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(int, rcu_cpu_kthread_cpu);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
139 DEFINE_PER_CPU(char, rcu_cpu_has_work);
141 #endif /* #ifdef CONFIG_RCU_BOOST */
143 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
144 static void invoke_rcu_core(void);
145 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
148 * Track the rcutorture test sequence number and the update version
149 * number within a given test. The rcutorture_testseq is incremented
150 * on every rcutorture module load and unload, so has an odd value
151 * when a test is running. The rcutorture_vernum is set to zero
152 * when rcutorture starts and is incremented on each rcutorture update.
153 * These variables enable correlating rcutorture output with the
154 * RCU tracing information.
156 unsigned long rcutorture_testseq;
157 unsigned long rcutorture_vernum;
160 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
161 * permit this function to be invoked without holding the root rcu_node
162 * structure's ->lock, but of course results can be subject to change.
164 static int rcu_gp_in_progress(struct rcu_state *rsp)
166 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
170 * Note a quiescent state. Because we do not need to know
171 * how many quiescent states passed, just if there was at least
172 * one since the start of the grace period, this just sets a flag.
173 * The caller must have disabled preemption.
175 void rcu_sched_qs(int cpu)
177 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
179 if (rdp->passed_quiesce == 0)
180 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
181 rdp->passed_quiesce = 1;
184 void rcu_bh_qs(int cpu)
186 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
188 if (rdp->passed_quiesce == 0)
189 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
190 rdp->passed_quiesce = 1;
194 * Note a context switch. This is a quiescent state for RCU-sched,
195 * and requires special handling for preemptible RCU.
196 * The caller must have disabled preemption.
198 void rcu_note_context_switch(int cpu)
200 trace_rcu_utilization("Start context switch");
202 rcu_preempt_note_context_switch(cpu);
203 trace_rcu_utilization("End context switch");
205 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
207 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
208 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
209 .dynticks = ATOMIC_INIT(1),
212 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
213 static int qhimark = 10000; /* If this many pending, ignore blimit. */
214 static int qlowmark = 100; /* Once only this many pending, use blimit. */
216 module_param(blimit, int, 0444);
217 module_param(qhimark, int, 0444);
218 module_param(qlowmark, int, 0444);
220 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
221 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
223 module_param(rcu_cpu_stall_suppress, int, 0644);
224 module_param(rcu_cpu_stall_timeout, int, 0644);
226 static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
227 static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
229 module_param(jiffies_till_first_fqs, ulong, 0644);
230 module_param(jiffies_till_next_fqs, ulong, 0644);
232 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
233 static void force_quiescent_state(struct rcu_state *rsp);
234 static int rcu_pending(int cpu);
237 * Return the number of RCU-sched batches processed thus far for debug & stats.
239 long rcu_batches_completed_sched(void)
241 return rcu_sched_state.completed;
243 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
246 * Return the number of RCU BH batches processed thus far for debug & stats.
248 long rcu_batches_completed_bh(void)
250 return rcu_bh_state.completed;
252 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
255 * Force a quiescent state for RCU BH.
257 void rcu_bh_force_quiescent_state(void)
259 force_quiescent_state(&rcu_bh_state);
261 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
264 * Record the number of times rcutorture tests have been initiated and
265 * terminated. This information allows the debugfs tracing stats to be
266 * correlated to the rcutorture messages, even when the rcutorture module
267 * is being repeatedly loaded and unloaded. In other words, we cannot
268 * store this state in rcutorture itself.
270 void rcutorture_record_test_transition(void)
272 rcutorture_testseq++;
273 rcutorture_vernum = 0;
275 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
278 * Record the number of writer passes through the current rcutorture test.
279 * This is also used to correlate debugfs tracing stats with the rcutorture
282 void rcutorture_record_progress(unsigned long vernum)
286 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
289 * Force a quiescent state for RCU-sched.
291 void rcu_sched_force_quiescent_state(void)
293 force_quiescent_state(&rcu_sched_state);
295 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
298 * Does the CPU have callbacks ready to be invoked?
301 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
303 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
307 * Does the current CPU require a yet-as-unscheduled grace period?
310 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
312 return *rdp->nxttail[RCU_DONE_TAIL +
313 ACCESS_ONCE(rsp->completed) != rdp->completed] &&
314 !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 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
328 * If the new value of the ->dynticks_nesting counter now is zero,
329 * we really have entered idle, and must do the appropriate accounting.
330 * The caller must have disabled interrupts.
332 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
334 trace_rcu_dyntick("Start", oldval, 0);
335 if (!is_idle_task(current)) {
336 struct task_struct *idle = idle_task(smp_processor_id());
338 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
339 ftrace_dump(DUMP_ORIG);
340 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
341 current->pid, current->comm,
342 idle->pid, idle->comm); /* must be idle task! */
344 rcu_prepare_for_idle(smp_processor_id());
345 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
346 smp_mb__before_atomic_inc(); /* See above. */
347 atomic_inc(&rdtp->dynticks);
348 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
349 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
352 * The idle task is not permitted to enter the idle loop while
353 * in an RCU read-side critical section.
355 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
356 "Illegal idle entry in RCU read-side critical section.");
357 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
358 "Illegal idle entry in RCU-bh read-side critical section.");
359 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
360 "Illegal idle entry in RCU-sched read-side critical section.");
364 * rcu_idle_enter - inform RCU that current CPU is entering idle
366 * Enter idle mode, in other words, -leave- the mode in which RCU
367 * read-side critical sections can occur. (Though RCU read-side
368 * critical sections can occur in irq handlers in idle, a possibility
369 * handled by irq_enter() and irq_exit().)
371 * We crowbar the ->dynticks_nesting field to zero to allow for
372 * the possibility of usermode upcalls having messed up our count
373 * of interrupt nesting level during the prior busy period.
375 void rcu_idle_enter(void)
379 struct rcu_dynticks *rdtp;
381 local_irq_save(flags);
382 rdtp = &__get_cpu_var(rcu_dynticks);
383 oldval = rdtp->dynticks_nesting;
384 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
385 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
386 rdtp->dynticks_nesting = 0;
388 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
389 rcu_idle_enter_common(rdtp, oldval);
390 local_irq_restore(flags);
392 EXPORT_SYMBOL_GPL(rcu_idle_enter);
395 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
397 * Exit from an interrupt handler, which might possibly result in entering
398 * idle mode, in other words, leaving the mode in which read-side critical
399 * sections can occur.
401 * This code assumes that the idle loop never does anything that might
402 * result in unbalanced calls to irq_enter() and irq_exit(). If your
403 * architecture violates this assumption, RCU will give you what you
404 * deserve, good and hard. But very infrequently and irreproducibly.
406 * Use things like work queues to work around this limitation.
408 * You have been warned.
410 void rcu_irq_exit(void)
414 struct rcu_dynticks *rdtp;
416 local_irq_save(flags);
417 rdtp = &__get_cpu_var(rcu_dynticks);
418 oldval = rdtp->dynticks_nesting;
419 rdtp->dynticks_nesting--;
420 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
421 if (rdtp->dynticks_nesting)
422 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
424 rcu_idle_enter_common(rdtp, oldval);
425 local_irq_restore(flags);
429 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
431 * If the new value of the ->dynticks_nesting counter was previously zero,
432 * we really have exited idle, and must do the appropriate accounting.
433 * The caller must have disabled interrupts.
435 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
437 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
438 atomic_inc(&rdtp->dynticks);
439 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
440 smp_mb__after_atomic_inc(); /* See above. */
441 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
442 rcu_cleanup_after_idle(smp_processor_id());
443 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
444 if (!is_idle_task(current)) {
445 struct task_struct *idle = idle_task(smp_processor_id());
447 trace_rcu_dyntick("Error on exit: not idle task",
448 oldval, rdtp->dynticks_nesting);
449 ftrace_dump(DUMP_ORIG);
450 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
451 current->pid, current->comm,
452 idle->pid, idle->comm); /* must be idle task! */
457 * rcu_idle_exit - inform RCU that current CPU is leaving idle
459 * Exit idle mode, in other words, -enter- the mode in which RCU
460 * read-side critical sections can occur.
462 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
463 * allow for the possibility of usermode upcalls messing up our count
464 * of interrupt nesting level during the busy period that is just
467 void rcu_idle_exit(void)
470 struct rcu_dynticks *rdtp;
473 local_irq_save(flags);
474 rdtp = &__get_cpu_var(rcu_dynticks);
475 oldval = rdtp->dynticks_nesting;
476 WARN_ON_ONCE(oldval < 0);
477 if (oldval & DYNTICK_TASK_NEST_MASK)
478 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
480 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
481 rcu_idle_exit_common(rdtp, oldval);
482 local_irq_restore(flags);
484 EXPORT_SYMBOL_GPL(rcu_idle_exit);
487 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
489 * Enter an interrupt handler, which might possibly result in exiting
490 * idle mode, in other words, entering the mode in which read-side critical
491 * sections can occur.
493 * Note that the Linux kernel is fully capable of entering an interrupt
494 * handler that it never exits, for example when doing upcalls to
495 * user mode! This code assumes that the idle loop never does upcalls to
496 * user mode. If your architecture does do upcalls from the idle loop (or
497 * does anything else that results in unbalanced calls to the irq_enter()
498 * and irq_exit() functions), RCU will give you what you deserve, good
499 * and hard. But very infrequently and irreproducibly.
501 * Use things like work queues to work around this limitation.
503 * You have been warned.
505 void rcu_irq_enter(void)
508 struct rcu_dynticks *rdtp;
511 local_irq_save(flags);
512 rdtp = &__get_cpu_var(rcu_dynticks);
513 oldval = rdtp->dynticks_nesting;
514 rdtp->dynticks_nesting++;
515 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
517 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
519 rcu_idle_exit_common(rdtp, oldval);
520 local_irq_restore(flags);
524 * rcu_nmi_enter - inform RCU of entry to NMI context
526 * If the CPU was idle with dynamic ticks active, and there is no
527 * irq handler running, this updates rdtp->dynticks_nmi to let the
528 * RCU grace-period handling know that the CPU is active.
530 void rcu_nmi_enter(void)
532 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
534 if (rdtp->dynticks_nmi_nesting == 0 &&
535 (atomic_read(&rdtp->dynticks) & 0x1))
537 rdtp->dynticks_nmi_nesting++;
538 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
539 atomic_inc(&rdtp->dynticks);
540 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
541 smp_mb__after_atomic_inc(); /* See above. */
542 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
546 * rcu_nmi_exit - inform RCU of exit from NMI context
548 * If the CPU was idle with dynamic ticks active, and there is no
549 * irq handler running, this updates rdtp->dynticks_nmi to let the
550 * RCU grace-period handling know that the CPU is no longer active.
552 void rcu_nmi_exit(void)
554 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
556 if (rdtp->dynticks_nmi_nesting == 0 ||
557 --rdtp->dynticks_nmi_nesting != 0)
559 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
560 smp_mb__before_atomic_inc(); /* See above. */
561 atomic_inc(&rdtp->dynticks);
562 smp_mb__after_atomic_inc(); /* Force delay to next write. */
563 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
567 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
569 * If the current CPU is in its idle loop and is neither in an interrupt
570 * or NMI handler, return true.
572 int rcu_is_cpu_idle(void)
577 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
581 EXPORT_SYMBOL(rcu_is_cpu_idle);
583 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
586 * Is the current CPU online? Disable preemption to avoid false positives
587 * that could otherwise happen due to the current CPU number being sampled,
588 * this task being preempted, its old CPU being taken offline, resuming
589 * on some other CPU, then determining that its old CPU is now offline.
590 * It is OK to use RCU on an offline processor during initial boot, hence
591 * the check for rcu_scheduler_fully_active. Note also that it is OK
592 * for a CPU coming online to use RCU for one jiffy prior to marking itself
593 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
594 * offline to continue to use RCU for one jiffy after marking itself
595 * offline in the cpu_online_mask. This leniency is necessary given the
596 * non-atomic nature of the online and offline processing, for example,
597 * the fact that a CPU enters the scheduler after completing the CPU_DYING
600 * This is also why RCU internally marks CPUs online during the
601 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
603 * Disable checking if in an NMI handler because we cannot safely report
604 * errors from NMI handlers anyway.
606 bool rcu_lockdep_current_cpu_online(void)
608 struct rcu_data *rdp;
609 struct rcu_node *rnp;
615 rdp = &__get_cpu_var(rcu_sched_data);
617 ret = (rdp->grpmask & rnp->qsmaskinit) ||
618 !rcu_scheduler_fully_active;
622 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
624 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
627 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
629 * If the current CPU is idle or running at a first-level (not nested)
630 * interrupt from idle, return true. The caller must have at least
631 * disabled preemption.
633 int rcu_is_cpu_rrupt_from_idle(void)
635 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
639 * Snapshot the specified CPU's dynticks counter so that we can later
640 * credit them with an implicit quiescent state. Return 1 if this CPU
641 * is in dynticks idle mode, which is an extended quiescent state.
643 static int dyntick_save_progress_counter(struct rcu_data *rdp)
645 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
646 return (rdp->dynticks_snap & 0x1) == 0;
650 * Return true if the specified CPU has passed through a quiescent
651 * state by virtue of being in or having passed through an dynticks
652 * idle state since the last call to dyntick_save_progress_counter()
653 * for this same CPU, or by virtue of having been offline.
655 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
660 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
661 snap = (unsigned int)rdp->dynticks_snap;
664 * If the CPU passed through or entered a dynticks idle phase with
665 * no active irq/NMI handlers, then we can safely pretend that the CPU
666 * already acknowledged the request to pass through a quiescent
667 * state. Either way, that CPU cannot possibly be in an RCU
668 * read-side critical section that started before the beginning
669 * of the current RCU grace period.
671 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
672 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
678 * Check for the CPU being offline, but only if the grace period
679 * is old enough. We don't need to worry about the CPU changing
680 * state: If we see it offline even once, it has been through a
683 * The reason for insisting that the grace period be at least
684 * one jiffy old is that CPUs that are not quite online and that
685 * have just gone offline can still execute RCU read-side critical
688 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
689 return 0; /* Grace period is not old enough. */
691 if (cpu_is_offline(rdp->cpu)) {
692 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
699 static int jiffies_till_stall_check(void)
701 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
704 * Limit check must be consistent with the Kconfig limits
705 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
707 if (till_stall_check < 3) {
708 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
709 till_stall_check = 3;
710 } else if (till_stall_check > 300) {
711 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
712 till_stall_check = 300;
714 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
717 static void record_gp_stall_check_time(struct rcu_state *rsp)
719 rsp->gp_start = jiffies;
720 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
723 static void print_other_cpu_stall(struct rcu_state *rsp)
729 struct rcu_node *rnp = rcu_get_root(rsp);
731 /* Only let one CPU complain about others per time interval. */
733 raw_spin_lock_irqsave(&rnp->lock, flags);
734 delta = jiffies - rsp->jiffies_stall;
735 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
736 raw_spin_unlock_irqrestore(&rnp->lock, flags);
739 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
740 raw_spin_unlock_irqrestore(&rnp->lock, flags);
743 * OK, time to rat on our buddy...
744 * See Documentation/RCU/stallwarn.txt for info on how to debug
745 * RCU CPU stall warnings.
747 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
749 print_cpu_stall_info_begin();
750 rcu_for_each_leaf_node(rsp, rnp) {
751 raw_spin_lock_irqsave(&rnp->lock, flags);
752 ndetected += rcu_print_task_stall(rnp);
753 if (rnp->qsmask != 0) {
754 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
755 if (rnp->qsmask & (1UL << cpu)) {
756 print_cpu_stall_info(rsp,
761 raw_spin_unlock_irqrestore(&rnp->lock, flags);
765 * Now rat on any tasks that got kicked up to the root rcu_node
766 * due to CPU offlining.
768 rnp = rcu_get_root(rsp);
769 raw_spin_lock_irqsave(&rnp->lock, flags);
770 ndetected += rcu_print_task_stall(rnp);
771 raw_spin_unlock_irqrestore(&rnp->lock, flags);
773 print_cpu_stall_info_end();
774 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
775 smp_processor_id(), (long)(jiffies - rsp->gp_start));
777 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
778 else if (!trigger_all_cpu_backtrace())
781 /* Complain about tasks blocking the grace period. */
783 rcu_print_detail_task_stall(rsp);
785 force_quiescent_state(rsp); /* Kick them all. */
788 static void print_cpu_stall(struct rcu_state *rsp)
791 struct rcu_node *rnp = rcu_get_root(rsp);
794 * OK, time to rat on ourselves...
795 * See Documentation/RCU/stallwarn.txt for info on how to debug
796 * RCU CPU stall warnings.
798 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
799 print_cpu_stall_info_begin();
800 print_cpu_stall_info(rsp, smp_processor_id());
801 print_cpu_stall_info_end();
802 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
803 if (!trigger_all_cpu_backtrace())
806 raw_spin_lock_irqsave(&rnp->lock, flags);
807 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
808 rsp->jiffies_stall = jiffies +
809 3 * jiffies_till_stall_check() + 3;
810 raw_spin_unlock_irqrestore(&rnp->lock, flags);
812 set_need_resched(); /* kick ourselves to get things going. */
815 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
819 struct rcu_node *rnp;
821 if (rcu_cpu_stall_suppress)
823 j = ACCESS_ONCE(jiffies);
824 js = ACCESS_ONCE(rsp->jiffies_stall);
826 if (rcu_gp_in_progress(rsp) &&
827 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
829 /* We haven't checked in, so go dump stack. */
830 print_cpu_stall(rsp);
832 } else if (rcu_gp_in_progress(rsp) &&
833 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
835 /* They had a few time units to dump stack, so complain. */
836 print_other_cpu_stall(rsp);
840 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
842 rcu_cpu_stall_suppress = 1;
847 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
849 * Set the stall-warning timeout way off into the future, thus preventing
850 * any RCU CPU stall-warning messages from appearing in the current set of
853 * The caller must disable hard irqs.
855 void rcu_cpu_stall_reset(void)
857 struct rcu_state *rsp;
859 for_each_rcu_flavor(rsp)
860 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
863 static struct notifier_block rcu_panic_block = {
864 .notifier_call = rcu_panic,
867 static void __init check_cpu_stall_init(void)
869 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
873 * Update CPU-local rcu_data state to record the newly noticed grace period.
874 * This is used both when we started the grace period and when we notice
875 * that someone else started the grace period. The caller must hold the
876 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
877 * and must have irqs disabled.
879 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
881 if (rdp->gpnum != rnp->gpnum) {
883 * If the current grace period is waiting for this CPU,
884 * set up to detect a quiescent state, otherwise don't
885 * go looking for one.
887 rdp->gpnum = rnp->gpnum;
888 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
889 rdp->passed_quiesce = 0;
890 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
891 zero_cpu_stall_ticks(rdp);
895 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
898 struct rcu_node *rnp;
900 local_irq_save(flags);
902 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
903 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
904 local_irq_restore(flags);
907 __note_new_gpnum(rsp, rnp, rdp);
908 raw_spin_unlock_irqrestore(&rnp->lock, flags);
912 * Did someone else start a new RCU grace period start since we last
913 * checked? Update local state appropriately if so. Must be called
914 * on the CPU corresponding to rdp.
917 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
922 local_irq_save(flags);
923 if (rdp->gpnum != rsp->gpnum) {
924 note_new_gpnum(rsp, rdp);
927 local_irq_restore(flags);
932 * Initialize the specified rcu_data structure's callback list to empty.
934 static void init_callback_list(struct rcu_data *rdp)
939 for (i = 0; i < RCU_NEXT_SIZE; i++)
940 rdp->nxttail[i] = &rdp->nxtlist;
944 * Advance this CPU's callbacks, but only if the current grace period
945 * has ended. This may be called only from the CPU to whom the rdp
946 * belongs. In addition, the corresponding leaf rcu_node structure's
947 * ->lock must be held by the caller, with irqs disabled.
950 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
952 /* Did another grace period end? */
953 if (rdp->completed != rnp->completed) {
955 /* Advance callbacks. No harm if list empty. */
956 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
957 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
958 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
960 /* Remember that we saw this grace-period completion. */
961 rdp->completed = rnp->completed;
962 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
965 * If we were in an extended quiescent state, we may have
966 * missed some grace periods that others CPUs handled on
967 * our behalf. Catch up with this state to avoid noting
968 * spurious new grace periods. If another grace period
969 * has started, then rnp->gpnum will have advanced, so
970 * we will detect this later on. Of course, any quiescent
971 * states we found for the old GP are now invalid.
973 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
974 rdp->gpnum = rdp->completed;
975 rdp->passed_quiesce = 0;
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);
1020 /* Set state so that this CPU will detect the next quiescent state. */
1021 __note_new_gpnum(rsp, rnp, rdp);
1025 * Initialize a new grace period.
1027 static int rcu_gp_init(struct rcu_state *rsp)
1029 struct rcu_data *rdp;
1030 struct rcu_node *rnp = rcu_get_root(rsp);
1032 raw_spin_lock_irq(&rnp->lock);
1033 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1035 if (rcu_gp_in_progress(rsp)) {
1036 /* Grace period already in progress, don't start another. */
1037 raw_spin_unlock_irq(&rnp->lock);
1041 /* Advance to a new grace period and initialize state. */
1043 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1044 record_gp_stall_check_time(rsp);
1045 raw_spin_unlock_irq(&rnp->lock);
1047 /* Exclude any concurrent CPU-hotplug operations. */
1051 * Set the quiescent-state-needed bits in all the rcu_node
1052 * structures for all currently online CPUs in breadth-first order,
1053 * starting from the root rcu_node structure, relying on the layout
1054 * of the tree within the rsp->node[] array. Note that other CPUs
1055 * will access only the leaves of the hierarchy, thus seeing that no
1056 * grace period is in progress, at least until the corresponding
1057 * leaf node has been initialized. In addition, we have excluded
1058 * CPU-hotplug operations.
1060 * The grace period cannot complete until the initialization
1061 * process finishes, because this kthread handles both.
1063 rcu_for_each_node_breadth_first(rsp, rnp) {
1064 raw_spin_lock_irq(&rnp->lock);
1065 rdp = this_cpu_ptr(rsp->rda);
1066 rcu_preempt_check_blocked_tasks(rnp);
1067 rnp->qsmask = rnp->qsmaskinit;
1068 rnp->gpnum = rsp->gpnum;
1069 WARN_ON_ONCE(rnp->completed != rsp->completed);
1070 rnp->completed = rsp->completed;
1071 if (rnp == rdp->mynode)
1072 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1073 rcu_preempt_boost_start_gp(rnp);
1074 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1075 rnp->level, rnp->grplo,
1076 rnp->grphi, rnp->qsmask);
1077 raw_spin_unlock_irq(&rnp->lock);
1078 #ifdef CONFIG_PROVE_RCU_DELAY
1079 if ((random32() % (rcu_num_nodes * 8)) == 0)
1080 schedule_timeout_uninterruptible(2);
1081 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1090 * Do one round of quiescent-state forcing.
1092 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1094 int fqs_state = fqs_state_in;
1095 struct rcu_node *rnp = rcu_get_root(rsp);
1098 if (fqs_state == RCU_SAVE_DYNTICK) {
1099 /* Collect dyntick-idle snapshots. */
1100 force_qs_rnp(rsp, dyntick_save_progress_counter);
1101 fqs_state = RCU_FORCE_QS;
1103 /* Handle dyntick-idle and offline CPUs. */
1104 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1106 /* Clear flag to prevent immediate re-entry. */
1107 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1108 raw_spin_lock_irq(&rnp->lock);
1109 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1110 raw_spin_unlock_irq(&rnp->lock);
1116 * Clean up after the old grace period.
1118 static void rcu_gp_cleanup(struct rcu_state *rsp)
1120 unsigned long gp_duration;
1121 struct rcu_data *rdp;
1122 struct rcu_node *rnp = rcu_get_root(rsp);
1124 raw_spin_lock_irq(&rnp->lock);
1125 gp_duration = jiffies - rsp->gp_start;
1126 if (gp_duration > rsp->gp_max)
1127 rsp->gp_max = gp_duration;
1130 * We know the grace period is complete, but to everyone else
1131 * it appears to still be ongoing. But it is also the case
1132 * that to everyone else it looks like there is nothing that
1133 * they can do to advance the grace period. It is therefore
1134 * safe for us to drop the lock in order to mark the grace
1135 * period as completed in all of the rcu_node structures.
1137 raw_spin_unlock_irq(&rnp->lock);
1140 * Propagate new ->completed value to rcu_node structures so
1141 * that other CPUs don't have to wait until the start of the next
1142 * grace period to process their callbacks. This also avoids
1143 * some nasty RCU grace-period initialization races by forcing
1144 * the end of the current grace period to be completely recorded in
1145 * all of the rcu_node structures before the beginning of the next
1146 * grace period is recorded in any of the rcu_node structures.
1148 rcu_for_each_node_breadth_first(rsp, rnp) {
1149 raw_spin_lock_irq(&rnp->lock);
1150 rnp->completed = rsp->gpnum;
1151 raw_spin_unlock_irq(&rnp->lock);
1154 rnp = rcu_get_root(rsp);
1155 raw_spin_lock_irq(&rnp->lock);
1157 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1158 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1159 rsp->fqs_state = RCU_GP_IDLE;
1160 rdp = this_cpu_ptr(rsp->rda);
1161 if (cpu_needs_another_gp(rsp, rdp))
1163 raw_spin_unlock_irq(&rnp->lock);
1167 * Body of kthread that handles grace periods.
1169 static int __noreturn rcu_gp_kthread(void *arg)
1174 struct rcu_state *rsp = arg;
1175 struct rcu_node *rnp = rcu_get_root(rsp);
1179 /* Handle grace-period start. */
1181 wait_event_interruptible(rsp->gp_wq,
1184 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1188 flush_signals(current);
1191 /* Handle quiescent-state forcing. */
1192 fqs_state = RCU_SAVE_DYNTICK;
1193 j = jiffies_till_first_fqs;
1196 jiffies_till_first_fqs = HZ;
1199 rsp->jiffies_force_qs = jiffies + j;
1200 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1201 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1202 (!ACCESS_ONCE(rnp->qsmask) &&
1203 !rcu_preempt_blocked_readers_cgp(rnp)),
1205 /* If grace period done, leave loop. */
1206 if (!ACCESS_ONCE(rnp->qsmask) &&
1207 !rcu_preempt_blocked_readers_cgp(rnp))
1209 /* If time for quiescent-state forcing, do it. */
1210 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1211 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1214 /* Deal with stray signal. */
1216 flush_signals(current);
1218 j = jiffies_till_next_fqs;
1221 jiffies_till_next_fqs = HZ;
1224 jiffies_till_next_fqs = 1;
1228 /* Handle grace-period end. */
1229 rcu_gp_cleanup(rsp);
1234 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1235 * in preparation for detecting the next grace period. The caller must hold
1236 * the root node's ->lock, which is released before return. Hard irqs must
1239 * Note that it is legal for a dying CPU (which is marked as offline) to
1240 * invoke this function. This can happen when the dying CPU reports its
1244 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1245 __releases(rcu_get_root(rsp)->lock)
1247 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1248 struct rcu_node *rnp = rcu_get_root(rsp);
1250 if (!rsp->gp_kthread ||
1251 !cpu_needs_another_gp(rsp, rdp)) {
1253 * Either we have not yet spawned the grace-period
1254 * task or this CPU does not need another grace period.
1255 * Either way, don't start a new grace period.
1257 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1261 rsp->gp_flags = RCU_GP_FLAG_INIT;
1262 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1263 wake_up(&rsp->gp_wq);
1267 * Report a full set of quiescent states to the specified rcu_state
1268 * data structure. This involves cleaning up after the prior grace
1269 * period and letting rcu_start_gp() start up the next grace period
1270 * if one is needed. Note that the caller must hold rnp->lock, as
1271 * required by rcu_start_gp(), which will release it.
1273 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1274 __releases(rcu_get_root(rsp)->lock)
1276 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1277 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1278 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1282 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1283 * Allows quiescent states for a group of CPUs to be reported at one go
1284 * to the specified rcu_node structure, though all the CPUs in the group
1285 * must be represented by the same rcu_node structure (which need not be
1286 * a leaf rcu_node structure, though it often will be). That structure's
1287 * lock must be held upon entry, and it is released before return.
1290 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1291 struct rcu_node *rnp, unsigned long flags)
1292 __releases(rnp->lock)
1294 struct rcu_node *rnp_c;
1296 /* Walk up the rcu_node hierarchy. */
1298 if (!(rnp->qsmask & mask)) {
1300 /* Our bit has already been cleared, so done. */
1301 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1304 rnp->qsmask &= ~mask;
1305 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1306 mask, rnp->qsmask, rnp->level,
1307 rnp->grplo, rnp->grphi,
1309 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1311 /* Other bits still set at this level, so done. */
1312 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1315 mask = rnp->grpmask;
1316 if (rnp->parent == NULL) {
1318 /* No more levels. Exit loop holding root lock. */
1322 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1325 raw_spin_lock_irqsave(&rnp->lock, flags);
1326 WARN_ON_ONCE(rnp_c->qsmask);
1330 * Get here if we are the last CPU to pass through a quiescent
1331 * state for this grace period. Invoke rcu_report_qs_rsp()
1332 * to clean up and start the next grace period if one is needed.
1334 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1338 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1339 * structure. This must be either called from the specified CPU, or
1340 * called when the specified CPU is known to be offline (and when it is
1341 * also known that no other CPU is concurrently trying to help the offline
1342 * CPU). The lastcomp argument is used to make sure we are still in the
1343 * grace period of interest. We don't want to end the current grace period
1344 * based on quiescent states detected in an earlier grace period!
1347 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1349 unsigned long flags;
1351 struct rcu_node *rnp;
1354 raw_spin_lock_irqsave(&rnp->lock, flags);
1355 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1356 rnp->completed == rnp->gpnum) {
1359 * The grace period in which this quiescent state was
1360 * recorded has ended, so don't report it upwards.
1361 * We will instead need a new quiescent state that lies
1362 * within the current grace period.
1364 rdp->passed_quiesce = 0; /* need qs for new gp. */
1365 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1368 mask = rdp->grpmask;
1369 if ((rnp->qsmask & mask) == 0) {
1370 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1372 rdp->qs_pending = 0;
1375 * This GP can't end until cpu checks in, so all of our
1376 * callbacks can be processed during the next GP.
1378 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1380 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1385 * Check to see if there is a new grace period of which this CPU
1386 * is not yet aware, and if so, set up local rcu_data state for it.
1387 * Otherwise, see if this CPU has just passed through its first
1388 * quiescent state for this grace period, and record that fact if so.
1391 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1393 /* If there is now a new grace period, record and return. */
1394 if (check_for_new_grace_period(rsp, rdp))
1398 * Does this CPU still need to do its part for current grace period?
1399 * If no, return and let the other CPUs do their part as well.
1401 if (!rdp->qs_pending)
1405 * Was there a quiescent state since the beginning of the grace
1406 * period? If no, then exit and wait for the next call.
1408 if (!rdp->passed_quiesce)
1412 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1415 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1418 #ifdef CONFIG_HOTPLUG_CPU
1421 * Send the specified CPU's RCU callbacks to the orphanage. The
1422 * specified CPU must be offline, and the caller must hold the
1426 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1427 struct rcu_node *rnp, struct rcu_data *rdp)
1430 * Orphan the callbacks. First adjust the counts. This is safe
1431 * because ->onofflock excludes _rcu_barrier()'s adoption of
1432 * the callbacks, thus no memory barrier is required.
1434 if (rdp->nxtlist != NULL) {
1435 rsp->qlen_lazy += rdp->qlen_lazy;
1436 rsp->qlen += rdp->qlen;
1437 rdp->n_cbs_orphaned += rdp->qlen;
1439 ACCESS_ONCE(rdp->qlen) = 0;
1443 * Next, move those callbacks still needing a grace period to
1444 * the orphanage, where some other CPU will pick them up.
1445 * Some of the callbacks might have gone partway through a grace
1446 * period, but that is too bad. They get to start over because we
1447 * cannot assume that grace periods are synchronized across CPUs.
1448 * We don't bother updating the ->nxttail[] array yet, instead
1449 * we just reset the whole thing later on.
1451 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1452 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1453 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1454 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1458 * Then move the ready-to-invoke callbacks to the orphanage,
1459 * where some other CPU will pick them up. These will not be
1460 * required to pass though another grace period: They are done.
1462 if (rdp->nxtlist != NULL) {
1463 *rsp->orphan_donetail = rdp->nxtlist;
1464 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1467 /* Finally, initialize the rcu_data structure's list to empty. */
1468 init_callback_list(rdp);
1472 * Adopt the RCU callbacks from the specified rcu_state structure's
1473 * orphanage. The caller must hold the ->onofflock.
1475 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1478 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1480 /* Do the accounting first. */
1481 rdp->qlen_lazy += rsp->qlen_lazy;
1482 rdp->qlen += rsp->qlen;
1483 rdp->n_cbs_adopted += rsp->qlen;
1484 if (rsp->qlen_lazy != rsp->qlen)
1485 rcu_idle_count_callbacks_posted();
1490 * We do not need a memory barrier here because the only way we
1491 * can get here if there is an rcu_barrier() in flight is if
1492 * we are the task doing the rcu_barrier().
1495 /* First adopt the ready-to-invoke callbacks. */
1496 if (rsp->orphan_donelist != NULL) {
1497 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1498 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1499 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1500 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1501 rdp->nxttail[i] = rsp->orphan_donetail;
1502 rsp->orphan_donelist = NULL;
1503 rsp->orphan_donetail = &rsp->orphan_donelist;
1506 /* And then adopt the callbacks that still need a grace period. */
1507 if (rsp->orphan_nxtlist != NULL) {
1508 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1509 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1510 rsp->orphan_nxtlist = NULL;
1511 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1516 * Trace the fact that this CPU is going offline.
1518 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1520 RCU_TRACE(unsigned long mask);
1521 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1522 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1524 RCU_TRACE(mask = rdp->grpmask);
1525 trace_rcu_grace_period(rsp->name,
1526 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1531 * The CPU has been completely removed, and some other CPU is reporting
1532 * this fact from process context. Do the remainder of the cleanup,
1533 * including orphaning the outgoing CPU's RCU callbacks, and also
1534 * adopting them. There can only be one CPU hotplug operation at a time,
1535 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1537 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1539 unsigned long flags;
1541 int need_report = 0;
1542 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1543 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1545 /* Adjust any no-longer-needed kthreads. */
1546 rcu_stop_cpu_kthread(cpu);
1547 rcu_node_kthread_setaffinity(rnp, -1);
1549 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1551 /* Exclude any attempts to start a new grace period. */
1552 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1554 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1555 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1556 rcu_adopt_orphan_cbs(rsp);
1558 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1559 mask = rdp->grpmask; /* rnp->grplo is constant. */
1561 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1562 rnp->qsmaskinit &= ~mask;
1563 if (rnp->qsmaskinit != 0) {
1564 if (rnp != rdp->mynode)
1565 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1568 if (rnp == rdp->mynode)
1569 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1571 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1572 mask = rnp->grpmask;
1574 } while (rnp != NULL);
1577 * We still hold the leaf rcu_node structure lock here, and
1578 * irqs are still disabled. The reason for this subterfuge is
1579 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1580 * held leads to deadlock.
1582 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1584 if (need_report & RCU_OFL_TASKS_NORM_GP)
1585 rcu_report_unblock_qs_rnp(rnp, flags);
1587 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1588 if (need_report & RCU_OFL_TASKS_EXP_GP)
1589 rcu_report_exp_rnp(rsp, rnp, true);
1590 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1591 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1592 cpu, rdp->qlen, rdp->nxtlist);
1593 init_callback_list(rdp);
1594 /* Disallow further callbacks on this CPU. */
1595 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1598 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1600 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1604 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1608 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1611 * Invoke any RCU callbacks that have made it to the end of their grace
1612 * period. Thottle as specified by rdp->blimit.
1614 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1616 unsigned long flags;
1617 struct rcu_head *next, *list, **tail;
1618 int bl, count, count_lazy, i;
1620 /* If no callbacks are ready, just return.*/
1621 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1622 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1623 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1624 need_resched(), is_idle_task(current),
1625 rcu_is_callbacks_kthread());
1630 * Extract the list of ready callbacks, disabling to prevent
1631 * races with call_rcu() from interrupt handlers.
1633 local_irq_save(flags);
1634 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1636 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1637 list = rdp->nxtlist;
1638 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1639 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1640 tail = rdp->nxttail[RCU_DONE_TAIL];
1641 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1642 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1643 rdp->nxttail[i] = &rdp->nxtlist;
1644 local_irq_restore(flags);
1646 /* Invoke callbacks. */
1647 count = count_lazy = 0;
1651 debug_rcu_head_unqueue(list);
1652 if (__rcu_reclaim(rsp->name, list))
1655 /* Stop only if limit reached and CPU has something to do. */
1656 if (++count >= bl &&
1658 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1662 local_irq_save(flags);
1663 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1664 is_idle_task(current),
1665 rcu_is_callbacks_kthread());
1667 /* Update count, and requeue any remaining callbacks. */
1669 *tail = rdp->nxtlist;
1670 rdp->nxtlist = list;
1671 for (i = 0; i < RCU_NEXT_SIZE; i++)
1672 if (&rdp->nxtlist == rdp->nxttail[i])
1673 rdp->nxttail[i] = tail;
1677 smp_mb(); /* List handling before counting for rcu_barrier(). */
1678 rdp->qlen_lazy -= count_lazy;
1679 ACCESS_ONCE(rdp->qlen) -= count;
1680 rdp->n_cbs_invoked += count;
1682 /* Reinstate batch limit if we have worked down the excess. */
1683 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1684 rdp->blimit = blimit;
1686 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1687 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1688 rdp->qlen_last_fqs_check = 0;
1689 rdp->n_force_qs_snap = rsp->n_force_qs;
1690 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1691 rdp->qlen_last_fqs_check = rdp->qlen;
1692 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1694 local_irq_restore(flags);
1696 /* Re-invoke RCU core processing if there are callbacks remaining. */
1697 if (cpu_has_callbacks_ready_to_invoke(rdp))
1702 * Check to see if this CPU is in a non-context-switch quiescent state
1703 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1704 * Also schedule RCU core processing.
1706 * This function must be called from hardirq context. It is normally
1707 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1708 * false, there is no point in invoking rcu_check_callbacks().
1710 void rcu_check_callbacks(int cpu, int user)
1712 trace_rcu_utilization("Start scheduler-tick");
1713 increment_cpu_stall_ticks();
1714 if (user || rcu_is_cpu_rrupt_from_idle()) {
1717 * Get here if this CPU took its interrupt from user
1718 * mode or from the idle loop, and if this is not a
1719 * nested interrupt. In this case, the CPU is in
1720 * a quiescent state, so note it.
1722 * No memory barrier is required here because both
1723 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1724 * variables that other CPUs neither access nor modify,
1725 * at least not while the corresponding CPU is online.
1731 } else if (!in_softirq()) {
1734 * Get here if this CPU did not take its interrupt from
1735 * softirq, in other words, if it is not interrupting
1736 * a rcu_bh read-side critical section. This is an _bh
1737 * critical section, so note it.
1742 rcu_preempt_check_callbacks(cpu);
1743 if (rcu_pending(cpu))
1745 trace_rcu_utilization("End scheduler-tick");
1749 * Scan the leaf rcu_node structures, processing dyntick state for any that
1750 * have not yet encountered a quiescent state, using the function specified.
1751 * Also initiate boosting for any threads blocked on the root rcu_node.
1753 * The caller must have suppressed start of new grace periods.
1755 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1759 unsigned long flags;
1761 struct rcu_node *rnp;
1763 rcu_for_each_leaf_node(rsp, rnp) {
1766 raw_spin_lock_irqsave(&rnp->lock, flags);
1767 if (!rcu_gp_in_progress(rsp)) {
1768 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1771 if (rnp->qsmask == 0) {
1772 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1777 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1778 if ((rnp->qsmask & bit) != 0 &&
1779 f(per_cpu_ptr(rsp->rda, cpu)))
1784 /* rcu_report_qs_rnp() releases rnp->lock. */
1785 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1788 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1790 rnp = rcu_get_root(rsp);
1791 if (rnp->qsmask == 0) {
1792 raw_spin_lock_irqsave(&rnp->lock, flags);
1793 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1798 * Force quiescent states on reluctant CPUs, and also detect which
1799 * CPUs are in dyntick-idle mode.
1801 static void force_quiescent_state(struct rcu_state *rsp)
1803 unsigned long flags;
1805 struct rcu_node *rnp;
1806 struct rcu_node *rnp_old = NULL;
1808 /* Funnel through hierarchy to reduce memory contention. */
1809 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1810 for (; rnp != NULL; rnp = rnp->parent) {
1811 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1812 !raw_spin_trylock(&rnp->fqslock);
1813 if (rnp_old != NULL)
1814 raw_spin_unlock(&rnp_old->fqslock);
1816 rsp->n_force_qs_lh++;
1821 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1823 /* Reached the root of the rcu_node tree, acquire lock. */
1824 raw_spin_lock_irqsave(&rnp_old->lock, flags);
1825 raw_spin_unlock(&rnp_old->fqslock);
1826 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1827 rsp->n_force_qs_lh++;
1828 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1829 return; /* Someone beat us to it. */
1831 rsp->gp_flags |= RCU_GP_FLAG_FQS;
1832 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1833 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1837 * This does the RCU core processing work for the specified rcu_state
1838 * and rcu_data structures. This may be called only from the CPU to
1839 * whom the rdp belongs.
1842 __rcu_process_callbacks(struct rcu_state *rsp)
1844 unsigned long flags;
1845 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1847 WARN_ON_ONCE(rdp->beenonline == 0);
1850 * Advance callbacks in response to end of earlier grace
1851 * period that some other CPU ended.
1853 rcu_process_gp_end(rsp, rdp);
1855 /* Update RCU state based on any recent quiescent states. */
1856 rcu_check_quiescent_state(rsp, rdp);
1858 /* Does this CPU require a not-yet-started grace period? */
1859 if (cpu_needs_another_gp(rsp, rdp)) {
1860 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1861 rcu_start_gp(rsp, flags); /* releases above lock */
1864 /* If there are callbacks ready, invoke them. */
1865 if (cpu_has_callbacks_ready_to_invoke(rdp))
1866 invoke_rcu_callbacks(rsp, rdp);
1870 * Do RCU core processing for the current CPU.
1872 static void rcu_process_callbacks(struct softirq_action *unused)
1874 struct rcu_state *rsp;
1876 if (cpu_is_offline(smp_processor_id()))
1878 trace_rcu_utilization("Start RCU core");
1879 for_each_rcu_flavor(rsp)
1880 __rcu_process_callbacks(rsp);
1881 trace_rcu_utilization("End RCU core");
1885 * Schedule RCU callback invocation. If the specified type of RCU
1886 * does not support RCU priority boosting, just do a direct call,
1887 * otherwise wake up the per-CPU kernel kthread. Note that because we
1888 * are running on the current CPU with interrupts disabled, the
1889 * rcu_cpu_kthread_task cannot disappear out from under us.
1891 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1893 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1895 if (likely(!rsp->boost)) {
1896 rcu_do_batch(rsp, rdp);
1899 invoke_rcu_callbacks_kthread();
1902 static void invoke_rcu_core(void)
1904 raise_softirq(RCU_SOFTIRQ);
1908 * Handle any core-RCU processing required by a call_rcu() invocation.
1910 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1911 struct rcu_head *head, unsigned long flags)
1914 * If called from an extended quiescent state, invoke the RCU
1915 * core in order to force a re-evaluation of RCU's idleness.
1917 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1920 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1921 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1925 * Force the grace period if too many callbacks or too long waiting.
1926 * Enforce hysteresis, and don't invoke force_quiescent_state()
1927 * if some other CPU has recently done so. Also, don't bother
1928 * invoking force_quiescent_state() if the newly enqueued callback
1929 * is the only one waiting for a grace period to complete.
1931 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1933 /* Are we ignoring a completed grace period? */
1934 rcu_process_gp_end(rsp, rdp);
1935 check_for_new_grace_period(rsp, rdp);
1937 /* Start a new grace period if one not already started. */
1938 if (!rcu_gp_in_progress(rsp)) {
1939 unsigned long nestflag;
1940 struct rcu_node *rnp_root = rcu_get_root(rsp);
1942 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1943 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1945 /* Give the grace period a kick. */
1946 rdp->blimit = LONG_MAX;
1947 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1948 *rdp->nxttail[RCU_DONE_TAIL] != head)
1949 force_quiescent_state(rsp);
1950 rdp->n_force_qs_snap = rsp->n_force_qs;
1951 rdp->qlen_last_fqs_check = rdp->qlen;
1957 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1958 struct rcu_state *rsp, bool lazy)
1960 unsigned long flags;
1961 struct rcu_data *rdp;
1963 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1964 debug_rcu_head_queue(head);
1969 * Opportunistically note grace-period endings and beginnings.
1970 * Note that we might see a beginning right after we see an
1971 * end, but never vice versa, since this CPU has to pass through
1972 * a quiescent state betweentimes.
1974 local_irq_save(flags);
1975 rdp = this_cpu_ptr(rsp->rda);
1977 /* Add the callback to our list. */
1978 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL)) {
1979 /* _call_rcu() is illegal on offline CPU; leak the callback. */
1981 local_irq_restore(flags);
1984 ACCESS_ONCE(rdp->qlen)++;
1988 rcu_idle_count_callbacks_posted();
1989 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1990 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1991 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1993 if (__is_kfree_rcu_offset((unsigned long)func))
1994 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1995 rdp->qlen_lazy, rdp->qlen);
1997 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1999 /* Go handle any RCU core processing required. */
2000 __call_rcu_core(rsp, rdp, head, flags);
2001 local_irq_restore(flags);
2005 * Queue an RCU-sched callback for invocation after a grace period.
2007 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2009 __call_rcu(head, func, &rcu_sched_state, 0);
2011 EXPORT_SYMBOL_GPL(call_rcu_sched);
2014 * Queue an RCU callback for invocation after a quicker grace period.
2016 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2018 __call_rcu(head, func, &rcu_bh_state, 0);
2020 EXPORT_SYMBOL_GPL(call_rcu_bh);
2023 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2024 * any blocking grace-period wait automatically implies a grace period
2025 * if there is only one CPU online at any point time during execution
2026 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2027 * occasionally incorrectly indicate that there are multiple CPUs online
2028 * when there was in fact only one the whole time, as this just adds
2029 * some overhead: RCU still operates correctly.
2031 static inline int rcu_blocking_is_gp(void)
2035 might_sleep(); /* Check for RCU read-side critical section. */
2037 ret = num_online_cpus() <= 1;
2043 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2045 * Control will return to the caller some time after a full rcu-sched
2046 * grace period has elapsed, in other words after all currently executing
2047 * rcu-sched read-side critical sections have completed. These read-side
2048 * critical sections are delimited by rcu_read_lock_sched() and
2049 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2050 * local_irq_disable(), and so on may be used in place of
2051 * rcu_read_lock_sched().
2053 * This means that all preempt_disable code sequences, including NMI and
2054 * hardware-interrupt handlers, in progress on entry will have completed
2055 * before this primitive returns. However, this does not guarantee that
2056 * softirq handlers will have completed, since in some kernels, these
2057 * handlers can run in process context, and can block.
2059 * This primitive provides the guarantees made by the (now removed)
2060 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2061 * guarantees that rcu_read_lock() sections will have completed.
2062 * In "classic RCU", these two guarantees happen to be one and
2063 * the same, but can differ in realtime RCU implementations.
2065 void synchronize_sched(void)
2067 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2068 !lock_is_held(&rcu_lock_map) &&
2069 !lock_is_held(&rcu_sched_lock_map),
2070 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2071 if (rcu_blocking_is_gp())
2073 wait_rcu_gp(call_rcu_sched);
2075 EXPORT_SYMBOL_GPL(synchronize_sched);
2078 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2080 * Control will return to the caller some time after a full rcu_bh grace
2081 * period has elapsed, in other words after all currently executing rcu_bh
2082 * read-side critical sections have completed. RCU read-side critical
2083 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2084 * and may be nested.
2086 void synchronize_rcu_bh(void)
2088 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2089 !lock_is_held(&rcu_lock_map) &&
2090 !lock_is_held(&rcu_sched_lock_map),
2091 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2092 if (rcu_blocking_is_gp())
2094 wait_rcu_gp(call_rcu_bh);
2096 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2098 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2099 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2101 static int synchronize_sched_expedited_cpu_stop(void *data)
2104 * There must be a full memory barrier on each affected CPU
2105 * between the time that try_stop_cpus() is called and the
2106 * time that it returns.
2108 * In the current initial implementation of cpu_stop, the
2109 * above condition is already met when the control reaches
2110 * this point and the following smp_mb() is not strictly
2111 * necessary. Do smp_mb() anyway for documentation and
2112 * robustness against future implementation changes.
2114 smp_mb(); /* See above comment block. */
2119 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2121 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2122 * approach to force the grace period to end quickly. This consumes
2123 * significant time on all CPUs and is unfriendly to real-time workloads,
2124 * so is thus not recommended for any sort of common-case code. In fact,
2125 * if you are using synchronize_sched_expedited() in a loop, please
2126 * restructure your code to batch your updates, and then use a single
2127 * synchronize_sched() instead.
2129 * Note that it is illegal to call this function while holding any lock
2130 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2131 * to call this function from a CPU-hotplug notifier. Failing to observe
2132 * these restriction will result in deadlock.
2134 * This implementation can be thought of as an application of ticket
2135 * locking to RCU, with sync_sched_expedited_started and
2136 * sync_sched_expedited_done taking on the roles of the halves
2137 * of the ticket-lock word. Each task atomically increments
2138 * sync_sched_expedited_started upon entry, snapshotting the old value,
2139 * then attempts to stop all the CPUs. If this succeeds, then each
2140 * CPU will have executed a context switch, resulting in an RCU-sched
2141 * grace period. We are then done, so we use atomic_cmpxchg() to
2142 * update sync_sched_expedited_done to match our snapshot -- but
2143 * only if someone else has not already advanced past our snapshot.
2145 * On the other hand, if try_stop_cpus() fails, we check the value
2146 * of sync_sched_expedited_done. If it has advanced past our
2147 * initial snapshot, then someone else must have forced a grace period
2148 * some time after we took our snapshot. In this case, our work is
2149 * done for us, and we can simply return. Otherwise, we try again,
2150 * but keep our initial snapshot for purposes of checking for someone
2151 * doing our work for us.
2153 * If we fail too many times in a row, we fall back to synchronize_sched().
2155 void synchronize_sched_expedited(void)
2157 int firstsnap, s, snap, trycount = 0;
2159 /* Note that atomic_inc_return() implies full memory barrier. */
2160 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2162 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2165 * Each pass through the following loop attempts to force a
2166 * context switch on each CPU.
2168 while (try_stop_cpus(cpu_online_mask,
2169 synchronize_sched_expedited_cpu_stop,
2173 /* No joy, try again later. Or just synchronize_sched(). */
2174 if (trycount++ < 10) {
2175 udelay(trycount * num_online_cpus());
2177 synchronize_sched();
2181 /* Check to see if someone else did our work for us. */
2182 s = atomic_read(&sync_sched_expedited_done);
2183 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2184 smp_mb(); /* ensure test happens before caller kfree */
2189 * Refetching sync_sched_expedited_started allows later
2190 * callers to piggyback on our grace period. We subtract
2191 * 1 to get the same token that the last incrementer got.
2192 * We retry after they started, so our grace period works
2193 * for them, and they started after our first try, so their
2194 * grace period works for us.
2197 snap = atomic_read(&sync_sched_expedited_started);
2198 smp_mb(); /* ensure read is before try_stop_cpus(). */
2202 * Everyone up to our most recent fetch is covered by our grace
2203 * period. Update the counter, but only if our work is still
2204 * relevant -- which it won't be if someone who started later
2205 * than we did beat us to the punch.
2208 s = atomic_read(&sync_sched_expedited_done);
2209 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2210 smp_mb(); /* ensure test happens before caller kfree */
2213 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2217 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2220 * Check to see if there is any immediate RCU-related work to be done
2221 * by the current CPU, for the specified type of RCU, returning 1 if so.
2222 * The checks are in order of increasing expense: checks that can be
2223 * carried out against CPU-local state are performed first. However,
2224 * we must check for CPU stalls first, else we might not get a chance.
2226 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2228 struct rcu_node *rnp = rdp->mynode;
2230 rdp->n_rcu_pending++;
2232 /* Check for CPU stalls, if enabled. */
2233 check_cpu_stall(rsp, rdp);
2235 /* Is the RCU core waiting for a quiescent state from this CPU? */
2236 if (rcu_scheduler_fully_active &&
2237 rdp->qs_pending && !rdp->passed_quiesce) {
2238 rdp->n_rp_qs_pending++;
2239 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2240 rdp->n_rp_report_qs++;
2244 /* Does this CPU have callbacks ready to invoke? */
2245 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2246 rdp->n_rp_cb_ready++;
2250 /* Has RCU gone idle with this CPU needing another grace period? */
2251 if (cpu_needs_another_gp(rsp, rdp)) {
2252 rdp->n_rp_cpu_needs_gp++;
2256 /* Has another RCU grace period completed? */
2257 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2258 rdp->n_rp_gp_completed++;
2262 /* Has a new RCU grace period started? */
2263 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2264 rdp->n_rp_gp_started++;
2269 rdp->n_rp_need_nothing++;
2274 * Check to see if there is any immediate RCU-related work to be done
2275 * by the current CPU, returning 1 if so. This function is part of the
2276 * RCU implementation; it is -not- an exported member of the RCU API.
2278 static int rcu_pending(int cpu)
2280 struct rcu_state *rsp;
2282 for_each_rcu_flavor(rsp)
2283 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2289 * Check to see if any future RCU-related work will need to be done
2290 * by the current CPU, even if none need be done immediately, returning
2293 static int rcu_cpu_has_callbacks(int cpu)
2295 struct rcu_state *rsp;
2297 /* RCU callbacks either ready or pending? */
2298 for_each_rcu_flavor(rsp)
2299 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2305 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2306 * the compiler is expected to optimize this away.
2308 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2309 int cpu, unsigned long done)
2311 trace_rcu_barrier(rsp->name, s, cpu,
2312 atomic_read(&rsp->barrier_cpu_count), done);
2316 * RCU callback function for _rcu_barrier(). If we are last, wake
2317 * up the task executing _rcu_barrier().
2319 static void rcu_barrier_callback(struct rcu_head *rhp)
2321 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2322 struct rcu_state *rsp = rdp->rsp;
2324 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2325 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2326 complete(&rsp->barrier_completion);
2328 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2333 * Called with preemption disabled, and from cross-cpu IRQ context.
2335 static void rcu_barrier_func(void *type)
2337 struct rcu_state *rsp = type;
2338 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2340 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2341 atomic_inc(&rsp->barrier_cpu_count);
2342 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2346 * Orchestrate the specified type of RCU barrier, waiting for all
2347 * RCU callbacks of the specified type to complete.
2349 static void _rcu_barrier(struct rcu_state *rsp)
2352 struct rcu_data *rdp;
2353 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2354 unsigned long snap_done;
2356 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2358 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2359 mutex_lock(&rsp->barrier_mutex);
2362 * Ensure that all prior references, including to ->n_barrier_done,
2363 * are ordered before the _rcu_barrier() machinery.
2365 smp_mb(); /* See above block comment. */
2368 * Recheck ->n_barrier_done to see if others did our work for us.
2369 * This means checking ->n_barrier_done for an even-to-odd-to-even
2370 * transition. The "if" expression below therefore rounds the old
2371 * value up to the next even number and adds two before comparing.
2373 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2374 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2375 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2376 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2377 smp_mb(); /* caller's subsequent code after above check. */
2378 mutex_unlock(&rsp->barrier_mutex);
2383 * Increment ->n_barrier_done to avoid duplicate work. Use
2384 * ACCESS_ONCE() to prevent the compiler from speculating
2385 * the increment to precede the early-exit check.
2387 ACCESS_ONCE(rsp->n_barrier_done)++;
2388 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2389 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2390 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2393 * Initialize the count to one rather than to zero in order to
2394 * avoid a too-soon return to zero in case of a short grace period
2395 * (or preemption of this task). Exclude CPU-hotplug operations
2396 * to ensure that no offline CPU has callbacks queued.
2398 init_completion(&rsp->barrier_completion);
2399 atomic_set(&rsp->barrier_cpu_count, 1);
2403 * Force each CPU with callbacks to register a new callback.
2404 * When that callback is invoked, we will know that all of the
2405 * corresponding CPU's preceding callbacks have been invoked.
2407 for_each_online_cpu(cpu) {
2408 rdp = per_cpu_ptr(rsp->rda, cpu);
2409 if (ACCESS_ONCE(rdp->qlen)) {
2410 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2411 rsp->n_barrier_done);
2412 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2414 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2415 rsp->n_barrier_done);
2421 * Now that we have an rcu_barrier_callback() callback on each
2422 * CPU, and thus each counted, remove the initial count.
2424 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2425 complete(&rsp->barrier_completion);
2427 /* Increment ->n_barrier_done to prevent duplicate work. */
2428 smp_mb(); /* Keep increment after above mechanism. */
2429 ACCESS_ONCE(rsp->n_barrier_done)++;
2430 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2431 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2432 smp_mb(); /* Keep increment before caller's subsequent code. */
2434 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2435 wait_for_completion(&rsp->barrier_completion);
2437 /* Other rcu_barrier() invocations can now safely proceed. */
2438 mutex_unlock(&rsp->barrier_mutex);
2442 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2444 void rcu_barrier_bh(void)
2446 _rcu_barrier(&rcu_bh_state);
2448 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2451 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2453 void rcu_barrier_sched(void)
2455 _rcu_barrier(&rcu_sched_state);
2457 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2460 * Do boot-time initialization of a CPU's per-CPU RCU data.
2463 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2465 unsigned long flags;
2466 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2467 struct rcu_node *rnp = rcu_get_root(rsp);
2469 /* Set up local state, ensuring consistent view of global state. */
2470 raw_spin_lock_irqsave(&rnp->lock, flags);
2471 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2472 init_callback_list(rdp);
2474 ACCESS_ONCE(rdp->qlen) = 0;
2475 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2476 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2477 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2480 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2484 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2485 * offline event can be happening at a given time. Note also that we
2486 * can accept some slop in the rsp->completed access due to the fact
2487 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2489 static void __cpuinit
2490 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2492 unsigned long flags;
2494 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2495 struct rcu_node *rnp = rcu_get_root(rsp);
2497 /* Set up local state, ensuring consistent view of global state. */
2498 raw_spin_lock_irqsave(&rnp->lock, flags);
2499 rdp->beenonline = 1; /* We have now been online. */
2500 rdp->preemptible = preemptible;
2501 rdp->qlen_last_fqs_check = 0;
2502 rdp->n_force_qs_snap = rsp->n_force_qs;
2503 rdp->blimit = blimit;
2504 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2505 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2506 atomic_set(&rdp->dynticks->dynticks,
2507 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2508 rcu_prepare_for_idle_init(cpu);
2509 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2512 * A new grace period might start here. If so, we won't be part
2513 * of it, but that is OK, as we are currently in a quiescent state.
2516 /* Exclude any attempts to start a new GP on large systems. */
2517 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2519 /* Add CPU to rcu_node bitmasks. */
2521 mask = rdp->grpmask;
2523 /* Exclude any attempts to start a new GP on small systems. */
2524 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2525 rnp->qsmaskinit |= mask;
2526 mask = rnp->grpmask;
2527 if (rnp == rdp->mynode) {
2529 * If there is a grace period in progress, we will
2530 * set up to wait for it next time we run the
2533 rdp->gpnum = rnp->completed;
2534 rdp->completed = rnp->completed;
2535 rdp->passed_quiesce = 0;
2536 rdp->qs_pending = 0;
2537 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2539 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2541 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2543 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2546 static void __cpuinit rcu_prepare_cpu(int cpu)
2548 struct rcu_state *rsp;
2550 for_each_rcu_flavor(rsp)
2551 rcu_init_percpu_data(cpu, rsp,
2552 strcmp(rsp->name, "rcu_preempt") == 0);
2556 * Handle CPU online/offline notification events.
2558 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2559 unsigned long action, void *hcpu)
2561 long cpu = (long)hcpu;
2562 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2563 struct rcu_node *rnp = rdp->mynode;
2564 struct rcu_state *rsp;
2566 trace_rcu_utilization("Start CPU hotplug");
2568 case CPU_UP_PREPARE:
2569 case CPU_UP_PREPARE_FROZEN:
2570 rcu_prepare_cpu(cpu);
2571 rcu_prepare_kthreads(cpu);
2574 case CPU_DOWN_FAILED:
2575 rcu_node_kthread_setaffinity(rnp, -1);
2576 rcu_cpu_kthread_setrt(cpu, 1);
2578 case CPU_DOWN_PREPARE:
2579 rcu_node_kthread_setaffinity(rnp, cpu);
2580 rcu_cpu_kthread_setrt(cpu, 0);
2583 case CPU_DYING_FROZEN:
2585 * The whole machine is "stopped" except this CPU, so we can
2586 * touch any data without introducing corruption. We send the
2587 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2589 for_each_rcu_flavor(rsp)
2590 rcu_cleanup_dying_cpu(rsp);
2591 rcu_cleanup_after_idle(cpu);
2594 case CPU_DEAD_FROZEN:
2595 case CPU_UP_CANCELED:
2596 case CPU_UP_CANCELED_FROZEN:
2597 for_each_rcu_flavor(rsp)
2598 rcu_cleanup_dead_cpu(cpu, rsp);
2603 trace_rcu_utilization("End CPU hotplug");
2608 * Spawn the kthread that handles this RCU flavor's grace periods.
2610 static int __init rcu_spawn_gp_kthread(void)
2612 unsigned long flags;
2613 struct rcu_node *rnp;
2614 struct rcu_state *rsp;
2615 struct task_struct *t;
2617 for_each_rcu_flavor(rsp) {
2618 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2620 rnp = rcu_get_root(rsp);
2621 raw_spin_lock_irqsave(&rnp->lock, flags);
2622 rsp->gp_kthread = t;
2623 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2627 early_initcall(rcu_spawn_gp_kthread);
2630 * This function is invoked towards the end of the scheduler's initialization
2631 * process. Before this is called, the idle task might contain
2632 * RCU read-side critical sections (during which time, this idle
2633 * task is booting the system). After this function is called, the
2634 * idle tasks are prohibited from containing RCU read-side critical
2635 * sections. This function also enables RCU lockdep checking.
2637 void rcu_scheduler_starting(void)
2639 WARN_ON(num_online_cpus() != 1);
2640 WARN_ON(nr_context_switches() > 0);
2641 rcu_scheduler_active = 1;
2645 * Compute the per-level fanout, either using the exact fanout specified
2646 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2648 #ifdef CONFIG_RCU_FANOUT_EXACT
2649 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2653 for (i = rcu_num_lvls - 1; i > 0; i--)
2654 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2655 rsp->levelspread[0] = rcu_fanout_leaf;
2657 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2658 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2665 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2666 ccur = rsp->levelcnt[i];
2667 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2671 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2674 * Helper function for rcu_init() that initializes one rcu_state structure.
2676 static void __init rcu_init_one(struct rcu_state *rsp,
2677 struct rcu_data __percpu *rda)
2679 static char *buf[] = { "rcu_node_0",
2682 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2683 static char *fqs[] = { "rcu_node_fqs_0",
2686 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2690 struct rcu_node *rnp;
2692 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2694 /* Initialize the level-tracking arrays. */
2696 for (i = 0; i < rcu_num_lvls; i++)
2697 rsp->levelcnt[i] = num_rcu_lvl[i];
2698 for (i = 1; i < rcu_num_lvls; i++)
2699 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2700 rcu_init_levelspread(rsp);
2702 /* Initialize the elements themselves, starting from the leaves. */
2704 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2705 cpustride *= rsp->levelspread[i];
2706 rnp = rsp->level[i];
2707 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2708 raw_spin_lock_init(&rnp->lock);
2709 lockdep_set_class_and_name(&rnp->lock,
2710 &rcu_node_class[i], buf[i]);
2711 raw_spin_lock_init(&rnp->fqslock);
2712 lockdep_set_class_and_name(&rnp->fqslock,
2713 &rcu_fqs_class[i], fqs[i]);
2714 rnp->gpnum = rsp->gpnum;
2715 rnp->completed = rsp->completed;
2717 rnp->qsmaskinit = 0;
2718 rnp->grplo = j * cpustride;
2719 rnp->grphi = (j + 1) * cpustride - 1;
2720 if (rnp->grphi >= NR_CPUS)
2721 rnp->grphi = NR_CPUS - 1;
2727 rnp->grpnum = j % rsp->levelspread[i - 1];
2728 rnp->grpmask = 1UL << rnp->grpnum;
2729 rnp->parent = rsp->level[i - 1] +
2730 j / rsp->levelspread[i - 1];
2733 INIT_LIST_HEAD(&rnp->blkd_tasks);
2738 init_waitqueue_head(&rsp->gp_wq);
2739 rnp = rsp->level[rcu_num_lvls - 1];
2740 for_each_possible_cpu(i) {
2741 while (i > rnp->grphi)
2743 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2744 rcu_boot_init_percpu_data(i, rsp);
2746 list_add(&rsp->flavors, &rcu_struct_flavors);
2750 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2751 * replace the definitions in rcutree.h because those are needed to size
2752 * the ->node array in the rcu_state structure.
2754 static void __init rcu_init_geometry(void)
2759 int rcu_capacity[MAX_RCU_LVLS + 1];
2761 /* If the compile-time values are accurate, just leave. */
2762 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
2763 nr_cpu_ids == NR_CPUS)
2767 * Compute number of nodes that can be handled an rcu_node tree
2768 * with the given number of levels. Setting rcu_capacity[0] makes
2769 * some of the arithmetic easier.
2771 rcu_capacity[0] = 1;
2772 rcu_capacity[1] = rcu_fanout_leaf;
2773 for (i = 2; i <= MAX_RCU_LVLS; i++)
2774 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2777 * The boot-time rcu_fanout_leaf parameter is only permitted
2778 * to increase the leaf-level fanout, not decrease it. Of course,
2779 * the leaf-level fanout cannot exceed the number of bits in
2780 * the rcu_node masks. Finally, the tree must be able to accommodate
2781 * the configured number of CPUs. Complain and fall back to the
2782 * compile-time values if these limits are exceeded.
2784 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2785 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2786 n > rcu_capacity[MAX_RCU_LVLS]) {
2791 /* Calculate the number of rcu_nodes at each level of the tree. */
2792 for (i = 1; i <= MAX_RCU_LVLS; i++)
2793 if (n <= rcu_capacity[i]) {
2794 for (j = 0; j <= i; j++)
2796 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2798 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2803 /* Calculate the total number of rcu_node structures. */
2805 for (i = 0; i <= MAX_RCU_LVLS; i++)
2806 rcu_num_nodes += num_rcu_lvl[i];
2810 void __init rcu_init(void)
2814 rcu_bootup_announce();
2815 rcu_init_geometry();
2816 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2817 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2818 __rcu_init_preempt();
2819 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2822 * We don't need protection against CPU-hotplug here because
2823 * this is called early in boot, before either interrupts
2824 * or the scheduler are operational.
2826 cpu_notifier(rcu_cpu_notify, 0);
2827 for_each_online_cpu(cpu)
2828 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2829 check_cpu_stall_init();
2832 #include "rcutree_plugin.h"