu_seq_end(&rcu_state.expedited_sequence);
41 smp_mb(); /* Ensure that consecutive grace periods serialize. */
45 * Take a snapshot of the expedited-grace-period counter, which is the
46 * earliest value that will indicate that a full grace period has
47 * elapsed since the current time.
49 static unsigned long rcu_exp_gp_seq_snap(void)
53 smp_mb(); /* Caller's modifications seen first by other CPUs. */
54 s = rcu_seq_snap(&rcu_state.expedited_sequence);
55 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
60 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
61 * if a full expedited grace period has elapsed since that snapshot
64 static bool rcu_exp_gp_seq_done(unsigned long s)
66 return rcu_seq_done(&rcu_state.expedited_sequence, s);
70 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
71 * recent CPU-online activity. Note that these masks are not cleared
72 * when CPUs go offline, so they reflect the union of all CPUs that have
73 * ever been online. This means that this function normally takes its
74 * no-work-to-do fastpath.
76 static void sync_exp_reset_tree_hotplug(void)
81 unsigned long oldmask;
82 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
84 struct rcu_node *rnp_up;
86 /* If no new CPUs onlined since last time, nothing to do. */
87 if (likely(ncpus == rcu_state.ncpus_snap))
89 rcu_state.ncpus_snap = ncpus;
92 * Each pass through the following loop propagates newly onlined
93 * CPUs for the current rcu_node structure up the rcu_node tree.
95 rcu_for_each_leaf_node(rnp) {
96 raw_spin_lock_irqsave_rcu_node(rnp, flags);
97 if (rnp->expmaskinit == rnp->expmaskinitnext) {
98 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
99 continue; /* No new CPUs, nothing to do. */
102 /* Update this node's mask, track old value for propagation. */
103 oldmask = rnp->expmaskinit;
104 rnp->expmaskinit = rnp->expmaskinitnext;
105 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
107 /* If was already nonzero, nothing to propagate. */
111 /* Propagate the new CPU up the tree. */
113 rnp_up = rnp->parent;
116 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
117 if (rnp_up->expmaskinit)
119 rnp_up->expmaskinit |= mask;
120 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
123 mask = rnp_up->grpmask;
124 rnp_up = rnp_up->parent;
130 * Reset the ->expmask values in the rcu_node tree in preparation for
131 * a new expedited grace period.
133 static void __maybe_unused sync_exp_reset_tree(void)
136 struct rcu_node *rnp;
138 sync_exp_reset_tree_hotplug();
139 rcu_for_each_node_breadth_first(rnp) {
140 raw_spin_lock_irqsave_rcu_node(rnp, flags);
141 WARN_ON_ONCE(rnp->expmask);
142 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
143 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
148 * Return non-zero if there is no RCU expedited grace period in progress
149 * for the specified rcu_node structure, in other words, if all CPUs and
150 * tasks covered by the specified rcu_node structure have done their bit
151 * for the current expedited grace period.
153 static bool sync_rcu_exp_done(struct rcu_node *rnp)
155 raw_lockdep_assert_held_rcu_node(rnp);
156 return READ_ONCE(rnp->exp_tasks) == NULL &&
157 READ_ONCE(rnp->expmask) == 0;
161 * Like sync_rcu_exp_done(), but where the caller does not hold the
164 static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
169 raw_spin_lock_irqsave_rcu_node(rnp, flags);
170 ret = sync_rcu_exp_done(rnp);
171 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
177 * Report the exit from RCU read-side critical section for the last task
178 * that queued itself during or before the current expedited preemptible-RCU
179 * grace period. This event is reported either to the rcu_node structure on
180 * which the task was queued or to one of that rcu_node structure's ancestors,
181 * recursively up the tree. (Calm down, calm down, we do the recursion
184 static void __rcu_report_exp_rnp(struct rcu_node *rnp,
185 bool wake, unsigned long flags)
186 __releases(rnp->lock)
190 raw_lockdep_assert_held_rcu_node(rnp);
192 if (!sync_rcu_exp_done(rnp)) {
194 rcu_initiate_boost(rnp, flags);
196 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
199 if (rnp->parent == NULL) {
200 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
202 smp_mb(); /* EGP done before wake_up(). */
203 swake_up_one_online(&rcu_state.expedited_wq);
208 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
210 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
211 WARN_ON_ONCE(!(rnp->expmask & mask));
212 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
217 * Report expedited quiescent state for specified node. This is a
218 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
220 static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
224 raw_spin_lock_irqsave_rcu_node(rnp, flags);
225 __rcu_report_exp_rnp(rnp, wake, flags);
229 * Report expedited quiescent state for multiple CPUs, all covered by the
230 * specified leaf rcu_node structure.
232 static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
233 unsigned long mask, bool wake)
237 struct rcu_data *rdp;
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 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
246 rdp = per_cpu_ptr(&rcu_data, cpu);
247 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
249 rdp->rcu_forced_tick_exp = false;
250 tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
252 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
256 * Report expedited quiescent state for specified rcu_data (CPU).
258 static void rcu_report_exp_rdp(struct rcu_data *rdp)
260 WRITE_ONCE(rdp->cpu_no_qs.b.exp, false);
261 rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
264 /* Common code for work-done checking. */
265 static bool sync_exp_work_done(unsigned long s)
267 if (rcu_exp_gp_seq_done(s)) {
268 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
269 smp_mb(); /* Ensure test happens before caller kfree(). */
276 * Funnel-lock acquisition for expedited grace periods. Returns true
277 * if some other task completed an expedited grace period that this task
278 * can piggy-back on, and with no mutex held. Otherwise, returns false
279 * with the mutex held, indicating that the caller must actually do the
280 * expedited grace period.
282 static bool exp_funnel_lock(unsigned long s)
284 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
285 struct rcu_node *rnp = rdp->mynode;
286 struct rcu_node *rnp_root = rcu_get_root();
288 /* Low-contention fastpath. */
289 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
291 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
292 mutex_trylock(&rcu_state.exp_mutex))
296 * Each pass through the following loop works its way up
297 * the rcu_node tree, returning if others have done the work or
298 * otherwise falls through to acquire ->exp_mutex. The mapping
299 * from CPU to rcu_node structure can be inexact, as it is just
300 * promoting locality and is not strictly needed for correctness.
302 for (; rnp != NULL; rnp = rnp->parent) {
303 if (sync_exp_work_done(s))
306 /* Work not done, either wait here or go up. */
307 spin_lock(&rnp->exp_lock);
308 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
310 /* Someone else doing GP, so wait for them. */
311 spin_unlock(&rnp->exp_lock);
312 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
313 rnp->grplo, rnp->grphi,
315 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
316 sync_exp_work_done(s));
319 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
320 spin_unlock(&rnp->exp_lock);
321 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
322 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
324 mutex_lock(&rcu_state.exp_mutex);
326 if (sync_exp_work_done(s)) {
327 mutex_unlock(&rcu_state.exp_mutex);
330 rcu_exp_gp_seq_start();
331 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
336 * Select the CPUs within the specified rcu_node that the upcoming
337 * expedited grace period needs to wait for.
339 static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
343 unsigned long mask_ofl_test;
344 unsigned long mask_ofl_ipi;
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(cpu);
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);
420 static void rcu_exp_sel_wait_wake(unsigned long s);
422 #ifdef CONFIG_RCU_EXP_KTHREAD
423 static void sync_rcu_exp_select_node_cpus(struct kthread_work *wp)
425 struct rcu_exp_work *rewp =
426 container_of(wp, struct rcu_exp_work, rew_work);
428 __sync_rcu_exp_select_node_cpus(rewp);
431 static inline bool rcu_gp_par_worker_started(void)
433 return !!READ_ONCE(rcu_exp_par_gp_kworker);
436 static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
438 kthread_init_work(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
440 * Use rcu_exp_par_gp_kworker, because flushing a work item from
441 * another work item on the same kthread worker can result in
444 kthread_queue_work(rcu_exp_par_gp_kworker, &rnp->rew.rew_work);
447 static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
449 kthread_flush_work(&rnp->rew.rew_work);
453 * Work-queue handler to drive an expedited grace period forward.
455 static void wait_rcu_exp_gp(struct kthread_work *wp)
457 struct rcu_exp_work *rewp;
459 rewp = container_of(wp, struct rcu_exp_work, rew_work);
460 rcu_exp_sel_wait_wake(rewp->rew_s);
463 static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
465 kthread_init_work(&rew->rew_work, wait_rcu_exp_gp);
466 kthread_queue_work(rcu_exp_gp_kworker, &rew->rew_work);
469 static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
472 #else /* !CONFIG_RCU_EXP_KTHREAD */
473 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
475 struct rcu_exp_work *rewp =
476 container_of(wp, struct rcu_exp_work, rew_work);
478 __sync_rcu_exp_select_node_cpus(rewp);
481 static inline bool rcu_gp_par_worker_started(void)
483 return !!READ_ONCE(rcu_par_gp_wq);
486 static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
488 int cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
490 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
491 /* If all offline, queue the work on an unbound CPU. */
492 if (unlikely(cpu > rnp->grphi - rnp->grplo))
493 cpu = WORK_CPU_UNBOUND;
496 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
499 static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
501 flush_work(&rnp->rew.rew_work);
505 * Work-queue handler to drive an expedited grace period forward.
507 static void wait_rcu_exp_gp(struct work_struct *wp)
509 struct rcu_exp_work *rewp;
511 rewp = container_of(wp, struct rcu_exp_work, rew_work);
512 rcu_exp_sel_wait_wake(rewp->rew_s);
515 static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
517 INIT_WORK_ONSTACK(&rew->rew_work, wait_rcu_exp_gp);
518 queue_work(rcu_gp_wq, &rew->rew_work);
521 static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
523 destroy_work_on_stack(&rew->rew_work);
525 #endif /* CONFIG_RCU_EXP_KTHREAD */
528 * Select the nodes that the upcoming expedited grace period needs
531 static void sync_rcu_exp_select_cpus(void)
533 struct rcu_node *rnp;
535 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
536 sync_exp_reset_tree();
537 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
539 /* Schedule work for each leaf rcu_node structure. */
540 rcu_for_each_leaf_node(rnp) {
541 rnp->exp_need_flush = false;
542 if (!READ_ONCE(rnp->expmask))
543 continue; /* Avoid early boot non-existent wq. */
544 if (!rcu_gp_par_worker_started() ||
545 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
546 rcu_is_last_leaf_node(rnp)) {
547 /* No worker started yet or last leaf, do direct call. */
548 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
551 sync_rcu_exp_select_cpus_queue_work(rnp);
552 rnp->exp_need_flush = true;
555 /* Wait for jobs (if any) to complete. */
556 rcu_for_each_leaf_node(rnp)
557 if (rnp->exp_need_flush)
558 sync_rcu_exp_select_cpus_flush_work(rnp);
562 * Wait for the expedited grace period to elapse, within time limit.
563 * If the time limit is exceeded without the grace period elapsing,
564 * return false, otherwise return true.
566 static bool synchronize_rcu_expedited_wait_once(long tlimit)
569 struct rcu_node *rnp_root = rcu_get_root();
571 t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
572 sync_rcu_exp_done_unlocked(rnp_root),
574 // Workqueues should not be signaled.
575 if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
577 WARN_ON(t < 0); /* workqueues should not be signaled. */
582 * Wait for the expedited grace period to elapse, issuing any needed
583 * RCU CPU stall warnings along the way.
585 static void synchronize_rcu_expedited_wait(void)
589 unsigned long jiffies_stall;
590 unsigned long jiffies_start;
593 struct rcu_data *rdp;
594 struct rcu_node *rnp;
595 struct rcu_node *rnp_root = rcu_get_root();
598 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
599 jiffies_stall = rcu_exp_jiffies_till_stall_check();
600 jiffies_start = jiffies;
601 if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
602 if (synchronize_rcu_expedited_wait_once(1))
604 rcu_for_each_leaf_node(rnp) {
605 raw_spin_lock_irqsave_rcu_node(rnp, flags);
606 mask = READ_ONCE(rnp->expmask);
607 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
608 rdp = per_cpu_ptr(&rcu_data, cpu);
609 if (rdp->rcu_forced_tick_exp)
611 rdp->rcu_forced_tick_exp = true;
613 tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
615 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
617 j = READ_ONCE(jiffies_till_first_fqs);
618 if (synchronize_rcu_expedited_wait_once(j + HZ))
623 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
625 if (rcu_stall_is_suppressed())
627 trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall"));
628 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
631 rcu_for_each_leaf_node(rnp) {
632 ndetected += rcu_print_task_exp_stall(rnp);
633 for_each_leaf_node_possible_cpu(rnp, cpu) {
634 struct rcu_data *rdp;
636 mask = leaf_node_cpu_bit(rnp, cpu);
637 if (!(READ_ONCE(rnp->expmask) & mask))
640 rdp = per_cpu_ptr(&rcu_data, cpu);
641 pr_cont(" %d-%c%c%c%c", cpu,
642 "O."[!!cpu_online(cpu)],
643 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
644 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)],
645 "D."[!!data_race(rdp->cpu_no_qs.b.exp)]);
648 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
649 jiffies - jiffies_start, rcu_state.expedited_sequence,
650 data_race(rnp_root->expmask),
651 ".T"[!!data_race(rnp_root->exp_tasks)]);
653 pr_err("blocking rcu_node structures (internal RCU debug):");
654 rcu_for_each_node_breadth_first(rnp) {
656 continue; /* printed unconditionally */
657 if (sync_rcu_exp_done_unlocked(rnp))
659 pr_cont(" l=%u:%d-%d:%#lx/%c",
660 rnp->level, rnp->grplo, rnp->grphi,
661 data_race(rnp->expmask),
662 ".T"[!!data_race(rnp->exp_tasks)]);
666 rcu_for_each_leaf_node(rnp) {
667 for_each_leaf_node_possible_cpu(rnp, cpu) {
668 mask = leaf_node_cpu_bit(rnp, cpu);
669 if (!(READ_ONCE(rnp->expmask) & mask))
671 preempt_disable(); // For smp_processor_id() in dump_cpu_task().
675 rcu_exp_print_detail_task_stall_rnp(rnp);
677 jiffies_stall = 3 * rcu_exp_jiffies_till_stall_check() + 3;
678 panic_on_rcu_stall();
683 * Wait for the current expedited grace period to complete, and then
684 * wake up everyone who piggybacked on the just-completed expedited
685 * grace period. Also update all the ->exp_seq_rq counters as needed
686 * in order to avoid counter-wrap problems.
688 static void rcu_exp_wait_wake(unsigned long s)
690 struct rcu_node *rnp;
692 synchronize_rcu_expedited_wait();
694 // Switch over to wakeup mode, allowing the next GP to proceed.
695 // End the previous grace period only after acquiring the mutex
696 // to ensure that only one GP runs concurrently with wakeups.
697 mutex_lock(&rcu_state.exp_wake_mutex);
698 rcu_exp_gp_seq_end();
699 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
701 rcu_for_each_node_breadth_first(rnp) {
702 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
703 spin_lock(&rnp->exp_lock);
704 /* Recheck, avoid hang in case someone just arrived. */
705 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
706 WRITE_ONCE(rnp->exp_seq_rq, s);
707 spin_unlock(&rnp->exp_lock);
709 smp_mb(); /* All above changes before wakeup. */
710 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
712 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
713 mutex_unlock(&rcu_state.exp_wake_mutex);
717 * Common code to drive an expedited grace period forward, used by
718 * workqueues and mid-boot-time tasks.
720 static void rcu_exp_sel_wait_wake(unsigned long s)
722 /* Initialize the rcu_node tree in preparation for the wait. */
723 sync_rcu_exp_select_cpus();
725 /* Wait and clean up, including waking everyone. */
726 rcu_exp_wait_wake(s);
729 #ifdef CONFIG_PREEMPT_RCU
732 * Remote handler for smp_call_function_single(). If there is an
733 * RCU read-side critical section in effect, request that the
734 * next rcu_read_unlock() record the quiescent state up the
735 * ->expmask fields in the rcu_node tree. Otherwise, immediately
736 * report the quiescent state.
738 static void rcu_exp_handler(void *unused)
740 int depth = rcu_preempt_depth();
742 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
743 struct rcu_node *rnp = rdp->mynode;
744 struct task_struct *t = current;
747 * First, the common case of not being in an RCU read-side
748 * critical section. If also enabled or idle, immediately
749 * report the quiescent state, otherwise defer.
752 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
753 rcu_is_cpu_rrupt_from_idle()) {
754 rcu_report_exp_rdp(rdp);
756 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
757 set_tsk_need_resched(t);
758 set_preempt_need_resched();
764 * Second, the less-common case of being in an RCU read-side
765 * critical section. In this case we can count on a future
766 * rcu_read_unlock(). However, this rcu_read_unlock() might
767 * execute on some other CPU, but in that case there will be
768 * a future context switch. Either way, if the expedited
769 * grace period is still waiting on this CPU, set ->deferred_qs
770 * so that the eventual quiescent state will be reported.
771 * Note that there is a large group of race conditions that
772 * can have caused this quiescent state to already have been
773 * reported, so we really do need to check ->expmask.
776 raw_spin_lock_irqsave_rcu_node(rnp, flags);
777 if (rnp->expmask & rdp->grpmask) {
778 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
779 t->rcu_read_unlock_special.b.exp_hint = true;
781 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
785 // Finally, negative nesting depth should not happen.
789 /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
790 static void sync_sched_exp_online_cleanup(int cpu)
795 * Scan the current list of tasks blocked within RCU read-side critical
796 * sections, printing out the tid of each that is blocking the current
797 * expedited grace period.
799 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
803 struct task_struct *t;
805 raw_spin_lock_irqsave_rcu_node(rnp, flags);
806 if (!rnp->exp_tasks) {
807 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
810 t = list_entry(rnp->exp_tasks->prev,
811 struct task_struct, rcu_node_entry);
812 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
813 pr_cont(" P%d", t->pid);
816 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
821 * Scan the current list of tasks blocked within RCU read-side critical
822 * sections, dumping the stack of each that is blocking the current
823 * expedited grace period.
825 static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
828 struct task_struct *t;
830 if (!rcu_exp_stall_task_details)
832 raw_spin_lock_irqsave_rcu_node(rnp, flags);
833 if (!READ_ONCE(rnp->exp_tasks)) {
834 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
837 t = list_entry(rnp->exp_tasks->prev,
838 struct task_struct, rcu_node_entry);
839 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
841 * We could be printing a lot while holding a spinlock.
842 * Avoid triggering hard lockup.
844 touch_nmi_watchdog();
847 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
850 #else /* #ifdef CONFIG_PREEMPT_RCU */
852 /* Request an expedited quiescent state. */
853 static void rcu_exp_need_qs(void)
855 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
856 /* Store .exp before .rcu_urgent_qs. */
857 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
858 set_tsk_need_resched(current);
859 set_preempt_need_resched();
862 /* Invoked on each online non-idle CPU for expedited quiescent state. */
863 static void rcu_exp_handler(void *unused)
865 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
866 struct rcu_node *rnp = rdp->mynode;
867 bool preempt_bh_enabled = !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
869 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
870 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
872 if (rcu_is_cpu_rrupt_from_idle() ||
873 (IS_ENABLED(CONFIG_PREEMPT_COUNT) && preempt_bh_enabled)) {
874 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
880 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
881 static void sync_sched_exp_online_cleanup(int cpu)
885 struct rcu_data *rdp;
887 struct rcu_node *rnp;
889 rdp = per_cpu_ptr(&rcu_data, cpu);
892 /* Quiescent state either not needed or already requested, leave. */
893 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
894 READ_ONCE(rdp->cpu_no_qs.b.exp)) {
898 /* Quiescent state needed on current CPU, so set it up locally. */
900 local_irq_save(flags);
902 local_irq_restore(flags);
906 /* Quiescent state needed on some other CPU, send IPI. */
907 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
913 * Because preemptible RCU does not exist, we never have to check for
914 * tasks blocked within RCU read-side critical sections that are
915 * blocking the current expedited grace period.
917 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
923 * Because preemptible RCU does not exist, we never have to print out
924 * tasks blocked within RCU read-side critical sections that are blocking
925 * the current expedited grace period.
927 static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
931 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
934 * synchronize_rcu_expedited - Brute-force RCU grace period
936 * Wait for an RCU grace period, but expedite it. The basic idea is to
937 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
938 * the CPU is in an RCU critical section, and if so, it sets a flag that
939 * causes the outermost rcu_read_unlock() to report the quiescent state
940 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
941 * other hand, if the CPU is not in an RCU read-side critical section,
942 * the IPI handler reports the quiescent state immediately.
944 * Although this is a great improvement over previous expedited
945 * implementations, it is still unfriendly to real-time workloads, so is
946 * thus not recommended for any sort of common-case code. In fact, if
947 * you are using synchronize_rcu_expedited() in a loop, please restructure
948 * your code to batch your updates, and then use a single synchronize_rcu()
951 * This has the same semantics as (but is more brutal than) synchronize_rcu().
953 void synchronize_rcu_expedited(void)
955 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
957 struct rcu_exp_work rew;
958 struct rcu_node *rnp;
961 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
962 lock_is_held(&rcu_lock_map) ||
963 lock_is_held(&rcu_sched_lock_map),
964 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
966 /* Is the state is such that the call is a grace period? */
967 if (rcu_blocking_is_gp()) {
968 // Note well that this code runs with !PREEMPT && !SMP.
969 // In addition, all code that advances grace periods runs
970 // at process level. Therefore, this expedited GP overlaps
971 // with other expedited GPs only by being fully nested within
972 // them, which allows reuse of ->gp_seq_polled_exp_snap.
973 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
974 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
976 local_irq_save(flags);
977 WARN_ON_ONCE(num_online_cpus() > 1);
978 rcu_state.expedited_sequence += (1 << RCU_SEQ_CTR_SHIFT);
979 local_irq_restore(flags);
980 return; // Context allows vacuous grace periods.
983 /* If expedited grace periods are prohibited, fall back to normal. */
984 if (rcu_gp_is_normal()) {
985 wait_rcu_gp(call_rcu_hurry);
989 /* Take a snapshot of the sequence number. */
990 s = rcu_exp_gp_seq_snap();
991 if (exp_funnel_lock(s))
992 return; /* Someone else did our work for us. */
994 /* Ensure that load happens before action based on it. */
995 if (unlikely(boottime)) {
996 /* Direct call during scheduler init and early_initcalls(). */
997 rcu_exp_sel_wait_wake(s);
999 /* Marshall arguments & schedule the expedited grace period. */
1001 synchronize_rcu_expedited_queue_work(&rew);
1004 /* Wait for expedited grace period to complete. */
1005 rnp = rcu_get_root();
1006 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
1007 sync_exp_work_done(s));
1008 smp_mb(); /* Work actions happen before return. */
1010 /* Let the next expedited grace period start. */
1011 mutex_unlock(&rcu_state.exp_mutex);
1013 if (likely(!boottime))
1014 synchronize_rcu_expedited_destroy_work(&rew);
1016 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1019 * Ensure that start_poll_synchronize_rcu_expedited() has the expedited
1020 * RCU grace periods that it needs.
1022 static void sync_rcu_do_polled_gp(struct work_struct *wp)
1024 unsigned long flags;
1026 struct rcu_node *rnp = container_of(wp, struct rcu_node, exp_poll_wq);
1029 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1030 s = rnp->exp_seq_poll_rq;
1031 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1032 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1033 if (s == RCU_GET_STATE_COMPLETED)
1035 while (!poll_state_synchronize_rcu(s)) {
1036 synchronize_rcu_expedited();
1037 if (i == 10 || i == 20)
1038 pr_info("%s: i = %d s = %lx gp_seq_polled = %lx\n", __func__, i, s, READ_ONCE(rcu_state.gp_seq_polled));
1041 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1042 s = rnp->exp_seq_poll_rq;
1043 if (poll_state_synchronize_rcu(s))
1044 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1045 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1049 * start_poll_synchronize_rcu_expedited - Snapshot current RCU state and start expedited grace period
1051 * Returns a cookie to pass to a call to cond_synchronize_rcu(),
1052 * cond_synchronize_rcu_expedited(), or poll_state_synchronize_rcu(),
1053 * allowing them to determine whether or not any sort of grace period has
1054 * elapsed in the meantime. If the needed expedited grace period is not
1055 * already slated to start, initiates that grace period.
1057 unsigned long start_poll_synchronize_rcu_expedited(void)
1059 unsigned long flags;
1060 struct rcu_data *rdp;
1061 struct rcu_node *rnp;
1064 s = get_state_synchronize_rcu();
1065 rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
1067 if (rcu_init_invoked())
1068 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1069 if (!poll_state_synchronize_rcu(s)) {
1070 if (rcu_init_invoked()) {
1071 rnp->exp_seq_poll_rq = s;
1072 queue_work(rcu_gp_wq, &rnp->exp_poll_wq);
1075 if (rcu_init_invoked())
1076 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1080 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited);
1083 * start_poll_synchronize_rcu_expedited_full - Take a full snapshot and start expedited grace period
1084 * @rgosp: Place to put snapshot of grace-period state
1086 * Places the normal and expedited grace-period states in rgosp. This
1087 * state value can be passed to a later call to cond_synchronize_rcu_full()
1088 * or poll_state_synchronize_rcu_full() to determine whether or not a
1089 * grace period (whether normal or expedited) has elapsed in the meantime.
1090 * If the needed expedited grace period is not already slated to start,
1091 * initiates that grace period.
1093 void start_poll_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1095 get_state_synchronize_rcu_full(rgosp);
1096 (void)start_poll_synchronize_rcu_expedited();
1098 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited_full);
1101 * cond_synchronize_rcu_expedited - Conditionally wait for an expedited RCU grace period
1103 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
1105 * If any type of full RCU grace period has elapsed since the earlier
1106 * call to get_state_synchronize_rcu(), start_poll_synchronize_rcu(),
1107 * or start_poll_synchronize_rcu_expedited(), just return. Otherwise,
1108 * invoke synchronize_rcu_expedited() to wait for a full grace period.
1110 * Yes, this function does not take counter wrap into account.
1111 * But counter wrap is harmless. If the counter wraps, we have waited for
1112 * more than 2 billion grace periods (and way more on a 64-bit system!),
1113 * so waiting for a couple of additional grace periods should be just fine.
1115 * This function provides the same memory-ordering guarantees that
1116 * would be provided by a synchronize_rcu() that was invoked at the call
1117 * to the function that provided @oldstate and that returned at the end
1120 void cond_synchronize_rcu_expedited(unsigned long oldstate)
1122 if (!poll_state_synchronize_rcu(oldstate))
1123 synchronize_rcu_expedited();
1125 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited);
1128 * cond_synchronize_rcu_expedited_full - Conditionally wait for an expedited RCU grace period
1129 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
1131 * If a full RCU grace period has elapsed since the call to
1132 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
1133 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
1134 * obtained, just return. Otherwise, invoke synchronize_rcu_expedited()
1135 * to wait for a full grace period.
1137 * Yes, this function does not take counter wrap into account.
1138 * But counter wrap is harmless. If the counter wraps, we have waited for
1139 * more than 2 billion grace periods (and way more on a 64-bit system!),
1140 * so waiting for a couple of additional grace periods should be just fine.
1142 * This function provides the same memory-ordering guarantees that
1143 * would be provided by a synchronize_rcu() that was invoked at the call
1144 * to the function that provided @rgosp and that returned at the end of
1147 void cond_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1149 if (!poll_state_synchronize_rcu_full(rgosp))
1150 synchronize_rcu_expedited();
1152 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited_full);