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/export.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##_state.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .fqs_state = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched);
83 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
85 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh);
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);
135 * Track the rcutorture test sequence number and the update version
136 * number within a given test. The rcutorture_testseq is incremented
137 * on every rcutorture module load and unload, so has an odd value
138 * when a test is running. The rcutorture_vernum is set to zero
139 * when rcutorture starts and is incremented on each rcutorture update.
140 * These variables enable correlating rcutorture output with the
141 * RCU tracing information.
143 unsigned long rcutorture_testseq;
144 unsigned long rcutorture_vernum;
147 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
148 * permit this function to be invoked without holding the root rcu_node
149 * structure's ->lock, but of course results can be subject to change.
151 static int rcu_gp_in_progress(struct rcu_state *rsp)
153 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
157 * Note a quiescent state. Because we do not need to know
158 * how many quiescent states passed, just if there was at least
159 * one since the start of the grace period, this just sets a flag.
160 * The caller must have disabled preemption.
162 void rcu_sched_qs(int cpu)
164 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
166 rdp->passed_quiesce_gpnum = rdp->gpnum;
168 if (rdp->passed_quiesce == 0)
169 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
170 rdp->passed_quiesce = 1;
173 void rcu_bh_qs(int cpu)
175 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
177 rdp->passed_quiesce_gpnum = rdp->gpnum;
179 if (rdp->passed_quiesce == 0)
180 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
181 rdp->passed_quiesce = 1;
185 * Note a context switch. This is a quiescent state for RCU-sched,
186 * and requires special handling for preemptible RCU.
187 * The caller must have disabled preemption.
189 void rcu_note_context_switch(int cpu)
191 trace_rcu_utilization("Start context switch");
193 rcu_preempt_note_context_switch(cpu);
194 trace_rcu_utilization("End context switch");
196 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
198 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
199 .dynticks_nesting = DYNTICK_TASK_NESTING,
200 .dynticks = ATOMIC_INIT(1),
203 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
204 static int qhimark = 10000; /* If this many pending, ignore blimit. */
205 static int qlowmark = 100; /* Once only this many pending, use blimit. */
207 module_param(blimit, int, 0);
208 module_param(qhimark, int, 0);
209 module_param(qlowmark, int, 0);
211 int rcu_cpu_stall_suppress __read_mostly;
212 module_param(rcu_cpu_stall_suppress, int, 0644);
214 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
215 static int rcu_pending(int cpu);
218 * Return the number of RCU-sched batches processed thus far for debug & stats.
220 long rcu_batches_completed_sched(void)
222 return rcu_sched_state.completed;
224 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
227 * Return the number of RCU BH batches processed thus far for debug & stats.
229 long rcu_batches_completed_bh(void)
231 return rcu_bh_state.completed;
233 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
236 * Force a quiescent state for RCU BH.
238 void rcu_bh_force_quiescent_state(void)
240 force_quiescent_state(&rcu_bh_state, 0);
242 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
245 * Record the number of times rcutorture tests have been initiated and
246 * terminated. This information allows the debugfs tracing stats to be
247 * correlated to the rcutorture messages, even when the rcutorture module
248 * is being repeatedly loaded and unloaded. In other words, we cannot
249 * store this state in rcutorture itself.
251 void rcutorture_record_test_transition(void)
253 rcutorture_testseq++;
254 rcutorture_vernum = 0;
256 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
259 * Record the number of writer passes through the current rcutorture test.
260 * This is also used to correlate debugfs tracing stats with the rcutorture
263 void rcutorture_record_progress(unsigned long vernum)
267 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
270 * Force a quiescent state for RCU-sched.
272 void rcu_sched_force_quiescent_state(void)
274 force_quiescent_state(&rcu_sched_state, 0);
276 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
279 * Does the CPU have callbacks ready to be invoked?
282 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
284 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
288 * Does the current CPU require a yet-as-unscheduled grace period?
291 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
293 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
297 * Return the root node of the specified rcu_state structure.
299 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
301 return &rsp->node[0];
307 * If the specified CPU is offline, tell the caller that it is in
308 * a quiescent state. Otherwise, whack it with a reschedule IPI.
309 * Grace periods can end up waiting on an offline CPU when that
310 * CPU is in the process of coming online -- it will be added to the
311 * rcu_node bitmasks before it actually makes it online. The same thing
312 * can happen while a CPU is in the process of coming online. Because this
313 * race is quite rare, we check for it after detecting that the grace
314 * period has been delayed rather than checking each and every CPU
315 * each and every time we start a new grace period.
317 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
320 * If the CPU is offline, it is in a quiescent state. We can
321 * trust its state not to change because interrupts are disabled.
323 if (cpu_is_offline(rdp->cpu)) {
324 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
330 * The CPU is online, so send it a reschedule IPI. This forces
331 * it through the scheduler, and (inefficiently) also handles cases
332 * where idle loops fail to inform RCU about the CPU being idle.
334 if (rdp->cpu != smp_processor_id())
335 smp_send_reschedule(rdp->cpu);
342 #endif /* #ifdef CONFIG_SMP */
345 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
347 * If the new value of the ->dynticks_nesting counter now is zero,
348 * we really have entered idle, and must do the appropriate accounting.
349 * The caller must have disabled interrupts.
351 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
353 trace_rcu_dyntick("Start", oldval, 0);
354 if (!is_idle_task(current)) {
355 struct task_struct *idle = idle_task(smp_processor_id());
357 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
358 ftrace_dump(DUMP_ALL);
359 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
360 current->pid, current->comm,
361 idle->pid, idle->comm); /* must be idle task! */
363 rcu_prepare_for_idle(smp_processor_id());
364 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
365 smp_mb__before_atomic_inc(); /* See above. */
366 atomic_inc(&rdtp->dynticks);
367 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
368 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
372 * rcu_idle_enter - inform RCU that current CPU is entering idle
374 * Enter idle mode, in other words, -leave- the mode in which RCU
375 * read-side critical sections can occur. (Though RCU read-side
376 * critical sections can occur in irq handlers in idle, a possibility
377 * handled by irq_enter() and irq_exit().)
379 * We crowbar the ->dynticks_nesting field to zero to allow for
380 * the possibility of usermode upcalls having messed up our count
381 * of interrupt nesting level during the prior busy period.
383 void rcu_idle_enter(void)
387 struct rcu_dynticks *rdtp;
389 local_irq_save(flags);
390 rdtp = &__get_cpu_var(rcu_dynticks);
391 oldval = rdtp->dynticks_nesting;
392 rdtp->dynticks_nesting = 0;
393 rcu_idle_enter_common(rdtp, oldval);
394 local_irq_restore(flags);
398 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
400 * Exit from an interrupt handler, which might possibly result in entering
401 * idle mode, in other words, leaving the mode in which read-side critical
402 * sections can occur.
404 * This code assumes that the idle loop never does anything that might
405 * result in unbalanced calls to irq_enter() and irq_exit(). If your
406 * architecture violates this assumption, RCU will give you what you
407 * deserve, good and hard. But very infrequently and irreproducibly.
409 * Use things like work queues to work around this limitation.
411 * You have been warned.
413 void rcu_irq_exit(void)
417 struct rcu_dynticks *rdtp;
419 local_irq_save(flags);
420 rdtp = &__get_cpu_var(rcu_dynticks);
421 oldval = rdtp->dynticks_nesting;
422 rdtp->dynticks_nesting--;
423 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
424 if (rdtp->dynticks_nesting)
425 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
427 rcu_idle_enter_common(rdtp, oldval);
428 local_irq_restore(flags);
432 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
434 * If the new value of the ->dynticks_nesting counter was previously zero,
435 * we really have exited idle, and must do the appropriate accounting.
436 * The caller must have disabled interrupts.
438 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
440 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
441 atomic_inc(&rdtp->dynticks);
442 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
443 smp_mb__after_atomic_inc(); /* See above. */
444 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
445 rcu_cleanup_after_idle(smp_processor_id());
446 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
447 if (!is_idle_task(current)) {
448 struct task_struct *idle = idle_task(smp_processor_id());
450 trace_rcu_dyntick("Error on exit: not idle task",
451 oldval, rdtp->dynticks_nesting);
452 ftrace_dump(DUMP_ALL);
453 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
454 current->pid, current->comm,
455 idle->pid, idle->comm); /* must be idle task! */
460 * rcu_idle_exit - inform RCU that current CPU is leaving idle
462 * Exit idle mode, in other words, -enter- the mode in which RCU
463 * read-side critical sections can occur.
465 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NESTING to
466 * allow for the possibility of usermode upcalls messing up our count
467 * of interrupt nesting level during the busy period that is just
470 void rcu_idle_exit(void)
473 struct rcu_dynticks *rdtp;
476 local_irq_save(flags);
477 rdtp = &__get_cpu_var(rcu_dynticks);
478 oldval = rdtp->dynticks_nesting;
479 WARN_ON_ONCE(oldval != 0);
480 rdtp->dynticks_nesting = DYNTICK_TASK_NESTING;
481 rcu_idle_exit_common(rdtp, oldval);
482 local_irq_restore(flags);
486 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
488 * Enter an interrupt handler, which might possibly result in exiting
489 * idle mode, in other words, entering the mode in which read-side critical
490 * sections can occur.
492 * Note that the Linux kernel is fully capable of entering an interrupt
493 * handler that it never exits, for example when doing upcalls to
494 * user mode! This code assumes that the idle loop never does upcalls to
495 * user mode. If your architecture does do upcalls from the idle loop (or
496 * does anything else that results in unbalanced calls to the irq_enter()
497 * and irq_exit() functions), RCU will give you what you deserve, good
498 * and hard. But very infrequently and irreproducibly.
500 * Use things like work queues to work around this limitation.
502 * You have been warned.
504 void rcu_irq_enter(void)
507 struct rcu_dynticks *rdtp;
510 local_irq_save(flags);
511 rdtp = &__get_cpu_var(rcu_dynticks);
512 oldval = rdtp->dynticks_nesting;
513 rdtp->dynticks_nesting++;
514 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
516 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
518 rcu_idle_exit_common(rdtp, oldval);
519 local_irq_restore(flags);
523 * rcu_nmi_enter - inform RCU of entry to NMI context
525 * If the CPU was idle with dynamic ticks active, and there is no
526 * irq handler running, this updates rdtp->dynticks_nmi to let the
527 * RCU grace-period handling know that the CPU is active.
529 void rcu_nmi_enter(void)
531 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
533 if (rdtp->dynticks_nmi_nesting == 0 &&
534 (atomic_read(&rdtp->dynticks) & 0x1))
536 rdtp->dynticks_nmi_nesting++;
537 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
538 atomic_inc(&rdtp->dynticks);
539 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
540 smp_mb__after_atomic_inc(); /* See above. */
541 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
545 * rcu_nmi_exit - inform RCU of exit from NMI context
547 * If the CPU was idle with dynamic ticks active, and there is no
548 * irq handler running, this updates rdtp->dynticks_nmi to let the
549 * RCU grace-period handling know that the CPU is no longer active.
551 void rcu_nmi_exit(void)
553 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
555 if (rdtp->dynticks_nmi_nesting == 0 ||
556 --rdtp->dynticks_nmi_nesting != 0)
558 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
559 smp_mb__before_atomic_inc(); /* See above. */
560 atomic_inc(&rdtp->dynticks);
561 smp_mb__after_atomic_inc(); /* Force delay to next write. */
562 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
565 #ifdef CONFIG_PROVE_RCU
568 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
570 * If the current CPU is in its idle loop and is neither in an interrupt
571 * or NMI handler, return true.
573 int rcu_is_cpu_idle(void)
578 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
582 EXPORT_SYMBOL(rcu_is_cpu_idle);
584 #endif /* #ifdef CONFIG_PROVE_RCU */
587 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
589 * If the current CPU is idle or running at a first-level (not nested)
590 * interrupt from idle, return true. The caller must have at least
591 * disabled preemption.
593 int rcu_is_cpu_rrupt_from_idle(void)
595 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
601 * Snapshot the specified CPU's dynticks counter so that we can later
602 * credit them with an implicit quiescent state. Return 1 if this CPU
603 * is in dynticks idle mode, which is an extended quiescent state.
605 static int dyntick_save_progress_counter(struct rcu_data *rdp)
607 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
608 return (rdp->dynticks_snap & 0x1) == 0;
612 * Return true if the specified CPU has passed through a quiescent
613 * state by virtue of being in or having passed through an dynticks
614 * idle state since the last call to dyntick_save_progress_counter()
617 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
622 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
623 snap = (unsigned int)rdp->dynticks_snap;
626 * If the CPU passed through or entered a dynticks idle phase with
627 * no active irq/NMI handlers, then we can safely pretend that the CPU
628 * already acknowledged the request to pass through a quiescent
629 * state. Either way, that CPU cannot possibly be in an RCU
630 * read-side critical section that started before the beginning
631 * of the current RCU grace period.
633 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
634 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
639 /* Go check for the CPU being offline. */
640 return rcu_implicit_offline_qs(rdp);
643 #endif /* #ifdef CONFIG_SMP */
645 static void record_gp_stall_check_time(struct rcu_state *rsp)
647 rsp->gp_start = jiffies;
648 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
651 static void print_other_cpu_stall(struct rcu_state *rsp)
657 struct rcu_node *rnp = rcu_get_root(rsp);
659 /* Only let one CPU complain about others per time interval. */
661 raw_spin_lock_irqsave(&rnp->lock, flags);
662 delta = jiffies - rsp->jiffies_stall;
663 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
664 raw_spin_unlock_irqrestore(&rnp->lock, flags);
667 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
670 * Now rat on any tasks that got kicked up to the root rcu_node
671 * due to CPU offlining.
673 ndetected = rcu_print_task_stall(rnp);
674 raw_spin_unlock_irqrestore(&rnp->lock, flags);
677 * OK, time to rat on our buddy...
678 * See Documentation/RCU/stallwarn.txt for info on how to debug
679 * RCU CPU stall warnings.
681 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
683 rcu_for_each_leaf_node(rsp, rnp) {
684 raw_spin_lock_irqsave(&rnp->lock, flags);
685 ndetected += rcu_print_task_stall(rnp);
686 raw_spin_unlock_irqrestore(&rnp->lock, flags);
687 if (rnp->qsmask == 0)
689 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
690 if (rnp->qsmask & (1UL << cpu)) {
691 printk(" %d", rnp->grplo + cpu);
695 printk("} (detected by %d, t=%ld jiffies)\n",
696 smp_processor_id(), (long)(jiffies - rsp->gp_start));
698 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
699 else if (!trigger_all_cpu_backtrace())
702 /* If so configured, complain about tasks blocking the grace period. */
704 rcu_print_detail_task_stall(rsp);
706 force_quiescent_state(rsp, 0); /* Kick them all. */
709 static void print_cpu_stall(struct rcu_state *rsp)
712 struct rcu_node *rnp = rcu_get_root(rsp);
715 * OK, time to rat on ourselves...
716 * See Documentation/RCU/stallwarn.txt for info on how to debug
717 * RCU CPU stall warnings.
719 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
720 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
721 if (!trigger_all_cpu_backtrace())
724 raw_spin_lock_irqsave(&rnp->lock, flags);
725 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
727 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
728 raw_spin_unlock_irqrestore(&rnp->lock, flags);
730 set_need_resched(); /* kick ourselves to get things going. */
733 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
737 struct rcu_node *rnp;
739 if (rcu_cpu_stall_suppress)
741 j = ACCESS_ONCE(jiffies);
742 js = ACCESS_ONCE(rsp->jiffies_stall);
744 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
746 /* We haven't checked in, so go dump stack. */
747 print_cpu_stall(rsp);
749 } else if (rcu_gp_in_progress(rsp) &&
750 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
752 /* They had a few time units to dump stack, so complain. */
753 print_other_cpu_stall(rsp);
757 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
759 rcu_cpu_stall_suppress = 1;
764 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
766 * Set the stall-warning timeout way off into the future, thus preventing
767 * any RCU CPU stall-warning messages from appearing in the current set of
770 * The caller must disable hard irqs.
772 void rcu_cpu_stall_reset(void)
774 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
775 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
776 rcu_preempt_stall_reset();
779 static struct notifier_block rcu_panic_block = {
780 .notifier_call = rcu_panic,
783 static void __init check_cpu_stall_init(void)
785 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
789 * Update CPU-local rcu_data state to record the newly noticed grace period.
790 * This is used both when we started the grace period and when we notice
791 * that someone else started the grace period. The caller must hold the
792 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
793 * and must have irqs disabled.
795 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
797 if (rdp->gpnum != rnp->gpnum) {
799 * If the current grace period is waiting for this CPU,
800 * set up to detect a quiescent state, otherwise don't
801 * go looking for one.
803 rdp->gpnum = rnp->gpnum;
804 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
805 if (rnp->qsmask & rdp->grpmask) {
807 rdp->passed_quiesce = 0;
813 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
816 struct rcu_node *rnp;
818 local_irq_save(flags);
820 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
821 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
822 local_irq_restore(flags);
825 __note_new_gpnum(rsp, rnp, rdp);
826 raw_spin_unlock_irqrestore(&rnp->lock, flags);
830 * Did someone else start a new RCU grace period start since we last
831 * checked? Update local state appropriately if so. Must be called
832 * on the CPU corresponding to rdp.
835 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
840 local_irq_save(flags);
841 if (rdp->gpnum != rsp->gpnum) {
842 note_new_gpnum(rsp, rdp);
845 local_irq_restore(flags);
850 * Advance this CPU's callbacks, but only if the current grace period
851 * has ended. This may be called only from the CPU to whom the rdp
852 * belongs. In addition, the corresponding leaf rcu_node structure's
853 * ->lock must be held by the caller, with irqs disabled.
856 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
858 /* Did another grace period end? */
859 if (rdp->completed != rnp->completed) {
861 /* Advance callbacks. No harm if list empty. */
862 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
863 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
864 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
866 /* Remember that we saw this grace-period completion. */
867 rdp->completed = rnp->completed;
868 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
871 * If we were in an extended quiescent state, we may have
872 * missed some grace periods that others CPUs handled on
873 * our behalf. Catch up with this state to avoid noting
874 * spurious new grace periods. If another grace period
875 * has started, then rnp->gpnum will have advanced, so
876 * we will detect this later on.
878 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
879 rdp->gpnum = rdp->completed;
882 * If RCU does not need a quiescent state from this CPU,
883 * then make sure that this CPU doesn't go looking for one.
885 if ((rnp->qsmask & rdp->grpmask) == 0)
891 * Advance this CPU's callbacks, but only if the current grace period
892 * has ended. This may be called only from the CPU to whom the rdp
896 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
899 struct rcu_node *rnp;
901 local_irq_save(flags);
903 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
904 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
905 local_irq_restore(flags);
908 __rcu_process_gp_end(rsp, rnp, rdp);
909 raw_spin_unlock_irqrestore(&rnp->lock, flags);
913 * Do per-CPU grace-period initialization for running CPU. The caller
914 * must hold the lock of the leaf rcu_node structure corresponding to
918 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
920 /* Prior grace period ended, so advance callbacks for current CPU. */
921 __rcu_process_gp_end(rsp, rnp, rdp);
924 * Because this CPU just now started the new grace period, we know
925 * that all of its callbacks will be covered by this upcoming grace
926 * period, even the ones that were registered arbitrarily recently.
927 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
929 * Other CPUs cannot be sure exactly when the grace period started.
930 * Therefore, their recently registered callbacks must pass through
931 * an additional RCU_NEXT_READY stage, so that they will be handled
932 * by the next RCU grace period.
934 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
935 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
937 /* Set state so that this CPU will detect the next quiescent state. */
938 __note_new_gpnum(rsp, rnp, rdp);
942 * Start a new RCU grace period if warranted, re-initializing the hierarchy
943 * in preparation for detecting the next grace period. The caller must hold
944 * the root node's ->lock, which is released before return. Hard irqs must
947 * Note that it is legal for a dying CPU (which is marked as offline) to
948 * invoke this function. This can happen when the dying CPU reports its
952 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
953 __releases(rcu_get_root(rsp)->lock)
955 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
956 struct rcu_node *rnp = rcu_get_root(rsp);
958 if (!rcu_scheduler_fully_active ||
959 !cpu_needs_another_gp(rsp, rdp)) {
961 * Either the scheduler hasn't yet spawned the first
962 * non-idle task or this CPU does not need another
963 * grace period. Either way, don't start a new grace
966 raw_spin_unlock_irqrestore(&rnp->lock, flags);
970 if (rsp->fqs_active) {
972 * This CPU needs a grace period, but force_quiescent_state()
973 * is running. Tell it to start one on this CPU's behalf.
975 rsp->fqs_need_gp = 1;
976 raw_spin_unlock_irqrestore(&rnp->lock, flags);
980 /* Advance to a new grace period and initialize state. */
982 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
983 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
984 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
985 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
986 record_gp_stall_check_time(rsp);
987 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
989 /* Exclude any concurrent CPU-hotplug operations. */
990 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
993 * Set the quiescent-state-needed bits in all the rcu_node
994 * structures for all currently online CPUs in breadth-first
995 * order, starting from the root rcu_node structure. This
996 * operation relies on the layout of the hierarchy within the
997 * rsp->node[] array. Note that other CPUs will access only
998 * the leaves of the hierarchy, which still indicate that no
999 * grace period is in progress, at least until the corresponding
1000 * leaf node has been initialized. In addition, we have excluded
1001 * CPU-hotplug operations.
1003 * Note that the grace period cannot complete until we finish
1004 * the initialization process, as there will be at least one
1005 * qsmask bit set in the root node until that time, namely the
1006 * one corresponding to this CPU, due to the fact that we have
1009 rcu_for_each_node_breadth_first(rsp, rnp) {
1010 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1011 rcu_preempt_check_blocked_tasks(rnp);
1012 rnp->qsmask = rnp->qsmaskinit;
1013 rnp->gpnum = rsp->gpnum;
1014 rnp->completed = rsp->completed;
1015 if (rnp == rdp->mynode)
1016 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1017 rcu_preempt_boost_start_gp(rnp);
1018 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1019 rnp->level, rnp->grplo,
1020 rnp->grphi, rnp->qsmask);
1021 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1024 rnp = rcu_get_root(rsp);
1025 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1026 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1027 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1028 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1032 * Report a full set of quiescent states to the specified rcu_state
1033 * data structure. This involves cleaning up after the prior grace
1034 * period and letting rcu_start_gp() start up the next grace period
1035 * if one is needed. Note that the caller must hold rnp->lock, as
1036 * required by rcu_start_gp(), which will release it.
1038 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1039 __releases(rcu_get_root(rsp)->lock)
1041 unsigned long gp_duration;
1042 struct rcu_node *rnp = rcu_get_root(rsp);
1043 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1045 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1048 * Ensure that all grace-period and pre-grace-period activity
1049 * is seen before the assignment to rsp->completed.
1051 smp_mb(); /* See above block comment. */
1052 gp_duration = jiffies - rsp->gp_start;
1053 if (gp_duration > rsp->gp_max)
1054 rsp->gp_max = gp_duration;
1057 * We know the grace period is complete, but to everyone else
1058 * it appears to still be ongoing. But it is also the case
1059 * that to everyone else it looks like there is nothing that
1060 * they can do to advance the grace period. It is therefore
1061 * safe for us to drop the lock in order to mark the grace
1062 * period as completed in all of the rcu_node structures.
1064 * But if this CPU needs another grace period, it will take
1065 * care of this while initializing the next grace period.
1066 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1067 * because the callbacks have not yet been advanced: Those
1068 * callbacks are waiting on the grace period that just now
1071 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1072 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1075 * Propagate new ->completed value to rcu_node structures
1076 * so that other CPUs don't have to wait until the start
1077 * of the next grace period to process their callbacks.
1079 rcu_for_each_node_breadth_first(rsp, rnp) {
1080 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1081 rnp->completed = rsp->gpnum;
1082 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1084 rnp = rcu_get_root(rsp);
1085 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1088 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1089 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1090 rsp->fqs_state = RCU_GP_IDLE;
1091 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1095 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1096 * Allows quiescent states for a group of CPUs to be reported at one go
1097 * to the specified rcu_node structure, though all the CPUs in the group
1098 * must be represented by the same rcu_node structure (which need not be
1099 * a leaf rcu_node structure, though it often will be). That structure's
1100 * lock must be held upon entry, and it is released before return.
1103 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1104 struct rcu_node *rnp, unsigned long flags)
1105 __releases(rnp->lock)
1107 struct rcu_node *rnp_c;
1109 /* Walk up the rcu_node hierarchy. */
1111 if (!(rnp->qsmask & mask)) {
1113 /* Our bit has already been cleared, so done. */
1114 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1117 rnp->qsmask &= ~mask;
1118 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1119 mask, rnp->qsmask, rnp->level,
1120 rnp->grplo, rnp->grphi,
1122 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1124 /* Other bits still set at this level, so done. */
1125 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1128 mask = rnp->grpmask;
1129 if (rnp->parent == NULL) {
1131 /* No more levels. Exit loop holding root lock. */
1135 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1138 raw_spin_lock_irqsave(&rnp->lock, flags);
1139 WARN_ON_ONCE(rnp_c->qsmask);
1143 * Get here if we are the last CPU to pass through a quiescent
1144 * state for this grace period. Invoke rcu_report_qs_rsp()
1145 * to clean up and start the next grace period if one is needed.
1147 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1151 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1152 * structure. This must be either called from the specified CPU, or
1153 * called when the specified CPU is known to be offline (and when it is
1154 * also known that no other CPU is concurrently trying to help the offline
1155 * CPU). The lastcomp argument is used to make sure we are still in the
1156 * grace period of interest. We don't want to end the current grace period
1157 * based on quiescent states detected in an earlier grace period!
1160 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1162 unsigned long flags;
1164 struct rcu_node *rnp;
1167 raw_spin_lock_irqsave(&rnp->lock, flags);
1168 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1171 * The grace period in which this quiescent state was
1172 * recorded has ended, so don't report it upwards.
1173 * We will instead need a new quiescent state that lies
1174 * within the current grace period.
1176 rdp->passed_quiesce = 0; /* need qs for new gp. */
1177 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1180 mask = rdp->grpmask;
1181 if ((rnp->qsmask & mask) == 0) {
1182 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1184 rdp->qs_pending = 0;
1187 * This GP can't end until cpu checks in, so all of our
1188 * callbacks can be processed during the next GP.
1190 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1192 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1197 * Check to see if there is a new grace period of which this CPU
1198 * is not yet aware, and if so, set up local rcu_data state for it.
1199 * Otherwise, see if this CPU has just passed through its first
1200 * quiescent state for this grace period, and record that fact if so.
1203 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1205 /* If there is now a new grace period, record and return. */
1206 if (check_for_new_grace_period(rsp, rdp))
1210 * Does this CPU still need to do its part for current grace period?
1211 * If no, return and let the other CPUs do their part as well.
1213 if (!rdp->qs_pending)
1217 * Was there a quiescent state since the beginning of the grace
1218 * period? If no, then exit and wait for the next call.
1220 if (!rdp->passed_quiesce)
1224 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1227 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1230 #ifdef CONFIG_HOTPLUG_CPU
1233 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1234 * Also record a quiescent state for this CPU for the current grace period.
1235 * Synchronization and interrupt disabling are not required because
1236 * this function executes in stop_machine() context. Therefore, cleanup
1237 * operations that might block must be done later from the CPU_DEAD
1240 * Note that the outgoing CPU's bit has already been cleared in the
1241 * cpu_online_mask. This allows us to randomly pick a callback
1242 * destination from the bits set in that mask.
1244 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1246 unsigned long flags;
1250 int receive_cpu = cpumask_any(cpu_online_mask);
1251 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1252 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1253 struct rcu_node *rnp = rdp->mynode; /* For dying CPU. */
1255 /* First, adjust the counts. */
1256 if (rdp->nxtlist != NULL) {
1257 receive_rdp->qlen_lazy += rdp->qlen_lazy;
1258 receive_rdp->qlen += rdp->qlen;
1264 * Next, move ready-to-invoke callbacks to be invoked on some
1265 * other CPU. These will not be required to pass through another
1266 * grace period: They are done, regardless of CPU.
1268 if (rdp->nxtlist != NULL &&
1269 rdp->nxttail[RCU_DONE_TAIL] != &rdp->nxtlist) {
1270 struct rcu_head *oldhead;
1271 struct rcu_head **oldtail;
1272 struct rcu_head **newtail;
1274 oldhead = rdp->nxtlist;
1275 oldtail = receive_rdp->nxttail[RCU_DONE_TAIL];
1276 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1277 *rdp->nxttail[RCU_DONE_TAIL] = *oldtail;
1278 *receive_rdp->nxttail[RCU_DONE_TAIL] = oldhead;
1279 newtail = rdp->nxttail[RCU_DONE_TAIL];
1280 for (i = RCU_DONE_TAIL; i < RCU_NEXT_SIZE; i++) {
1281 if (receive_rdp->nxttail[i] == oldtail)
1282 receive_rdp->nxttail[i] = newtail;
1283 if (rdp->nxttail[i] == newtail)
1284 rdp->nxttail[i] = &rdp->nxtlist;
1289 * Finally, put the rest of the callbacks at the end of the list.
1290 * The ones that made it partway through get to start over: We
1291 * cannot assume that grace periods are synchronized across CPUs.
1292 * (We could splice RCU_WAIT_TAIL into RCU_NEXT_READY_TAIL, but
1293 * this does not seem compelling. Not yet, anyway.)
1295 if (rdp->nxtlist != NULL) {
1296 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1297 receive_rdp->nxttail[RCU_NEXT_TAIL] =
1298 rdp->nxttail[RCU_NEXT_TAIL];
1299 receive_rdp->n_cbs_adopted += rdp->qlen;
1300 rdp->n_cbs_orphaned += rdp->qlen;
1302 rdp->nxtlist = NULL;
1303 for (i = 0; i < RCU_NEXT_SIZE; i++)
1304 rdp->nxttail[i] = &rdp->nxtlist;
1308 * Record a quiescent state for the dying CPU. This is safe
1309 * only because we have already cleared out the callbacks.
1310 * (Otherwise, the RCU core might try to schedule the invocation
1311 * of callbacks on this now-offline CPU, which would be bad.)
1313 mask = rdp->grpmask; /* rnp->grplo is constant. */
1314 trace_rcu_grace_period(rsp->name,
1315 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1317 rcu_report_qs_rdp(smp_processor_id(), rsp, rdp, rsp->gpnum);
1318 /* Note that rcu_report_qs_rdp() might call trace_rcu_grace_period(). */
1321 * Remove the dying CPU from the bitmasks in the rcu_node
1322 * hierarchy. Because we are in stop_machine() context, we
1323 * automatically exclude ->onofflock critical sections.
1326 raw_spin_lock_irqsave(&rnp->lock, flags);
1327 rnp->qsmaskinit &= ~mask;
1328 if (rnp->qsmaskinit != 0) {
1329 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1332 if (rnp == rdp->mynode) {
1333 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1334 if (need_report & RCU_OFL_TASKS_NORM_GP)
1335 rcu_report_unblock_qs_rnp(rnp, flags);
1337 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1338 if (need_report & RCU_OFL_TASKS_EXP_GP)
1339 rcu_report_exp_rnp(rsp, rnp, true);
1341 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1342 mask = rnp->grpmask;
1344 } while (rnp != NULL);
1348 * The CPU has been completely removed, and some other CPU is reporting
1349 * this fact from process context. Do the remainder of the cleanup.
1350 * There can only be one CPU hotplug operation at a time, so no other
1351 * CPU can be attempting to update rcu_cpu_kthread_task.
1353 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1355 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1356 struct rcu_node *rnp = rdp->mynode;
1358 rcu_stop_cpu_kthread(cpu);
1359 rcu_node_kthread_setaffinity(rnp, -1);
1362 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1364 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1368 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1372 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1375 * Invoke any RCU callbacks that have made it to the end of their grace
1376 * period. Thottle as specified by rdp->blimit.
1378 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1380 unsigned long flags;
1381 struct rcu_head *next, *list, **tail;
1382 int bl, count, count_lazy;
1384 /* If no callbacks are ready, just return.*/
1385 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1386 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1387 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1388 need_resched(), is_idle_task(current),
1389 rcu_is_callbacks_kthread());
1394 * Extract the list of ready callbacks, disabling to prevent
1395 * races with call_rcu() from interrupt handlers.
1397 local_irq_save(flags);
1398 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1400 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1401 list = rdp->nxtlist;
1402 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1403 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1404 tail = rdp->nxttail[RCU_DONE_TAIL];
1405 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1406 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1407 rdp->nxttail[count] = &rdp->nxtlist;
1408 local_irq_restore(flags);
1410 /* Invoke callbacks. */
1411 count = count_lazy = 0;
1415 debug_rcu_head_unqueue(list);
1416 if (__rcu_reclaim(rsp->name, list))
1419 /* Stop only if limit reached and CPU has something to do. */
1420 if (++count >= bl &&
1422 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1426 local_irq_save(flags);
1427 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1428 is_idle_task(current),
1429 rcu_is_callbacks_kthread());
1431 /* Update count, and requeue any remaining callbacks. */
1432 rdp->qlen_lazy -= count_lazy;
1434 rdp->n_cbs_invoked += count;
1436 *tail = rdp->nxtlist;
1437 rdp->nxtlist = list;
1438 for (count = 0; count < RCU_NEXT_SIZE; count++)
1439 if (&rdp->nxtlist == rdp->nxttail[count])
1440 rdp->nxttail[count] = tail;
1445 /* Reinstate batch limit if we have worked down the excess. */
1446 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1447 rdp->blimit = blimit;
1449 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1450 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1451 rdp->qlen_last_fqs_check = 0;
1452 rdp->n_force_qs_snap = rsp->n_force_qs;
1453 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1454 rdp->qlen_last_fqs_check = rdp->qlen;
1456 local_irq_restore(flags);
1458 /* Re-invoke RCU core processing if there are callbacks remaining. */
1459 if (cpu_has_callbacks_ready_to_invoke(rdp))
1464 * Check to see if this CPU is in a non-context-switch quiescent state
1465 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1466 * Also schedule RCU core processing.
1468 * This function must be called from hardirq context. It is normally
1469 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1470 * false, there is no point in invoking rcu_check_callbacks().
1472 void rcu_check_callbacks(int cpu, int user)
1474 trace_rcu_utilization("Start scheduler-tick");
1475 if (user || rcu_is_cpu_rrupt_from_idle()) {
1478 * Get here if this CPU took its interrupt from user
1479 * mode or from the idle loop, and if this is not a
1480 * nested interrupt. In this case, the CPU is in
1481 * a quiescent state, so note it.
1483 * No memory barrier is required here because both
1484 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1485 * variables that other CPUs neither access nor modify,
1486 * at least not while the corresponding CPU is online.
1492 } else if (!in_softirq()) {
1495 * Get here if this CPU did not take its interrupt from
1496 * softirq, in other words, if it is not interrupting
1497 * a rcu_bh read-side critical section. This is an _bh
1498 * critical section, so note it.
1503 rcu_preempt_check_callbacks(cpu);
1504 if (rcu_pending(cpu))
1506 trace_rcu_utilization("End scheduler-tick");
1512 * Scan the leaf rcu_node structures, processing dyntick state for any that
1513 * have not yet encountered a quiescent state, using the function specified.
1514 * Also initiate boosting for any threads blocked on the root rcu_node.
1516 * The caller must have suppressed start of new grace periods.
1518 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1522 unsigned long flags;
1524 struct rcu_node *rnp;
1526 rcu_for_each_leaf_node(rsp, rnp) {
1528 raw_spin_lock_irqsave(&rnp->lock, flags);
1529 if (!rcu_gp_in_progress(rsp)) {
1530 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1533 if (rnp->qsmask == 0) {
1534 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1539 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1540 if ((rnp->qsmask & bit) != 0 &&
1541 f(per_cpu_ptr(rsp->rda, cpu)))
1546 /* rcu_report_qs_rnp() releases rnp->lock. */
1547 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1550 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1552 rnp = rcu_get_root(rsp);
1553 if (rnp->qsmask == 0) {
1554 raw_spin_lock_irqsave(&rnp->lock, flags);
1555 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1560 * Force quiescent states on reluctant CPUs, and also detect which
1561 * CPUs are in dyntick-idle mode.
1563 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1565 unsigned long flags;
1566 struct rcu_node *rnp = rcu_get_root(rsp);
1568 trace_rcu_utilization("Start fqs");
1569 if (!rcu_gp_in_progress(rsp)) {
1570 trace_rcu_utilization("End fqs");
1571 return; /* No grace period in progress, nothing to force. */
1573 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1574 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1575 trace_rcu_utilization("End fqs");
1576 return; /* Someone else is already on the job. */
1578 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1579 goto unlock_fqs_ret; /* no emergency and done recently. */
1581 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1582 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1583 if(!rcu_gp_in_progress(rsp)) {
1584 rsp->n_force_qs_ngp++;
1585 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1586 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1588 rsp->fqs_active = 1;
1589 switch (rsp->fqs_state) {
1593 break; /* grace period idle or initializing, ignore. */
1595 case RCU_SAVE_DYNTICK:
1596 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1597 break; /* So gcc recognizes the dead code. */
1599 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1601 /* Record dyntick-idle state. */
1602 force_qs_rnp(rsp, dyntick_save_progress_counter);
1603 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1604 if (rcu_gp_in_progress(rsp))
1605 rsp->fqs_state = RCU_FORCE_QS;
1610 /* Check dyntick-idle state, send IPI to laggarts. */
1611 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1612 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1614 /* Leave state in case more forcing is required. */
1616 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1619 rsp->fqs_active = 0;
1620 if (rsp->fqs_need_gp) {
1621 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1622 rsp->fqs_need_gp = 0;
1623 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1624 trace_rcu_utilization("End fqs");
1627 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1629 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1630 trace_rcu_utilization("End fqs");
1633 #else /* #ifdef CONFIG_SMP */
1635 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1640 #endif /* #else #ifdef CONFIG_SMP */
1643 * This does the RCU core processing work for the specified rcu_state
1644 * and rcu_data structures. This may be called only from the CPU to
1645 * whom the rdp belongs.
1648 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1650 unsigned long flags;
1652 WARN_ON_ONCE(rdp->beenonline == 0);
1655 * If an RCU GP has gone long enough, go check for dyntick
1656 * idle CPUs and, if needed, send resched IPIs.
1658 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1659 force_quiescent_state(rsp, 1);
1662 * Advance callbacks in response to end of earlier grace
1663 * period that some other CPU ended.
1665 rcu_process_gp_end(rsp, rdp);
1667 /* Update RCU state based on any recent quiescent states. */
1668 rcu_check_quiescent_state(rsp, rdp);
1670 /* Does this CPU require a not-yet-started grace period? */
1671 if (cpu_needs_another_gp(rsp, rdp)) {
1672 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1673 rcu_start_gp(rsp, flags); /* releases above lock */
1676 /* If there are callbacks ready, invoke them. */
1677 if (cpu_has_callbacks_ready_to_invoke(rdp))
1678 invoke_rcu_callbacks(rsp, rdp);
1682 * Do RCU core processing for the current CPU.
1684 static void rcu_process_callbacks(struct softirq_action *unused)
1686 trace_rcu_utilization("Start RCU core");
1687 __rcu_process_callbacks(&rcu_sched_state,
1688 &__get_cpu_var(rcu_sched_data));
1689 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1690 rcu_preempt_process_callbacks();
1691 trace_rcu_utilization("End RCU core");
1695 * Schedule RCU callback invocation. If the specified type of RCU
1696 * does not support RCU priority boosting, just do a direct call,
1697 * otherwise wake up the per-CPU kernel kthread. Note that because we
1698 * are running on the current CPU with interrupts disabled, the
1699 * rcu_cpu_kthread_task cannot disappear out from under us.
1701 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1703 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1705 if (likely(!rsp->boost)) {
1706 rcu_do_batch(rsp, rdp);
1709 invoke_rcu_callbacks_kthread();
1712 static void invoke_rcu_core(void)
1714 raise_softirq(RCU_SOFTIRQ);
1718 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1719 struct rcu_state *rsp, bool lazy)
1721 unsigned long flags;
1722 struct rcu_data *rdp;
1724 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1725 debug_rcu_head_queue(head);
1729 smp_mb(); /* Ensure RCU update seen before callback registry. */
1732 * Opportunistically note grace-period endings and beginnings.
1733 * Note that we might see a beginning right after we see an
1734 * end, but never vice versa, since this CPU has to pass through
1735 * a quiescent state betweentimes.
1737 local_irq_save(flags);
1738 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1739 rdp = this_cpu_ptr(rsp->rda);
1741 /* Add the callback to our list. */
1742 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1743 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1748 if (__is_kfree_rcu_offset((unsigned long)func))
1749 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1750 rdp->qlen_lazy, rdp->qlen);
1752 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1754 /* If interrupts were disabled, don't dive into RCU core. */
1755 if (irqs_disabled_flags(flags)) {
1756 local_irq_restore(flags);
1761 * Force the grace period if too many callbacks or too long waiting.
1762 * Enforce hysteresis, and don't invoke force_quiescent_state()
1763 * if some other CPU has recently done so. Also, don't bother
1764 * invoking force_quiescent_state() if the newly enqueued callback
1765 * is the only one waiting for a grace period to complete.
1767 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1769 /* Are we ignoring a completed grace period? */
1770 rcu_process_gp_end(rsp, rdp);
1771 check_for_new_grace_period(rsp, rdp);
1773 /* Start a new grace period if one not already started. */
1774 if (!rcu_gp_in_progress(rsp)) {
1775 unsigned long nestflag;
1776 struct rcu_node *rnp_root = rcu_get_root(rsp);
1778 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1779 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1781 /* Give the grace period a kick. */
1782 rdp->blimit = LONG_MAX;
1783 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1784 *rdp->nxttail[RCU_DONE_TAIL] != head)
1785 force_quiescent_state(rsp, 0);
1786 rdp->n_force_qs_snap = rsp->n_force_qs;
1787 rdp->qlen_last_fqs_check = rdp->qlen;
1789 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1790 force_quiescent_state(rsp, 1);
1791 local_irq_restore(flags);
1795 * Queue an RCU-sched callback for invocation after a grace period.
1797 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1799 __call_rcu(head, func, &rcu_sched_state, 0);
1801 EXPORT_SYMBOL_GPL(call_rcu_sched);
1804 * Queue an RCU callback for invocation after a quicker grace period.
1806 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1808 __call_rcu(head, func, &rcu_bh_state, 0);
1810 EXPORT_SYMBOL_GPL(call_rcu_bh);
1813 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1815 * Control will return to the caller some time after a full rcu-sched
1816 * grace period has elapsed, in other words after all currently executing
1817 * rcu-sched read-side critical sections have completed. These read-side
1818 * critical sections are delimited by rcu_read_lock_sched() and
1819 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1820 * local_irq_disable(), and so on may be used in place of
1821 * rcu_read_lock_sched().
1823 * This means that all preempt_disable code sequences, including NMI and
1824 * hardware-interrupt handlers, in progress on entry will have completed
1825 * before this primitive returns. However, this does not guarantee that
1826 * softirq handlers will have completed, since in some kernels, these
1827 * handlers can run in process context, and can block.
1829 * This primitive provides the guarantees made by the (now removed)
1830 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1831 * guarantees that rcu_read_lock() sections will have completed.
1832 * In "classic RCU", these two guarantees happen to be one and
1833 * the same, but can differ in realtime RCU implementations.
1835 void synchronize_sched(void)
1837 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1838 !lock_is_held(&rcu_lock_map) &&
1839 !lock_is_held(&rcu_sched_lock_map),
1840 "Illegal synchronize_sched() in RCU-sched read-side critical section");
1841 if (rcu_blocking_is_gp())
1843 wait_rcu_gp(call_rcu_sched);
1845 EXPORT_SYMBOL_GPL(synchronize_sched);
1848 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1850 * Control will return to the caller some time after a full rcu_bh grace
1851 * period has elapsed, in other words after all currently executing rcu_bh
1852 * read-side critical sections have completed. RCU read-side critical
1853 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1854 * and may be nested.
1856 void synchronize_rcu_bh(void)
1858 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1859 !lock_is_held(&rcu_lock_map) &&
1860 !lock_is_held(&rcu_sched_lock_map),
1861 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
1862 if (rcu_blocking_is_gp())
1864 wait_rcu_gp(call_rcu_bh);
1866 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1869 * Check to see if there is any immediate RCU-related work to be done
1870 * by the current CPU, for the specified type of RCU, returning 1 if so.
1871 * The checks are in order of increasing expense: checks that can be
1872 * carried out against CPU-local state are performed first. However,
1873 * we must check for CPU stalls first, else we might not get a chance.
1875 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1877 struct rcu_node *rnp = rdp->mynode;
1879 rdp->n_rcu_pending++;
1881 /* Check for CPU stalls, if enabled. */
1882 check_cpu_stall(rsp, rdp);
1884 /* Is the RCU core waiting for a quiescent state from this CPU? */
1885 if (rcu_scheduler_fully_active &&
1886 rdp->qs_pending && !rdp->passed_quiesce) {
1889 * If force_quiescent_state() coming soon and this CPU
1890 * needs a quiescent state, and this is either RCU-sched
1891 * or RCU-bh, force a local reschedule.
1893 rdp->n_rp_qs_pending++;
1894 if (!rdp->preemptible &&
1895 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1898 } else if (rdp->qs_pending && rdp->passed_quiesce) {
1899 rdp->n_rp_report_qs++;
1903 /* Does this CPU have callbacks ready to invoke? */
1904 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1905 rdp->n_rp_cb_ready++;
1909 /* Has RCU gone idle with this CPU needing another grace period? */
1910 if (cpu_needs_another_gp(rsp, rdp)) {
1911 rdp->n_rp_cpu_needs_gp++;
1915 /* Has another RCU grace period completed? */
1916 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1917 rdp->n_rp_gp_completed++;
1921 /* Has a new RCU grace period started? */
1922 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1923 rdp->n_rp_gp_started++;
1927 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1928 if (rcu_gp_in_progress(rsp) &&
1929 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1930 rdp->n_rp_need_fqs++;
1935 rdp->n_rp_need_nothing++;
1940 * Check to see if there is any immediate RCU-related work to be done
1941 * by the current CPU, returning 1 if so. This function is part of the
1942 * RCU implementation; it is -not- an exported member of the RCU API.
1944 static int rcu_pending(int cpu)
1946 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1947 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1948 rcu_preempt_pending(cpu);
1952 * Check to see if any future RCU-related work will need to be done
1953 * by the current CPU, even if none need be done immediately, returning
1956 static int rcu_cpu_has_callbacks(int cpu)
1958 /* RCU callbacks either ready or pending? */
1959 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1960 per_cpu(rcu_bh_data, cpu).nxtlist ||
1961 rcu_preempt_cpu_has_callbacks(cpu);
1964 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1965 static atomic_t rcu_barrier_cpu_count;
1966 static DEFINE_MUTEX(rcu_barrier_mutex);
1967 static struct completion rcu_barrier_completion;
1969 static void rcu_barrier_callback(struct rcu_head *notused)
1971 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1972 complete(&rcu_barrier_completion);
1976 * Called with preemption disabled, and from cross-cpu IRQ context.
1978 static void rcu_barrier_func(void *type)
1980 int cpu = smp_processor_id();
1981 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1982 void (*call_rcu_func)(struct rcu_head *head,
1983 void (*func)(struct rcu_head *head));
1985 atomic_inc(&rcu_barrier_cpu_count);
1986 call_rcu_func = type;
1987 call_rcu_func(head, rcu_barrier_callback);
1991 * Orchestrate the specified type of RCU barrier, waiting for all
1992 * RCU callbacks of the specified type to complete.
1994 static void _rcu_barrier(struct rcu_state *rsp,
1995 void (*call_rcu_func)(struct rcu_head *head,
1996 void (*func)(struct rcu_head *head)))
1998 BUG_ON(in_interrupt());
1999 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2000 mutex_lock(&rcu_barrier_mutex);
2001 init_completion(&rcu_barrier_completion);
2003 * Initialize rcu_barrier_cpu_count to 1, then invoke
2004 * rcu_barrier_func() on each CPU, so that each CPU also has
2005 * incremented rcu_barrier_cpu_count. Only then is it safe to
2006 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2007 * might complete its grace period before all of the other CPUs
2008 * did their increment, causing this function to return too
2009 * early. Note that on_each_cpu() disables irqs, which prevents
2010 * any CPUs from coming online or going offline until each online
2011 * CPU has queued its RCU-barrier callback.
2013 atomic_set(&rcu_barrier_cpu_count, 1);
2014 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
2015 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2016 complete(&rcu_barrier_completion);
2017 wait_for_completion(&rcu_barrier_completion);
2018 mutex_unlock(&rcu_barrier_mutex);
2022 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2024 void rcu_barrier_bh(void)
2026 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
2028 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2031 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2033 void rcu_barrier_sched(void)
2035 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
2037 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2040 * Do boot-time initialization of a CPU's per-CPU RCU data.
2043 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2045 unsigned long flags;
2047 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2048 struct rcu_node *rnp = rcu_get_root(rsp);
2050 /* Set up local state, ensuring consistent view of global state. */
2051 raw_spin_lock_irqsave(&rnp->lock, flags);
2052 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2053 rdp->nxtlist = NULL;
2054 for (i = 0; i < RCU_NEXT_SIZE; i++)
2055 rdp->nxttail[i] = &rdp->nxtlist;
2058 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2059 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_NESTING);
2060 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2063 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2067 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2068 * offline event can be happening at a given time. Note also that we
2069 * can accept some slop in the rsp->completed access due to the fact
2070 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2072 static void __cpuinit
2073 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2075 unsigned long flags;
2077 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2078 struct rcu_node *rnp = rcu_get_root(rsp);
2080 /* Set up local state, ensuring consistent view of global state. */
2081 raw_spin_lock_irqsave(&rnp->lock, flags);
2082 rdp->beenonline = 1; /* We have now been online. */
2083 rdp->preemptible = preemptible;
2084 rdp->qlen_last_fqs_check = 0;
2085 rdp->n_force_qs_snap = rsp->n_force_qs;
2086 rdp->blimit = blimit;
2087 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_NESTING;
2088 atomic_set(&rdp->dynticks->dynticks,
2089 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2090 rcu_prepare_for_idle_init(cpu);
2091 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2094 * A new grace period might start here. If so, we won't be part
2095 * of it, but that is OK, as we are currently in a quiescent state.
2098 /* Exclude any attempts to start a new GP on large systems. */
2099 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2101 /* Add CPU to rcu_node bitmasks. */
2103 mask = rdp->grpmask;
2105 /* Exclude any attempts to start a new GP on small systems. */
2106 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2107 rnp->qsmaskinit |= mask;
2108 mask = rnp->grpmask;
2109 if (rnp == rdp->mynode) {
2111 * If there is a grace period in progress, we will
2112 * set up to wait for it next time we run the
2115 rdp->gpnum = rnp->completed;
2116 rdp->completed = rnp->completed;
2117 rdp->passed_quiesce = 0;
2118 rdp->qs_pending = 0;
2119 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2120 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2122 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2124 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2126 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2129 static void __cpuinit rcu_prepare_cpu(int cpu)
2131 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2132 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2133 rcu_preempt_init_percpu_data(cpu);
2137 * Handle CPU online/offline notification events.
2139 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2140 unsigned long action, void *hcpu)
2142 long cpu = (long)hcpu;
2143 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2144 struct rcu_node *rnp = rdp->mynode;
2146 trace_rcu_utilization("Start CPU hotplug");
2148 case CPU_UP_PREPARE:
2149 case CPU_UP_PREPARE_FROZEN:
2150 rcu_prepare_cpu(cpu);
2151 rcu_prepare_kthreads(cpu);
2154 case CPU_DOWN_FAILED:
2155 rcu_node_kthread_setaffinity(rnp, -1);
2156 rcu_cpu_kthread_setrt(cpu, 1);
2158 case CPU_DOWN_PREPARE:
2159 rcu_node_kthread_setaffinity(rnp, cpu);
2160 rcu_cpu_kthread_setrt(cpu, 0);
2163 case CPU_DYING_FROZEN:
2165 * The whole machine is "stopped" except this CPU, so we can
2166 * touch any data without introducing corruption. We send the
2167 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2169 rcu_cleanup_dying_cpu(&rcu_bh_state);
2170 rcu_cleanup_dying_cpu(&rcu_sched_state);
2171 rcu_preempt_cleanup_dying_cpu();
2172 rcu_cleanup_after_idle(cpu);
2175 case CPU_DEAD_FROZEN:
2176 case CPU_UP_CANCELED:
2177 case CPU_UP_CANCELED_FROZEN:
2178 rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2179 rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2180 rcu_preempt_cleanup_dead_cpu(cpu);
2185 trace_rcu_utilization("End CPU hotplug");
2190 * This function is invoked towards the end of the scheduler's initialization
2191 * process. Before this is called, the idle task might contain
2192 * RCU read-side critical sections (during which time, this idle
2193 * task is booting the system). After this function is called, the
2194 * idle tasks are prohibited from containing RCU read-side critical
2195 * sections. This function also enables RCU lockdep checking.
2197 void rcu_scheduler_starting(void)
2199 WARN_ON(num_online_cpus() != 1);
2200 WARN_ON(nr_context_switches() > 0);
2201 rcu_scheduler_active = 1;
2205 * Compute the per-level fanout, either using the exact fanout specified
2206 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2208 #ifdef CONFIG_RCU_FANOUT_EXACT
2209 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2213 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2214 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2215 rsp->levelspread[0] = RCU_FANOUT_LEAF;
2217 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2218 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2225 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2226 ccur = rsp->levelcnt[i];
2227 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2231 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2234 * Helper function for rcu_init() that initializes one rcu_state structure.
2236 static void __init rcu_init_one(struct rcu_state *rsp,
2237 struct rcu_data __percpu *rda)
2239 static char *buf[] = { "rcu_node_level_0",
2242 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2246 struct rcu_node *rnp;
2248 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2250 /* Initialize the level-tracking arrays. */
2252 for (i = 1; i < NUM_RCU_LVLS; i++)
2253 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2254 rcu_init_levelspread(rsp);
2256 /* Initialize the elements themselves, starting from the leaves. */
2258 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2259 cpustride *= rsp->levelspread[i];
2260 rnp = rsp->level[i];
2261 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2262 raw_spin_lock_init(&rnp->lock);
2263 lockdep_set_class_and_name(&rnp->lock,
2264 &rcu_node_class[i], buf[i]);
2267 rnp->qsmaskinit = 0;
2268 rnp->grplo = j * cpustride;
2269 rnp->grphi = (j + 1) * cpustride - 1;
2270 if (rnp->grphi >= NR_CPUS)
2271 rnp->grphi = NR_CPUS - 1;
2277 rnp->grpnum = j % rsp->levelspread[i - 1];
2278 rnp->grpmask = 1UL << rnp->grpnum;
2279 rnp->parent = rsp->level[i - 1] +
2280 j / rsp->levelspread[i - 1];
2283 INIT_LIST_HEAD(&rnp->blkd_tasks);
2288 rnp = rsp->level[NUM_RCU_LVLS - 1];
2289 for_each_possible_cpu(i) {
2290 while (i > rnp->grphi)
2292 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2293 rcu_boot_init_percpu_data(i, rsp);
2297 void __init rcu_init(void)
2301 rcu_bootup_announce();
2302 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2303 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2304 __rcu_init_preempt();
2305 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2308 * We don't need protection against CPU-hotplug here because
2309 * this is called early in boot, before either interrupts
2310 * or the scheduler are operational.
2312 cpu_notifier(rcu_cpu_notify, 0);
2313 for_each_online_cpu(cpu)
2314 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2315 check_cpu_stall_init();
2318 #include "rcutree_plugin.h"