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>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
67 #define RCU_STATE_INITIALIZER(sname, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
73 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
80 struct rcu_state rcu_sched_state =
81 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
82 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
84 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
85 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
87 static struct rcu_state *rcu_state;
88 LIST_HEAD(rcu_struct_flavors);
90 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
91 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
92 module_param(rcu_fanout_leaf, int, 0444);
93 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
94 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
101 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
104 * The rcu_scheduler_active variable transitions from zero to one just
105 * before the first task is spawned. So when this variable is zero, RCU
106 * can assume that there is but one task, allowing RCU to (for example)
107 * optimized synchronize_sched() to a simple barrier(). When this variable
108 * is one, RCU must actually do all the hard work required to detect real
109 * grace periods. This variable is also used to suppress boot-time false
110 * positives from lockdep-RCU error checking.
112 int rcu_scheduler_active __read_mostly;
113 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
116 * The rcu_scheduler_fully_active variable transitions from zero to one
117 * during the early_initcall() processing, which is after the scheduler
118 * is capable of creating new tasks. So RCU processing (for example,
119 * creating tasks for RCU priority boosting) must be delayed until after
120 * rcu_scheduler_fully_active transitions from zero to one. We also
121 * currently delay invocation of any RCU callbacks until after this point.
123 * It might later prove better for people registering RCU callbacks during
124 * early boot to take responsibility for these callbacks, but one step at
127 static int rcu_scheduler_fully_active __read_mostly;
129 #ifdef CONFIG_RCU_BOOST
132 * Control variables for per-CPU and per-rcu_node kthreads. These
133 * handle all flavors of RCU.
135 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
136 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
138 DEFINE_PER_CPU(char, rcu_cpu_has_work);
140 #endif /* #ifdef CONFIG_RCU_BOOST */
142 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
143 static void invoke_rcu_core(void);
144 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
147 * Track the rcutorture test sequence number and the update version
148 * number within a given test. The rcutorture_testseq is incremented
149 * on every rcutorture module load and unload, so has an odd value
150 * when a test is running. The rcutorture_vernum is set to zero
151 * when rcutorture starts and is incremented on each rcutorture update.
152 * These variables enable correlating rcutorture output with the
153 * RCU tracing information.
155 unsigned long rcutorture_testseq;
156 unsigned long rcutorture_vernum;
159 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
160 * permit this function to be invoked without holding the root rcu_node
161 * structure's ->lock, but of course results can be subject to change.
163 static int rcu_gp_in_progress(struct rcu_state *rsp)
165 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
169 * Note a quiescent state. Because we do not need to know
170 * how many quiescent states passed, just if there was at least
171 * one since the start of the grace period, this just sets a flag.
172 * The caller must have disabled preemption.
174 void rcu_sched_qs(int cpu)
176 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
178 if (rdp->passed_quiesce == 0)
179 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
180 rdp->passed_quiesce = 1;
183 void rcu_bh_qs(int cpu)
185 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
187 if (rdp->passed_quiesce == 0)
188 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
189 rdp->passed_quiesce = 1;
193 * Note a context switch. This is a quiescent state for RCU-sched,
194 * and requires special handling for preemptible RCU.
195 * The caller must have disabled preemption.
197 void rcu_note_context_switch(int cpu)
199 trace_rcu_utilization("Start context switch");
201 rcu_preempt_note_context_switch(cpu);
202 trace_rcu_utilization("End context switch");
204 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
206 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
207 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
208 .dynticks = ATOMIC_INIT(1),
209 #ifdef CONFIG_RCU_USER_QS
210 .ignore_user_qs = true,
214 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
215 static int qhimark = 10000; /* If this many pending, ignore blimit. */
216 static int qlowmark = 100; /* Once only this many pending, use blimit. */
218 module_param(blimit, int, 0444);
219 module_param(qhimark, int, 0444);
220 module_param(qlowmark, int, 0444);
222 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
223 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
225 module_param(rcu_cpu_stall_suppress, int, 0644);
226 module_param(rcu_cpu_stall_timeout, int, 0644);
228 static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
229 static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
231 module_param(jiffies_till_first_fqs, ulong, 0644);
232 module_param(jiffies_till_next_fqs, ulong, 0644);
234 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
235 static void force_quiescent_state(struct rcu_state *rsp);
236 static int rcu_pending(int cpu);
239 * Return the number of RCU-sched batches processed thus far for debug & stats.
241 long rcu_batches_completed_sched(void)
243 return rcu_sched_state.completed;
245 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
248 * Return the number of RCU BH batches processed thus far for debug & stats.
250 long rcu_batches_completed_bh(void)
252 return rcu_bh_state.completed;
254 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
257 * Force a quiescent state for RCU BH.
259 void rcu_bh_force_quiescent_state(void)
261 force_quiescent_state(&rcu_bh_state);
263 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
266 * Record the number of times rcutorture tests have been initiated and
267 * terminated. This information allows the debugfs tracing stats to be
268 * correlated to the rcutorture messages, even when the rcutorture module
269 * is being repeatedly loaded and unloaded. In other words, we cannot
270 * store this state in rcutorture itself.
272 void rcutorture_record_test_transition(void)
274 rcutorture_testseq++;
275 rcutorture_vernum = 0;
277 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
280 * Record the number of writer passes through the current rcutorture test.
281 * This is also used to correlate debugfs tracing stats with the rcutorture
284 void rcutorture_record_progress(unsigned long vernum)
288 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
291 * Force a quiescent state for RCU-sched.
293 void rcu_sched_force_quiescent_state(void)
295 force_quiescent_state(&rcu_sched_state);
297 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
300 * Does the CPU have callbacks ready to be invoked?
303 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
305 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
309 * Does the current CPU require a yet-as-unscheduled grace period?
312 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
314 return *rdp->nxttail[RCU_DONE_TAIL +
315 ACCESS_ONCE(rsp->completed) != rdp->completed] &&
316 !rcu_gp_in_progress(rsp);
320 * Return the root node of the specified rcu_state structure.
322 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
324 return &rsp->node[0];
328 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
330 * If the new value of the ->dynticks_nesting counter now is zero,
331 * we really have entered idle, and must do the appropriate accounting.
332 * The caller must have disabled interrupts.
334 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
337 trace_rcu_dyntick("Start", oldval, 0);
338 if (!is_idle_task(current) && !user) {
339 struct task_struct *idle = idle_task(smp_processor_id());
341 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
342 ftrace_dump(DUMP_ORIG);
343 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
344 current->pid, current->comm,
345 idle->pid, idle->comm); /* must be idle task! */
347 rcu_prepare_for_idle(smp_processor_id());
348 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
349 smp_mb__before_atomic_inc(); /* See above. */
350 atomic_inc(&rdtp->dynticks);
351 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
352 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
355 * It is illegal to enter an extended quiescent state while
356 * in an RCU read-side critical section.
358 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
359 "Illegal idle entry in RCU read-side critical section.");
360 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
361 "Illegal idle entry in RCU-bh read-side critical section.");
362 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
363 "Illegal idle entry in RCU-sched read-side critical section.");
367 * Enter an RCU extended quiescent state, which can be either the
368 * idle loop or adaptive-tickless usermode execution.
370 static void rcu_eqs_enter(bool user)
373 struct rcu_dynticks *rdtp;
375 rdtp = &__get_cpu_var(rcu_dynticks);
376 oldval = rdtp->dynticks_nesting;
377 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
378 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
379 rdtp->dynticks_nesting = 0;
381 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
382 rcu_eqs_enter_common(rdtp, oldval, user);
386 * rcu_idle_enter - inform RCU that current CPU is entering idle
388 * Enter idle mode, in other words, -leave- the mode in which RCU
389 * read-side critical sections can occur. (Though RCU read-side
390 * critical sections can occur in irq handlers in idle, a possibility
391 * handled by irq_enter() and irq_exit().)
393 * We crowbar the ->dynticks_nesting field to zero to allow for
394 * the possibility of usermode upcalls having messed up our count
395 * of interrupt nesting level during the prior busy period.
397 void rcu_idle_enter(void)
401 local_irq_save(flags);
403 local_irq_restore(flags);
405 EXPORT_SYMBOL_GPL(rcu_idle_enter);
407 #ifdef CONFIG_RCU_USER_QS
409 * rcu_user_enter - inform RCU that we are resuming userspace.
411 * Enter RCU idle mode right before resuming userspace. No use of RCU
412 * is permitted between this call and rcu_user_exit(). This way the
413 * CPU doesn't need to maintain the tick for RCU maintenance purposes
414 * when the CPU runs in userspace.
416 void rcu_user_enter(void)
419 struct rcu_dynticks *rdtp;
422 * Some contexts may involve an exception occuring in an irq,
423 * leading to that nesting:
424 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
425 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
426 * helpers are enough to protect RCU uses inside the exception. So
427 * just return immediately if we detect we are in an IRQ.
432 WARN_ON_ONCE(!current->mm);
434 local_irq_save(flags);
435 rdtp = &__get_cpu_var(rcu_dynticks);
436 if (!rdtp->ignore_user_qs && !rdtp->in_user) {
437 rdtp->in_user = true;
440 local_irq_restore(flags);
444 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
445 * after the current irq returns.
447 * This is similar to rcu_user_enter() but in the context of a non-nesting
448 * irq. After this call, RCU enters into idle mode when the interrupt
451 void rcu_user_enter_after_irq(void)
454 struct rcu_dynticks *rdtp;
456 local_irq_save(flags);
457 rdtp = &__get_cpu_var(rcu_dynticks);
458 /* Ensure this irq is interrupting a non-idle RCU state. */
459 WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
460 rdtp->dynticks_nesting = 1;
461 local_irq_restore(flags);
463 #endif /* CONFIG_RCU_USER_QS */
466 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
468 * Exit from an interrupt handler, which might possibly result in entering
469 * idle mode, in other words, leaving the mode in which read-side critical
470 * sections can occur.
472 * This code assumes that the idle loop never does anything that might
473 * result in unbalanced calls to irq_enter() and irq_exit(). If your
474 * architecture violates this assumption, RCU will give you what you
475 * deserve, good and hard. But very infrequently and irreproducibly.
477 * Use things like work queues to work around this limitation.
479 * You have been warned.
481 void rcu_irq_exit(void)
485 struct rcu_dynticks *rdtp;
487 local_irq_save(flags);
488 rdtp = &__get_cpu_var(rcu_dynticks);
489 oldval = rdtp->dynticks_nesting;
490 rdtp->dynticks_nesting--;
491 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
492 if (rdtp->dynticks_nesting)
493 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
495 rcu_eqs_enter_common(rdtp, oldval, 1);
496 local_irq_restore(flags);
500 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
502 * If the new value of the ->dynticks_nesting counter was previously zero,
503 * we really have exited idle, and must do the appropriate accounting.
504 * The caller must have disabled interrupts.
506 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
509 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
510 atomic_inc(&rdtp->dynticks);
511 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
512 smp_mb__after_atomic_inc(); /* See above. */
513 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
514 rcu_cleanup_after_idle(smp_processor_id());
515 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
516 if (!is_idle_task(current) && !user) {
517 struct task_struct *idle = idle_task(smp_processor_id());
519 trace_rcu_dyntick("Error on exit: not idle task",
520 oldval, rdtp->dynticks_nesting);
521 ftrace_dump(DUMP_ORIG);
522 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
523 current->pid, current->comm,
524 idle->pid, idle->comm); /* must be idle task! */
529 * Exit an RCU extended quiescent state, which can be either the
530 * idle loop or adaptive-tickless usermode execution.
532 static void rcu_eqs_exit(bool user)
534 struct rcu_dynticks *rdtp;
537 rdtp = &__get_cpu_var(rcu_dynticks);
538 oldval = rdtp->dynticks_nesting;
539 WARN_ON_ONCE(oldval < 0);
540 if (oldval & DYNTICK_TASK_NEST_MASK)
541 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
543 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
544 rcu_eqs_exit_common(rdtp, oldval, user);
548 * rcu_idle_exit - inform RCU that current CPU is leaving idle
550 * Exit idle mode, in other words, -enter- the mode in which RCU
551 * read-side critical sections can occur.
553 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
554 * allow for the possibility of usermode upcalls messing up our count
555 * of interrupt nesting level during the busy period that is just
558 void rcu_idle_exit(void)
562 local_irq_save(flags);
564 local_irq_restore(flags);
566 EXPORT_SYMBOL_GPL(rcu_idle_exit);
568 #ifdef CONFIG_RCU_USER_QS
570 * rcu_user_exit - inform RCU that we are exiting userspace.
572 * Exit RCU idle mode while entering the kernel because it can
573 * run a RCU read side critical section anytime.
575 void rcu_user_exit(void)
578 struct rcu_dynticks *rdtp;
581 * Some contexts may involve an exception occuring in an irq,
582 * leading to that nesting:
583 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
584 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
585 * helpers are enough to protect RCU uses inside the exception. So
586 * just return immediately if we detect we are in an IRQ.
591 local_irq_save(flags);
592 rdtp = &__get_cpu_var(rcu_dynticks);
594 rdtp->in_user = false;
597 local_irq_restore(flags);
601 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
602 * idle mode after the current non-nesting irq returns.
604 * This is similar to rcu_user_exit() but in the context of an irq.
605 * This is called when the irq has interrupted a userspace RCU idle mode
606 * context. When the current non-nesting interrupt returns after this call,
607 * the CPU won't restore the RCU idle mode.
609 void rcu_user_exit_after_irq(void)
612 struct rcu_dynticks *rdtp;
614 local_irq_save(flags);
615 rdtp = &__get_cpu_var(rcu_dynticks);
616 /* Ensure we are interrupting an RCU idle mode. */
617 WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
618 rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
619 local_irq_restore(flags);
621 #endif /* CONFIG_RCU_USER_QS */
624 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
626 * Enter an interrupt handler, which might possibly result in exiting
627 * idle mode, in other words, entering the mode in which read-side critical
628 * sections can occur.
630 * Note that the Linux kernel is fully capable of entering an interrupt
631 * handler that it never exits, for example when doing upcalls to
632 * user mode! This code assumes that the idle loop never does upcalls to
633 * user mode. If your architecture does do upcalls from the idle loop (or
634 * does anything else that results in unbalanced calls to the irq_enter()
635 * and irq_exit() functions), RCU will give you what you deserve, good
636 * and hard. But very infrequently and irreproducibly.
638 * Use things like work queues to work around this limitation.
640 * You have been warned.
642 void rcu_irq_enter(void)
645 struct rcu_dynticks *rdtp;
648 local_irq_save(flags);
649 rdtp = &__get_cpu_var(rcu_dynticks);
650 oldval = rdtp->dynticks_nesting;
651 rdtp->dynticks_nesting++;
652 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
654 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
656 rcu_eqs_exit_common(rdtp, oldval, 1);
657 local_irq_restore(flags);
661 * rcu_nmi_enter - inform RCU of entry to NMI context
663 * If the CPU was idle with dynamic ticks active, and there is no
664 * irq handler running, this updates rdtp->dynticks_nmi to let the
665 * RCU grace-period handling know that the CPU is active.
667 void rcu_nmi_enter(void)
669 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
671 if (rdtp->dynticks_nmi_nesting == 0 &&
672 (atomic_read(&rdtp->dynticks) & 0x1))
674 rdtp->dynticks_nmi_nesting++;
675 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
676 atomic_inc(&rdtp->dynticks);
677 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
678 smp_mb__after_atomic_inc(); /* See above. */
679 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
683 * rcu_nmi_exit - inform RCU of exit from NMI context
685 * If the CPU was idle with dynamic ticks active, and there is no
686 * irq handler running, this updates rdtp->dynticks_nmi to let the
687 * RCU grace-period handling know that the CPU is no longer active.
689 void rcu_nmi_exit(void)
691 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
693 if (rdtp->dynticks_nmi_nesting == 0 ||
694 --rdtp->dynticks_nmi_nesting != 0)
696 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
697 smp_mb__before_atomic_inc(); /* See above. */
698 atomic_inc(&rdtp->dynticks);
699 smp_mb__after_atomic_inc(); /* Force delay to next write. */
700 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
704 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
706 * If the current CPU is in its idle loop and is neither in an interrupt
707 * or NMI handler, return true.
709 int rcu_is_cpu_idle(void)
714 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
718 EXPORT_SYMBOL(rcu_is_cpu_idle);
720 #ifdef CONFIG_RCU_USER_QS
721 void rcu_user_hooks_switch(struct task_struct *prev,
722 struct task_struct *next)
724 struct rcu_dynticks *rdtp;
726 /* Interrupts are disabled in context switch */
727 rdtp = &__get_cpu_var(rcu_dynticks);
728 if (!rdtp->ignore_user_qs) {
729 clear_tsk_thread_flag(prev, TIF_NOHZ);
730 set_tsk_thread_flag(next, TIF_NOHZ);
733 #endif /* #ifdef CONFIG_RCU_USER_QS */
735 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
738 * Is the current CPU online? Disable preemption to avoid false positives
739 * that could otherwise happen due to the current CPU number being sampled,
740 * this task being preempted, its old CPU being taken offline, resuming
741 * on some other CPU, then determining that its old CPU is now offline.
742 * It is OK to use RCU on an offline processor during initial boot, hence
743 * the check for rcu_scheduler_fully_active. Note also that it is OK
744 * for a CPU coming online to use RCU for one jiffy prior to marking itself
745 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
746 * offline to continue to use RCU for one jiffy after marking itself
747 * offline in the cpu_online_mask. This leniency is necessary given the
748 * non-atomic nature of the online and offline processing, for example,
749 * the fact that a CPU enters the scheduler after completing the CPU_DYING
752 * This is also why RCU internally marks CPUs online during the
753 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
755 * Disable checking if in an NMI handler because we cannot safely report
756 * errors from NMI handlers anyway.
758 bool rcu_lockdep_current_cpu_online(void)
760 struct rcu_data *rdp;
761 struct rcu_node *rnp;
767 rdp = &__get_cpu_var(rcu_sched_data);
769 ret = (rdp->grpmask & rnp->qsmaskinit) ||
770 !rcu_scheduler_fully_active;
774 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
776 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
779 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
781 * If the current CPU is idle or running at a first-level (not nested)
782 * interrupt from idle, return true. The caller must have at least
783 * disabled preemption.
785 int rcu_is_cpu_rrupt_from_idle(void)
787 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
791 * Snapshot the specified CPU's dynticks counter so that we can later
792 * credit them with an implicit quiescent state. Return 1 if this CPU
793 * is in dynticks idle mode, which is an extended quiescent state.
795 static int dyntick_save_progress_counter(struct rcu_data *rdp)
797 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
798 return (rdp->dynticks_snap & 0x1) == 0;
802 * Return true if the specified CPU has passed through a quiescent
803 * state by virtue of being in or having passed through an dynticks
804 * idle state since the last call to dyntick_save_progress_counter()
805 * for this same CPU, or by virtue of having been offline.
807 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
812 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
813 snap = (unsigned int)rdp->dynticks_snap;
816 * If the CPU passed through or entered a dynticks idle phase with
817 * no active irq/NMI handlers, then we can safely pretend that the CPU
818 * already acknowledged the request to pass through a quiescent
819 * state. Either way, that CPU cannot possibly be in an RCU
820 * read-side critical section that started before the beginning
821 * of the current RCU grace period.
823 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
824 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
830 * Check for the CPU being offline, but only if the grace period
831 * is old enough. We don't need to worry about the CPU changing
832 * state: If we see it offline even once, it has been through a
835 * The reason for insisting that the grace period be at least
836 * one jiffy old is that CPUs that are not quite online and that
837 * have just gone offline can still execute RCU read-side critical
840 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
841 return 0; /* Grace period is not old enough. */
843 if (cpu_is_offline(rdp->cpu)) {
844 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
851 static int jiffies_till_stall_check(void)
853 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
856 * Limit check must be consistent with the Kconfig limits
857 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
859 if (till_stall_check < 3) {
860 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
861 till_stall_check = 3;
862 } else if (till_stall_check > 300) {
863 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
864 till_stall_check = 300;
866 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
869 static void record_gp_stall_check_time(struct rcu_state *rsp)
871 rsp->gp_start = jiffies;
872 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
875 static void print_other_cpu_stall(struct rcu_state *rsp)
881 struct rcu_node *rnp = rcu_get_root(rsp);
883 /* Only let one CPU complain about others per time interval. */
885 raw_spin_lock_irqsave(&rnp->lock, flags);
886 delta = jiffies - rsp->jiffies_stall;
887 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
888 raw_spin_unlock_irqrestore(&rnp->lock, flags);
891 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
892 raw_spin_unlock_irqrestore(&rnp->lock, flags);
895 * OK, time to rat on our buddy...
896 * See Documentation/RCU/stallwarn.txt for info on how to debug
897 * RCU CPU stall warnings.
899 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
901 print_cpu_stall_info_begin();
902 rcu_for_each_leaf_node(rsp, rnp) {
903 raw_spin_lock_irqsave(&rnp->lock, flags);
904 ndetected += rcu_print_task_stall(rnp);
905 if (rnp->qsmask != 0) {
906 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
907 if (rnp->qsmask & (1UL << cpu)) {
908 print_cpu_stall_info(rsp,
913 raw_spin_unlock_irqrestore(&rnp->lock, flags);
917 * Now rat on any tasks that got kicked up to the root rcu_node
918 * due to CPU offlining.
920 rnp = rcu_get_root(rsp);
921 raw_spin_lock_irqsave(&rnp->lock, flags);
922 ndetected += rcu_print_task_stall(rnp);
923 raw_spin_unlock_irqrestore(&rnp->lock, flags);
925 print_cpu_stall_info_end();
926 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
927 smp_processor_id(), (long)(jiffies - rsp->gp_start));
929 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
930 else if (!trigger_all_cpu_backtrace())
933 /* Complain about tasks blocking the grace period. */
935 rcu_print_detail_task_stall(rsp);
937 force_quiescent_state(rsp); /* Kick them all. */
940 static void print_cpu_stall(struct rcu_state *rsp)
943 struct rcu_node *rnp = rcu_get_root(rsp);
946 * OK, time to rat on ourselves...
947 * See Documentation/RCU/stallwarn.txt for info on how to debug
948 * RCU CPU stall warnings.
950 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
951 print_cpu_stall_info_begin();
952 print_cpu_stall_info(rsp, smp_processor_id());
953 print_cpu_stall_info_end();
954 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
955 if (!trigger_all_cpu_backtrace())
958 raw_spin_lock_irqsave(&rnp->lock, flags);
959 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
960 rsp->jiffies_stall = jiffies +
961 3 * jiffies_till_stall_check() + 3;
962 raw_spin_unlock_irqrestore(&rnp->lock, flags);
964 set_need_resched(); /* kick ourselves to get things going. */
967 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
971 struct rcu_node *rnp;
973 if (rcu_cpu_stall_suppress)
975 j = ACCESS_ONCE(jiffies);
976 js = ACCESS_ONCE(rsp->jiffies_stall);
978 if (rcu_gp_in_progress(rsp) &&
979 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
981 /* We haven't checked in, so go dump stack. */
982 print_cpu_stall(rsp);
984 } else if (rcu_gp_in_progress(rsp) &&
985 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
987 /* They had a few time units to dump stack, so complain. */
988 print_other_cpu_stall(rsp);
992 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
994 rcu_cpu_stall_suppress = 1;
999 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1001 * Set the stall-warning timeout way off into the future, thus preventing
1002 * any RCU CPU stall-warning messages from appearing in the current set of
1003 * RCU grace periods.
1005 * The caller must disable hard irqs.
1007 void rcu_cpu_stall_reset(void)
1009 struct rcu_state *rsp;
1011 for_each_rcu_flavor(rsp)
1012 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1015 static struct notifier_block rcu_panic_block = {
1016 .notifier_call = rcu_panic,
1019 static void __init check_cpu_stall_init(void)
1021 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
1025 * Update CPU-local rcu_data state to record the newly noticed grace period.
1026 * This is used both when we started the grace period and when we notice
1027 * that someone else started the grace period. The caller must hold the
1028 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
1029 * and must have irqs disabled.
1031 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1033 if (rdp->gpnum != rnp->gpnum) {
1035 * If the current grace period is waiting for this CPU,
1036 * set up to detect a quiescent state, otherwise don't
1037 * go looking for one.
1039 rdp->gpnum = rnp->gpnum;
1040 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
1041 rdp->passed_quiesce = 0;
1042 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1043 zero_cpu_stall_ticks(rdp);
1047 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1049 unsigned long flags;
1050 struct rcu_node *rnp;
1052 local_irq_save(flags);
1054 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1055 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1056 local_irq_restore(flags);
1059 __note_new_gpnum(rsp, rnp, rdp);
1060 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1064 * Did someone else start a new RCU grace period start since we last
1065 * checked? Update local state appropriately if so. Must be called
1066 * on the CPU corresponding to rdp.
1069 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1071 unsigned long flags;
1074 local_irq_save(flags);
1075 if (rdp->gpnum != rsp->gpnum) {
1076 note_new_gpnum(rsp, rdp);
1079 local_irq_restore(flags);
1084 * Initialize the specified rcu_data structure's callback list to empty.
1086 static void init_callback_list(struct rcu_data *rdp)
1090 rdp->nxtlist = NULL;
1091 for (i = 0; i < RCU_NEXT_SIZE; i++)
1092 rdp->nxttail[i] = &rdp->nxtlist;
1096 * Advance this CPU's callbacks, but only if the current grace period
1097 * has ended. This may be called only from the CPU to whom the rdp
1098 * belongs. In addition, the corresponding leaf rcu_node structure's
1099 * ->lock must be held by the caller, with irqs disabled.
1102 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1104 /* Did another grace period end? */
1105 if (rdp->completed != rnp->completed) {
1107 /* Advance callbacks. No harm if list empty. */
1108 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
1109 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
1110 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1112 /* Remember that we saw this grace-period completion. */
1113 rdp->completed = rnp->completed;
1114 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1117 * If we were in an extended quiescent state, we may have
1118 * missed some grace periods that others CPUs handled on
1119 * our behalf. Catch up with this state to avoid noting
1120 * spurious new grace periods. If another grace period
1121 * has started, then rnp->gpnum will have advanced, so
1122 * we will detect this later on. Of course, any quiescent
1123 * states we found for the old GP are now invalid.
1125 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1126 rdp->gpnum = rdp->completed;
1127 rdp->passed_quiesce = 0;
1131 * If RCU does not need a quiescent state from this CPU,
1132 * then make sure that this CPU doesn't go looking for one.
1134 if ((rnp->qsmask & rdp->grpmask) == 0)
1135 rdp->qs_pending = 0;
1140 * Advance this CPU's callbacks, but only if the current grace period
1141 * has ended. This may be called only from the CPU to whom the rdp
1145 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1147 unsigned long flags;
1148 struct rcu_node *rnp;
1150 local_irq_save(flags);
1152 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1153 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1154 local_irq_restore(flags);
1157 __rcu_process_gp_end(rsp, rnp, rdp);
1158 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1162 * Do per-CPU grace-period initialization for running CPU. The caller
1163 * must hold the lock of the leaf rcu_node structure corresponding to
1167 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1169 /* Prior grace period ended, so advance callbacks for current CPU. */
1170 __rcu_process_gp_end(rsp, rnp, rdp);
1172 /* Set state so that this CPU will detect the next quiescent state. */
1173 __note_new_gpnum(rsp, rnp, rdp);
1177 * Initialize a new grace period.
1179 static int rcu_gp_init(struct rcu_state *rsp)
1181 struct rcu_data *rdp;
1182 struct rcu_node *rnp = rcu_get_root(rsp);
1184 raw_spin_lock_irq(&rnp->lock);
1185 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1187 if (rcu_gp_in_progress(rsp)) {
1188 /* Grace period already in progress, don't start another. */
1189 raw_spin_unlock_irq(&rnp->lock);
1193 /* Advance to a new grace period and initialize state. */
1195 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1196 record_gp_stall_check_time(rsp);
1197 raw_spin_unlock_irq(&rnp->lock);
1199 /* Exclude any concurrent CPU-hotplug operations. */
1203 * Set the quiescent-state-needed bits in all the rcu_node
1204 * structures for all currently online CPUs in breadth-first order,
1205 * starting from the root rcu_node structure, relying on the layout
1206 * of the tree within the rsp->node[] array. Note that other CPUs
1207 * will access only the leaves of the hierarchy, thus seeing that no
1208 * grace period is in progress, at least until the corresponding
1209 * leaf node has been initialized. In addition, we have excluded
1210 * CPU-hotplug operations.
1212 * The grace period cannot complete until the initialization
1213 * process finishes, because this kthread handles both.
1215 rcu_for_each_node_breadth_first(rsp, rnp) {
1216 raw_spin_lock_irq(&rnp->lock);
1217 rdp = this_cpu_ptr(rsp->rda);
1218 rcu_preempt_check_blocked_tasks(rnp);
1219 rnp->qsmask = rnp->qsmaskinit;
1220 rnp->gpnum = rsp->gpnum;
1221 WARN_ON_ONCE(rnp->completed != rsp->completed);
1222 rnp->completed = rsp->completed;
1223 if (rnp == rdp->mynode)
1224 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1225 rcu_preempt_boost_start_gp(rnp);
1226 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1227 rnp->level, rnp->grplo,
1228 rnp->grphi, rnp->qsmask);
1229 raw_spin_unlock_irq(&rnp->lock);
1230 #ifdef CONFIG_PROVE_RCU_DELAY
1231 if ((random32() % (rcu_num_nodes * 8)) == 0)
1232 schedule_timeout_uninterruptible(2);
1233 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1242 * Do one round of quiescent-state forcing.
1244 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1246 int fqs_state = fqs_state_in;
1247 struct rcu_node *rnp = rcu_get_root(rsp);
1250 if (fqs_state == RCU_SAVE_DYNTICK) {
1251 /* Collect dyntick-idle snapshots. */
1252 force_qs_rnp(rsp, dyntick_save_progress_counter);
1253 fqs_state = RCU_FORCE_QS;
1255 /* Handle dyntick-idle and offline CPUs. */
1256 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1258 /* Clear flag to prevent immediate re-entry. */
1259 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1260 raw_spin_lock_irq(&rnp->lock);
1261 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1262 raw_spin_unlock_irq(&rnp->lock);
1268 * Clean up after the old grace period.
1270 static void rcu_gp_cleanup(struct rcu_state *rsp)
1272 unsigned long gp_duration;
1273 struct rcu_data *rdp;
1274 struct rcu_node *rnp = rcu_get_root(rsp);
1276 raw_spin_lock_irq(&rnp->lock);
1277 gp_duration = jiffies - rsp->gp_start;
1278 if (gp_duration > rsp->gp_max)
1279 rsp->gp_max = gp_duration;
1282 * We know the grace period is complete, but to everyone else
1283 * it appears to still be ongoing. But it is also the case
1284 * that to everyone else it looks like there is nothing that
1285 * they can do to advance the grace period. It is therefore
1286 * safe for us to drop the lock in order to mark the grace
1287 * period as completed in all of the rcu_node structures.
1289 raw_spin_unlock_irq(&rnp->lock);
1292 * Propagate new ->completed value to rcu_node structures so
1293 * that other CPUs don't have to wait until the start of the next
1294 * grace period to process their callbacks. This also avoids
1295 * some nasty RCU grace-period initialization races by forcing
1296 * the end of the current grace period to be completely recorded in
1297 * all of the rcu_node structures before the beginning of the next
1298 * grace period is recorded in any of the rcu_node structures.
1300 rcu_for_each_node_breadth_first(rsp, rnp) {
1301 raw_spin_lock_irq(&rnp->lock);
1302 rnp->completed = rsp->gpnum;
1303 raw_spin_unlock_irq(&rnp->lock);
1306 rnp = rcu_get_root(rsp);
1307 raw_spin_lock_irq(&rnp->lock);
1309 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1310 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1311 rsp->fqs_state = RCU_GP_IDLE;
1312 rdp = this_cpu_ptr(rsp->rda);
1313 if (cpu_needs_another_gp(rsp, rdp))
1315 raw_spin_unlock_irq(&rnp->lock);
1319 * Body of kthread that handles grace periods.
1321 static int __noreturn rcu_gp_kthread(void *arg)
1326 struct rcu_state *rsp = arg;
1327 struct rcu_node *rnp = rcu_get_root(rsp);
1331 /* Handle grace-period start. */
1333 wait_event_interruptible(rsp->gp_wq,
1336 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1340 flush_signals(current);
1343 /* Handle quiescent-state forcing. */
1344 fqs_state = RCU_SAVE_DYNTICK;
1345 j = jiffies_till_first_fqs;
1348 jiffies_till_first_fqs = HZ;
1351 rsp->jiffies_force_qs = jiffies + j;
1352 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1353 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1354 (!ACCESS_ONCE(rnp->qsmask) &&
1355 !rcu_preempt_blocked_readers_cgp(rnp)),
1357 /* If grace period done, leave loop. */
1358 if (!ACCESS_ONCE(rnp->qsmask) &&
1359 !rcu_preempt_blocked_readers_cgp(rnp))
1361 /* If time for quiescent-state forcing, do it. */
1362 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1363 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1366 /* Deal with stray signal. */
1368 flush_signals(current);
1370 j = jiffies_till_next_fqs;
1373 jiffies_till_next_fqs = HZ;
1376 jiffies_till_next_fqs = 1;
1380 /* Handle grace-period end. */
1381 rcu_gp_cleanup(rsp);
1386 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1387 * in preparation for detecting the next grace period. The caller must hold
1388 * the root node's ->lock, which is released before return. Hard irqs must
1391 * Note that it is legal for a dying CPU (which is marked as offline) to
1392 * invoke this function. This can happen when the dying CPU reports its
1396 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1397 __releases(rcu_get_root(rsp)->lock)
1399 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1400 struct rcu_node *rnp = rcu_get_root(rsp);
1402 if (!rsp->gp_kthread ||
1403 !cpu_needs_another_gp(rsp, rdp)) {
1405 * Either we have not yet spawned the grace-period
1406 * task or this CPU does not need another grace period.
1407 * Either way, don't start a new grace period.
1409 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1413 rsp->gp_flags = RCU_GP_FLAG_INIT;
1414 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1415 wake_up(&rsp->gp_wq);
1419 * Report a full set of quiescent states to the specified rcu_state
1420 * data structure. This involves cleaning up after the prior grace
1421 * period and letting rcu_start_gp() start up the next grace period
1422 * if one is needed. Note that the caller must hold rnp->lock, as
1423 * required by rcu_start_gp(), which will release it.
1425 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1426 __releases(rcu_get_root(rsp)->lock)
1428 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1429 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1430 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1434 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1435 * Allows quiescent states for a group of CPUs to be reported at one go
1436 * to the specified rcu_node structure, though all the CPUs in the group
1437 * must be represented by the same rcu_node structure (which need not be
1438 * a leaf rcu_node structure, though it often will be). That structure's
1439 * lock must be held upon entry, and it is released before return.
1442 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1443 struct rcu_node *rnp, unsigned long flags)
1444 __releases(rnp->lock)
1446 struct rcu_node *rnp_c;
1448 /* Walk up the rcu_node hierarchy. */
1450 if (!(rnp->qsmask & mask)) {
1452 /* Our bit has already been cleared, so done. */
1453 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1456 rnp->qsmask &= ~mask;
1457 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1458 mask, rnp->qsmask, rnp->level,
1459 rnp->grplo, rnp->grphi,
1461 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1463 /* Other bits still set at this level, so done. */
1464 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1467 mask = rnp->grpmask;
1468 if (rnp->parent == NULL) {
1470 /* No more levels. Exit loop holding root lock. */
1474 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1477 raw_spin_lock_irqsave(&rnp->lock, flags);
1478 WARN_ON_ONCE(rnp_c->qsmask);
1482 * Get here if we are the last CPU to pass through a quiescent
1483 * state for this grace period. Invoke rcu_report_qs_rsp()
1484 * to clean up and start the next grace period if one is needed.
1486 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1490 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1491 * structure. This must be either called from the specified CPU, or
1492 * called when the specified CPU is known to be offline (and when it is
1493 * also known that no other CPU is concurrently trying to help the offline
1494 * CPU). The lastcomp argument is used to make sure we are still in the
1495 * grace period of interest. We don't want to end the current grace period
1496 * based on quiescent states detected in an earlier grace period!
1499 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1501 unsigned long flags;
1503 struct rcu_node *rnp;
1506 raw_spin_lock_irqsave(&rnp->lock, flags);
1507 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1508 rnp->completed == rnp->gpnum) {
1511 * The grace period in which this quiescent state was
1512 * recorded has ended, so don't report it upwards.
1513 * We will instead need a new quiescent state that lies
1514 * within the current grace period.
1516 rdp->passed_quiesce = 0; /* need qs for new gp. */
1517 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1520 mask = rdp->grpmask;
1521 if ((rnp->qsmask & mask) == 0) {
1522 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1524 rdp->qs_pending = 0;
1527 * This GP can't end until cpu checks in, so all of our
1528 * callbacks can be processed during the next GP.
1530 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1532 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1537 * Check to see if there is a new grace period of which this CPU
1538 * is not yet aware, and if so, set up local rcu_data state for it.
1539 * Otherwise, see if this CPU has just passed through its first
1540 * quiescent state for this grace period, and record that fact if so.
1543 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1545 /* If there is now a new grace period, record and return. */
1546 if (check_for_new_grace_period(rsp, rdp))
1550 * Does this CPU still need to do its part for current grace period?
1551 * If no, return and let the other CPUs do their part as well.
1553 if (!rdp->qs_pending)
1557 * Was there a quiescent state since the beginning of the grace
1558 * period? If no, then exit and wait for the next call.
1560 if (!rdp->passed_quiesce)
1564 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1567 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1570 #ifdef CONFIG_HOTPLUG_CPU
1573 * Send the specified CPU's RCU callbacks to the orphanage. The
1574 * specified CPU must be offline, and the caller must hold the
1578 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1579 struct rcu_node *rnp, struct rcu_data *rdp)
1582 * Orphan the callbacks. First adjust the counts. This is safe
1583 * because ->onofflock excludes _rcu_barrier()'s adoption of
1584 * the callbacks, thus no memory barrier is required.
1586 if (rdp->nxtlist != NULL) {
1587 rsp->qlen_lazy += rdp->qlen_lazy;
1588 rsp->qlen += rdp->qlen;
1589 rdp->n_cbs_orphaned += rdp->qlen;
1591 ACCESS_ONCE(rdp->qlen) = 0;
1595 * Next, move those callbacks still needing a grace period to
1596 * the orphanage, where some other CPU will pick them up.
1597 * Some of the callbacks might have gone partway through a grace
1598 * period, but that is too bad. They get to start over because we
1599 * cannot assume that grace periods are synchronized across CPUs.
1600 * We don't bother updating the ->nxttail[] array yet, instead
1601 * we just reset the whole thing later on.
1603 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1604 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1605 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1606 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1610 * Then move the ready-to-invoke callbacks to the orphanage,
1611 * where some other CPU will pick them up. These will not be
1612 * required to pass though another grace period: They are done.
1614 if (rdp->nxtlist != NULL) {
1615 *rsp->orphan_donetail = rdp->nxtlist;
1616 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1619 /* Finally, initialize the rcu_data structure's list to empty. */
1620 init_callback_list(rdp);
1624 * Adopt the RCU callbacks from the specified rcu_state structure's
1625 * orphanage. The caller must hold the ->onofflock.
1627 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1630 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1632 /* Do the accounting first. */
1633 rdp->qlen_lazy += rsp->qlen_lazy;
1634 rdp->qlen += rsp->qlen;
1635 rdp->n_cbs_adopted += rsp->qlen;
1636 if (rsp->qlen_lazy != rsp->qlen)
1637 rcu_idle_count_callbacks_posted();
1642 * We do not need a memory barrier here because the only way we
1643 * can get here if there is an rcu_barrier() in flight is if
1644 * we are the task doing the rcu_barrier().
1647 /* First adopt the ready-to-invoke callbacks. */
1648 if (rsp->orphan_donelist != NULL) {
1649 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1650 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1651 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1652 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1653 rdp->nxttail[i] = rsp->orphan_donetail;
1654 rsp->orphan_donelist = NULL;
1655 rsp->orphan_donetail = &rsp->orphan_donelist;
1658 /* And then adopt the callbacks that still need a grace period. */
1659 if (rsp->orphan_nxtlist != NULL) {
1660 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1661 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1662 rsp->orphan_nxtlist = NULL;
1663 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1668 * Trace the fact that this CPU is going offline.
1670 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1672 RCU_TRACE(unsigned long mask);
1673 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1674 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1676 RCU_TRACE(mask = rdp->grpmask);
1677 trace_rcu_grace_period(rsp->name,
1678 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1683 * The CPU has been completely removed, and some other CPU is reporting
1684 * this fact from process context. Do the remainder of the cleanup,
1685 * including orphaning the outgoing CPU's RCU callbacks, and also
1686 * adopting them. There can only be one CPU hotplug operation at a time,
1687 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1689 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1691 unsigned long flags;
1693 int need_report = 0;
1694 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1695 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1697 /* Adjust any no-longer-needed kthreads. */
1698 rcu_boost_kthread_setaffinity(rnp, -1);
1700 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1702 /* Exclude any attempts to start a new grace period. */
1703 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1705 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1706 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1707 rcu_adopt_orphan_cbs(rsp);
1709 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1710 mask = rdp->grpmask; /* rnp->grplo is constant. */
1712 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1713 rnp->qsmaskinit &= ~mask;
1714 if (rnp->qsmaskinit != 0) {
1715 if (rnp != rdp->mynode)
1716 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1719 if (rnp == rdp->mynode)
1720 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1722 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1723 mask = rnp->grpmask;
1725 } while (rnp != NULL);
1728 * We still hold the leaf rcu_node structure lock here, and
1729 * irqs are still disabled. The reason for this subterfuge is
1730 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1731 * held leads to deadlock.
1733 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1735 if (need_report & RCU_OFL_TASKS_NORM_GP)
1736 rcu_report_unblock_qs_rnp(rnp, flags);
1738 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1739 if (need_report & RCU_OFL_TASKS_EXP_GP)
1740 rcu_report_exp_rnp(rsp, rnp, true);
1741 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1742 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1743 cpu, rdp->qlen, rdp->nxtlist);
1744 init_callback_list(rdp);
1745 /* Disallow further callbacks on this CPU. */
1746 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1749 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1751 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1755 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1759 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1762 * Invoke any RCU callbacks that have made it to the end of their grace
1763 * period. Thottle as specified by rdp->blimit.
1765 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1767 unsigned long flags;
1768 struct rcu_head *next, *list, **tail;
1769 int bl, count, count_lazy, i;
1771 /* If no callbacks are ready, just return.*/
1772 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1773 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1774 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1775 need_resched(), is_idle_task(current),
1776 rcu_is_callbacks_kthread());
1781 * Extract the list of ready callbacks, disabling to prevent
1782 * races with call_rcu() from interrupt handlers.
1784 local_irq_save(flags);
1785 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1787 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1788 list = rdp->nxtlist;
1789 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1790 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1791 tail = rdp->nxttail[RCU_DONE_TAIL];
1792 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1793 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1794 rdp->nxttail[i] = &rdp->nxtlist;
1795 local_irq_restore(flags);
1797 /* Invoke callbacks. */
1798 count = count_lazy = 0;
1802 debug_rcu_head_unqueue(list);
1803 if (__rcu_reclaim(rsp->name, list))
1806 /* Stop only if limit reached and CPU has something to do. */
1807 if (++count >= bl &&
1809 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1813 local_irq_save(flags);
1814 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1815 is_idle_task(current),
1816 rcu_is_callbacks_kthread());
1818 /* Update count, and requeue any remaining callbacks. */
1820 *tail = rdp->nxtlist;
1821 rdp->nxtlist = list;
1822 for (i = 0; i < RCU_NEXT_SIZE; i++)
1823 if (&rdp->nxtlist == rdp->nxttail[i])
1824 rdp->nxttail[i] = tail;
1828 smp_mb(); /* List handling before counting for rcu_barrier(). */
1829 rdp->qlen_lazy -= count_lazy;
1830 ACCESS_ONCE(rdp->qlen) -= count;
1831 rdp->n_cbs_invoked += count;
1833 /* Reinstate batch limit if we have worked down the excess. */
1834 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1835 rdp->blimit = blimit;
1837 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1838 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1839 rdp->qlen_last_fqs_check = 0;
1840 rdp->n_force_qs_snap = rsp->n_force_qs;
1841 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1842 rdp->qlen_last_fqs_check = rdp->qlen;
1843 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1845 local_irq_restore(flags);
1847 /* Re-invoke RCU core processing if there are callbacks remaining. */
1848 if (cpu_has_callbacks_ready_to_invoke(rdp))
1853 * Check to see if this CPU is in a non-context-switch quiescent state
1854 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1855 * Also schedule RCU core processing.
1857 * This function must be called from hardirq context. It is normally
1858 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1859 * false, there is no point in invoking rcu_check_callbacks().
1861 void rcu_check_callbacks(int cpu, int user)
1863 trace_rcu_utilization("Start scheduler-tick");
1864 increment_cpu_stall_ticks();
1865 if (user || rcu_is_cpu_rrupt_from_idle()) {
1868 * Get here if this CPU took its interrupt from user
1869 * mode or from the idle loop, and if this is not a
1870 * nested interrupt. In this case, the CPU is in
1871 * a quiescent state, so note it.
1873 * No memory barrier is required here because both
1874 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1875 * variables that other CPUs neither access nor modify,
1876 * at least not while the corresponding CPU is online.
1882 } else if (!in_softirq()) {
1885 * Get here if this CPU did not take its interrupt from
1886 * softirq, in other words, if it is not interrupting
1887 * a rcu_bh read-side critical section. This is an _bh
1888 * critical section, so note it.
1893 rcu_preempt_check_callbacks(cpu);
1894 if (rcu_pending(cpu))
1896 trace_rcu_utilization("End scheduler-tick");
1900 * Scan the leaf rcu_node structures, processing dyntick state for any that
1901 * have not yet encountered a quiescent state, using the function specified.
1902 * Also initiate boosting for any threads blocked on the root rcu_node.
1904 * The caller must have suppressed start of new grace periods.
1906 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1910 unsigned long flags;
1912 struct rcu_node *rnp;
1914 rcu_for_each_leaf_node(rsp, rnp) {
1917 raw_spin_lock_irqsave(&rnp->lock, flags);
1918 if (!rcu_gp_in_progress(rsp)) {
1919 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1922 if (rnp->qsmask == 0) {
1923 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1928 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1929 if ((rnp->qsmask & bit) != 0 &&
1930 f(per_cpu_ptr(rsp->rda, cpu)))
1935 /* rcu_report_qs_rnp() releases rnp->lock. */
1936 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1939 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1941 rnp = rcu_get_root(rsp);
1942 if (rnp->qsmask == 0) {
1943 raw_spin_lock_irqsave(&rnp->lock, flags);
1944 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1949 * Force quiescent states on reluctant CPUs, and also detect which
1950 * CPUs are in dyntick-idle mode.
1952 static void force_quiescent_state(struct rcu_state *rsp)
1954 unsigned long flags;
1956 struct rcu_node *rnp;
1957 struct rcu_node *rnp_old = NULL;
1959 /* Funnel through hierarchy to reduce memory contention. */
1960 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1961 for (; rnp != NULL; rnp = rnp->parent) {
1962 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1963 !raw_spin_trylock(&rnp->fqslock);
1964 if (rnp_old != NULL)
1965 raw_spin_unlock(&rnp_old->fqslock);
1967 rsp->n_force_qs_lh++;
1972 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1974 /* Reached the root of the rcu_node tree, acquire lock. */
1975 raw_spin_lock_irqsave(&rnp_old->lock, flags);
1976 raw_spin_unlock(&rnp_old->fqslock);
1977 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1978 rsp->n_force_qs_lh++;
1979 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1980 return; /* Someone beat us to it. */
1982 rsp->gp_flags |= RCU_GP_FLAG_FQS;
1983 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1984 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1988 * This does the RCU core processing work for the specified rcu_state
1989 * and rcu_data structures. This may be called only from the CPU to
1990 * whom the rdp belongs.
1993 __rcu_process_callbacks(struct rcu_state *rsp)
1995 unsigned long flags;
1996 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1998 WARN_ON_ONCE(rdp->beenonline == 0);
2001 * Advance callbacks in response to end of earlier grace
2002 * period that some other CPU ended.
2004 rcu_process_gp_end(rsp, rdp);
2006 /* Update RCU state based on any recent quiescent states. */
2007 rcu_check_quiescent_state(rsp, rdp);
2009 /* Does this CPU require a not-yet-started grace period? */
2010 if (cpu_needs_another_gp(rsp, rdp)) {
2011 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
2012 rcu_start_gp(rsp, flags); /* releases above lock */
2015 /* If there are callbacks ready, invoke them. */
2016 if (cpu_has_callbacks_ready_to_invoke(rdp))
2017 invoke_rcu_callbacks(rsp, rdp);
2021 * Do RCU core processing for the current CPU.
2023 static void rcu_process_callbacks(struct softirq_action *unused)
2025 struct rcu_state *rsp;
2027 if (cpu_is_offline(smp_processor_id()))
2029 trace_rcu_utilization("Start RCU core");
2030 for_each_rcu_flavor(rsp)
2031 __rcu_process_callbacks(rsp);
2032 trace_rcu_utilization("End RCU core");
2036 * Schedule RCU callback invocation. If the specified type of RCU
2037 * does not support RCU priority boosting, just do a direct call,
2038 * otherwise wake up the per-CPU kernel kthread. Note that because we
2039 * are running on the current CPU with interrupts disabled, the
2040 * rcu_cpu_kthread_task cannot disappear out from under us.
2042 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2044 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2046 if (likely(!rsp->boost)) {
2047 rcu_do_batch(rsp, rdp);
2050 invoke_rcu_callbacks_kthread();
2053 static void invoke_rcu_core(void)
2055 raise_softirq(RCU_SOFTIRQ);
2059 * Handle any core-RCU processing required by a call_rcu() invocation.
2061 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2062 struct rcu_head *head, unsigned long flags)
2065 * If called from an extended quiescent state, invoke the RCU
2066 * core in order to force a re-evaluation of RCU's idleness.
2068 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2071 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2072 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2076 * Force the grace period if too many callbacks or too long waiting.
2077 * Enforce hysteresis, and don't invoke force_quiescent_state()
2078 * if some other CPU has recently done so. Also, don't bother
2079 * invoking force_quiescent_state() if the newly enqueued callback
2080 * is the only one waiting for a grace period to complete.
2082 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2084 /* Are we ignoring a completed grace period? */
2085 rcu_process_gp_end(rsp, rdp);
2086 check_for_new_grace_period(rsp, rdp);
2088 /* Start a new grace period if one not already started. */
2089 if (!rcu_gp_in_progress(rsp)) {
2090 unsigned long nestflag;
2091 struct rcu_node *rnp_root = rcu_get_root(rsp);
2093 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2094 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
2096 /* Give the grace period a kick. */
2097 rdp->blimit = LONG_MAX;
2098 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2099 *rdp->nxttail[RCU_DONE_TAIL] != head)
2100 force_quiescent_state(rsp);
2101 rdp->n_force_qs_snap = rsp->n_force_qs;
2102 rdp->qlen_last_fqs_check = rdp->qlen;
2108 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2109 struct rcu_state *rsp, bool lazy)
2111 unsigned long flags;
2112 struct rcu_data *rdp;
2114 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2115 debug_rcu_head_queue(head);
2120 * Opportunistically note grace-period endings and beginnings.
2121 * Note that we might see a beginning right after we see an
2122 * end, but never vice versa, since this CPU has to pass through
2123 * a quiescent state betweentimes.
2125 local_irq_save(flags);
2126 rdp = this_cpu_ptr(rsp->rda);
2128 /* Add the callback to our list. */
2129 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL)) {
2130 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2132 local_irq_restore(flags);
2135 ACCESS_ONCE(rdp->qlen)++;
2139 rcu_idle_count_callbacks_posted();
2140 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2141 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2142 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2144 if (__is_kfree_rcu_offset((unsigned long)func))
2145 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2146 rdp->qlen_lazy, rdp->qlen);
2148 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2150 /* Go handle any RCU core processing required. */
2151 __call_rcu_core(rsp, rdp, head, flags);
2152 local_irq_restore(flags);
2156 * Queue an RCU-sched callback for invocation after a grace period.
2158 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2160 __call_rcu(head, func, &rcu_sched_state, 0);
2162 EXPORT_SYMBOL_GPL(call_rcu_sched);
2165 * Queue an RCU callback for invocation after a quicker grace period.
2167 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2169 __call_rcu(head, func, &rcu_bh_state, 0);
2171 EXPORT_SYMBOL_GPL(call_rcu_bh);
2174 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2175 * any blocking grace-period wait automatically implies a grace period
2176 * if there is only one CPU online at any point time during execution
2177 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2178 * occasionally incorrectly indicate that there are multiple CPUs online
2179 * when there was in fact only one the whole time, as this just adds
2180 * some overhead: RCU still operates correctly.
2182 static inline int rcu_blocking_is_gp(void)
2186 might_sleep(); /* Check for RCU read-side critical section. */
2188 ret = num_online_cpus() <= 1;
2194 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2196 * Control will return to the caller some time after a full rcu-sched
2197 * grace period has elapsed, in other words after all currently executing
2198 * rcu-sched read-side critical sections have completed. These read-side
2199 * critical sections are delimited by rcu_read_lock_sched() and
2200 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2201 * local_irq_disable(), and so on may be used in place of
2202 * rcu_read_lock_sched().
2204 * This means that all preempt_disable code sequences, including NMI and
2205 * hardware-interrupt handlers, in progress on entry will have completed
2206 * before this primitive returns. However, this does not guarantee that
2207 * softirq handlers will have completed, since in some kernels, these
2208 * handlers can run in process context, and can block.
2210 * This primitive provides the guarantees made by the (now removed)
2211 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2212 * guarantees that rcu_read_lock() sections will have completed.
2213 * In "classic RCU", these two guarantees happen to be one and
2214 * the same, but can differ in realtime RCU implementations.
2216 void synchronize_sched(void)
2218 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2219 !lock_is_held(&rcu_lock_map) &&
2220 !lock_is_held(&rcu_sched_lock_map),
2221 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2222 if (rcu_blocking_is_gp())
2224 wait_rcu_gp(call_rcu_sched);
2226 EXPORT_SYMBOL_GPL(synchronize_sched);
2229 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2231 * Control will return to the caller some time after a full rcu_bh grace
2232 * period has elapsed, in other words after all currently executing rcu_bh
2233 * read-side critical sections have completed. RCU read-side critical
2234 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2235 * and may be nested.
2237 void synchronize_rcu_bh(void)
2239 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2240 !lock_is_held(&rcu_lock_map) &&
2241 !lock_is_held(&rcu_sched_lock_map),
2242 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2243 if (rcu_blocking_is_gp())
2245 wait_rcu_gp(call_rcu_bh);
2247 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2249 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2250 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2252 static int synchronize_sched_expedited_cpu_stop(void *data)
2255 * There must be a full memory barrier on each affected CPU
2256 * between the time that try_stop_cpus() is called and the
2257 * time that it returns.
2259 * In the current initial implementation of cpu_stop, the
2260 * above condition is already met when the control reaches
2261 * this point and the following smp_mb() is not strictly
2262 * necessary. Do smp_mb() anyway for documentation and
2263 * robustness against future implementation changes.
2265 smp_mb(); /* See above comment block. */
2270 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2272 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2273 * approach to force the grace period to end quickly. This consumes
2274 * significant time on all CPUs and is unfriendly to real-time workloads,
2275 * so is thus not recommended for any sort of common-case code. In fact,
2276 * if you are using synchronize_sched_expedited() in a loop, please
2277 * restructure your code to batch your updates, and then use a single
2278 * synchronize_sched() instead.
2280 * Note that it is illegal to call this function while holding any lock
2281 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2282 * to call this function from a CPU-hotplug notifier. Failing to observe
2283 * these restriction will result in deadlock.
2285 * This implementation can be thought of as an application of ticket
2286 * locking to RCU, with sync_sched_expedited_started and
2287 * sync_sched_expedited_done taking on the roles of the halves
2288 * of the ticket-lock word. Each task atomically increments
2289 * sync_sched_expedited_started upon entry, snapshotting the old value,
2290 * then attempts to stop all the CPUs. If this succeeds, then each
2291 * CPU will have executed a context switch, resulting in an RCU-sched
2292 * grace period. We are then done, so we use atomic_cmpxchg() to
2293 * update sync_sched_expedited_done to match our snapshot -- but
2294 * only if someone else has not already advanced past our snapshot.
2296 * On the other hand, if try_stop_cpus() fails, we check the value
2297 * of sync_sched_expedited_done. If it has advanced past our
2298 * initial snapshot, then someone else must have forced a grace period
2299 * some time after we took our snapshot. In this case, our work is
2300 * done for us, and we can simply return. Otherwise, we try again,
2301 * but keep our initial snapshot for purposes of checking for someone
2302 * doing our work for us.
2304 * If we fail too many times in a row, we fall back to synchronize_sched().
2306 void synchronize_sched_expedited(void)
2308 int firstsnap, s, snap, trycount = 0;
2310 /* Note that atomic_inc_return() implies full memory barrier. */
2311 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2313 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2316 * Each pass through the following loop attempts to force a
2317 * context switch on each CPU.
2319 while (try_stop_cpus(cpu_online_mask,
2320 synchronize_sched_expedited_cpu_stop,
2324 /* No joy, try again later. Or just synchronize_sched(). */
2325 if (trycount++ < 10) {
2326 udelay(trycount * num_online_cpus());
2328 synchronize_sched();
2332 /* Check to see if someone else did our work for us. */
2333 s = atomic_read(&sync_sched_expedited_done);
2334 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2335 smp_mb(); /* ensure test happens before caller kfree */
2340 * Refetching sync_sched_expedited_started allows later
2341 * callers to piggyback on our grace period. We subtract
2342 * 1 to get the same token that the last incrementer got.
2343 * We retry after they started, so our grace period works
2344 * for them, and they started after our first try, so their
2345 * grace period works for us.
2348 snap = atomic_read(&sync_sched_expedited_started);
2349 smp_mb(); /* ensure read is before try_stop_cpus(). */
2353 * Everyone up to our most recent fetch is covered by our grace
2354 * period. Update the counter, but only if our work is still
2355 * relevant -- which it won't be if someone who started later
2356 * than we did beat us to the punch.
2359 s = atomic_read(&sync_sched_expedited_done);
2360 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2361 smp_mb(); /* ensure test happens before caller kfree */
2364 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2368 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2371 * Check to see if there is any immediate RCU-related work to be done
2372 * by the current CPU, for the specified type of RCU, returning 1 if so.
2373 * The checks are in order of increasing expense: checks that can be
2374 * carried out against CPU-local state are performed first. However,
2375 * we must check for CPU stalls first, else we might not get a chance.
2377 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2379 struct rcu_node *rnp = rdp->mynode;
2381 rdp->n_rcu_pending++;
2383 /* Check for CPU stalls, if enabled. */
2384 check_cpu_stall(rsp, rdp);
2386 /* Is the RCU core waiting for a quiescent state from this CPU? */
2387 if (rcu_scheduler_fully_active &&
2388 rdp->qs_pending && !rdp->passed_quiesce) {
2389 rdp->n_rp_qs_pending++;
2390 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2391 rdp->n_rp_report_qs++;
2395 /* Does this CPU have callbacks ready to invoke? */
2396 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2397 rdp->n_rp_cb_ready++;
2401 /* Has RCU gone idle with this CPU needing another grace period? */
2402 if (cpu_needs_another_gp(rsp, rdp)) {
2403 rdp->n_rp_cpu_needs_gp++;
2407 /* Has another RCU grace period completed? */
2408 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2409 rdp->n_rp_gp_completed++;
2413 /* Has a new RCU grace period started? */
2414 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2415 rdp->n_rp_gp_started++;
2420 rdp->n_rp_need_nothing++;
2425 * Check to see if there is any immediate RCU-related work to be done
2426 * by the current CPU, returning 1 if so. This function is part of the
2427 * RCU implementation; it is -not- an exported member of the RCU API.
2429 static int rcu_pending(int cpu)
2431 struct rcu_state *rsp;
2433 for_each_rcu_flavor(rsp)
2434 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2440 * Check to see if any future RCU-related work will need to be done
2441 * by the current CPU, even if none need be done immediately, returning
2444 static int rcu_cpu_has_callbacks(int cpu)
2446 struct rcu_state *rsp;
2448 /* RCU callbacks either ready or pending? */
2449 for_each_rcu_flavor(rsp)
2450 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2456 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2457 * the compiler is expected to optimize this away.
2459 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2460 int cpu, unsigned long done)
2462 trace_rcu_barrier(rsp->name, s, cpu,
2463 atomic_read(&rsp->barrier_cpu_count), done);
2467 * RCU callback function for _rcu_barrier(). If we are last, wake
2468 * up the task executing _rcu_barrier().
2470 static void rcu_barrier_callback(struct rcu_head *rhp)
2472 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2473 struct rcu_state *rsp = rdp->rsp;
2475 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2476 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2477 complete(&rsp->barrier_completion);
2479 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2484 * Called with preemption disabled, and from cross-cpu IRQ context.
2486 static void rcu_barrier_func(void *type)
2488 struct rcu_state *rsp = type;
2489 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2491 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2492 atomic_inc(&rsp->barrier_cpu_count);
2493 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2497 * Orchestrate the specified type of RCU barrier, waiting for all
2498 * RCU callbacks of the specified type to complete.
2500 static void _rcu_barrier(struct rcu_state *rsp)
2503 struct rcu_data *rdp;
2504 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2505 unsigned long snap_done;
2507 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2509 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2510 mutex_lock(&rsp->barrier_mutex);
2513 * Ensure that all prior references, including to ->n_barrier_done,
2514 * are ordered before the _rcu_barrier() machinery.
2516 smp_mb(); /* See above block comment. */
2519 * Recheck ->n_barrier_done to see if others did our work for us.
2520 * This means checking ->n_barrier_done for an even-to-odd-to-even
2521 * transition. The "if" expression below therefore rounds the old
2522 * value up to the next even number and adds two before comparing.
2524 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2525 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2526 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2527 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2528 smp_mb(); /* caller's subsequent code after above check. */
2529 mutex_unlock(&rsp->barrier_mutex);
2534 * Increment ->n_barrier_done to avoid duplicate work. Use
2535 * ACCESS_ONCE() to prevent the compiler from speculating
2536 * the increment to precede the early-exit check.
2538 ACCESS_ONCE(rsp->n_barrier_done)++;
2539 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2540 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2541 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2544 * Initialize the count to one rather than to zero in order to
2545 * avoid a too-soon return to zero in case of a short grace period
2546 * (or preemption of this task). Exclude CPU-hotplug operations
2547 * to ensure that no offline CPU has callbacks queued.
2549 init_completion(&rsp->barrier_completion);
2550 atomic_set(&rsp->barrier_cpu_count, 1);
2554 * Force each CPU with callbacks to register a new callback.
2555 * When that callback is invoked, we will know that all of the
2556 * corresponding CPU's preceding callbacks have been invoked.
2558 for_each_online_cpu(cpu) {
2559 rdp = per_cpu_ptr(rsp->rda, cpu);
2560 if (ACCESS_ONCE(rdp->qlen)) {
2561 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2562 rsp->n_barrier_done);
2563 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2565 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2566 rsp->n_barrier_done);
2572 * Now that we have an rcu_barrier_callback() callback on each
2573 * CPU, and thus each counted, remove the initial count.
2575 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2576 complete(&rsp->barrier_completion);
2578 /* Increment ->n_barrier_done to prevent duplicate work. */
2579 smp_mb(); /* Keep increment after above mechanism. */
2580 ACCESS_ONCE(rsp->n_barrier_done)++;
2581 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2582 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2583 smp_mb(); /* Keep increment before caller's subsequent code. */
2585 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2586 wait_for_completion(&rsp->barrier_completion);
2588 /* Other rcu_barrier() invocations can now safely proceed. */
2589 mutex_unlock(&rsp->barrier_mutex);
2593 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2595 void rcu_barrier_bh(void)
2597 _rcu_barrier(&rcu_bh_state);
2599 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2602 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2604 void rcu_barrier_sched(void)
2606 _rcu_barrier(&rcu_sched_state);
2608 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2611 * Do boot-time initialization of a CPU's per-CPU RCU data.
2614 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2616 unsigned long flags;
2617 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2618 struct rcu_node *rnp = rcu_get_root(rsp);
2620 /* Set up local state, ensuring consistent view of global state. */
2621 raw_spin_lock_irqsave(&rnp->lock, flags);
2622 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2623 init_callback_list(rdp);
2625 ACCESS_ONCE(rdp->qlen) = 0;
2626 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2627 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2628 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2629 #ifdef CONFIG_RCU_USER_QS
2630 WARN_ON_ONCE(rdp->dynticks->in_user);
2634 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2638 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2639 * offline event can be happening at a given time. Note also that we
2640 * can accept some slop in the rsp->completed access due to the fact
2641 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2643 static void __cpuinit
2644 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2646 unsigned long flags;
2648 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2649 struct rcu_node *rnp = rcu_get_root(rsp);
2651 /* Set up local state, ensuring consistent view of global state. */
2652 raw_spin_lock_irqsave(&rnp->lock, flags);
2653 rdp->beenonline = 1; /* We have now been online. */
2654 rdp->preemptible = preemptible;
2655 rdp->qlen_last_fqs_check = 0;
2656 rdp->n_force_qs_snap = rsp->n_force_qs;
2657 rdp->blimit = blimit;
2658 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2659 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2660 atomic_set(&rdp->dynticks->dynticks,
2661 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2662 rcu_prepare_for_idle_init(cpu);
2663 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2666 * A new grace period might start here. If so, we won't be part
2667 * of it, but that is OK, as we are currently in a quiescent state.
2670 /* Exclude any attempts to start a new GP on large systems. */
2671 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2673 /* Add CPU to rcu_node bitmasks. */
2675 mask = rdp->grpmask;
2677 /* Exclude any attempts to start a new GP on small systems. */
2678 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2679 rnp->qsmaskinit |= mask;
2680 mask = rnp->grpmask;
2681 if (rnp == rdp->mynode) {
2683 * If there is a grace period in progress, we will
2684 * set up to wait for it next time we run the
2687 rdp->gpnum = rnp->completed;
2688 rdp->completed = rnp->completed;
2689 rdp->passed_quiesce = 0;
2690 rdp->qs_pending = 0;
2691 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2693 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2695 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2697 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2700 static void __cpuinit rcu_prepare_cpu(int cpu)
2702 struct rcu_state *rsp;
2704 for_each_rcu_flavor(rsp)
2705 rcu_init_percpu_data(cpu, rsp,
2706 strcmp(rsp->name, "rcu_preempt") == 0);
2710 * Handle CPU online/offline notification events.
2712 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2713 unsigned long action, void *hcpu)
2715 long cpu = (long)hcpu;
2716 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2717 struct rcu_node *rnp = rdp->mynode;
2718 struct rcu_state *rsp;
2720 trace_rcu_utilization("Start CPU hotplug");
2722 case CPU_UP_PREPARE:
2723 case CPU_UP_PREPARE_FROZEN:
2724 rcu_prepare_cpu(cpu);
2725 rcu_prepare_kthreads(cpu);
2728 case CPU_DOWN_FAILED:
2729 rcu_boost_kthread_setaffinity(rnp, -1);
2731 case CPU_DOWN_PREPARE:
2732 rcu_boost_kthread_setaffinity(rnp, cpu);
2735 case CPU_DYING_FROZEN:
2737 * The whole machine is "stopped" except this CPU, so we can
2738 * touch any data without introducing corruption. We send the
2739 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2741 for_each_rcu_flavor(rsp)
2742 rcu_cleanup_dying_cpu(rsp);
2743 rcu_cleanup_after_idle(cpu);
2746 case CPU_DEAD_FROZEN:
2747 case CPU_UP_CANCELED:
2748 case CPU_UP_CANCELED_FROZEN:
2749 for_each_rcu_flavor(rsp)
2750 rcu_cleanup_dead_cpu(cpu, rsp);
2755 trace_rcu_utilization("End CPU hotplug");
2760 * Spawn the kthread that handles this RCU flavor's grace periods.
2762 static int __init rcu_spawn_gp_kthread(void)
2764 unsigned long flags;
2765 struct rcu_node *rnp;
2766 struct rcu_state *rsp;
2767 struct task_struct *t;
2769 for_each_rcu_flavor(rsp) {
2770 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2772 rnp = rcu_get_root(rsp);
2773 raw_spin_lock_irqsave(&rnp->lock, flags);
2774 rsp->gp_kthread = t;
2775 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2779 early_initcall(rcu_spawn_gp_kthread);
2782 * This function is invoked towards the end of the scheduler's initialization
2783 * process. Before this is called, the idle task might contain
2784 * RCU read-side critical sections (during which time, this idle
2785 * task is booting the system). After this function is called, the
2786 * idle tasks are prohibited from containing RCU read-side critical
2787 * sections. This function also enables RCU lockdep checking.
2789 void rcu_scheduler_starting(void)
2791 WARN_ON(num_online_cpus() != 1);
2792 WARN_ON(nr_context_switches() > 0);
2793 rcu_scheduler_active = 1;
2797 * Compute the per-level fanout, either using the exact fanout specified
2798 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2800 #ifdef CONFIG_RCU_FANOUT_EXACT
2801 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2805 for (i = rcu_num_lvls - 1; i > 0; i--)
2806 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2807 rsp->levelspread[0] = rcu_fanout_leaf;
2809 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2810 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2817 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2818 ccur = rsp->levelcnt[i];
2819 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2823 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2826 * Helper function for rcu_init() that initializes one rcu_state structure.
2828 static void __init rcu_init_one(struct rcu_state *rsp,
2829 struct rcu_data __percpu *rda)
2831 static char *buf[] = { "rcu_node_0",
2834 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2835 static char *fqs[] = { "rcu_node_fqs_0",
2838 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2842 struct rcu_node *rnp;
2844 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2846 /* Initialize the level-tracking arrays. */
2848 for (i = 0; i < rcu_num_lvls; i++)
2849 rsp->levelcnt[i] = num_rcu_lvl[i];
2850 for (i = 1; i < rcu_num_lvls; i++)
2851 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2852 rcu_init_levelspread(rsp);
2854 /* Initialize the elements themselves, starting from the leaves. */
2856 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2857 cpustride *= rsp->levelspread[i];
2858 rnp = rsp->level[i];
2859 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2860 raw_spin_lock_init(&rnp->lock);
2861 lockdep_set_class_and_name(&rnp->lock,
2862 &rcu_node_class[i], buf[i]);
2863 raw_spin_lock_init(&rnp->fqslock);
2864 lockdep_set_class_and_name(&rnp->fqslock,
2865 &rcu_fqs_class[i], fqs[i]);
2866 rnp->gpnum = rsp->gpnum;
2867 rnp->completed = rsp->completed;
2869 rnp->qsmaskinit = 0;
2870 rnp->grplo = j * cpustride;
2871 rnp->grphi = (j + 1) * cpustride - 1;
2872 if (rnp->grphi >= NR_CPUS)
2873 rnp->grphi = NR_CPUS - 1;
2879 rnp->grpnum = j % rsp->levelspread[i - 1];
2880 rnp->grpmask = 1UL << rnp->grpnum;
2881 rnp->parent = rsp->level[i - 1] +
2882 j / rsp->levelspread[i - 1];
2885 INIT_LIST_HEAD(&rnp->blkd_tasks);
2890 init_waitqueue_head(&rsp->gp_wq);
2891 rnp = rsp->level[rcu_num_lvls - 1];
2892 for_each_possible_cpu(i) {
2893 while (i > rnp->grphi)
2895 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2896 rcu_boot_init_percpu_data(i, rsp);
2898 list_add(&rsp->flavors, &rcu_struct_flavors);
2902 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2903 * replace the definitions in rcutree.h because those are needed to size
2904 * the ->node array in the rcu_state structure.
2906 static void __init rcu_init_geometry(void)
2911 int rcu_capacity[MAX_RCU_LVLS + 1];
2913 /* If the compile-time values are accurate, just leave. */
2914 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
2915 nr_cpu_ids == NR_CPUS)
2919 * Compute number of nodes that can be handled an rcu_node tree
2920 * with the given number of levels. Setting rcu_capacity[0] makes
2921 * some of the arithmetic easier.
2923 rcu_capacity[0] = 1;
2924 rcu_capacity[1] = rcu_fanout_leaf;
2925 for (i = 2; i <= MAX_RCU_LVLS; i++)
2926 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2929 * The boot-time rcu_fanout_leaf parameter is only permitted
2930 * to increase the leaf-level fanout, not decrease it. Of course,
2931 * the leaf-level fanout cannot exceed the number of bits in
2932 * the rcu_node masks. Finally, the tree must be able to accommodate
2933 * the configured number of CPUs. Complain and fall back to the
2934 * compile-time values if these limits are exceeded.
2936 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2937 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2938 n > rcu_capacity[MAX_RCU_LVLS]) {
2943 /* Calculate the number of rcu_nodes at each level of the tree. */
2944 for (i = 1; i <= MAX_RCU_LVLS; i++)
2945 if (n <= rcu_capacity[i]) {
2946 for (j = 0; j <= i; j++)
2948 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2950 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2955 /* Calculate the total number of rcu_node structures. */
2957 for (i = 0; i <= MAX_RCU_LVLS; i++)
2958 rcu_num_nodes += num_rcu_lvl[i];
2962 void __init rcu_init(void)
2966 rcu_bootup_announce();
2967 rcu_init_geometry();
2968 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2969 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2970 __rcu_init_preempt();
2971 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2974 * We don't need protection against CPU-hotplug here because
2975 * this is called early in boot, before either interrupts
2976 * or the scheduler are operational.
2978 cpu_notifier(rcu_cpu_notify, 0);
2979 for_each_online_cpu(cpu)
2980 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2981 check_cpu_stall_init();
2984 #include "rcutree_plugin.h"