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>
57 #include <trace/events/rcu.h>
61 /* Data structures. */
63 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 #define RCU_STATE_INITIALIZER(sname, cr) { \
66 .level = { &sname##_state.node[0] }, \
68 .fqs_state = RCU_GP_IDLE, \
71 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
72 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
73 .orphan_donetail = &sname##_state.orphan_donelist, \
74 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \
76 .n_force_qs_ngp = 0, \
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;
89 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
90 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
91 module_param(rcu_fanout_leaf, int, 0);
92 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
93 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
100 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
103 * The rcu_scheduler_active variable transitions from zero to one just
104 * before the first task is spawned. So when this variable is zero, RCU
105 * can assume that there is but one task, allowing RCU to (for example)
106 * optimized synchronize_sched() to a simple barrier(). When this variable
107 * is one, RCU must actually do all the hard work required to detect real
108 * grace periods. This variable is also used to suppress boot-time false
109 * positives from lockdep-RCU error checking.
111 int rcu_scheduler_active __read_mostly;
112 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
115 * The rcu_scheduler_fully_active variable transitions from zero to one
116 * during the early_initcall() processing, which is after the scheduler
117 * is capable of creating new tasks. So RCU processing (for example,
118 * creating tasks for RCU priority boosting) must be delayed until after
119 * rcu_scheduler_fully_active transitions from zero to one. We also
120 * currently delay invocation of any RCU callbacks until after this point.
122 * It might later prove better for people registering RCU callbacks during
123 * early boot to take responsibility for these callbacks, but one step at
126 static int rcu_scheduler_fully_active __read_mostly;
128 #ifdef CONFIG_RCU_BOOST
131 * Control variables for per-CPU and per-rcu_node kthreads. These
132 * handle all flavors of RCU.
134 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
135 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
136 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
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_node_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;
158 /* State information for rcu_barrier() and friends. */
160 static atomic_t rcu_barrier_cpu_count;
161 static DEFINE_MUTEX(rcu_barrier_mutex);
162 static struct completion rcu_barrier_completion;
165 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
166 * permit this function to be invoked without holding the root rcu_node
167 * structure's ->lock, but of course results can be subject to change.
169 static int rcu_gp_in_progress(struct rcu_state *rsp)
171 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
175 * Note a quiescent state. Because we do not need to know
176 * how many quiescent states passed, just if there was at least
177 * one since the start of the grace period, this just sets a flag.
178 * The caller must have disabled preemption.
180 void rcu_sched_qs(int cpu)
182 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
184 rdp->passed_quiesce_gpnum = rdp->gpnum;
186 if (rdp->passed_quiesce == 0)
187 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
188 rdp->passed_quiesce = 1;
191 void rcu_bh_qs(int cpu)
193 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
195 rdp->passed_quiesce_gpnum = rdp->gpnum;
197 if (rdp->passed_quiesce == 0)
198 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
199 rdp->passed_quiesce = 1;
203 * Note a context switch. This is a quiescent state for RCU-sched,
204 * and requires special handling for preemptible RCU.
205 * The caller must have disabled preemption.
207 void rcu_note_context_switch(int cpu)
209 trace_rcu_utilization("Start context switch");
211 rcu_preempt_note_context_switch(cpu);
212 trace_rcu_utilization("End context switch");
214 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
216 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
217 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
218 .dynticks = ATOMIC_INIT(1),
221 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
222 static int qhimark = 10000; /* If this many pending, ignore blimit. */
223 static int qlowmark = 100; /* Once only this many pending, use blimit. */
225 module_param(blimit, int, 0);
226 module_param(qhimark, int, 0);
227 module_param(qlowmark, int, 0);
229 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
230 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
232 module_param(rcu_cpu_stall_suppress, int, 0644);
233 module_param(rcu_cpu_stall_timeout, int, 0644);
235 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
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, 0);
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, 0);
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] && !rcu_gp_in_progress(rsp);
318 * Return the root node of the specified rcu_state structure.
320 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
322 return &rsp->node[0];
326 * If the specified CPU is offline, tell the caller that it is in
327 * a quiescent state. Otherwise, whack it with a reschedule IPI.
328 * Grace periods can end up waiting on an offline CPU when that
329 * CPU is in the process of coming online -- it will be added to the
330 * rcu_node bitmasks before it actually makes it online. The same thing
331 * can happen while a CPU is in the process of coming online. Because this
332 * race is quite rare, we check for it after detecting that the grace
333 * period has been delayed rather than checking each and every CPU
334 * each and every time we start a new grace period.
336 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
339 * If the CPU is offline for more than a jiffy, it is in a quiescent
340 * state. We can trust its state not to change because interrupts
341 * are disabled. The reason for the jiffy's worth of slack is to
342 * handle CPUs initializing on the way up and finding their way
343 * to the idle loop on the way down.
345 if (cpu_is_offline(rdp->cpu) &&
346 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
347 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
355 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
357 * If the new value of the ->dynticks_nesting counter now is zero,
358 * we really have entered idle, and must do the appropriate accounting.
359 * The caller must have disabled interrupts.
361 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
363 trace_rcu_dyntick("Start", oldval, 0);
364 if (!is_idle_task(current)) {
365 struct task_struct *idle = idle_task(smp_processor_id());
367 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
368 ftrace_dump(DUMP_ALL);
369 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
370 current->pid, current->comm,
371 idle->pid, idle->comm); /* must be idle task! */
373 rcu_prepare_for_idle(smp_processor_id());
374 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
375 smp_mb__before_atomic_inc(); /* See above. */
376 atomic_inc(&rdtp->dynticks);
377 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
378 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
381 * The idle task is not permitted to enter the idle loop while
382 * in an RCU read-side critical section.
384 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
385 "Illegal idle entry in RCU read-side critical section.");
386 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
387 "Illegal idle entry in RCU-bh read-side critical section.");
388 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
389 "Illegal idle entry in RCU-sched read-side critical section.");
393 * rcu_idle_enter - inform RCU that current CPU is entering idle
395 * Enter idle mode, in other words, -leave- the mode in which RCU
396 * read-side critical sections can occur. (Though RCU read-side
397 * critical sections can occur in irq handlers in idle, a possibility
398 * handled by irq_enter() and irq_exit().)
400 * We crowbar the ->dynticks_nesting field to zero to allow for
401 * the possibility of usermode upcalls having messed up our count
402 * of interrupt nesting level during the prior busy period.
404 void rcu_idle_enter(void)
408 struct rcu_dynticks *rdtp;
410 local_irq_save(flags);
411 rdtp = &__get_cpu_var(rcu_dynticks);
412 oldval = rdtp->dynticks_nesting;
413 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
414 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
415 rdtp->dynticks_nesting = 0;
417 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
418 rcu_idle_enter_common(rdtp, oldval);
419 local_irq_restore(flags);
421 EXPORT_SYMBOL_GPL(rcu_idle_enter);
424 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
426 * Exit from an interrupt handler, which might possibly result in entering
427 * idle mode, in other words, leaving the mode in which read-side critical
428 * sections can occur.
430 * This code assumes that the idle loop never does anything that might
431 * result in unbalanced calls to irq_enter() and irq_exit(). If your
432 * architecture violates this assumption, RCU will give you what you
433 * deserve, good and hard. But very infrequently and irreproducibly.
435 * Use things like work queues to work around this limitation.
437 * You have been warned.
439 void rcu_irq_exit(void)
443 struct rcu_dynticks *rdtp;
445 local_irq_save(flags);
446 rdtp = &__get_cpu_var(rcu_dynticks);
447 oldval = rdtp->dynticks_nesting;
448 rdtp->dynticks_nesting--;
449 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
450 if (rdtp->dynticks_nesting)
451 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
453 rcu_idle_enter_common(rdtp, oldval);
454 local_irq_restore(flags);
458 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
460 * If the new value of the ->dynticks_nesting counter was previously zero,
461 * we really have exited idle, and must do the appropriate accounting.
462 * The caller must have disabled interrupts.
464 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
466 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
467 atomic_inc(&rdtp->dynticks);
468 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
469 smp_mb__after_atomic_inc(); /* See above. */
470 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
471 rcu_cleanup_after_idle(smp_processor_id());
472 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
473 if (!is_idle_task(current)) {
474 struct task_struct *idle = idle_task(smp_processor_id());
476 trace_rcu_dyntick("Error on exit: not idle task",
477 oldval, rdtp->dynticks_nesting);
478 ftrace_dump(DUMP_ALL);
479 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
480 current->pid, current->comm,
481 idle->pid, idle->comm); /* must be idle task! */
486 * rcu_idle_exit - inform RCU that current CPU is leaving idle
488 * Exit idle mode, in other words, -enter- the mode in which RCU
489 * read-side critical sections can occur.
491 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
492 * allow for the possibility of usermode upcalls messing up our count
493 * of interrupt nesting level during the busy period that is just
496 void rcu_idle_exit(void)
499 struct rcu_dynticks *rdtp;
502 local_irq_save(flags);
503 rdtp = &__get_cpu_var(rcu_dynticks);
504 oldval = rdtp->dynticks_nesting;
505 WARN_ON_ONCE(oldval < 0);
506 if (oldval & DYNTICK_TASK_NEST_MASK)
507 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
509 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
510 rcu_idle_exit_common(rdtp, oldval);
511 local_irq_restore(flags);
513 EXPORT_SYMBOL_GPL(rcu_idle_exit);
516 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
518 * Enter an interrupt handler, which might possibly result in exiting
519 * idle mode, in other words, entering the mode in which read-side critical
520 * sections can occur.
522 * Note that the Linux kernel is fully capable of entering an interrupt
523 * handler that it never exits, for example when doing upcalls to
524 * user mode! This code assumes that the idle loop never does upcalls to
525 * user mode. If your architecture does do upcalls from the idle loop (or
526 * does anything else that results in unbalanced calls to the irq_enter()
527 * and irq_exit() functions), RCU will give you what you deserve, good
528 * and hard. But very infrequently and irreproducibly.
530 * Use things like work queues to work around this limitation.
532 * You have been warned.
534 void rcu_irq_enter(void)
537 struct rcu_dynticks *rdtp;
540 local_irq_save(flags);
541 rdtp = &__get_cpu_var(rcu_dynticks);
542 oldval = rdtp->dynticks_nesting;
543 rdtp->dynticks_nesting++;
544 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
546 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
548 rcu_idle_exit_common(rdtp, oldval);
549 local_irq_restore(flags);
553 * rcu_nmi_enter - inform RCU of entry to NMI context
555 * If the CPU was idle with dynamic ticks active, and there is no
556 * irq handler running, this updates rdtp->dynticks_nmi to let the
557 * RCU grace-period handling know that the CPU is active.
559 void rcu_nmi_enter(void)
561 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
563 if (rdtp->dynticks_nmi_nesting == 0 &&
564 (atomic_read(&rdtp->dynticks) & 0x1))
566 rdtp->dynticks_nmi_nesting++;
567 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
568 atomic_inc(&rdtp->dynticks);
569 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
570 smp_mb__after_atomic_inc(); /* See above. */
571 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
575 * rcu_nmi_exit - inform RCU of exit from NMI context
577 * If the CPU was idle with dynamic ticks active, and there is no
578 * irq handler running, this updates rdtp->dynticks_nmi to let the
579 * RCU grace-period handling know that the CPU is no longer active.
581 void rcu_nmi_exit(void)
583 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
585 if (rdtp->dynticks_nmi_nesting == 0 ||
586 --rdtp->dynticks_nmi_nesting != 0)
588 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
589 smp_mb__before_atomic_inc(); /* See above. */
590 atomic_inc(&rdtp->dynticks);
591 smp_mb__after_atomic_inc(); /* Force delay to next write. */
592 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
595 #ifdef CONFIG_PROVE_RCU
598 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
600 * If the current CPU is in its idle loop and is neither in an interrupt
601 * or NMI handler, return true.
603 int rcu_is_cpu_idle(void)
608 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
612 EXPORT_SYMBOL(rcu_is_cpu_idle);
614 #ifdef CONFIG_HOTPLUG_CPU
617 * Is the current CPU online? Disable preemption to avoid false positives
618 * that could otherwise happen due to the current CPU number being sampled,
619 * this task being preempted, its old CPU being taken offline, resuming
620 * on some other CPU, then determining that its old CPU is now offline.
621 * It is OK to use RCU on an offline processor during initial boot, hence
622 * the check for rcu_scheduler_fully_active. Note also that it is OK
623 * for a CPU coming online to use RCU for one jiffy prior to marking itself
624 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
625 * offline to continue to use RCU for one jiffy after marking itself
626 * offline in the cpu_online_mask. This leniency is necessary given the
627 * non-atomic nature of the online and offline processing, for example,
628 * the fact that a CPU enters the scheduler after completing the CPU_DYING
631 * This is also why RCU internally marks CPUs online during the
632 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
634 * Disable checking if in an NMI handler because we cannot safely report
635 * errors from NMI handlers anyway.
637 bool rcu_lockdep_current_cpu_online(void)
639 struct rcu_data *rdp;
640 struct rcu_node *rnp;
646 rdp = &__get_cpu_var(rcu_sched_data);
648 ret = (rdp->grpmask & rnp->qsmaskinit) ||
649 !rcu_scheduler_fully_active;
653 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
655 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
657 #endif /* #ifdef CONFIG_PROVE_RCU */
660 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
662 * If the current CPU is idle or running at a first-level (not nested)
663 * interrupt from idle, return true. The caller must have at least
664 * disabled preemption.
666 int rcu_is_cpu_rrupt_from_idle(void)
668 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
672 * Snapshot the specified CPU's dynticks counter so that we can later
673 * credit them with an implicit quiescent state. Return 1 if this CPU
674 * is in dynticks idle mode, which is an extended quiescent state.
676 static int dyntick_save_progress_counter(struct rcu_data *rdp)
678 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
679 return (rdp->dynticks_snap & 0x1) == 0;
683 * Return true if the specified CPU has passed through a quiescent
684 * state by virtue of being in or having passed through an dynticks
685 * idle state since the last call to dyntick_save_progress_counter()
688 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
693 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
694 snap = (unsigned int)rdp->dynticks_snap;
697 * If the CPU passed through or entered a dynticks idle phase with
698 * no active irq/NMI handlers, then we can safely pretend that the CPU
699 * already acknowledged the request to pass through a quiescent
700 * state. Either way, that CPU cannot possibly be in an RCU
701 * read-side critical section that started before the beginning
702 * of the current RCU grace period.
704 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
705 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
710 /* Go check for the CPU being offline. */
711 return rcu_implicit_offline_qs(rdp);
714 static int jiffies_till_stall_check(void)
716 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
719 * Limit check must be consistent with the Kconfig limits
720 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
722 if (till_stall_check < 3) {
723 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
724 till_stall_check = 3;
725 } else if (till_stall_check > 300) {
726 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
727 till_stall_check = 300;
729 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
732 static void record_gp_stall_check_time(struct rcu_state *rsp)
734 rsp->gp_start = jiffies;
735 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
738 static void print_other_cpu_stall(struct rcu_state *rsp)
744 struct rcu_node *rnp = rcu_get_root(rsp);
746 /* Only let one CPU complain about others per time interval. */
748 raw_spin_lock_irqsave(&rnp->lock, flags);
749 delta = jiffies - rsp->jiffies_stall;
750 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
751 raw_spin_unlock_irqrestore(&rnp->lock, flags);
754 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
755 raw_spin_unlock_irqrestore(&rnp->lock, flags);
758 * OK, time to rat on our buddy...
759 * See Documentation/RCU/stallwarn.txt for info on how to debug
760 * RCU CPU stall warnings.
762 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
764 print_cpu_stall_info_begin();
765 rcu_for_each_leaf_node(rsp, rnp) {
766 raw_spin_lock_irqsave(&rnp->lock, flags);
767 ndetected += rcu_print_task_stall(rnp);
768 raw_spin_unlock_irqrestore(&rnp->lock, flags);
769 if (rnp->qsmask == 0)
771 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
772 if (rnp->qsmask & (1UL << cpu)) {
773 print_cpu_stall_info(rsp, rnp->grplo + cpu);
779 * Now rat on any tasks that got kicked up to the root rcu_node
780 * due to CPU offlining.
782 rnp = rcu_get_root(rsp);
783 raw_spin_lock_irqsave(&rnp->lock, flags);
784 ndetected = rcu_print_task_stall(rnp);
785 raw_spin_unlock_irqrestore(&rnp->lock, flags);
787 print_cpu_stall_info_end();
788 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
789 smp_processor_id(), (long)(jiffies - rsp->gp_start));
791 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
792 else if (!trigger_all_cpu_backtrace())
795 /* If so configured, complain about tasks blocking the grace period. */
797 rcu_print_detail_task_stall(rsp);
799 force_quiescent_state(rsp, 0); /* Kick them all. */
802 static void print_cpu_stall(struct rcu_state *rsp)
805 struct rcu_node *rnp = rcu_get_root(rsp);
808 * OK, time to rat on ourselves...
809 * See Documentation/RCU/stallwarn.txt for info on how to debug
810 * RCU CPU stall warnings.
812 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
813 print_cpu_stall_info_begin();
814 print_cpu_stall_info(rsp, smp_processor_id());
815 print_cpu_stall_info_end();
816 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
817 if (!trigger_all_cpu_backtrace())
820 raw_spin_lock_irqsave(&rnp->lock, flags);
821 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
822 rsp->jiffies_stall = jiffies +
823 3 * jiffies_till_stall_check() + 3;
824 raw_spin_unlock_irqrestore(&rnp->lock, flags);
826 set_need_resched(); /* kick ourselves to get things going. */
829 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
833 struct rcu_node *rnp;
835 if (rcu_cpu_stall_suppress)
837 j = ACCESS_ONCE(jiffies);
838 js = ACCESS_ONCE(rsp->jiffies_stall);
840 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
842 /* We haven't checked in, so go dump stack. */
843 print_cpu_stall(rsp);
845 } else if (rcu_gp_in_progress(rsp) &&
846 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
848 /* They had a few time units to dump stack, so complain. */
849 print_other_cpu_stall(rsp);
853 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
855 rcu_cpu_stall_suppress = 1;
860 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
862 * Set the stall-warning timeout way off into the future, thus preventing
863 * any RCU CPU stall-warning messages from appearing in the current set of
866 * The caller must disable hard irqs.
868 void rcu_cpu_stall_reset(void)
870 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
871 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
872 rcu_preempt_stall_reset();
875 static struct notifier_block rcu_panic_block = {
876 .notifier_call = rcu_panic,
879 static void __init check_cpu_stall_init(void)
881 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
885 * Update CPU-local rcu_data state to record the newly noticed grace period.
886 * This is used both when we started the grace period and when we notice
887 * that someone else started the grace period. The caller must hold the
888 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
889 * and must have irqs disabled.
891 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
893 if (rdp->gpnum != rnp->gpnum) {
895 * If the current grace period is waiting for this CPU,
896 * set up to detect a quiescent state, otherwise don't
897 * go looking for one.
899 rdp->gpnum = rnp->gpnum;
900 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
901 if (rnp->qsmask & rdp->grpmask) {
903 rdp->passed_quiesce = 0;
906 zero_cpu_stall_ticks(rdp);
910 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
913 struct rcu_node *rnp;
915 local_irq_save(flags);
917 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
918 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
919 local_irq_restore(flags);
922 __note_new_gpnum(rsp, rnp, rdp);
923 raw_spin_unlock_irqrestore(&rnp->lock, flags);
927 * Did someone else start a new RCU grace period start since we last
928 * checked? Update local state appropriately if so. Must be called
929 * on the CPU corresponding to rdp.
932 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
937 local_irq_save(flags);
938 if (rdp->gpnum != rsp->gpnum) {
939 note_new_gpnum(rsp, rdp);
942 local_irq_restore(flags);
947 * Advance this CPU's callbacks, but only if the current grace period
948 * has ended. This may be called only from the CPU to whom the rdp
949 * belongs. In addition, the corresponding leaf rcu_node structure's
950 * ->lock must be held by the caller, with irqs disabled.
953 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
955 /* Did another grace period end? */
956 if (rdp->completed != rnp->completed) {
958 /* Advance callbacks. No harm if list empty. */
959 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
960 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
961 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
963 /* Remember that we saw this grace-period completion. */
964 rdp->completed = rnp->completed;
965 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
968 * If we were in an extended quiescent state, we may have
969 * missed some grace periods that others CPUs handled on
970 * our behalf. Catch up with this state to avoid noting
971 * spurious new grace periods. If another grace period
972 * has started, then rnp->gpnum will have advanced, so
973 * we will detect this later on.
975 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
976 rdp->gpnum = rdp->completed;
979 * If RCU does not need a quiescent state from this CPU,
980 * then make sure that this CPU doesn't go looking for one.
982 if ((rnp->qsmask & rdp->grpmask) == 0)
988 * Advance this CPU's callbacks, but only if the current grace period
989 * has ended. This may be called only from the CPU to whom the rdp
993 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
996 struct rcu_node *rnp;
998 local_irq_save(flags);
1000 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1001 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1002 local_irq_restore(flags);
1005 __rcu_process_gp_end(rsp, rnp, rdp);
1006 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1010 * Do per-CPU grace-period initialization for running CPU. The caller
1011 * must hold the lock of the leaf rcu_node structure corresponding to
1015 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1017 /* Prior grace period ended, so advance callbacks for current CPU. */
1018 __rcu_process_gp_end(rsp, rnp, rdp);
1021 * Because this CPU just now started the new grace period, we know
1022 * that all of its callbacks will be covered by this upcoming grace
1023 * period, even the ones that were registered arbitrarily recently.
1024 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1026 * Other CPUs cannot be sure exactly when the grace period started.
1027 * Therefore, their recently registered callbacks must pass through
1028 * an additional RCU_NEXT_READY stage, so that they will be handled
1029 * by the next RCU grace period.
1031 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1032 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1034 /* Set state so that this CPU will detect the next quiescent state. */
1035 __note_new_gpnum(rsp, rnp, rdp);
1039 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1040 * in preparation for detecting the next grace period. The caller must hold
1041 * the root node's ->lock, which is released before return. Hard irqs must
1044 * Note that it is legal for a dying CPU (which is marked as offline) to
1045 * invoke this function. This can happen when the dying CPU reports its
1049 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1050 __releases(rcu_get_root(rsp)->lock)
1052 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1053 struct rcu_node *rnp = rcu_get_root(rsp);
1055 if (!rcu_scheduler_fully_active ||
1056 !cpu_needs_another_gp(rsp, rdp)) {
1058 * Either the scheduler hasn't yet spawned the first
1059 * non-idle task or this CPU does not need another
1060 * grace period. Either way, don't start a new grace
1063 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1067 if (rsp->fqs_active) {
1069 * This CPU needs a grace period, but force_quiescent_state()
1070 * is running. Tell it to start one on this CPU's behalf.
1072 rsp->fqs_need_gp = 1;
1073 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1077 /* Advance to a new grace period and initialize state. */
1079 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1080 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1081 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1082 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1083 record_gp_stall_check_time(rsp);
1084 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1086 /* Exclude any concurrent CPU-hotplug operations. */
1087 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1090 * Set the quiescent-state-needed bits in all the rcu_node
1091 * structures for all currently online CPUs in breadth-first
1092 * order, starting from the root rcu_node structure. This
1093 * operation relies on the layout of the hierarchy within the
1094 * rsp->node[] array. Note that other CPUs will access only
1095 * the leaves of the hierarchy, which still indicate that no
1096 * grace period is in progress, at least until the corresponding
1097 * leaf node has been initialized. In addition, we have excluded
1098 * CPU-hotplug operations.
1100 * Note that the grace period cannot complete until we finish
1101 * the initialization process, as there will be at least one
1102 * qsmask bit set in the root node until that time, namely the
1103 * one corresponding to this CPU, due to the fact that we have
1106 rcu_for_each_node_breadth_first(rsp, rnp) {
1107 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1108 rcu_preempt_check_blocked_tasks(rnp);
1109 rnp->qsmask = rnp->qsmaskinit;
1110 rnp->gpnum = rsp->gpnum;
1111 rnp->completed = rsp->completed;
1112 if (rnp == rdp->mynode)
1113 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1114 rcu_preempt_boost_start_gp(rnp);
1115 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1116 rnp->level, rnp->grplo,
1117 rnp->grphi, rnp->qsmask);
1118 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1121 rnp = rcu_get_root(rsp);
1122 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1123 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1124 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1125 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1129 * Report a full set of quiescent states to the specified rcu_state
1130 * data structure. This involves cleaning up after the prior grace
1131 * period and letting rcu_start_gp() start up the next grace period
1132 * if one is needed. Note that the caller must hold rnp->lock, as
1133 * required by rcu_start_gp(), which will release it.
1135 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1136 __releases(rcu_get_root(rsp)->lock)
1138 unsigned long gp_duration;
1139 struct rcu_node *rnp = rcu_get_root(rsp);
1140 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1142 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1145 * Ensure that all grace-period and pre-grace-period activity
1146 * is seen before the assignment to rsp->completed.
1148 smp_mb(); /* See above block comment. */
1149 gp_duration = jiffies - rsp->gp_start;
1150 if (gp_duration > rsp->gp_max)
1151 rsp->gp_max = gp_duration;
1154 * We know the grace period is complete, but to everyone else
1155 * it appears to still be ongoing. But it is also the case
1156 * that to everyone else it looks like there is nothing that
1157 * they can do to advance the grace period. It is therefore
1158 * safe for us to drop the lock in order to mark the grace
1159 * period as completed in all of the rcu_node structures.
1161 * But if this CPU needs another grace period, it will take
1162 * care of this while initializing the next grace period.
1163 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1164 * because the callbacks have not yet been advanced: Those
1165 * callbacks are waiting on the grace period that just now
1168 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1169 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1172 * Propagate new ->completed value to rcu_node structures
1173 * so that other CPUs don't have to wait until the start
1174 * of the next grace period to process their callbacks.
1176 rcu_for_each_node_breadth_first(rsp, rnp) {
1177 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1178 rnp->completed = rsp->gpnum;
1179 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1181 rnp = rcu_get_root(rsp);
1182 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1185 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1186 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1187 rsp->fqs_state = RCU_GP_IDLE;
1188 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1192 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1193 * Allows quiescent states for a group of CPUs to be reported at one go
1194 * to the specified rcu_node structure, though all the CPUs in the group
1195 * must be represented by the same rcu_node structure (which need not be
1196 * a leaf rcu_node structure, though it often will be). That structure's
1197 * lock must be held upon entry, and it is released before return.
1200 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1201 struct rcu_node *rnp, unsigned long flags)
1202 __releases(rnp->lock)
1204 struct rcu_node *rnp_c;
1206 /* Walk up the rcu_node hierarchy. */
1208 if (!(rnp->qsmask & mask)) {
1210 /* Our bit has already been cleared, so done. */
1211 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1214 rnp->qsmask &= ~mask;
1215 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1216 mask, rnp->qsmask, rnp->level,
1217 rnp->grplo, rnp->grphi,
1219 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1221 /* Other bits still set at this level, so done. */
1222 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1225 mask = rnp->grpmask;
1226 if (rnp->parent == NULL) {
1228 /* No more levels. Exit loop holding root lock. */
1232 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1235 raw_spin_lock_irqsave(&rnp->lock, flags);
1236 WARN_ON_ONCE(rnp_c->qsmask);
1240 * Get here if we are the last CPU to pass through a quiescent
1241 * state for this grace period. Invoke rcu_report_qs_rsp()
1242 * to clean up and start the next grace period if one is needed.
1244 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1248 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1249 * structure. This must be either called from the specified CPU, or
1250 * called when the specified CPU is known to be offline (and when it is
1251 * also known that no other CPU is concurrently trying to help the offline
1252 * CPU). The lastcomp argument is used to make sure we are still in the
1253 * grace period of interest. We don't want to end the current grace period
1254 * based on quiescent states detected in an earlier grace period!
1257 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1259 unsigned long flags;
1261 struct rcu_node *rnp;
1264 raw_spin_lock_irqsave(&rnp->lock, flags);
1265 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1268 * The grace period in which this quiescent state was
1269 * recorded has ended, so don't report it upwards.
1270 * We will instead need a new quiescent state that lies
1271 * within the current grace period.
1273 rdp->passed_quiesce = 0; /* need qs for new gp. */
1274 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1277 mask = rdp->grpmask;
1278 if ((rnp->qsmask & mask) == 0) {
1279 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1281 rdp->qs_pending = 0;
1284 * This GP can't end until cpu checks in, so all of our
1285 * callbacks can be processed during the next GP.
1287 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1289 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1294 * Check to see if there is a new grace period of which this CPU
1295 * is not yet aware, and if so, set up local rcu_data state for it.
1296 * Otherwise, see if this CPU has just passed through its first
1297 * quiescent state for this grace period, and record that fact if so.
1300 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1302 /* If there is now a new grace period, record and return. */
1303 if (check_for_new_grace_period(rsp, rdp))
1307 * Does this CPU still need to do its part for current grace period?
1308 * If no, return and let the other CPUs do their part as well.
1310 if (!rdp->qs_pending)
1314 * Was there a quiescent state since the beginning of the grace
1315 * period? If no, then exit and wait for the next call.
1317 if (!rdp->passed_quiesce)
1321 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1324 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1327 #ifdef CONFIG_HOTPLUG_CPU
1330 * Send the specified CPU's RCU callbacks to the orphanage. The
1331 * specified CPU must be offline, and the caller must hold the
1335 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1336 struct rcu_node *rnp, struct rcu_data *rdp)
1341 * Orphan the callbacks. First adjust the counts. This is safe
1342 * because ->onofflock excludes _rcu_barrier()'s adoption of
1343 * the callbacks, thus no memory barrier is required.
1345 if (rdp->nxtlist != NULL) {
1346 rsp->qlen_lazy += rdp->qlen_lazy;
1347 rsp->qlen += rdp->qlen;
1348 rdp->n_cbs_orphaned += rdp->qlen;
1354 * Next, move those callbacks still needing a grace period to
1355 * the orphanage, where some other CPU will pick them up.
1356 * Some of the callbacks might have gone partway through a grace
1357 * period, but that is too bad. They get to start over because we
1358 * cannot assume that grace periods are synchronized across CPUs.
1359 * We don't bother updating the ->nxttail[] array yet, instead
1360 * we just reset the whole thing later on.
1362 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1363 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1364 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1365 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1369 * Then move the ready-to-invoke callbacks to the orphanage,
1370 * where some other CPU will pick them up. These will not be
1371 * required to pass though another grace period: They are done.
1373 if (rdp->nxtlist != NULL) {
1374 *rsp->orphan_donetail = rdp->nxtlist;
1375 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1378 /* Finally, initialize the rcu_data structure's list to empty. */
1379 rdp->nxtlist = NULL;
1380 for (i = 0; i < RCU_NEXT_SIZE; i++)
1381 rdp->nxttail[i] = &rdp->nxtlist;
1385 * Adopt the RCU callbacks from the specified rcu_state structure's
1386 * orphanage. The caller must hold the ->onofflock.
1388 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1391 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1394 * If there is an rcu_barrier() operation in progress, then
1395 * only the task doing that operation is permitted to adopt
1396 * callbacks. To do otherwise breaks rcu_barrier() and friends
1397 * by causing them to fail to wait for the callbacks in the
1400 if (rsp->rcu_barrier_in_progress &&
1401 rsp->rcu_barrier_in_progress != current)
1404 /* Do the accounting first. */
1405 rdp->qlen_lazy += rsp->qlen_lazy;
1406 rdp->qlen += rsp->qlen;
1407 rdp->n_cbs_adopted += rsp->qlen;
1408 if (rsp->qlen_lazy != rsp->qlen)
1409 rcu_idle_count_callbacks_posted();
1414 * We do not need a memory barrier here because the only way we
1415 * can get here if there is an rcu_barrier() in flight is if
1416 * we are the task doing the rcu_barrier().
1419 /* First adopt the ready-to-invoke callbacks. */
1420 if (rsp->orphan_donelist != NULL) {
1421 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1422 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1423 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1424 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1425 rdp->nxttail[i] = rsp->orphan_donetail;
1426 rsp->orphan_donelist = NULL;
1427 rsp->orphan_donetail = &rsp->orphan_donelist;
1430 /* And then adopt the callbacks that still need a grace period. */
1431 if (rsp->orphan_nxtlist != NULL) {
1432 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1433 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1434 rsp->orphan_nxtlist = NULL;
1435 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1440 * Trace the fact that this CPU is going offline.
1442 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1444 RCU_TRACE(unsigned long mask);
1445 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1446 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1448 RCU_TRACE(mask = rdp->grpmask);
1449 trace_rcu_grace_period(rsp->name,
1450 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1455 * The CPU has been completely removed, and some other CPU is reporting
1456 * this fact from process context. Do the remainder of the cleanup,
1457 * including orphaning the outgoing CPU's RCU callbacks, and also
1458 * adopting them, if there is no _rcu_barrier() instance running.
1459 * There can only be one CPU hotplug operation at a time, so no other
1460 * CPU can be attempting to update rcu_cpu_kthread_task.
1462 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1464 unsigned long flags;
1466 int need_report = 0;
1467 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1468 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1470 /* Adjust any no-longer-needed kthreads. */
1471 rcu_stop_cpu_kthread(cpu);
1472 rcu_node_kthread_setaffinity(rnp, -1);
1474 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1476 /* Exclude any attempts to start a new grace period. */
1477 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1479 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1480 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1481 rcu_adopt_orphan_cbs(rsp);
1483 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1484 mask = rdp->grpmask; /* rnp->grplo is constant. */
1486 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1487 rnp->qsmaskinit &= ~mask;
1488 if (rnp->qsmaskinit != 0) {
1489 if (rnp != rdp->mynode)
1490 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1493 if (rnp == rdp->mynode)
1494 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1496 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1497 mask = rnp->grpmask;
1499 } while (rnp != NULL);
1502 * We still hold the leaf rcu_node structure lock here, and
1503 * irqs are still disabled. The reason for this subterfuge is
1504 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1505 * held leads to deadlock.
1507 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1509 if (need_report & RCU_OFL_TASKS_NORM_GP)
1510 rcu_report_unblock_qs_rnp(rnp, flags);
1512 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1513 if (need_report & RCU_OFL_TASKS_EXP_GP)
1514 rcu_report_exp_rnp(rsp, rnp, true);
1517 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1519 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1523 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1527 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1531 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1534 * Invoke any RCU callbacks that have made it to the end of their grace
1535 * period. Thottle as specified by rdp->blimit.
1537 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1539 unsigned long flags;
1540 struct rcu_head *next, *list, **tail;
1541 int bl, count, count_lazy, i;
1543 /* If no callbacks are ready, just return.*/
1544 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1545 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1546 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1547 need_resched(), is_idle_task(current),
1548 rcu_is_callbacks_kthread());
1553 * Extract the list of ready callbacks, disabling to prevent
1554 * races with call_rcu() from interrupt handlers.
1556 local_irq_save(flags);
1557 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1559 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1560 list = rdp->nxtlist;
1561 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1562 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1563 tail = rdp->nxttail[RCU_DONE_TAIL];
1564 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1565 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1566 rdp->nxttail[i] = &rdp->nxtlist;
1567 local_irq_restore(flags);
1569 /* Invoke callbacks. */
1570 count = count_lazy = 0;
1574 debug_rcu_head_unqueue(list);
1575 if (__rcu_reclaim(rsp->name, list))
1578 /* Stop only if limit reached and CPU has something to do. */
1579 if (++count >= bl &&
1581 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1585 local_irq_save(flags);
1586 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1587 is_idle_task(current),
1588 rcu_is_callbacks_kthread());
1590 /* Update count, and requeue any remaining callbacks. */
1592 *tail = rdp->nxtlist;
1593 rdp->nxtlist = list;
1594 for (i = 0; i < RCU_NEXT_SIZE; i++)
1595 if (&rdp->nxtlist == rdp->nxttail[i])
1596 rdp->nxttail[i] = tail;
1600 smp_mb(); /* List handling before counting for rcu_barrier(). */
1601 rdp->qlen_lazy -= count_lazy;
1603 rdp->n_cbs_invoked += count;
1605 /* Reinstate batch limit if we have worked down the excess. */
1606 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1607 rdp->blimit = blimit;
1609 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1610 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1611 rdp->qlen_last_fqs_check = 0;
1612 rdp->n_force_qs_snap = rsp->n_force_qs;
1613 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1614 rdp->qlen_last_fqs_check = rdp->qlen;
1616 local_irq_restore(flags);
1618 /* Re-invoke RCU core processing if there are callbacks remaining. */
1619 if (cpu_has_callbacks_ready_to_invoke(rdp))
1624 * Check to see if this CPU is in a non-context-switch quiescent state
1625 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1626 * Also schedule RCU core processing.
1628 * This function must be called from hardirq context. It is normally
1629 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1630 * false, there is no point in invoking rcu_check_callbacks().
1632 void rcu_check_callbacks(int cpu, int user)
1634 trace_rcu_utilization("Start scheduler-tick");
1635 increment_cpu_stall_ticks();
1636 if (user || rcu_is_cpu_rrupt_from_idle()) {
1639 * Get here if this CPU took its interrupt from user
1640 * mode or from the idle loop, and if this is not a
1641 * nested interrupt. In this case, the CPU is in
1642 * a quiescent state, so note it.
1644 * No memory barrier is required here because both
1645 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1646 * variables that other CPUs neither access nor modify,
1647 * at least not while the corresponding CPU is online.
1653 } else if (!in_softirq()) {
1656 * Get here if this CPU did not take its interrupt from
1657 * softirq, in other words, if it is not interrupting
1658 * a rcu_bh read-side critical section. This is an _bh
1659 * critical section, so note it.
1664 rcu_preempt_check_callbacks(cpu);
1665 if (rcu_pending(cpu))
1667 trace_rcu_utilization("End scheduler-tick");
1671 * Scan the leaf rcu_node structures, processing dyntick state for any that
1672 * have not yet encountered a quiescent state, using the function specified.
1673 * Also initiate boosting for any threads blocked on the root rcu_node.
1675 * The caller must have suppressed start of new grace periods.
1677 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1681 unsigned long flags;
1683 struct rcu_node *rnp;
1685 rcu_for_each_leaf_node(rsp, rnp) {
1687 raw_spin_lock_irqsave(&rnp->lock, flags);
1688 if (!rcu_gp_in_progress(rsp)) {
1689 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1692 if (rnp->qsmask == 0) {
1693 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1698 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1699 if ((rnp->qsmask & bit) != 0 &&
1700 f(per_cpu_ptr(rsp->rda, cpu)))
1705 /* rcu_report_qs_rnp() releases rnp->lock. */
1706 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1709 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1711 rnp = rcu_get_root(rsp);
1712 if (rnp->qsmask == 0) {
1713 raw_spin_lock_irqsave(&rnp->lock, flags);
1714 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1719 * Force quiescent states on reluctant CPUs, and also detect which
1720 * CPUs are in dyntick-idle mode.
1722 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1724 unsigned long flags;
1725 struct rcu_node *rnp = rcu_get_root(rsp);
1727 trace_rcu_utilization("Start fqs");
1728 if (!rcu_gp_in_progress(rsp)) {
1729 trace_rcu_utilization("End fqs");
1730 return; /* No grace period in progress, nothing to force. */
1732 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1733 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1734 trace_rcu_utilization("End fqs");
1735 return; /* Someone else is already on the job. */
1737 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1738 goto unlock_fqs_ret; /* no emergency and done recently. */
1740 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1741 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1742 if(!rcu_gp_in_progress(rsp)) {
1743 rsp->n_force_qs_ngp++;
1744 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1745 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1747 rsp->fqs_active = 1;
1748 switch (rsp->fqs_state) {
1752 break; /* grace period idle or initializing, ignore. */
1754 case RCU_SAVE_DYNTICK:
1755 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1756 break; /* So gcc recognizes the dead code. */
1758 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1760 /* Record dyntick-idle state. */
1761 force_qs_rnp(rsp, dyntick_save_progress_counter);
1762 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1763 if (rcu_gp_in_progress(rsp))
1764 rsp->fqs_state = RCU_FORCE_QS;
1769 /* Check dyntick-idle state, send IPI to laggarts. */
1770 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1771 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1773 /* Leave state in case more forcing is required. */
1775 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1778 rsp->fqs_active = 0;
1779 if (rsp->fqs_need_gp) {
1780 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1781 rsp->fqs_need_gp = 0;
1782 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1783 trace_rcu_utilization("End fqs");
1786 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1788 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1789 trace_rcu_utilization("End fqs");
1793 * This does the RCU core processing work for the specified rcu_state
1794 * and rcu_data structures. This may be called only from the CPU to
1795 * whom the rdp belongs.
1798 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1800 unsigned long flags;
1802 WARN_ON_ONCE(rdp->beenonline == 0);
1805 * If an RCU GP has gone long enough, go check for dyntick
1806 * idle CPUs and, if needed, send resched IPIs.
1808 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1809 force_quiescent_state(rsp, 1);
1812 * Advance callbacks in response to end of earlier grace
1813 * period that some other CPU ended.
1815 rcu_process_gp_end(rsp, rdp);
1817 /* Update RCU state based on any recent quiescent states. */
1818 rcu_check_quiescent_state(rsp, rdp);
1820 /* Does this CPU require a not-yet-started grace period? */
1821 if (cpu_needs_another_gp(rsp, rdp)) {
1822 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1823 rcu_start_gp(rsp, flags); /* releases above lock */
1826 /* If there are callbacks ready, invoke them. */
1827 if (cpu_has_callbacks_ready_to_invoke(rdp))
1828 invoke_rcu_callbacks(rsp, rdp);
1832 * Do RCU core processing for the current CPU.
1834 static void rcu_process_callbacks(struct softirq_action *unused)
1836 trace_rcu_utilization("Start RCU core");
1837 __rcu_process_callbacks(&rcu_sched_state,
1838 &__get_cpu_var(rcu_sched_data));
1839 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1840 rcu_preempt_process_callbacks();
1841 trace_rcu_utilization("End RCU core");
1845 * Schedule RCU callback invocation. If the specified type of RCU
1846 * does not support RCU priority boosting, just do a direct call,
1847 * otherwise wake up the per-CPU kernel kthread. Note that because we
1848 * are running on the current CPU with interrupts disabled, the
1849 * rcu_cpu_kthread_task cannot disappear out from under us.
1851 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1853 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1855 if (likely(!rsp->boost)) {
1856 rcu_do_batch(rsp, rdp);
1859 invoke_rcu_callbacks_kthread();
1862 static void invoke_rcu_core(void)
1864 raise_softirq(RCU_SOFTIRQ);
1868 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1869 struct rcu_state *rsp, bool lazy)
1871 unsigned long flags;
1872 struct rcu_data *rdp;
1874 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1875 debug_rcu_head_queue(head);
1879 smp_mb(); /* Ensure RCU update seen before callback registry. */
1882 * Opportunistically note grace-period endings and beginnings.
1883 * Note that we might see a beginning right after we see an
1884 * end, but never vice versa, since this CPU has to pass through
1885 * a quiescent state betweentimes.
1887 local_irq_save(flags);
1888 rdp = this_cpu_ptr(rsp->rda);
1890 /* Add the callback to our list. */
1895 rcu_idle_count_callbacks_posted();
1896 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1897 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1898 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1900 if (__is_kfree_rcu_offset((unsigned long)func))
1901 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1902 rdp->qlen_lazy, rdp->qlen);
1904 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1906 /* If interrupts were disabled, don't dive into RCU core. */
1907 if (irqs_disabled_flags(flags)) {
1908 local_irq_restore(flags);
1913 * Force the grace period if too many callbacks or too long waiting.
1914 * Enforce hysteresis, and don't invoke force_quiescent_state()
1915 * if some other CPU has recently done so. Also, don't bother
1916 * invoking force_quiescent_state() if the newly enqueued callback
1917 * is the only one waiting for a grace period to complete.
1919 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1921 /* Are we ignoring a completed grace period? */
1922 rcu_process_gp_end(rsp, rdp);
1923 check_for_new_grace_period(rsp, rdp);
1925 /* Start a new grace period if one not already started. */
1926 if (!rcu_gp_in_progress(rsp)) {
1927 unsigned long nestflag;
1928 struct rcu_node *rnp_root = rcu_get_root(rsp);
1930 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1931 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1933 /* Give the grace period a kick. */
1934 rdp->blimit = LONG_MAX;
1935 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1936 *rdp->nxttail[RCU_DONE_TAIL] != head)
1937 force_quiescent_state(rsp, 0);
1938 rdp->n_force_qs_snap = rsp->n_force_qs;
1939 rdp->qlen_last_fqs_check = rdp->qlen;
1941 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1942 force_quiescent_state(rsp, 1);
1943 local_irq_restore(flags);
1947 * Queue an RCU-sched callback for invocation after a grace period.
1949 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1951 __call_rcu(head, func, &rcu_sched_state, 0);
1953 EXPORT_SYMBOL_GPL(call_rcu_sched);
1956 * Queue an RCU callback for invocation after a quicker grace period.
1958 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1960 __call_rcu(head, func, &rcu_bh_state, 0);
1962 EXPORT_SYMBOL_GPL(call_rcu_bh);
1965 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1966 * any blocking grace-period wait automatically implies a grace period
1967 * if there is only one CPU online at any point time during execution
1968 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
1969 * occasionally incorrectly indicate that there are multiple CPUs online
1970 * when there was in fact only one the whole time, as this just adds
1971 * some overhead: RCU still operates correctly.
1973 * Of course, sampling num_online_cpus() with preemption enabled can
1974 * give erroneous results if there are concurrent CPU-hotplug operations.
1975 * For example, given a demonic sequence of preemptions in num_online_cpus()
1976 * and CPU-hotplug operations, there could be two or more CPUs online at
1977 * all times, but num_online_cpus() might well return one (or even zero).
1979 * However, all such demonic sequences require at least one CPU-offline
1980 * operation. Furthermore, rcu_blocking_is_gp() giving the wrong answer
1981 * is only a problem if there is an RCU read-side critical section executing
1982 * throughout. But RCU-sched and RCU-bh read-side critical sections
1983 * disable either preemption or bh, which prevents a CPU from going offline.
1984 * Therefore, the only way that rcu_blocking_is_gp() can incorrectly return
1985 * that there is only one CPU when in fact there was more than one throughout
1986 * is when there were no RCU readers in the system. If there are no
1987 * RCU readers, the grace period by definition can be of zero length,
1988 * regardless of the number of online CPUs.
1990 static inline int rcu_blocking_is_gp(void)
1992 might_sleep(); /* Check for RCU read-side critical section. */
1993 return num_online_cpus() <= 1;
1997 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1999 * Control will return to the caller some time after a full rcu-sched
2000 * grace period has elapsed, in other words after all currently executing
2001 * rcu-sched read-side critical sections have completed. These read-side
2002 * critical sections are delimited by rcu_read_lock_sched() and
2003 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2004 * local_irq_disable(), and so on may be used in place of
2005 * rcu_read_lock_sched().
2007 * This means that all preempt_disable code sequences, including NMI and
2008 * hardware-interrupt handlers, in progress on entry will have completed
2009 * before this primitive returns. However, this does not guarantee that
2010 * softirq handlers will have completed, since in some kernels, these
2011 * handlers can run in process context, and can block.
2013 * This primitive provides the guarantees made by the (now removed)
2014 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2015 * guarantees that rcu_read_lock() sections will have completed.
2016 * In "classic RCU", these two guarantees happen to be one and
2017 * the same, but can differ in realtime RCU implementations.
2019 void synchronize_sched(void)
2021 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2022 !lock_is_held(&rcu_lock_map) &&
2023 !lock_is_held(&rcu_sched_lock_map),
2024 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2025 if (rcu_blocking_is_gp())
2027 wait_rcu_gp(call_rcu_sched);
2029 EXPORT_SYMBOL_GPL(synchronize_sched);
2032 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2034 * Control will return to the caller some time after a full rcu_bh grace
2035 * period has elapsed, in other words after all currently executing rcu_bh
2036 * read-side critical sections have completed. RCU read-side critical
2037 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2038 * and may be nested.
2040 void synchronize_rcu_bh(void)
2042 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2043 !lock_is_held(&rcu_lock_map) &&
2044 !lock_is_held(&rcu_sched_lock_map),
2045 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2046 if (rcu_blocking_is_gp())
2048 wait_rcu_gp(call_rcu_bh);
2050 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2052 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2053 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2055 static int synchronize_sched_expedited_cpu_stop(void *data)
2058 * There must be a full memory barrier on each affected CPU
2059 * between the time that try_stop_cpus() is called and the
2060 * time that it returns.
2062 * In the current initial implementation of cpu_stop, the
2063 * above condition is already met when the control reaches
2064 * this point and the following smp_mb() is not strictly
2065 * necessary. Do smp_mb() anyway for documentation and
2066 * robustness against future implementation changes.
2068 smp_mb(); /* See above comment block. */
2073 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2075 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2076 * approach to force the grace period to end quickly. This consumes
2077 * significant time on all CPUs and is unfriendly to real-time workloads,
2078 * so is thus not recommended for any sort of common-case code. In fact,
2079 * if you are using synchronize_sched_expedited() in a loop, please
2080 * restructure your code to batch your updates, and then use a single
2081 * synchronize_sched() instead.
2083 * Note that it is illegal to call this function while holding any lock
2084 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2085 * to call this function from a CPU-hotplug notifier. Failing to observe
2086 * these restriction will result in deadlock.
2088 * This implementation can be thought of as an application of ticket
2089 * locking to RCU, with sync_sched_expedited_started and
2090 * sync_sched_expedited_done taking on the roles of the halves
2091 * of the ticket-lock word. Each task atomically increments
2092 * sync_sched_expedited_started upon entry, snapshotting the old value,
2093 * then attempts to stop all the CPUs. If this succeeds, then each
2094 * CPU will have executed a context switch, resulting in an RCU-sched
2095 * grace period. We are then done, so we use atomic_cmpxchg() to
2096 * update sync_sched_expedited_done to match our snapshot -- but
2097 * only if someone else has not already advanced past our snapshot.
2099 * On the other hand, if try_stop_cpus() fails, we check the value
2100 * of sync_sched_expedited_done. If it has advanced past our
2101 * initial snapshot, then someone else must have forced a grace period
2102 * some time after we took our snapshot. In this case, our work is
2103 * done for us, and we can simply return. Otherwise, we try again,
2104 * but keep our initial snapshot for purposes of checking for someone
2105 * doing our work for us.
2107 * If we fail too many times in a row, we fall back to synchronize_sched().
2109 void synchronize_sched_expedited(void)
2111 int firstsnap, s, snap, trycount = 0;
2113 /* Note that atomic_inc_return() implies full memory barrier. */
2114 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2116 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2119 * Each pass through the following loop attempts to force a
2120 * context switch on each CPU.
2122 while (try_stop_cpus(cpu_online_mask,
2123 synchronize_sched_expedited_cpu_stop,
2127 /* No joy, try again later. Or just synchronize_sched(). */
2128 if (trycount++ < 10)
2129 udelay(trycount * num_online_cpus());
2131 synchronize_sched();
2135 /* Check to see if someone else did our work for us. */
2136 s = atomic_read(&sync_sched_expedited_done);
2137 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2138 smp_mb(); /* ensure test happens before caller kfree */
2143 * Refetching sync_sched_expedited_started allows later
2144 * callers to piggyback on our grace period. We subtract
2145 * 1 to get the same token that the last incrementer got.
2146 * We retry after they started, so our grace period works
2147 * for them, and they started after our first try, so their
2148 * grace period works for us.
2151 snap = atomic_read(&sync_sched_expedited_started);
2152 smp_mb(); /* ensure read is before try_stop_cpus(). */
2156 * Everyone up to our most recent fetch is covered by our grace
2157 * period. Update the counter, but only if our work is still
2158 * relevant -- which it won't be if someone who started later
2159 * than we did beat us to the punch.
2162 s = atomic_read(&sync_sched_expedited_done);
2163 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2164 smp_mb(); /* ensure test happens before caller kfree */
2167 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2171 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2174 * Check to see if there is any immediate RCU-related work to be done
2175 * by the current CPU, for the specified type of RCU, returning 1 if so.
2176 * The checks are in order of increasing expense: checks that can be
2177 * carried out against CPU-local state are performed first. However,
2178 * we must check for CPU stalls first, else we might not get a chance.
2180 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2182 struct rcu_node *rnp = rdp->mynode;
2184 rdp->n_rcu_pending++;
2186 /* Check for CPU stalls, if enabled. */
2187 check_cpu_stall(rsp, rdp);
2189 /* Is the RCU core waiting for a quiescent state from this CPU? */
2190 if (rcu_scheduler_fully_active &&
2191 rdp->qs_pending && !rdp->passed_quiesce) {
2194 * If force_quiescent_state() coming soon and this CPU
2195 * needs a quiescent state, and this is either RCU-sched
2196 * or RCU-bh, force a local reschedule.
2198 rdp->n_rp_qs_pending++;
2199 if (!rdp->preemptible &&
2200 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2203 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2204 rdp->n_rp_report_qs++;
2208 /* Does this CPU have callbacks ready to invoke? */
2209 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2210 rdp->n_rp_cb_ready++;
2214 /* Has RCU gone idle with this CPU needing another grace period? */
2215 if (cpu_needs_another_gp(rsp, rdp)) {
2216 rdp->n_rp_cpu_needs_gp++;
2220 /* Has another RCU grace period completed? */
2221 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2222 rdp->n_rp_gp_completed++;
2226 /* Has a new RCU grace period started? */
2227 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2228 rdp->n_rp_gp_started++;
2232 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2233 if (rcu_gp_in_progress(rsp) &&
2234 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2235 rdp->n_rp_need_fqs++;
2240 rdp->n_rp_need_nothing++;
2245 * Check to see if there is any immediate RCU-related work to be done
2246 * by the current CPU, returning 1 if so. This function is part of the
2247 * RCU implementation; it is -not- an exported member of the RCU API.
2249 static int rcu_pending(int cpu)
2251 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
2252 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2253 rcu_preempt_pending(cpu);
2257 * Check to see if any future RCU-related work will need to be done
2258 * by the current CPU, even if none need be done immediately, returning
2261 static int rcu_cpu_has_callbacks(int cpu)
2263 /* RCU callbacks either ready or pending? */
2264 return per_cpu(rcu_sched_data, cpu).nxtlist ||
2265 per_cpu(rcu_bh_data, cpu).nxtlist ||
2266 rcu_preempt_cpu_has_callbacks(cpu);
2270 * RCU callback function for _rcu_barrier(). If we are last, wake
2271 * up the task executing _rcu_barrier().
2273 static void rcu_barrier_callback(struct rcu_head *notused)
2275 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2276 complete(&rcu_barrier_completion);
2280 * Called with preemption disabled, and from cross-cpu IRQ context.
2282 static void rcu_barrier_func(void *type)
2284 struct rcu_state *rsp = type;
2285 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2287 atomic_inc(&rcu_barrier_cpu_count);
2288 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2292 * Orchestrate the specified type of RCU barrier, waiting for all
2293 * RCU callbacks of the specified type to complete.
2295 static void _rcu_barrier(struct rcu_state *rsp)
2298 unsigned long flags;
2299 struct rcu_data *rdp;
2302 init_rcu_head_on_stack(&rh);
2304 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2305 mutex_lock(&rcu_barrier_mutex);
2307 smp_mb(); /* Prevent any prior operations from leaking in. */
2310 * Initialize the count to one rather than to zero in order to
2311 * avoid a too-soon return to zero in case of a short grace period
2312 * (or preemption of this task). Also flag this task as doing
2313 * an rcu_barrier(). This will prevent anyone else from adopting
2314 * orphaned callbacks, which could cause otherwise failure if a
2315 * CPU went offline and quickly came back online. To see this,
2316 * consider the following sequence of events:
2318 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2319 * 2. CPU 1 goes offline, orphaning its callbacks.
2320 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2321 * 4. CPU 1 comes back online.
2322 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2323 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2324 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2326 init_completion(&rcu_barrier_completion);
2327 atomic_set(&rcu_barrier_cpu_count, 1);
2328 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2329 rsp->rcu_barrier_in_progress = current;
2330 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2333 * Force every CPU with callbacks to register a new callback
2334 * that will tell us when all the preceding callbacks have
2335 * been invoked. If an offline CPU has callbacks, wait for
2336 * it to either come back online or to finish orphaning those
2339 for_each_possible_cpu(cpu) {
2341 rdp = per_cpu_ptr(rsp->rda, cpu);
2342 if (cpu_is_offline(cpu)) {
2344 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2345 schedule_timeout_interruptible(1);
2346 } else if (ACCESS_ONCE(rdp->qlen)) {
2347 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2355 * Now that all online CPUs have rcu_barrier_callback() callbacks
2356 * posted, we can adopt all of the orphaned callbacks and place
2357 * an rcu_barrier_callback() callback after them. When that is done,
2358 * we are guaranteed to have an rcu_barrier_callback() callback
2359 * following every callback that could possibly have been
2360 * registered before _rcu_barrier() was called.
2362 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2363 rcu_adopt_orphan_cbs(rsp);
2364 rsp->rcu_barrier_in_progress = NULL;
2365 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2366 atomic_inc(&rcu_barrier_cpu_count);
2367 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2368 rsp->call(&rh, rcu_barrier_callback);
2371 * Now that we have an rcu_barrier_callback() callback on each
2372 * CPU, and thus each counted, remove the initial count.
2374 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2375 complete(&rcu_barrier_completion);
2377 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2378 wait_for_completion(&rcu_barrier_completion);
2380 /* Other rcu_barrier() invocations can now safely proceed. */
2381 mutex_unlock(&rcu_barrier_mutex);
2383 destroy_rcu_head_on_stack(&rh);
2387 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2389 void rcu_barrier_bh(void)
2391 _rcu_barrier(&rcu_bh_state);
2393 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2396 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2398 void rcu_barrier_sched(void)
2400 _rcu_barrier(&rcu_sched_state);
2402 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2405 * Do boot-time initialization of a CPU's per-CPU RCU data.
2408 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2410 unsigned long flags;
2412 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2413 struct rcu_node *rnp = rcu_get_root(rsp);
2415 /* Set up local state, ensuring consistent view of global state. */
2416 raw_spin_lock_irqsave(&rnp->lock, flags);
2417 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2418 rdp->nxtlist = NULL;
2419 for (i = 0; i < RCU_NEXT_SIZE; i++)
2420 rdp->nxttail[i] = &rdp->nxtlist;
2423 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2424 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2425 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2428 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2432 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2433 * offline event can be happening at a given time. Note also that we
2434 * can accept some slop in the rsp->completed access due to the fact
2435 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2437 static void __cpuinit
2438 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2440 unsigned long flags;
2442 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2443 struct rcu_node *rnp = rcu_get_root(rsp);
2445 /* Set up local state, ensuring consistent view of global state. */
2446 raw_spin_lock_irqsave(&rnp->lock, flags);
2447 rdp->beenonline = 1; /* We have now been online. */
2448 rdp->preemptible = preemptible;
2449 rdp->qlen_last_fqs_check = 0;
2450 rdp->n_force_qs_snap = rsp->n_force_qs;
2451 rdp->blimit = blimit;
2452 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2453 atomic_set(&rdp->dynticks->dynticks,
2454 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2455 rcu_prepare_for_idle_init(cpu);
2456 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2459 * A new grace period might start here. If so, we won't be part
2460 * of it, but that is OK, as we are currently in a quiescent state.
2463 /* Exclude any attempts to start a new GP on large systems. */
2464 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2466 /* Add CPU to rcu_node bitmasks. */
2468 mask = rdp->grpmask;
2470 /* Exclude any attempts to start a new GP on small systems. */
2471 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2472 rnp->qsmaskinit |= mask;
2473 mask = rnp->grpmask;
2474 if (rnp == rdp->mynode) {
2476 * If there is a grace period in progress, we will
2477 * set up to wait for it next time we run the
2480 rdp->gpnum = rnp->completed;
2481 rdp->completed = rnp->completed;
2482 rdp->passed_quiesce = 0;
2483 rdp->qs_pending = 0;
2484 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2485 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2487 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2489 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2491 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2494 static void __cpuinit rcu_prepare_cpu(int cpu)
2496 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2497 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2498 rcu_preempt_init_percpu_data(cpu);
2502 * Handle CPU online/offline notification events.
2504 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2505 unsigned long action, void *hcpu)
2507 long cpu = (long)hcpu;
2508 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2509 struct rcu_node *rnp = rdp->mynode;
2511 trace_rcu_utilization("Start CPU hotplug");
2513 case CPU_UP_PREPARE:
2514 case CPU_UP_PREPARE_FROZEN:
2515 rcu_prepare_cpu(cpu);
2516 rcu_prepare_kthreads(cpu);
2519 case CPU_DOWN_FAILED:
2520 rcu_node_kthread_setaffinity(rnp, -1);
2521 rcu_cpu_kthread_setrt(cpu, 1);
2523 case CPU_DOWN_PREPARE:
2524 rcu_node_kthread_setaffinity(rnp, cpu);
2525 rcu_cpu_kthread_setrt(cpu, 0);
2528 case CPU_DYING_FROZEN:
2530 * The whole machine is "stopped" except this CPU, so we can
2531 * touch any data without introducing corruption. We send the
2532 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2534 rcu_cleanup_dying_cpu(&rcu_bh_state);
2535 rcu_cleanup_dying_cpu(&rcu_sched_state);
2536 rcu_preempt_cleanup_dying_cpu();
2537 rcu_cleanup_after_idle(cpu);
2540 case CPU_DEAD_FROZEN:
2541 case CPU_UP_CANCELED:
2542 case CPU_UP_CANCELED_FROZEN:
2543 rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2544 rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2545 rcu_preempt_cleanup_dead_cpu(cpu);
2550 trace_rcu_utilization("End CPU hotplug");
2555 * This function is invoked towards the end of the scheduler's initialization
2556 * process. Before this is called, the idle task might contain
2557 * RCU read-side critical sections (during which time, this idle
2558 * task is booting the system). After this function is called, the
2559 * idle tasks are prohibited from containing RCU read-side critical
2560 * sections. This function also enables RCU lockdep checking.
2562 void rcu_scheduler_starting(void)
2564 WARN_ON(num_online_cpus() != 1);
2565 WARN_ON(nr_context_switches() > 0);
2566 rcu_scheduler_active = 1;
2570 * Compute the per-level fanout, either using the exact fanout specified
2571 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2573 #ifdef CONFIG_RCU_FANOUT_EXACT
2574 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2578 for (i = rcu_num_lvls - 1; i > 0; i--)
2579 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2580 rsp->levelspread[0] = rcu_fanout_leaf;
2582 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2583 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2590 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2591 ccur = rsp->levelcnt[i];
2592 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2596 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2599 * Helper function for rcu_init() that initializes one rcu_state structure.
2601 static void __init rcu_init_one(struct rcu_state *rsp,
2602 struct rcu_data __percpu *rda)
2604 static char *buf[] = { "rcu_node_level_0",
2607 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2611 struct rcu_node *rnp;
2613 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2615 /* Initialize the level-tracking arrays. */
2617 for (i = 0; i < rcu_num_lvls; i++)
2618 rsp->levelcnt[i] = num_rcu_lvl[i];
2619 for (i = 1; i < rcu_num_lvls; i++)
2620 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2621 rcu_init_levelspread(rsp);
2623 /* Initialize the elements themselves, starting from the leaves. */
2625 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2626 cpustride *= rsp->levelspread[i];
2627 rnp = rsp->level[i];
2628 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2629 raw_spin_lock_init(&rnp->lock);
2630 lockdep_set_class_and_name(&rnp->lock,
2631 &rcu_node_class[i], buf[i]);
2634 rnp->qsmaskinit = 0;
2635 rnp->grplo = j * cpustride;
2636 rnp->grphi = (j + 1) * cpustride - 1;
2637 if (rnp->grphi >= NR_CPUS)
2638 rnp->grphi = NR_CPUS - 1;
2644 rnp->grpnum = j % rsp->levelspread[i - 1];
2645 rnp->grpmask = 1UL << rnp->grpnum;
2646 rnp->parent = rsp->level[i - 1] +
2647 j / rsp->levelspread[i - 1];
2650 INIT_LIST_HEAD(&rnp->blkd_tasks);
2655 rnp = rsp->level[rcu_num_lvls - 1];
2656 for_each_possible_cpu(i) {
2657 while (i > rnp->grphi)
2659 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2660 rcu_boot_init_percpu_data(i, rsp);
2665 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2666 * replace the definitions in rcutree.h because those are needed to size
2667 * the ->node array in the rcu_state structure.
2669 static void __init rcu_init_geometry(void)
2674 int rcu_capacity[MAX_RCU_LVLS + 1];
2676 /* If the compile-time values are accurate, just leave. */
2677 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2681 * Compute number of nodes that can be handled an rcu_node tree
2682 * with the given number of levels. Setting rcu_capacity[0] makes
2683 * some of the arithmetic easier.
2685 rcu_capacity[0] = 1;
2686 rcu_capacity[1] = rcu_fanout_leaf;
2687 for (i = 2; i <= MAX_RCU_LVLS; i++)
2688 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2691 * The boot-time rcu_fanout_leaf parameter is only permitted
2692 * to increase the leaf-level fanout, not decrease it. Of course,
2693 * the leaf-level fanout cannot exceed the number of bits in
2694 * the rcu_node masks. Finally, the tree must be able to accommodate
2695 * the configured number of CPUs. Complain and fall back to the
2696 * compile-time values if these limits are exceeded.
2698 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2699 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2700 n > rcu_capacity[MAX_RCU_LVLS]) {
2705 /* Calculate the number of rcu_nodes at each level of the tree. */
2706 for (i = 1; i <= MAX_RCU_LVLS; i++)
2707 if (n <= rcu_capacity[i]) {
2708 for (j = 0; j <= i; j++)
2710 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2712 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2717 /* Calculate the total number of rcu_node structures. */
2719 for (i = 0; i <= MAX_RCU_LVLS; i++)
2720 rcu_num_nodes += num_rcu_lvl[i];
2724 void __init rcu_init(void)
2728 rcu_bootup_announce();
2729 rcu_init_geometry();
2730 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2731 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2732 __rcu_init_preempt();
2733 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2736 * We don't need protection against CPU-hotplug here because
2737 * this is called early in boot, before either interrupts
2738 * or the scheduler are operational.
2740 cpu_notifier(rcu_cpu_notify, 0);
2741 for_each_online_cpu(cpu)
2742 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2743 check_cpu_stall_init();
2746 #include "rcutree_plugin.h"