2 * Read-Copy Update mechanism for mutual exclusion
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, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
55 #include <trace/events/rcu.h>
59 /* Data structures. */
61 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
63 #define RCU_STATE_INITIALIZER(structname) { \
64 .level = { &structname.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .signaled = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
83 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
85 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
86 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
88 static struct rcu_state *rcu_state;
91 * The rcu_scheduler_active variable transitions from zero to one just
92 * before the first task is spawned. So when this variable is zero, RCU
93 * can assume that there is but one task, allowing RCU to (for example)
94 * optimized synchronize_sched() to a simple barrier(). When this variable
95 * is one, RCU must actually do all the hard work required to detect real
96 * grace periods. This variable is also used to suppress boot-time false
97 * positives from lockdep-RCU error checking.
99 int rcu_scheduler_active __read_mostly;
100 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
103 * The rcu_scheduler_fully_active variable transitions from zero to one
104 * during the early_initcall() processing, which is after the scheduler
105 * is capable of creating new tasks. So RCU processing (for example,
106 * creating tasks for RCU priority boosting) must be delayed until after
107 * rcu_scheduler_fully_active transitions from zero to one. We also
108 * currently delay invocation of any RCU callbacks until after this point.
110 * It might later prove better for people registering RCU callbacks during
111 * early boot to take responsibility for these callbacks, but one step at
114 static int rcu_scheduler_fully_active __read_mostly;
116 #ifdef CONFIG_RCU_BOOST
119 * Control variables for per-CPU and per-rcu_node kthreads. These
120 * handle all flavors of RCU.
122 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
123 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
124 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
126 DEFINE_PER_CPU(char, rcu_cpu_has_work);
128 #endif /* #ifdef CONFIG_RCU_BOOST */
130 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
131 static void invoke_rcu_core(void);
132 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
134 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
137 * Track the rcutorture test sequence number and the update version
138 * number within a given test. The rcutorture_testseq is incremented
139 * on every rcutorture module load and unload, so has an odd value
140 * when a test is running. The rcutorture_vernum is set to zero
141 * when rcutorture starts and is incremented on each rcutorture update.
142 * These variables enable correlating rcutorture output with the
143 * RCU tracing information.
145 unsigned long rcutorture_testseq;
146 unsigned long rcutorture_vernum;
149 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
150 * permit this function to be invoked without holding the root rcu_node
151 * structure's ->lock, but of course results can be subject to change.
153 static int rcu_gp_in_progress(struct rcu_state *rsp)
155 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
159 * Note a quiescent state. Because we do not need to know
160 * how many quiescent states passed, just if there was at least
161 * one since the start of the grace period, this just sets a flag.
163 void rcu_sched_qs(int cpu)
165 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
167 rdp->passed_quiesc_completed = rdp->gpnum - 1;
169 rdp->passed_quiesc = 1;
172 void rcu_bh_qs(int cpu)
174 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
176 rdp->passed_quiesc_completed = rdp->gpnum - 1;
178 rdp->passed_quiesc = 1;
182 * Note a context switch. This is a quiescent state for RCU-sched,
183 * and requires special handling for preemptible RCU.
185 void rcu_note_context_switch(int cpu)
188 rcu_preempt_note_context_switch(cpu);
190 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
193 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
194 .dynticks_nesting = 1,
195 .dynticks = ATOMIC_INIT(1),
197 #endif /* #ifdef CONFIG_NO_HZ */
199 static int blimit = 10; /* Maximum callbacks per softirq. */
200 static int qhimark = 10000; /* If this many pending, ignore blimit. */
201 static int qlowmark = 100; /* Once only this many pending, use blimit. */
203 module_param(blimit, int, 0);
204 module_param(qhimark, int, 0);
205 module_param(qlowmark, int, 0);
207 int rcu_cpu_stall_suppress __read_mostly;
208 module_param(rcu_cpu_stall_suppress, int, 0644);
210 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
211 static int rcu_pending(int cpu);
214 * Return the number of RCU-sched batches processed thus far for debug & stats.
216 long rcu_batches_completed_sched(void)
218 return rcu_sched_state.completed;
220 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
223 * Return the number of RCU BH batches processed thus far for debug & stats.
225 long rcu_batches_completed_bh(void)
227 return rcu_bh_state.completed;
229 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
232 * Force a quiescent state for RCU BH.
234 void rcu_bh_force_quiescent_state(void)
236 force_quiescent_state(&rcu_bh_state, 0);
238 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
241 * Record the number of times rcutorture tests have been initiated and
242 * terminated. This information allows the debugfs tracing stats to be
243 * correlated to the rcutorture messages, even when the rcutorture module
244 * is being repeatedly loaded and unloaded. In other words, we cannot
245 * store this state in rcutorture itself.
247 void rcutorture_record_test_transition(void)
249 rcutorture_testseq++;
250 rcutorture_vernum = 0;
252 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
255 * Record the number of writer passes through the current rcutorture test.
256 * This is also used to correlate debugfs tracing stats with the rcutorture
259 void rcutorture_record_progress(unsigned long vernum)
263 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
266 * Force a quiescent state for RCU-sched.
268 void rcu_sched_force_quiescent_state(void)
270 force_quiescent_state(&rcu_sched_state, 0);
272 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
275 * Does the CPU have callbacks ready to be invoked?
278 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
280 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
284 * Does the current CPU require a yet-as-unscheduled grace period?
287 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
289 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
293 * Return the root node of the specified rcu_state structure.
295 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
297 return &rsp->node[0];
303 * If the specified CPU is offline, tell the caller that it is in
304 * a quiescent state. Otherwise, whack it with a reschedule IPI.
305 * Grace periods can end up waiting on an offline CPU when that
306 * CPU is in the process of coming online -- it will be added to the
307 * rcu_node bitmasks before it actually makes it online. The same thing
308 * can happen while a CPU is in the process of coming online. Because this
309 * race is quite rare, we check for it after detecting that the grace
310 * period has been delayed rather than checking each and every CPU
311 * each and every time we start a new grace period.
313 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
316 * If the CPU is offline, it is in a quiescent state. We can
317 * trust its state not to change because interrupts are disabled.
319 if (cpu_is_offline(rdp->cpu)) {
324 /* If preemptible RCU, no point in sending reschedule IPI. */
325 if (rdp->preemptible)
328 /* The CPU is online, so send it a reschedule IPI. */
329 if (rdp->cpu != smp_processor_id())
330 smp_send_reschedule(rdp->cpu);
337 #endif /* #ifdef CONFIG_SMP */
342 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
344 * Enter nohz mode, in other words, -leave- the mode in which RCU
345 * read-side critical sections can occur. (Though RCU read-side
346 * critical sections can occur in irq handlers in nohz mode, a possibility
347 * handled by rcu_irq_enter() and rcu_irq_exit()).
349 void rcu_enter_nohz(void)
352 struct rcu_dynticks *rdtp;
354 local_irq_save(flags);
355 rdtp = &__get_cpu_var(rcu_dynticks);
356 if (--rdtp->dynticks_nesting) {
357 local_irq_restore(flags);
360 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
361 smp_mb__before_atomic_inc(); /* See above. */
362 atomic_inc(&rdtp->dynticks);
363 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
364 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
365 local_irq_restore(flags);
367 /* If the interrupt queued a callback, get out of dyntick mode. */
369 (__get_cpu_var(rcu_sched_data).nxtlist ||
370 __get_cpu_var(rcu_bh_data).nxtlist ||
371 rcu_preempt_needs_cpu(smp_processor_id())))
376 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
378 * Exit nohz mode, in other words, -enter- the mode in which RCU
379 * read-side critical sections normally occur.
381 void rcu_exit_nohz(void)
384 struct rcu_dynticks *rdtp;
386 local_irq_save(flags);
387 rdtp = &__get_cpu_var(rcu_dynticks);
388 if (rdtp->dynticks_nesting++) {
389 local_irq_restore(flags);
392 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
393 atomic_inc(&rdtp->dynticks);
394 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
395 smp_mb__after_atomic_inc(); /* See above. */
396 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
397 local_irq_restore(flags);
401 * rcu_nmi_enter - inform RCU of entry to NMI context
403 * If the CPU was idle with dynamic ticks active, and there is no
404 * irq handler running, this updates rdtp->dynticks_nmi to let the
405 * RCU grace-period handling know that the CPU is active.
407 void rcu_nmi_enter(void)
409 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
411 if (rdtp->dynticks_nmi_nesting == 0 &&
412 (atomic_read(&rdtp->dynticks) & 0x1))
414 rdtp->dynticks_nmi_nesting++;
415 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
416 atomic_inc(&rdtp->dynticks);
417 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
418 smp_mb__after_atomic_inc(); /* See above. */
419 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
423 * rcu_nmi_exit - inform RCU of exit from NMI context
425 * If the CPU was idle with dynamic ticks active, and there is no
426 * irq handler running, this updates rdtp->dynticks_nmi to let the
427 * RCU grace-period handling know that the CPU is no longer active.
429 void rcu_nmi_exit(void)
431 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
433 if (rdtp->dynticks_nmi_nesting == 0 ||
434 --rdtp->dynticks_nmi_nesting != 0)
436 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
437 smp_mb__before_atomic_inc(); /* See above. */
438 atomic_inc(&rdtp->dynticks);
439 smp_mb__after_atomic_inc(); /* Force delay to next write. */
440 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
444 * rcu_irq_enter - inform RCU of entry to hard irq context
446 * If the CPU was idle with dynamic ticks active, this updates the
447 * rdtp->dynticks to let the RCU handling know that the CPU is active.
449 void rcu_irq_enter(void)
455 * rcu_irq_exit - inform RCU of exit from hard irq context
457 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
458 * to put let the RCU handling be aware that the CPU is going back to idle
461 void rcu_irq_exit(void)
469 * Snapshot the specified CPU's dynticks counter so that we can later
470 * credit them with an implicit quiescent state. Return 1 if this CPU
471 * is in dynticks idle mode, which is an extended quiescent state.
473 static int dyntick_save_progress_counter(struct rcu_data *rdp)
475 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
480 * Return true if the specified CPU has passed through a quiescent
481 * state by virtue of being in or having passed through an dynticks
482 * idle state since the last call to dyntick_save_progress_counter()
485 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
490 curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
491 snap = (unsigned long)rdp->dynticks_snap;
494 * If the CPU passed through or entered a dynticks idle phase with
495 * no active irq/NMI handlers, then we can safely pretend that the CPU
496 * already acknowledged the request to pass through a quiescent
497 * state. Either way, that CPU cannot possibly be in an RCU
498 * read-side critical section that started before the beginning
499 * of the current RCU grace period.
501 if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
506 /* Go check for the CPU being offline. */
507 return rcu_implicit_offline_qs(rdp);
510 #endif /* #ifdef CONFIG_SMP */
512 #else /* #ifdef CONFIG_NO_HZ */
516 static int dyntick_save_progress_counter(struct rcu_data *rdp)
521 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
523 return rcu_implicit_offline_qs(rdp);
526 #endif /* #ifdef CONFIG_SMP */
528 #endif /* #else #ifdef CONFIG_NO_HZ */
530 int rcu_cpu_stall_suppress __read_mostly;
532 static void record_gp_stall_check_time(struct rcu_state *rsp)
534 rsp->gp_start = jiffies;
535 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
538 static void print_other_cpu_stall(struct rcu_state *rsp)
543 struct rcu_node *rnp = rcu_get_root(rsp);
545 /* Only let one CPU complain about others per time interval. */
547 raw_spin_lock_irqsave(&rnp->lock, flags);
548 delta = jiffies - rsp->jiffies_stall;
549 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
550 raw_spin_unlock_irqrestore(&rnp->lock, flags);
553 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
556 * Now rat on any tasks that got kicked up to the root rcu_node
557 * due to CPU offlining.
559 rcu_print_task_stall(rnp);
560 raw_spin_unlock_irqrestore(&rnp->lock, flags);
563 * OK, time to rat on our buddy...
564 * See Documentation/RCU/stallwarn.txt for info on how to debug
565 * RCU CPU stall warnings.
567 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
569 rcu_for_each_leaf_node(rsp, rnp) {
570 raw_spin_lock_irqsave(&rnp->lock, flags);
571 rcu_print_task_stall(rnp);
572 raw_spin_unlock_irqrestore(&rnp->lock, flags);
573 if (rnp->qsmask == 0)
575 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
576 if (rnp->qsmask & (1UL << cpu))
577 printk(" %d", rnp->grplo + cpu);
579 printk("} (detected by %d, t=%ld jiffies)\n",
580 smp_processor_id(), (long)(jiffies - rsp->gp_start));
581 trigger_all_cpu_backtrace();
583 /* If so configured, complain about tasks blocking the grace period. */
585 rcu_print_detail_task_stall(rsp);
587 force_quiescent_state(rsp, 0); /* Kick them all. */
590 static void print_cpu_stall(struct rcu_state *rsp)
593 struct rcu_node *rnp = rcu_get_root(rsp);
596 * OK, time to rat on ourselves...
597 * See Documentation/RCU/stallwarn.txt for info on how to debug
598 * RCU CPU stall warnings.
600 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
601 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
602 trigger_all_cpu_backtrace();
604 raw_spin_lock_irqsave(&rnp->lock, flags);
605 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
607 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
608 raw_spin_unlock_irqrestore(&rnp->lock, flags);
610 set_need_resched(); /* kick ourselves to get things going. */
613 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
617 struct rcu_node *rnp;
619 if (rcu_cpu_stall_suppress)
621 j = ACCESS_ONCE(jiffies);
622 js = ACCESS_ONCE(rsp->jiffies_stall);
624 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
626 /* We haven't checked in, so go dump stack. */
627 print_cpu_stall(rsp);
629 } else if (rcu_gp_in_progress(rsp) &&
630 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
632 /* They had a few time units to dump stack, so complain. */
633 print_other_cpu_stall(rsp);
637 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
639 rcu_cpu_stall_suppress = 1;
644 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
646 * Set the stall-warning timeout way off into the future, thus preventing
647 * any RCU CPU stall-warning messages from appearing in the current set of
650 * The caller must disable hard irqs.
652 void rcu_cpu_stall_reset(void)
654 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
655 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
656 rcu_preempt_stall_reset();
659 static struct notifier_block rcu_panic_block = {
660 .notifier_call = rcu_panic,
663 static void __init check_cpu_stall_init(void)
665 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
669 * Update CPU-local rcu_data state to record the newly noticed grace period.
670 * This is used both when we started the grace period and when we notice
671 * that someone else started the grace period. The caller must hold the
672 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
673 * and must have irqs disabled.
675 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
677 if (rdp->gpnum != rnp->gpnum) {
679 * If the current grace period is waiting for this CPU,
680 * set up to detect a quiescent state, otherwise don't
681 * go looking for one.
683 rdp->gpnum = rnp->gpnum;
684 if (rnp->qsmask & rdp->grpmask) {
686 rdp->passed_quiesc = 0;
692 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
695 struct rcu_node *rnp;
697 local_irq_save(flags);
699 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
700 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
701 local_irq_restore(flags);
704 __note_new_gpnum(rsp, rnp, rdp);
705 raw_spin_unlock_irqrestore(&rnp->lock, flags);
709 * Did someone else start a new RCU grace period start since we last
710 * checked? Update local state appropriately if so. Must be called
711 * on the CPU corresponding to rdp.
714 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
719 local_irq_save(flags);
720 if (rdp->gpnum != rsp->gpnum) {
721 note_new_gpnum(rsp, rdp);
724 local_irq_restore(flags);
729 * Advance this CPU's callbacks, but only if the current grace period
730 * has ended. This may be called only from the CPU to whom the rdp
731 * belongs. In addition, the corresponding leaf rcu_node structure's
732 * ->lock must be held by the caller, with irqs disabled.
735 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
737 /* Did another grace period end? */
738 if (rdp->completed != rnp->completed) {
740 /* Advance callbacks. No harm if list empty. */
741 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
742 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
743 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
745 /* Remember that we saw this grace-period completion. */
746 rdp->completed = rnp->completed;
749 * If we were in an extended quiescent state, we may have
750 * missed some grace periods that others CPUs handled on
751 * our behalf. Catch up with this state to avoid noting
752 * spurious new grace periods. If another grace period
753 * has started, then rnp->gpnum will have advanced, so
754 * we will detect this later on.
756 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
757 rdp->gpnum = rdp->completed;
760 * If RCU does not need a quiescent state from this CPU,
761 * then make sure that this CPU doesn't go looking for one.
763 if ((rnp->qsmask & rdp->grpmask) == 0)
769 * Advance this CPU's callbacks, but only if the current grace period
770 * has ended. This may be called only from the CPU to whom the rdp
774 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
777 struct rcu_node *rnp;
779 local_irq_save(flags);
781 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
782 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
783 local_irq_restore(flags);
786 __rcu_process_gp_end(rsp, rnp, rdp);
787 raw_spin_unlock_irqrestore(&rnp->lock, flags);
791 * Do per-CPU grace-period initialization for running CPU. The caller
792 * must hold the lock of the leaf rcu_node structure corresponding to
796 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
798 /* Prior grace period ended, so advance callbacks for current CPU. */
799 __rcu_process_gp_end(rsp, rnp, rdp);
802 * Because this CPU just now started the new grace period, we know
803 * that all of its callbacks will be covered by this upcoming grace
804 * period, even the ones that were registered arbitrarily recently.
805 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
807 * Other CPUs cannot be sure exactly when the grace period started.
808 * Therefore, their recently registered callbacks must pass through
809 * an additional RCU_NEXT_READY stage, so that they will be handled
810 * by the next RCU grace period.
812 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
813 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
815 /* Set state so that this CPU will detect the next quiescent state. */
816 __note_new_gpnum(rsp, rnp, rdp);
820 * Start a new RCU grace period if warranted, re-initializing the hierarchy
821 * in preparation for detecting the next grace period. The caller must hold
822 * the root node's ->lock, which is released before return. Hard irqs must
826 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
827 __releases(rcu_get_root(rsp)->lock)
829 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
830 struct rcu_node *rnp = rcu_get_root(rsp);
832 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
833 if (cpu_needs_another_gp(rsp, rdp))
834 rsp->fqs_need_gp = 1;
835 if (rnp->completed == rsp->completed) {
836 raw_spin_unlock_irqrestore(&rnp->lock, flags);
839 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
842 * Propagate new ->completed value to rcu_node structures
843 * so that other CPUs don't have to wait until the start
844 * of the next grace period to process their callbacks.
846 rcu_for_each_node_breadth_first(rsp, rnp) {
847 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
848 rnp->completed = rsp->completed;
849 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
851 local_irq_restore(flags);
855 /* Advance to a new grace period and initialize state. */
857 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
858 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
859 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
860 record_gp_stall_check_time(rsp);
862 /* Special-case the common single-level case. */
863 if (NUM_RCU_NODES == 1) {
864 rcu_preempt_check_blocked_tasks(rnp);
865 rnp->qsmask = rnp->qsmaskinit;
866 rnp->gpnum = rsp->gpnum;
867 rnp->completed = rsp->completed;
868 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
869 rcu_start_gp_per_cpu(rsp, rnp, rdp);
870 rcu_preempt_boost_start_gp(rnp);
871 raw_spin_unlock_irqrestore(&rnp->lock, flags);
875 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
878 /* Exclude any concurrent CPU-hotplug operations. */
879 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
882 * Set the quiescent-state-needed bits in all the rcu_node
883 * structures for all currently online CPUs in breadth-first
884 * order, starting from the root rcu_node structure. This
885 * operation relies on the layout of the hierarchy within the
886 * rsp->node[] array. Note that other CPUs will access only
887 * the leaves of the hierarchy, which still indicate that no
888 * grace period is in progress, at least until the corresponding
889 * leaf node has been initialized. In addition, we have excluded
890 * CPU-hotplug operations.
892 * Note that the grace period cannot complete until we finish
893 * the initialization process, as there will be at least one
894 * qsmask bit set in the root node until that time, namely the
895 * one corresponding to this CPU, due to the fact that we have
898 rcu_for_each_node_breadth_first(rsp, rnp) {
899 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
900 rcu_preempt_check_blocked_tasks(rnp);
901 rnp->qsmask = rnp->qsmaskinit;
902 rnp->gpnum = rsp->gpnum;
903 rnp->completed = rsp->completed;
904 if (rnp == rdp->mynode)
905 rcu_start_gp_per_cpu(rsp, rnp, rdp);
906 rcu_preempt_boost_start_gp(rnp);
907 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
910 rnp = rcu_get_root(rsp);
911 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
912 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
913 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
914 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
918 * Report a full set of quiescent states to the specified rcu_state
919 * data structure. This involves cleaning up after the prior grace
920 * period and letting rcu_start_gp() start up the next grace period
921 * if one is needed. Note that the caller must hold rnp->lock, as
922 * required by rcu_start_gp(), which will release it.
924 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
925 __releases(rcu_get_root(rsp)->lock)
927 unsigned long gp_duration;
929 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
932 * Ensure that all grace-period and pre-grace-period activity
933 * is seen before the assignment to rsp->completed.
935 smp_mb(); /* See above block comment. */
936 gp_duration = jiffies - rsp->gp_start;
937 if (gp_duration > rsp->gp_max)
938 rsp->gp_max = gp_duration;
939 rsp->completed = rsp->gpnum;
940 rsp->signaled = RCU_GP_IDLE;
941 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
945 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
946 * Allows quiescent states for a group of CPUs to be reported at one go
947 * to the specified rcu_node structure, though all the CPUs in the group
948 * must be represented by the same rcu_node structure (which need not be
949 * a leaf rcu_node structure, though it often will be). That structure's
950 * lock must be held upon entry, and it is released before return.
953 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
954 struct rcu_node *rnp, unsigned long flags)
955 __releases(rnp->lock)
957 struct rcu_node *rnp_c;
959 /* Walk up the rcu_node hierarchy. */
961 if (!(rnp->qsmask & mask)) {
963 /* Our bit has already been cleared, so done. */
964 raw_spin_unlock_irqrestore(&rnp->lock, flags);
967 rnp->qsmask &= ~mask;
968 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
970 /* Other bits still set at this level, so done. */
971 raw_spin_unlock_irqrestore(&rnp->lock, flags);
975 if (rnp->parent == NULL) {
977 /* No more levels. Exit loop holding root lock. */
981 raw_spin_unlock_irqrestore(&rnp->lock, flags);
984 raw_spin_lock_irqsave(&rnp->lock, flags);
985 WARN_ON_ONCE(rnp_c->qsmask);
989 * Get here if we are the last CPU to pass through a quiescent
990 * state for this grace period. Invoke rcu_report_qs_rsp()
991 * to clean up and start the next grace period if one is needed.
993 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
997 * Record a quiescent state for the specified CPU to that CPU's rcu_data
998 * structure. This must be either called from the specified CPU, or
999 * called when the specified CPU is known to be offline (and when it is
1000 * also known that no other CPU is concurrently trying to help the offline
1001 * CPU). The lastcomp argument is used to make sure we are still in the
1002 * grace period of interest. We don't want to end the current grace period
1003 * based on quiescent states detected in an earlier grace period!
1006 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
1008 unsigned long flags;
1010 struct rcu_node *rnp;
1013 raw_spin_lock_irqsave(&rnp->lock, flags);
1014 if (lastcomp != rnp->completed) {
1017 * Someone beat us to it for this grace period, so leave.
1018 * The race with GP start is resolved by the fact that we
1019 * hold the leaf rcu_node lock, so that the per-CPU bits
1020 * cannot yet be initialized -- so we would simply find our
1021 * CPU's bit already cleared in rcu_report_qs_rnp() if this
1024 rdp->passed_quiesc = 0; /* try again later! */
1025 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1028 mask = rdp->grpmask;
1029 if ((rnp->qsmask & mask) == 0) {
1030 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1032 rdp->qs_pending = 0;
1035 * This GP can't end until cpu checks in, so all of our
1036 * callbacks can be processed during the next GP.
1038 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1040 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1045 * Check to see if there is a new grace period of which this CPU
1046 * is not yet aware, and if so, set up local rcu_data state for it.
1047 * Otherwise, see if this CPU has just passed through its first
1048 * quiescent state for this grace period, and record that fact if so.
1051 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1053 /* If there is now a new grace period, record and return. */
1054 if (check_for_new_grace_period(rsp, rdp))
1058 * Does this CPU still need to do its part for current grace period?
1059 * If no, return and let the other CPUs do their part as well.
1061 if (!rdp->qs_pending)
1065 * Was there a quiescent state since the beginning of the grace
1066 * period? If no, then exit and wait for the next call.
1068 if (!rdp->passed_quiesc)
1072 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1075 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
1078 #ifdef CONFIG_HOTPLUG_CPU
1081 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1082 * Synchronization is not required because this function executes
1083 * in stop_machine() context.
1085 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1088 /* current DYING CPU is cleared in the cpu_online_mask */
1089 int receive_cpu = cpumask_any(cpu_online_mask);
1090 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1091 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1093 if (rdp->nxtlist == NULL)
1094 return; /* irqs disabled, so comparison is stable. */
1096 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1097 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1098 receive_rdp->qlen += rdp->qlen;
1099 receive_rdp->n_cbs_adopted += rdp->qlen;
1100 rdp->n_cbs_orphaned += rdp->qlen;
1102 rdp->nxtlist = NULL;
1103 for (i = 0; i < RCU_NEXT_SIZE; i++)
1104 rdp->nxttail[i] = &rdp->nxtlist;
1109 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1110 * and move all callbacks from the outgoing CPU to the current one.
1111 * There can only be one CPU hotplug operation at a time, so no other
1112 * CPU can be attempting to update rcu_cpu_kthread_task.
1114 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1116 unsigned long flags;
1118 int need_report = 0;
1119 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1120 struct rcu_node *rnp;
1122 rcu_stop_cpu_kthread(cpu);
1124 /* Exclude any attempts to start a new grace period. */
1125 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1127 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1128 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1129 mask = rdp->grpmask; /* rnp->grplo is constant. */
1131 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1132 rnp->qsmaskinit &= ~mask;
1133 if (rnp->qsmaskinit != 0) {
1134 if (rnp != rdp->mynode)
1135 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1138 if (rnp == rdp->mynode)
1139 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1141 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1142 mask = rnp->grpmask;
1144 } while (rnp != NULL);
1147 * We still hold the leaf rcu_node structure lock here, and
1148 * irqs are still disabled. The reason for this subterfuge is
1149 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1150 * held leads to deadlock.
1152 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1154 if (need_report & RCU_OFL_TASKS_NORM_GP)
1155 rcu_report_unblock_qs_rnp(rnp, flags);
1157 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1158 if (need_report & RCU_OFL_TASKS_EXP_GP)
1159 rcu_report_exp_rnp(rsp, rnp);
1160 rcu_node_kthread_setaffinity(rnp, -1);
1164 * Remove the specified CPU from the RCU hierarchy and move any pending
1165 * callbacks that it might have to the current CPU. This code assumes
1166 * that at least one CPU in the system will remain running at all times.
1167 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1169 static void rcu_offline_cpu(int cpu)
1171 __rcu_offline_cpu(cpu, &rcu_sched_state);
1172 __rcu_offline_cpu(cpu, &rcu_bh_state);
1173 rcu_preempt_offline_cpu(cpu);
1176 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1178 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1182 static void rcu_offline_cpu(int cpu)
1186 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1189 * Invoke any RCU callbacks that have made it to the end of their grace
1190 * period. Thottle as specified by rdp->blimit.
1192 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1194 unsigned long flags;
1195 struct rcu_head *next, *list, **tail;
1198 /* If no callbacks are ready, just return.*/
1199 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1200 trace_rcu_batch_start(0, 0);
1201 trace_rcu_batch_end(0);
1206 * Extract the list of ready callbacks, disabling to prevent
1207 * races with call_rcu() from interrupt handlers.
1209 local_irq_save(flags);
1211 trace_rcu_batch_start(rdp->qlen, bl);
1212 list = rdp->nxtlist;
1213 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1214 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1215 tail = rdp->nxttail[RCU_DONE_TAIL];
1216 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1217 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1218 rdp->nxttail[count] = &rdp->nxtlist;
1219 local_irq_restore(flags);
1221 /* Invoke callbacks. */
1226 debug_rcu_head_unqueue(list);
1227 __rcu_reclaim(list);
1233 local_irq_save(flags);
1234 trace_rcu_batch_end(count);
1236 /* Update count, and requeue any remaining callbacks. */
1238 rdp->n_cbs_invoked += count;
1240 *tail = rdp->nxtlist;
1241 rdp->nxtlist = list;
1242 for (count = 0; count < RCU_NEXT_SIZE; count++)
1243 if (&rdp->nxtlist == rdp->nxttail[count])
1244 rdp->nxttail[count] = tail;
1249 /* Reinstate batch limit if we have worked down the excess. */
1250 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1251 rdp->blimit = blimit;
1253 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1254 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1255 rdp->qlen_last_fqs_check = 0;
1256 rdp->n_force_qs_snap = rsp->n_force_qs;
1257 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1258 rdp->qlen_last_fqs_check = rdp->qlen;
1260 local_irq_restore(flags);
1262 /* Re-raise the RCU softirq if there are callbacks remaining. */
1263 if (cpu_has_callbacks_ready_to_invoke(rdp))
1268 * Check to see if this CPU is in a non-context-switch quiescent state
1269 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1270 * Also schedule the RCU softirq handler.
1272 * This function must be called with hardirqs disabled. It is normally
1273 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1274 * false, there is no point in invoking rcu_check_callbacks().
1276 void rcu_check_callbacks(int cpu, int user)
1279 (idle_cpu(cpu) && rcu_scheduler_active &&
1280 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1283 * Get here if this CPU took its interrupt from user
1284 * mode or from the idle loop, and if this is not a
1285 * nested interrupt. In this case, the CPU is in
1286 * a quiescent state, so note it.
1288 * No memory barrier is required here because both
1289 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1290 * variables that other CPUs neither access nor modify,
1291 * at least not while the corresponding CPU is online.
1297 } else if (!in_softirq()) {
1300 * Get here if this CPU did not take its interrupt from
1301 * softirq, in other words, if it is not interrupting
1302 * a rcu_bh read-side critical section. This is an _bh
1303 * critical section, so note it.
1308 rcu_preempt_check_callbacks(cpu);
1309 if (rcu_pending(cpu))
1316 * Scan the leaf rcu_node structures, processing dyntick state for any that
1317 * have not yet encountered a quiescent state, using the function specified.
1318 * Also initiate boosting for any threads blocked on the root rcu_node.
1320 * The caller must have suppressed start of new grace periods.
1322 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1326 unsigned long flags;
1328 struct rcu_node *rnp;
1330 rcu_for_each_leaf_node(rsp, rnp) {
1332 raw_spin_lock_irqsave(&rnp->lock, flags);
1333 if (!rcu_gp_in_progress(rsp)) {
1334 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1337 if (rnp->qsmask == 0) {
1338 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1343 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1344 if ((rnp->qsmask & bit) != 0 &&
1345 f(per_cpu_ptr(rsp->rda, cpu)))
1350 /* rcu_report_qs_rnp() releases rnp->lock. */
1351 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1354 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1356 rnp = rcu_get_root(rsp);
1357 if (rnp->qsmask == 0) {
1358 raw_spin_lock_irqsave(&rnp->lock, flags);
1359 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1364 * Force quiescent states on reluctant CPUs, and also detect which
1365 * CPUs are in dyntick-idle mode.
1367 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1369 unsigned long flags;
1370 struct rcu_node *rnp = rcu_get_root(rsp);
1372 if (!rcu_gp_in_progress(rsp))
1373 return; /* No grace period in progress, nothing to force. */
1374 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1375 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1376 return; /* Someone else is already on the job. */
1378 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1379 goto unlock_fqs_ret; /* no emergency and done recently. */
1381 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1382 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1383 if(!rcu_gp_in_progress(rsp)) {
1384 rsp->n_force_qs_ngp++;
1385 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1386 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1388 rsp->fqs_active = 1;
1389 switch (rsp->signaled) {
1393 break; /* grace period idle or initializing, ignore. */
1395 case RCU_SAVE_DYNTICK:
1396 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1397 break; /* So gcc recognizes the dead code. */
1399 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1401 /* Record dyntick-idle state. */
1402 force_qs_rnp(rsp, dyntick_save_progress_counter);
1403 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1404 if (rcu_gp_in_progress(rsp))
1405 rsp->signaled = RCU_FORCE_QS;
1410 /* Check dyntick-idle state, send IPI to laggarts. */
1411 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1412 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1414 /* Leave state in case more forcing is required. */
1416 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1419 rsp->fqs_active = 0;
1420 if (rsp->fqs_need_gp) {
1421 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1422 rsp->fqs_need_gp = 0;
1423 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1426 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1428 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1431 #else /* #ifdef CONFIG_SMP */
1433 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1438 #endif /* #else #ifdef CONFIG_SMP */
1441 * This does the RCU processing work from softirq context for the
1442 * specified rcu_state and rcu_data structures. This may be called
1443 * only from the CPU to whom the rdp belongs.
1446 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1448 unsigned long flags;
1450 WARN_ON_ONCE(rdp->beenonline == 0);
1453 * If an RCU GP has gone long enough, go check for dyntick
1454 * idle CPUs and, if needed, send resched IPIs.
1456 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1457 force_quiescent_state(rsp, 1);
1460 * Advance callbacks in response to end of earlier grace
1461 * period that some other CPU ended.
1463 rcu_process_gp_end(rsp, rdp);
1465 /* Update RCU state based on any recent quiescent states. */
1466 rcu_check_quiescent_state(rsp, rdp);
1468 /* Does this CPU require a not-yet-started grace period? */
1469 if (cpu_needs_another_gp(rsp, rdp)) {
1470 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1471 rcu_start_gp(rsp, flags); /* releases above lock */
1474 /* If there are callbacks ready, invoke them. */
1475 if (cpu_has_callbacks_ready_to_invoke(rdp))
1476 invoke_rcu_callbacks(rsp, rdp);
1480 * Do softirq processing for the current CPU.
1482 static void rcu_process_callbacks(struct softirq_action *unused)
1484 __rcu_process_callbacks(&rcu_sched_state,
1485 &__get_cpu_var(rcu_sched_data));
1486 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1487 rcu_preempt_process_callbacks();
1489 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1490 rcu_needs_cpu_flush();
1494 * Wake up the current CPU's kthread. This replaces raise_softirq()
1495 * in earlier versions of RCU. Note that because we are running on
1496 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1497 * cannot disappear out from under us.
1499 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1501 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1503 if (likely(!rsp->boost)) {
1504 rcu_do_batch(rsp, rdp);
1507 invoke_rcu_callbacks_kthread();
1510 static void invoke_rcu_core(void)
1512 raise_softirq(RCU_SOFTIRQ);
1516 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1517 struct rcu_state *rsp)
1519 unsigned long flags;
1520 struct rcu_data *rdp;
1522 debug_rcu_head_queue(head);
1526 smp_mb(); /* Ensure RCU update seen before callback registry. */
1529 * Opportunistically note grace-period endings and beginnings.
1530 * Note that we might see a beginning right after we see an
1531 * end, but never vice versa, since this CPU has to pass through
1532 * a quiescent state betweentimes.
1534 local_irq_save(flags);
1535 rdp = this_cpu_ptr(rsp->rda);
1537 /* Add the callback to our list. */
1538 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1539 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1542 /* If interrupts were disabled, don't dive into RCU core. */
1543 if (irqs_disabled_flags(flags)) {
1544 local_irq_restore(flags);
1549 * Force the grace period if too many callbacks or too long waiting.
1550 * Enforce hysteresis, and don't invoke force_quiescent_state()
1551 * if some other CPU has recently done so. Also, don't bother
1552 * invoking force_quiescent_state() if the newly enqueued callback
1553 * is the only one waiting for a grace period to complete.
1555 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1557 /* Are we ignoring a completed grace period? */
1558 rcu_process_gp_end(rsp, rdp);
1559 check_for_new_grace_period(rsp, rdp);
1561 /* Start a new grace period if one not already started. */
1562 if (!rcu_gp_in_progress(rsp)) {
1563 unsigned long nestflag;
1564 struct rcu_node *rnp_root = rcu_get_root(rsp);
1566 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1567 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1569 /* Give the grace period a kick. */
1570 rdp->blimit = LONG_MAX;
1571 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1572 *rdp->nxttail[RCU_DONE_TAIL] != head)
1573 force_quiescent_state(rsp, 0);
1574 rdp->n_force_qs_snap = rsp->n_force_qs;
1575 rdp->qlen_last_fqs_check = rdp->qlen;
1577 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1578 force_quiescent_state(rsp, 1);
1579 local_irq_restore(flags);
1583 * Queue an RCU-sched callback for invocation after a grace period.
1585 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1587 __call_rcu(head, func, &rcu_sched_state);
1589 EXPORT_SYMBOL_GPL(call_rcu_sched);
1592 * Queue an RCU for invocation after a quicker grace period.
1594 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1596 __call_rcu(head, func, &rcu_bh_state);
1598 EXPORT_SYMBOL_GPL(call_rcu_bh);
1601 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1603 * Control will return to the caller some time after a full rcu-sched
1604 * grace period has elapsed, in other words after all currently executing
1605 * rcu-sched read-side critical sections have completed. These read-side
1606 * critical sections are delimited by rcu_read_lock_sched() and
1607 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1608 * local_irq_disable(), and so on may be used in place of
1609 * rcu_read_lock_sched().
1611 * This means that all preempt_disable code sequences, including NMI and
1612 * hardware-interrupt handlers, in progress on entry will have completed
1613 * before this primitive returns. However, this does not guarantee that
1614 * softirq handlers will have completed, since in some kernels, these
1615 * handlers can run in process context, and can block.
1617 * This primitive provides the guarantees made by the (now removed)
1618 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1619 * guarantees that rcu_read_lock() sections will have completed.
1620 * In "classic RCU", these two guarantees happen to be one and
1621 * the same, but can differ in realtime RCU implementations.
1623 void synchronize_sched(void)
1625 if (rcu_blocking_is_gp())
1627 wait_rcu_gp(call_rcu_sched);
1629 EXPORT_SYMBOL_GPL(synchronize_sched);
1632 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1634 * Control will return to the caller some time after a full rcu_bh grace
1635 * period has elapsed, in other words after all currently executing rcu_bh
1636 * read-side critical sections have completed. RCU read-side critical
1637 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1638 * and may be nested.
1640 void synchronize_rcu_bh(void)
1642 if (rcu_blocking_is_gp())
1644 wait_rcu_gp(call_rcu_bh);
1646 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1649 * Check to see if there is any immediate RCU-related work to be done
1650 * by the current CPU, for the specified type of RCU, returning 1 if so.
1651 * The checks are in order of increasing expense: checks that can be
1652 * carried out against CPU-local state are performed first. However,
1653 * we must check for CPU stalls first, else we might not get a chance.
1655 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1657 struct rcu_node *rnp = rdp->mynode;
1659 rdp->n_rcu_pending++;
1661 /* Check for CPU stalls, if enabled. */
1662 check_cpu_stall(rsp, rdp);
1664 /* Is the RCU core waiting for a quiescent state from this CPU? */
1665 if (rdp->qs_pending && !rdp->passed_quiesc) {
1668 * If force_quiescent_state() coming soon and this CPU
1669 * needs a quiescent state, and this is either RCU-sched
1670 * or RCU-bh, force a local reschedule.
1672 rdp->n_rp_qs_pending++;
1673 if (!rdp->preemptible &&
1674 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1677 } else if (rdp->qs_pending && rdp->passed_quiesc) {
1678 rdp->n_rp_report_qs++;
1682 /* Does this CPU have callbacks ready to invoke? */
1683 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1684 rdp->n_rp_cb_ready++;
1688 /* Has RCU gone idle with this CPU needing another grace period? */
1689 if (cpu_needs_another_gp(rsp, rdp)) {
1690 rdp->n_rp_cpu_needs_gp++;
1694 /* Has another RCU grace period completed? */
1695 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1696 rdp->n_rp_gp_completed++;
1700 /* Has a new RCU grace period started? */
1701 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1702 rdp->n_rp_gp_started++;
1706 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1707 if (rcu_gp_in_progress(rsp) &&
1708 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1709 rdp->n_rp_need_fqs++;
1714 rdp->n_rp_need_nothing++;
1719 * Check to see if there is any immediate RCU-related work to be done
1720 * by the current CPU, returning 1 if so. This function is part of the
1721 * RCU implementation; it is -not- an exported member of the RCU API.
1723 static int rcu_pending(int cpu)
1725 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1726 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1727 rcu_preempt_pending(cpu);
1731 * Check to see if any future RCU-related work will need to be done
1732 * by the current CPU, even if none need be done immediately, returning
1735 static int rcu_needs_cpu_quick_check(int cpu)
1737 /* RCU callbacks either ready or pending? */
1738 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1739 per_cpu(rcu_bh_data, cpu).nxtlist ||
1740 rcu_preempt_needs_cpu(cpu);
1743 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1744 static atomic_t rcu_barrier_cpu_count;
1745 static DEFINE_MUTEX(rcu_barrier_mutex);
1746 static struct completion rcu_barrier_completion;
1748 static void rcu_barrier_callback(struct rcu_head *notused)
1750 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1751 complete(&rcu_barrier_completion);
1755 * Called with preemption disabled, and from cross-cpu IRQ context.
1757 static void rcu_barrier_func(void *type)
1759 int cpu = smp_processor_id();
1760 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1761 void (*call_rcu_func)(struct rcu_head *head,
1762 void (*func)(struct rcu_head *head));
1764 atomic_inc(&rcu_barrier_cpu_count);
1765 call_rcu_func = type;
1766 call_rcu_func(head, rcu_barrier_callback);
1770 * Orchestrate the specified type of RCU barrier, waiting for all
1771 * RCU callbacks of the specified type to complete.
1773 static void _rcu_barrier(struct rcu_state *rsp,
1774 void (*call_rcu_func)(struct rcu_head *head,
1775 void (*func)(struct rcu_head *head)))
1777 BUG_ON(in_interrupt());
1778 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1779 mutex_lock(&rcu_barrier_mutex);
1780 init_completion(&rcu_barrier_completion);
1782 * Initialize rcu_barrier_cpu_count to 1, then invoke
1783 * rcu_barrier_func() on each CPU, so that each CPU also has
1784 * incremented rcu_barrier_cpu_count. Only then is it safe to
1785 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1786 * might complete its grace period before all of the other CPUs
1787 * did their increment, causing this function to return too
1788 * early. Note that on_each_cpu() disables irqs, which prevents
1789 * any CPUs from coming online or going offline until each online
1790 * CPU has queued its RCU-barrier callback.
1792 atomic_set(&rcu_barrier_cpu_count, 1);
1793 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1794 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1795 complete(&rcu_barrier_completion);
1796 wait_for_completion(&rcu_barrier_completion);
1797 mutex_unlock(&rcu_barrier_mutex);
1801 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1803 void rcu_barrier_bh(void)
1805 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1807 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1810 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1812 void rcu_barrier_sched(void)
1814 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1816 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1819 * Do boot-time initialization of a CPU's per-CPU RCU data.
1822 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1824 unsigned long flags;
1826 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1827 struct rcu_node *rnp = rcu_get_root(rsp);
1829 /* Set up local state, ensuring consistent view of global state. */
1830 raw_spin_lock_irqsave(&rnp->lock, flags);
1831 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1832 rdp->nxtlist = NULL;
1833 for (i = 0; i < RCU_NEXT_SIZE; i++)
1834 rdp->nxttail[i] = &rdp->nxtlist;
1837 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1838 #endif /* #ifdef CONFIG_NO_HZ */
1840 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1844 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1845 * offline event can be happening at a given time. Note also that we
1846 * can accept some slop in the rsp->completed access due to the fact
1847 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1849 static void __cpuinit
1850 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
1852 unsigned long flags;
1854 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1855 struct rcu_node *rnp = rcu_get_root(rsp);
1857 /* Set up local state, ensuring consistent view of global state. */
1858 raw_spin_lock_irqsave(&rnp->lock, flags);
1859 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1860 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1861 rdp->beenonline = 1; /* We have now been online. */
1862 rdp->preemptible = preemptible;
1863 rdp->qlen_last_fqs_check = 0;
1864 rdp->n_force_qs_snap = rsp->n_force_qs;
1865 rdp->blimit = blimit;
1866 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1869 * A new grace period might start here. If so, we won't be part
1870 * of it, but that is OK, as we are currently in a quiescent state.
1873 /* Exclude any attempts to start a new GP on large systems. */
1874 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1876 /* Add CPU to rcu_node bitmasks. */
1878 mask = rdp->grpmask;
1880 /* Exclude any attempts to start a new GP on small systems. */
1881 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1882 rnp->qsmaskinit |= mask;
1883 mask = rnp->grpmask;
1884 if (rnp == rdp->mynode) {
1885 rdp->gpnum = rnp->completed; /* if GP in progress... */
1886 rdp->completed = rnp->completed;
1887 rdp->passed_quiesc_completed = rnp->completed - 1;
1889 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1891 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1893 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1896 static void __cpuinit rcu_prepare_cpu(int cpu)
1898 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1899 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1900 rcu_preempt_init_percpu_data(cpu);
1904 * Handle CPU online/offline notification events.
1906 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1907 unsigned long action, void *hcpu)
1909 long cpu = (long)hcpu;
1910 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1911 struct rcu_node *rnp = rdp->mynode;
1914 case CPU_UP_PREPARE:
1915 case CPU_UP_PREPARE_FROZEN:
1916 rcu_prepare_cpu(cpu);
1917 rcu_prepare_kthreads(cpu);
1920 case CPU_DOWN_FAILED:
1921 rcu_node_kthread_setaffinity(rnp, -1);
1922 rcu_cpu_kthread_setrt(cpu, 1);
1924 case CPU_DOWN_PREPARE:
1925 rcu_node_kthread_setaffinity(rnp, cpu);
1926 rcu_cpu_kthread_setrt(cpu, 0);
1929 case CPU_DYING_FROZEN:
1931 * The whole machine is "stopped" except this CPU, so we can
1932 * touch any data without introducing corruption. We send the
1933 * dying CPU's callbacks to an arbitrarily chosen online CPU.
1935 rcu_send_cbs_to_online(&rcu_bh_state);
1936 rcu_send_cbs_to_online(&rcu_sched_state);
1937 rcu_preempt_send_cbs_to_online();
1940 case CPU_DEAD_FROZEN:
1941 case CPU_UP_CANCELED:
1942 case CPU_UP_CANCELED_FROZEN:
1943 rcu_offline_cpu(cpu);
1952 * This function is invoked towards the end of the scheduler's initialization
1953 * process. Before this is called, the idle task might contain
1954 * RCU read-side critical sections (during which time, this idle
1955 * task is booting the system). After this function is called, the
1956 * idle tasks are prohibited from containing RCU read-side critical
1957 * sections. This function also enables RCU lockdep checking.
1959 void rcu_scheduler_starting(void)
1961 WARN_ON(num_online_cpus() != 1);
1962 WARN_ON(nr_context_switches() > 0);
1963 rcu_scheduler_active = 1;
1967 * Compute the per-level fanout, either using the exact fanout specified
1968 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1970 #ifdef CONFIG_RCU_FANOUT_EXACT
1971 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1975 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
1976 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1977 rsp->levelspread[0] = RCU_FANOUT_LEAF;
1979 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1980 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1987 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1988 ccur = rsp->levelcnt[i];
1989 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1993 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1996 * Helper function for rcu_init() that initializes one rcu_state structure.
1998 static void __init rcu_init_one(struct rcu_state *rsp,
1999 struct rcu_data __percpu *rda)
2001 static char *buf[] = { "rcu_node_level_0",
2004 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2008 struct rcu_node *rnp;
2010 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2012 /* Initialize the level-tracking arrays. */
2014 for (i = 1; i < NUM_RCU_LVLS; i++)
2015 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2016 rcu_init_levelspread(rsp);
2018 /* Initialize the elements themselves, starting from the leaves. */
2020 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2021 cpustride *= rsp->levelspread[i];
2022 rnp = rsp->level[i];
2023 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2024 raw_spin_lock_init(&rnp->lock);
2025 lockdep_set_class_and_name(&rnp->lock,
2026 &rcu_node_class[i], buf[i]);
2029 rnp->qsmaskinit = 0;
2030 rnp->grplo = j * cpustride;
2031 rnp->grphi = (j + 1) * cpustride - 1;
2032 if (rnp->grphi >= NR_CPUS)
2033 rnp->grphi = NR_CPUS - 1;
2039 rnp->grpnum = j % rsp->levelspread[i - 1];
2040 rnp->grpmask = 1UL << rnp->grpnum;
2041 rnp->parent = rsp->level[i - 1] +
2042 j / rsp->levelspread[i - 1];
2045 INIT_LIST_HEAD(&rnp->blkd_tasks);
2050 rnp = rsp->level[NUM_RCU_LVLS - 1];
2051 for_each_possible_cpu(i) {
2052 while (i > rnp->grphi)
2054 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2055 rcu_boot_init_percpu_data(i, rsp);
2059 void __init rcu_init(void)
2063 rcu_bootup_announce();
2064 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2065 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2066 __rcu_init_preempt();
2067 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2070 * We don't need protection against CPU-hotplug here because
2071 * this is called early in boot, before either interrupts
2072 * or the scheduler are operational.
2074 cpu_notifier(rcu_cpu_notify, 0);
2075 for_each_online_cpu(cpu)
2076 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2077 check_cpu_stall_init();
2080 #include "rcutree_plugin.h"