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 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
75 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \
79 struct rcu_state rcu_sched_state =
80 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
81 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
83 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
84 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
86 static struct rcu_state *rcu_state;
87 LIST_HEAD(rcu_struct_flavors);
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;
159 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
160 * permit this function to be invoked without holding the root rcu_node
161 * structure's ->lock, but of course results can be subject to change.
163 static int rcu_gp_in_progress(struct rcu_state *rsp)
165 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
169 * Note a quiescent state. Because we do not need to know
170 * how many quiescent states passed, just if there was at least
171 * one since the start of the grace period, this just sets a flag.
172 * The caller must have disabled preemption.
174 void rcu_sched_qs(int cpu)
176 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
178 rdp->passed_quiesce_gpnum = rdp->gpnum;
180 if (rdp->passed_quiesce == 0)
181 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
182 rdp->passed_quiesce = 1;
185 void rcu_bh_qs(int cpu)
187 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
189 rdp->passed_quiesce_gpnum = rdp->gpnum;
191 if (rdp->passed_quiesce == 0)
192 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
193 rdp->passed_quiesce = 1;
197 * Note a context switch. This is a quiescent state for RCU-sched,
198 * and requires special handling for preemptible RCU.
199 * The caller must have disabled preemption.
201 void rcu_note_context_switch(int cpu)
203 trace_rcu_utilization("Start context switch");
205 rcu_preempt_note_context_switch(cpu);
206 trace_rcu_utilization("End context switch");
208 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
210 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
211 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
212 .dynticks = ATOMIC_INIT(1),
215 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
216 static int qhimark = 10000; /* If this many pending, ignore blimit. */
217 static int qlowmark = 100; /* Once only this many pending, use blimit. */
219 module_param(blimit, int, 0);
220 module_param(qhimark, int, 0);
221 module_param(qlowmark, int, 0);
223 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
224 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
226 module_param(rcu_cpu_stall_suppress, int, 0644);
227 module_param(rcu_cpu_stall_timeout, int, 0644);
229 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
230 static int rcu_pending(int cpu);
233 * Return the number of RCU-sched batches processed thus far for debug & stats.
235 long rcu_batches_completed_sched(void)
237 return rcu_sched_state.completed;
239 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
242 * Return the number of RCU BH batches processed thus far for debug & stats.
244 long rcu_batches_completed_bh(void)
246 return rcu_bh_state.completed;
248 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
251 * Force a quiescent state for RCU BH.
253 void rcu_bh_force_quiescent_state(void)
255 force_quiescent_state(&rcu_bh_state, 0);
257 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
260 * Record the number of times rcutorture tests have been initiated and
261 * terminated. This information allows the debugfs tracing stats to be
262 * correlated to the rcutorture messages, even when the rcutorture module
263 * is being repeatedly loaded and unloaded. In other words, we cannot
264 * store this state in rcutorture itself.
266 void rcutorture_record_test_transition(void)
268 rcutorture_testseq++;
269 rcutorture_vernum = 0;
271 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
274 * Record the number of writer passes through the current rcutorture test.
275 * This is also used to correlate debugfs tracing stats with the rcutorture
278 void rcutorture_record_progress(unsigned long vernum)
282 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
285 * Force a quiescent state for RCU-sched.
287 void rcu_sched_force_quiescent_state(void)
289 force_quiescent_state(&rcu_sched_state, 0);
291 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
294 * Does the CPU have callbacks ready to be invoked?
297 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
299 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
303 * Does the current CPU require a yet-as-unscheduled grace period?
306 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
308 return *rdp->nxttail[RCU_DONE_TAIL +
309 ACCESS_ONCE(rsp->completed) != rdp->completed] &&
310 !rcu_gp_in_progress(rsp);
314 * Return the root node of the specified rcu_state structure.
316 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
318 return &rsp->node[0];
322 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
324 * If the new value of the ->dynticks_nesting counter now is zero,
325 * we really have entered idle, and must do the appropriate accounting.
326 * The caller must have disabled interrupts.
328 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
330 trace_rcu_dyntick("Start", oldval, 0);
331 if (!is_idle_task(current)) {
332 struct task_struct *idle = idle_task(smp_processor_id());
334 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
335 ftrace_dump(DUMP_ORIG);
336 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
337 current->pid, current->comm,
338 idle->pid, idle->comm); /* must be idle task! */
340 rcu_prepare_for_idle(smp_processor_id());
341 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
342 smp_mb__before_atomic_inc(); /* See above. */
343 atomic_inc(&rdtp->dynticks);
344 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
345 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
348 * The idle task is not permitted to enter the idle loop while
349 * in an RCU read-side critical section.
351 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
352 "Illegal idle entry in RCU read-side critical section.");
353 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
354 "Illegal idle entry in RCU-bh read-side critical section.");
355 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
356 "Illegal idle entry in RCU-sched read-side critical section.");
360 * rcu_idle_enter - inform RCU that current CPU is entering idle
362 * Enter idle mode, in other words, -leave- the mode in which RCU
363 * read-side critical sections can occur. (Though RCU read-side
364 * critical sections can occur in irq handlers in idle, a possibility
365 * handled by irq_enter() and irq_exit().)
367 * We crowbar the ->dynticks_nesting field to zero to allow for
368 * the possibility of usermode upcalls having messed up our count
369 * of interrupt nesting level during the prior busy period.
371 void rcu_idle_enter(void)
375 struct rcu_dynticks *rdtp;
377 local_irq_save(flags);
378 rdtp = &__get_cpu_var(rcu_dynticks);
379 oldval = rdtp->dynticks_nesting;
380 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
381 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
382 rdtp->dynticks_nesting = 0;
384 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
385 rcu_idle_enter_common(rdtp, oldval);
386 local_irq_restore(flags);
388 EXPORT_SYMBOL_GPL(rcu_idle_enter);
391 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
393 * Exit from an interrupt handler, which might possibly result in entering
394 * idle mode, in other words, leaving the mode in which read-side critical
395 * sections can occur.
397 * This code assumes that the idle loop never does anything that might
398 * result in unbalanced calls to irq_enter() and irq_exit(). If your
399 * architecture violates this assumption, RCU will give you what you
400 * deserve, good and hard. But very infrequently and irreproducibly.
402 * Use things like work queues to work around this limitation.
404 * You have been warned.
406 void rcu_irq_exit(void)
410 struct rcu_dynticks *rdtp;
412 local_irq_save(flags);
413 rdtp = &__get_cpu_var(rcu_dynticks);
414 oldval = rdtp->dynticks_nesting;
415 rdtp->dynticks_nesting--;
416 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
417 if (rdtp->dynticks_nesting)
418 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
420 rcu_idle_enter_common(rdtp, oldval);
421 local_irq_restore(flags);
425 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
427 * If the new value of the ->dynticks_nesting counter was previously zero,
428 * we really have exited idle, and must do the appropriate accounting.
429 * The caller must have disabled interrupts.
431 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
433 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
434 atomic_inc(&rdtp->dynticks);
435 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
436 smp_mb__after_atomic_inc(); /* See above. */
437 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
438 rcu_cleanup_after_idle(smp_processor_id());
439 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
440 if (!is_idle_task(current)) {
441 struct task_struct *idle = idle_task(smp_processor_id());
443 trace_rcu_dyntick("Error on exit: not idle task",
444 oldval, rdtp->dynticks_nesting);
445 ftrace_dump(DUMP_ORIG);
446 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
447 current->pid, current->comm,
448 idle->pid, idle->comm); /* must be idle task! */
453 * rcu_idle_exit - inform RCU that current CPU is leaving idle
455 * Exit idle mode, in other words, -enter- the mode in which RCU
456 * read-side critical sections can occur.
458 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
459 * allow for the possibility of usermode upcalls messing up our count
460 * of interrupt nesting level during the busy period that is just
463 void rcu_idle_exit(void)
466 struct rcu_dynticks *rdtp;
469 local_irq_save(flags);
470 rdtp = &__get_cpu_var(rcu_dynticks);
471 oldval = rdtp->dynticks_nesting;
472 WARN_ON_ONCE(oldval < 0);
473 if (oldval & DYNTICK_TASK_NEST_MASK)
474 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
476 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
477 rcu_idle_exit_common(rdtp, oldval);
478 local_irq_restore(flags);
480 EXPORT_SYMBOL_GPL(rcu_idle_exit);
483 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
485 * Enter an interrupt handler, which might possibly result in exiting
486 * idle mode, in other words, entering the mode in which read-side critical
487 * sections can occur.
489 * Note that the Linux kernel is fully capable of entering an interrupt
490 * handler that it never exits, for example when doing upcalls to
491 * user mode! This code assumes that the idle loop never does upcalls to
492 * user mode. If your architecture does do upcalls from the idle loop (or
493 * does anything else that results in unbalanced calls to the irq_enter()
494 * and irq_exit() functions), RCU will give you what you deserve, good
495 * and hard. But very infrequently and irreproducibly.
497 * Use things like work queues to work around this limitation.
499 * You have been warned.
501 void rcu_irq_enter(void)
504 struct rcu_dynticks *rdtp;
507 local_irq_save(flags);
508 rdtp = &__get_cpu_var(rcu_dynticks);
509 oldval = rdtp->dynticks_nesting;
510 rdtp->dynticks_nesting++;
511 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
513 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
515 rcu_idle_exit_common(rdtp, oldval);
516 local_irq_restore(flags);
520 * rcu_nmi_enter - inform RCU of entry to NMI context
522 * If the CPU was idle with dynamic ticks active, and there is no
523 * irq handler running, this updates rdtp->dynticks_nmi to let the
524 * RCU grace-period handling know that the CPU is active.
526 void rcu_nmi_enter(void)
528 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
530 if (rdtp->dynticks_nmi_nesting == 0 &&
531 (atomic_read(&rdtp->dynticks) & 0x1))
533 rdtp->dynticks_nmi_nesting++;
534 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
535 atomic_inc(&rdtp->dynticks);
536 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
537 smp_mb__after_atomic_inc(); /* See above. */
538 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
542 * rcu_nmi_exit - inform RCU of exit from NMI context
544 * If the CPU was idle with dynamic ticks active, and there is no
545 * irq handler running, this updates rdtp->dynticks_nmi to let the
546 * RCU grace-period handling know that the CPU is no longer active.
548 void rcu_nmi_exit(void)
550 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
552 if (rdtp->dynticks_nmi_nesting == 0 ||
553 --rdtp->dynticks_nmi_nesting != 0)
555 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
556 smp_mb__before_atomic_inc(); /* See above. */
557 atomic_inc(&rdtp->dynticks);
558 smp_mb__after_atomic_inc(); /* Force delay to next write. */
559 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
563 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
565 * If the current CPU is in its idle loop and is neither in an interrupt
566 * or NMI handler, return true.
568 int rcu_is_cpu_idle(void)
573 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
577 EXPORT_SYMBOL(rcu_is_cpu_idle);
579 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
582 * Is the current CPU online? Disable preemption to avoid false positives
583 * that could otherwise happen due to the current CPU number being sampled,
584 * this task being preempted, its old CPU being taken offline, resuming
585 * on some other CPU, then determining that its old CPU is now offline.
586 * It is OK to use RCU on an offline processor during initial boot, hence
587 * the check for rcu_scheduler_fully_active. Note also that it is OK
588 * for a CPU coming online to use RCU for one jiffy prior to marking itself
589 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
590 * offline to continue to use RCU for one jiffy after marking itself
591 * offline in the cpu_online_mask. This leniency is necessary given the
592 * non-atomic nature of the online and offline processing, for example,
593 * the fact that a CPU enters the scheduler after completing the CPU_DYING
596 * This is also why RCU internally marks CPUs online during the
597 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
599 * Disable checking if in an NMI handler because we cannot safely report
600 * errors from NMI handlers anyway.
602 bool rcu_lockdep_current_cpu_online(void)
604 struct rcu_data *rdp;
605 struct rcu_node *rnp;
611 rdp = &__get_cpu_var(rcu_sched_data);
613 ret = (rdp->grpmask & rnp->qsmaskinit) ||
614 !rcu_scheduler_fully_active;
618 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
620 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
623 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
625 * If the current CPU is idle or running at a first-level (not nested)
626 * interrupt from idle, return true. The caller must have at least
627 * disabled preemption.
629 int rcu_is_cpu_rrupt_from_idle(void)
631 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
635 * Snapshot the specified CPU's dynticks counter so that we can later
636 * credit them with an implicit quiescent state. Return 1 if this CPU
637 * is in dynticks idle mode, which is an extended quiescent state.
639 static int dyntick_save_progress_counter(struct rcu_data *rdp)
641 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
642 return (rdp->dynticks_snap & 0x1) == 0;
646 * Return true if the specified CPU has passed through a quiescent
647 * state by virtue of being in or having passed through an dynticks
648 * idle state since the last call to dyntick_save_progress_counter()
649 * for this same CPU, or by virtue of having been offline.
651 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
656 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
657 snap = (unsigned int)rdp->dynticks_snap;
660 * If the CPU passed through or entered a dynticks idle phase with
661 * no active irq/NMI handlers, then we can safely pretend that the CPU
662 * already acknowledged the request to pass through a quiescent
663 * state. Either way, that CPU cannot possibly be in an RCU
664 * read-side critical section that started before the beginning
665 * of the current RCU grace period.
667 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
668 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
674 * Check for the CPU being offline, but only if the grace period
675 * is old enough. We don't need to worry about the CPU changing
676 * state: If we see it offline even once, it has been through a
679 * The reason for insisting that the grace period be at least
680 * one jiffy old is that CPUs that are not quite online and that
681 * have just gone offline can still execute RCU read-side critical
684 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
685 return 0; /* Grace period is not old enough. */
687 if (cpu_is_offline(rdp->cpu)) {
688 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
695 static int jiffies_till_stall_check(void)
697 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
700 * Limit check must be consistent with the Kconfig limits
701 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
703 if (till_stall_check < 3) {
704 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
705 till_stall_check = 3;
706 } else if (till_stall_check > 300) {
707 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
708 till_stall_check = 300;
710 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
713 static void record_gp_stall_check_time(struct rcu_state *rsp)
715 rsp->gp_start = jiffies;
716 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
719 static void print_other_cpu_stall(struct rcu_state *rsp)
725 struct rcu_node *rnp = rcu_get_root(rsp);
727 /* Only let one CPU complain about others per time interval. */
729 raw_spin_lock_irqsave(&rnp->lock, flags);
730 delta = jiffies - rsp->jiffies_stall;
731 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
732 raw_spin_unlock_irqrestore(&rnp->lock, flags);
735 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
736 raw_spin_unlock_irqrestore(&rnp->lock, flags);
739 * OK, time to rat on our buddy...
740 * See Documentation/RCU/stallwarn.txt for info on how to debug
741 * RCU CPU stall warnings.
743 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
745 print_cpu_stall_info_begin();
746 rcu_for_each_leaf_node(rsp, rnp) {
747 raw_spin_lock_irqsave(&rnp->lock, flags);
748 ndetected += rcu_print_task_stall(rnp);
749 if (rnp->qsmask != 0) {
750 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
751 if (rnp->qsmask & (1UL << cpu)) {
752 print_cpu_stall_info(rsp,
757 raw_spin_unlock_irqrestore(&rnp->lock, flags);
761 * Now rat on any tasks that got kicked up to the root rcu_node
762 * due to CPU offlining.
764 rnp = rcu_get_root(rsp);
765 raw_spin_lock_irqsave(&rnp->lock, flags);
766 ndetected += rcu_print_task_stall(rnp);
767 raw_spin_unlock_irqrestore(&rnp->lock, flags);
769 print_cpu_stall_info_end();
770 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
771 smp_processor_id(), (long)(jiffies - rsp->gp_start));
773 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
774 else if (!trigger_all_cpu_backtrace())
777 /* If so configured, complain about tasks blocking the grace period. */
779 rcu_print_detail_task_stall(rsp);
781 force_quiescent_state(rsp, 0); /* Kick them all. */
784 static void print_cpu_stall(struct rcu_state *rsp)
787 struct rcu_node *rnp = rcu_get_root(rsp);
790 * OK, time to rat on ourselves...
791 * See Documentation/RCU/stallwarn.txt for info on how to debug
792 * RCU CPU stall warnings.
794 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
795 print_cpu_stall_info_begin();
796 print_cpu_stall_info(rsp, smp_processor_id());
797 print_cpu_stall_info_end();
798 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
799 if (!trigger_all_cpu_backtrace())
802 raw_spin_lock_irqsave(&rnp->lock, flags);
803 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
804 rsp->jiffies_stall = jiffies +
805 3 * jiffies_till_stall_check() + 3;
806 raw_spin_unlock_irqrestore(&rnp->lock, flags);
808 set_need_resched(); /* kick ourselves to get things going. */
811 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
815 struct rcu_node *rnp;
817 if (rcu_cpu_stall_suppress)
819 j = ACCESS_ONCE(jiffies);
820 js = ACCESS_ONCE(rsp->jiffies_stall);
822 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
824 /* We haven't checked in, so go dump stack. */
825 print_cpu_stall(rsp);
827 } else if (rcu_gp_in_progress(rsp) &&
828 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
830 /* They had a few time units to dump stack, so complain. */
831 print_other_cpu_stall(rsp);
835 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
837 rcu_cpu_stall_suppress = 1;
842 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
844 * Set the stall-warning timeout way off into the future, thus preventing
845 * any RCU CPU stall-warning messages from appearing in the current set of
848 * The caller must disable hard irqs.
850 void rcu_cpu_stall_reset(void)
852 struct rcu_state *rsp;
854 for_each_rcu_flavor(rsp)
855 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
858 static struct notifier_block rcu_panic_block = {
859 .notifier_call = rcu_panic,
862 static void __init check_cpu_stall_init(void)
864 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
868 * Update CPU-local rcu_data state to record the newly noticed grace period.
869 * This is used both when we started the grace period and when we notice
870 * that someone else started the grace period. The caller must hold the
871 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
872 * and must have irqs disabled.
874 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
876 if (rdp->gpnum != rnp->gpnum) {
878 * If the current grace period is waiting for this CPU,
879 * set up to detect a quiescent state, otherwise don't
880 * go looking for one.
882 rdp->gpnum = rnp->gpnum;
883 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
884 if (rnp->qsmask & rdp->grpmask) {
886 rdp->passed_quiesce = 0;
890 zero_cpu_stall_ticks(rdp);
894 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
897 struct rcu_node *rnp;
899 local_irq_save(flags);
901 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
902 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
903 local_irq_restore(flags);
906 __note_new_gpnum(rsp, rnp, rdp);
907 raw_spin_unlock_irqrestore(&rnp->lock, flags);
911 * Did someone else start a new RCU grace period start since we last
912 * checked? Update local state appropriately if so. Must be called
913 * on the CPU corresponding to rdp.
916 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
921 local_irq_save(flags);
922 if (rdp->gpnum != rsp->gpnum) {
923 note_new_gpnum(rsp, rdp);
926 local_irq_restore(flags);
931 * Initialize the specified rcu_data structure's callback list to empty.
933 static void init_callback_list(struct rcu_data *rdp)
938 for (i = 0; i < RCU_NEXT_SIZE; i++)
939 rdp->nxttail[i] = &rdp->nxtlist;
943 * Advance this CPU's callbacks, but only if the current grace period
944 * has ended. This may be called only from the CPU to whom the rdp
945 * belongs. In addition, the corresponding leaf rcu_node structure's
946 * ->lock must be held by the caller, with irqs disabled.
949 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
951 /* Did another grace period end? */
952 if (rdp->completed != rnp->completed) {
954 /* Advance callbacks. No harm if list empty. */
955 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
956 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
957 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
959 /* Remember that we saw this grace-period completion. */
960 rdp->completed = rnp->completed;
961 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
964 * If we were in an extended quiescent state, we may have
965 * missed some grace periods that others CPUs handled on
966 * our behalf. Catch up with this state to avoid noting
967 * spurious new grace periods. If another grace period
968 * has started, then rnp->gpnum will have advanced, so
969 * we will detect this later on.
971 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
972 rdp->gpnum = rdp->completed;
975 * If RCU does not need a quiescent state from this CPU,
976 * then make sure that this CPU doesn't go looking for one.
978 if ((rnp->qsmask & rdp->grpmask) == 0)
984 * Advance this CPU's callbacks, but only if the current grace period
985 * has ended. This may be called only from the CPU to whom the rdp
989 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
992 struct rcu_node *rnp;
994 local_irq_save(flags);
996 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
997 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
998 local_irq_restore(flags);
1001 __rcu_process_gp_end(rsp, rnp, rdp);
1002 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1006 * Do per-CPU grace-period initialization for running CPU. The caller
1007 * must hold the lock of the leaf rcu_node structure corresponding to
1011 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1013 /* Prior grace period ended, so advance callbacks for current CPU. */
1014 __rcu_process_gp_end(rsp, rnp, rdp);
1017 * Because this CPU just now started the new grace period, we know
1018 * that all of its callbacks will be covered by this upcoming grace
1019 * period, even the ones that were registered arbitrarily recently.
1020 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1022 * Other CPUs cannot be sure exactly when the grace period started.
1023 * Therefore, their recently registered callbacks must pass through
1024 * an additional RCU_NEXT_READY stage, so that they will be handled
1025 * by the next RCU grace period.
1027 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1028 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1030 /* Set state so that this CPU will detect the next quiescent state. */
1031 __note_new_gpnum(rsp, rnp, rdp);
1035 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1036 * in preparation for detecting the next grace period. The caller must hold
1037 * the root node's ->lock, which is released before return. Hard irqs must
1040 * Note that it is legal for a dying CPU (which is marked as offline) to
1041 * invoke this function. This can happen when the dying CPU reports its
1045 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1046 __releases(rcu_get_root(rsp)->lock)
1048 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1049 struct rcu_node *rnp = rcu_get_root(rsp);
1051 if (!rcu_scheduler_fully_active ||
1052 !cpu_needs_another_gp(rsp, rdp)) {
1054 * Either the scheduler hasn't yet spawned the first
1055 * non-idle task or this CPU does not need another
1056 * grace period. Either way, don't start a new grace
1059 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1063 if (rsp->fqs_active) {
1065 * This CPU needs a grace period, but force_quiescent_state()
1066 * is running. Tell it to start one on this CPU's behalf.
1068 rsp->fqs_need_gp = 1;
1069 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1073 /* Advance to a new grace period and initialize state. */
1075 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1076 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1077 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1078 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1079 record_gp_stall_check_time(rsp);
1080 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1082 /* Exclude any concurrent CPU-hotplug operations. */
1083 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1086 * Set the quiescent-state-needed bits in all the rcu_node
1087 * structures for all currently online CPUs in breadth-first
1088 * order, starting from the root rcu_node structure. This
1089 * operation relies on the layout of the hierarchy within the
1090 * rsp->node[] array. Note that other CPUs will access only
1091 * the leaves of the hierarchy, which still indicate that no
1092 * grace period is in progress, at least until the corresponding
1093 * leaf node has been initialized. In addition, we have excluded
1094 * CPU-hotplug operations.
1096 * Note that the grace period cannot complete until we finish
1097 * the initialization process, as there will be at least one
1098 * qsmask bit set in the root node until that time, namely the
1099 * one corresponding to this CPU, due to the fact that we have
1102 rcu_for_each_node_breadth_first(rsp, rnp) {
1103 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1104 rcu_preempt_check_blocked_tasks(rnp);
1105 rnp->qsmask = rnp->qsmaskinit;
1106 rnp->gpnum = rsp->gpnum;
1107 rnp->completed = rsp->completed;
1108 if (rnp == rdp->mynode)
1109 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1110 rcu_preempt_boost_start_gp(rnp);
1111 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1112 rnp->level, rnp->grplo,
1113 rnp->grphi, rnp->qsmask);
1114 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1117 rnp = rcu_get_root(rsp);
1118 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1119 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1120 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1121 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1125 * Report a full set of quiescent states to the specified rcu_state
1126 * data structure. This involves cleaning up after the prior grace
1127 * period and letting rcu_start_gp() start up the next grace period
1128 * if one is needed. Note that the caller must hold rnp->lock, as
1129 * required by rcu_start_gp(), which will release it.
1131 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1132 __releases(rcu_get_root(rsp)->lock)
1134 unsigned long gp_duration;
1135 struct rcu_node *rnp = rcu_get_root(rsp);
1136 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1138 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1141 * Ensure that all grace-period and pre-grace-period activity
1142 * is seen before the assignment to rsp->completed.
1144 smp_mb(); /* See above block comment. */
1145 gp_duration = jiffies - rsp->gp_start;
1146 if (gp_duration > rsp->gp_max)
1147 rsp->gp_max = gp_duration;
1150 * We know the grace period is complete, but to everyone else
1151 * it appears to still be ongoing. But it is also the case
1152 * that to everyone else it looks like there is nothing that
1153 * they can do to advance the grace period. It is therefore
1154 * safe for us to drop the lock in order to mark the grace
1155 * period as completed in all of the rcu_node structures.
1157 * But if this CPU needs another grace period, it will take
1158 * care of this while initializing the next grace period.
1159 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1160 * because the callbacks have not yet been advanced: Those
1161 * callbacks are waiting on the grace period that just now
1164 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1165 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1168 * Propagate new ->completed value to rcu_node structures
1169 * so that other CPUs don't have to wait until the start
1170 * of the next grace period to process their callbacks.
1172 rcu_for_each_node_breadth_first(rsp, rnp) {
1173 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1174 rnp->completed = rsp->gpnum;
1175 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1177 rnp = rcu_get_root(rsp);
1178 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1181 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1182 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1183 rsp->fqs_state = RCU_GP_IDLE;
1184 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1188 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1189 * Allows quiescent states for a group of CPUs to be reported at one go
1190 * to the specified rcu_node structure, though all the CPUs in the group
1191 * must be represented by the same rcu_node structure (which need not be
1192 * a leaf rcu_node structure, though it often will be). That structure's
1193 * lock must be held upon entry, and it is released before return.
1196 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1197 struct rcu_node *rnp, unsigned long flags)
1198 __releases(rnp->lock)
1200 struct rcu_node *rnp_c;
1202 /* Walk up the rcu_node hierarchy. */
1204 if (!(rnp->qsmask & mask)) {
1206 /* Our bit has already been cleared, so done. */
1207 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1210 rnp->qsmask &= ~mask;
1211 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1212 mask, rnp->qsmask, rnp->level,
1213 rnp->grplo, rnp->grphi,
1215 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1217 /* Other bits still set at this level, so done. */
1218 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1221 mask = rnp->grpmask;
1222 if (rnp->parent == NULL) {
1224 /* No more levels. Exit loop holding root lock. */
1228 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1231 raw_spin_lock_irqsave(&rnp->lock, flags);
1232 WARN_ON_ONCE(rnp_c->qsmask);
1236 * Get here if we are the last CPU to pass through a quiescent
1237 * state for this grace period. Invoke rcu_report_qs_rsp()
1238 * to clean up and start the next grace period if one is needed.
1240 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1244 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1245 * structure. This must be either called from the specified CPU, or
1246 * called when the specified CPU is known to be offline (and when it is
1247 * also known that no other CPU is concurrently trying to help the offline
1248 * CPU). The lastcomp argument is used to make sure we are still in the
1249 * grace period of interest. We don't want to end the current grace period
1250 * based on quiescent states detected in an earlier grace period!
1253 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1255 unsigned long flags;
1257 struct rcu_node *rnp;
1260 raw_spin_lock_irqsave(&rnp->lock, flags);
1261 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1264 * The grace period in which this quiescent state was
1265 * recorded has ended, so don't report it upwards.
1266 * We will instead need a new quiescent state that lies
1267 * within the current grace period.
1269 rdp->passed_quiesce = 0; /* need qs for new gp. */
1270 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1273 mask = rdp->grpmask;
1274 if ((rnp->qsmask & mask) == 0) {
1275 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1277 rdp->qs_pending = 0;
1280 * This GP can't end until cpu checks in, so all of our
1281 * callbacks can be processed during the next GP.
1283 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1285 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1290 * Check to see if there is a new grace period of which this CPU
1291 * is not yet aware, and if so, set up local rcu_data state for it.
1292 * Otherwise, see if this CPU has just passed through its first
1293 * quiescent state for this grace period, and record that fact if so.
1296 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1298 /* If there is now a new grace period, record and return. */
1299 if (check_for_new_grace_period(rsp, rdp))
1303 * Does this CPU still need to do its part for current grace period?
1304 * If no, return and let the other CPUs do their part as well.
1306 if (!rdp->qs_pending)
1310 * Was there a quiescent state since the beginning of the grace
1311 * period? If no, then exit and wait for the next call.
1313 if (!rdp->passed_quiesce)
1317 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1320 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1323 #ifdef CONFIG_HOTPLUG_CPU
1326 * Send the specified CPU's RCU callbacks to the orphanage. The
1327 * specified CPU must be offline, and the caller must hold the
1331 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1332 struct rcu_node *rnp, struct rcu_data *rdp)
1335 * Orphan the callbacks. First adjust the counts. This is safe
1336 * because ->onofflock excludes _rcu_barrier()'s adoption of
1337 * the callbacks, thus no memory barrier is required.
1339 if (rdp->nxtlist != NULL) {
1340 rsp->qlen_lazy += rdp->qlen_lazy;
1341 rsp->qlen += rdp->qlen;
1342 rdp->n_cbs_orphaned += rdp->qlen;
1344 ACCESS_ONCE(rdp->qlen) = 0;
1348 * Next, move those callbacks still needing a grace period to
1349 * the orphanage, where some other CPU will pick them up.
1350 * Some of the callbacks might have gone partway through a grace
1351 * period, but that is too bad. They get to start over because we
1352 * cannot assume that grace periods are synchronized across CPUs.
1353 * We don't bother updating the ->nxttail[] array yet, instead
1354 * we just reset the whole thing later on.
1356 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1357 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1358 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1359 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1363 * Then move the ready-to-invoke callbacks to the orphanage,
1364 * where some other CPU will pick them up. These will not be
1365 * required to pass though another grace period: They are done.
1367 if (rdp->nxtlist != NULL) {
1368 *rsp->orphan_donetail = rdp->nxtlist;
1369 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1372 /* Finally, initialize the rcu_data structure's list to empty. */
1373 init_callback_list(rdp);
1377 * Adopt the RCU callbacks from the specified rcu_state structure's
1378 * orphanage. The caller must hold the ->onofflock.
1380 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1383 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1386 * If there is an rcu_barrier() operation in progress, then
1387 * only the task doing that operation is permitted to adopt
1388 * callbacks. To do otherwise breaks rcu_barrier() and friends
1389 * by causing them to fail to wait for the callbacks in the
1392 if (rsp->rcu_barrier_in_progress &&
1393 rsp->rcu_barrier_in_progress != current)
1396 /* Do the accounting first. */
1397 rdp->qlen_lazy += rsp->qlen_lazy;
1398 rdp->qlen += rsp->qlen;
1399 rdp->n_cbs_adopted += rsp->qlen;
1400 if (rsp->qlen_lazy != rsp->qlen)
1401 rcu_idle_count_callbacks_posted();
1406 * We do not need a memory barrier here because the only way we
1407 * can get here if there is an rcu_barrier() in flight is if
1408 * we are the task doing the rcu_barrier().
1411 /* First adopt the ready-to-invoke callbacks. */
1412 if (rsp->orphan_donelist != NULL) {
1413 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1414 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1415 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1416 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1417 rdp->nxttail[i] = rsp->orphan_donetail;
1418 rsp->orphan_donelist = NULL;
1419 rsp->orphan_donetail = &rsp->orphan_donelist;
1422 /* And then adopt the callbacks that still need a grace period. */
1423 if (rsp->orphan_nxtlist != NULL) {
1424 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1425 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1426 rsp->orphan_nxtlist = NULL;
1427 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1432 * Trace the fact that this CPU is going offline.
1434 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1436 RCU_TRACE(unsigned long mask);
1437 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1438 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1440 RCU_TRACE(mask = rdp->grpmask);
1441 trace_rcu_grace_period(rsp->name,
1442 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1447 * The CPU has been completely removed, and some other CPU is reporting
1448 * this fact from process context. Do the remainder of the cleanup,
1449 * including orphaning the outgoing CPU's RCU callbacks, and also
1450 * adopting them, if there is no _rcu_barrier() instance running.
1451 * There can only be one CPU hotplug operation at a time, so no other
1452 * CPU can be attempting to update rcu_cpu_kthread_task.
1454 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1456 unsigned long flags;
1458 int need_report = 0;
1459 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1460 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1462 /* Adjust any no-longer-needed kthreads. */
1463 rcu_stop_cpu_kthread(cpu);
1464 rcu_node_kthread_setaffinity(rnp, -1);
1466 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1468 /* Exclude any attempts to start a new grace period. */
1469 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1471 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1472 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1473 rcu_adopt_orphan_cbs(rsp);
1475 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1476 mask = rdp->grpmask; /* rnp->grplo is constant. */
1478 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1479 rnp->qsmaskinit &= ~mask;
1480 if (rnp->qsmaskinit != 0) {
1481 if (rnp != rdp->mynode)
1482 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1485 if (rnp == rdp->mynode)
1486 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1488 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1489 mask = rnp->grpmask;
1491 } while (rnp != NULL);
1494 * We still hold the leaf rcu_node structure lock here, and
1495 * irqs are still disabled. The reason for this subterfuge is
1496 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1497 * held leads to deadlock.
1499 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1501 if (need_report & RCU_OFL_TASKS_NORM_GP)
1502 rcu_report_unblock_qs_rnp(rnp, flags);
1504 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1505 if (need_report & RCU_OFL_TASKS_EXP_GP)
1506 rcu_report_exp_rnp(rsp, rnp, true);
1507 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1508 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1509 cpu, rdp->qlen, rdp->nxtlist);
1512 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1514 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1518 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1522 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1526 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1529 * Invoke any RCU callbacks that have made it to the end of their grace
1530 * period. Thottle as specified by rdp->blimit.
1532 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1534 unsigned long flags;
1535 struct rcu_head *next, *list, **tail;
1536 int bl, count, count_lazy, i;
1538 /* If no callbacks are ready, just return.*/
1539 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1540 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1541 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1542 need_resched(), is_idle_task(current),
1543 rcu_is_callbacks_kthread());
1548 * Extract the list of ready callbacks, disabling to prevent
1549 * races with call_rcu() from interrupt handlers.
1551 local_irq_save(flags);
1552 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1554 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1555 list = rdp->nxtlist;
1556 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1557 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1558 tail = rdp->nxttail[RCU_DONE_TAIL];
1559 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1560 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1561 rdp->nxttail[i] = &rdp->nxtlist;
1562 local_irq_restore(flags);
1564 /* Invoke callbacks. */
1565 count = count_lazy = 0;
1569 debug_rcu_head_unqueue(list);
1570 if (__rcu_reclaim(rsp->name, list))
1573 /* Stop only if limit reached and CPU has something to do. */
1574 if (++count >= bl &&
1576 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1580 local_irq_save(flags);
1581 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1582 is_idle_task(current),
1583 rcu_is_callbacks_kthread());
1585 /* Update count, and requeue any remaining callbacks. */
1587 *tail = rdp->nxtlist;
1588 rdp->nxtlist = list;
1589 for (i = 0; i < RCU_NEXT_SIZE; i++)
1590 if (&rdp->nxtlist == rdp->nxttail[i])
1591 rdp->nxttail[i] = tail;
1595 smp_mb(); /* List handling before counting for rcu_barrier(). */
1596 rdp->qlen_lazy -= count_lazy;
1597 ACCESS_ONCE(rdp->qlen) -= count;
1598 rdp->n_cbs_invoked += count;
1600 /* Reinstate batch limit if we have worked down the excess. */
1601 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1602 rdp->blimit = blimit;
1604 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1605 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1606 rdp->qlen_last_fqs_check = 0;
1607 rdp->n_force_qs_snap = rsp->n_force_qs;
1608 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1609 rdp->qlen_last_fqs_check = rdp->qlen;
1610 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1612 local_irq_restore(flags);
1614 /* Re-invoke RCU core processing if there are callbacks remaining. */
1615 if (cpu_has_callbacks_ready_to_invoke(rdp))
1620 * Check to see if this CPU is in a non-context-switch quiescent state
1621 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1622 * Also schedule RCU core processing.
1624 * This function must be called from hardirq context. It is normally
1625 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1626 * false, there is no point in invoking rcu_check_callbacks().
1628 void rcu_check_callbacks(int cpu, int user)
1630 trace_rcu_utilization("Start scheduler-tick");
1631 increment_cpu_stall_ticks();
1632 if (user || rcu_is_cpu_rrupt_from_idle()) {
1635 * Get here if this CPU took its interrupt from user
1636 * mode or from the idle loop, and if this is not a
1637 * nested interrupt. In this case, the CPU is in
1638 * a quiescent state, so note it.
1640 * No memory barrier is required here because both
1641 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1642 * variables that other CPUs neither access nor modify,
1643 * at least not while the corresponding CPU is online.
1649 } else if (!in_softirq()) {
1652 * Get here if this CPU did not take its interrupt from
1653 * softirq, in other words, if it is not interrupting
1654 * a rcu_bh read-side critical section. This is an _bh
1655 * critical section, so note it.
1660 rcu_preempt_check_callbacks(cpu);
1661 if (rcu_pending(cpu))
1663 trace_rcu_utilization("End scheduler-tick");
1667 * Scan the leaf rcu_node structures, processing dyntick state for any that
1668 * have not yet encountered a quiescent state, using the function specified.
1669 * Also initiate boosting for any threads blocked on the root rcu_node.
1671 * The caller must have suppressed start of new grace periods.
1673 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1677 unsigned long flags;
1679 struct rcu_node *rnp;
1681 rcu_for_each_leaf_node(rsp, rnp) {
1683 raw_spin_lock_irqsave(&rnp->lock, flags);
1684 if (!rcu_gp_in_progress(rsp)) {
1685 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1688 if (rnp->qsmask == 0) {
1689 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1694 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1695 if ((rnp->qsmask & bit) != 0 &&
1696 f(per_cpu_ptr(rsp->rda, cpu)))
1701 /* rcu_report_qs_rnp() releases rnp->lock. */
1702 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1705 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1707 rnp = rcu_get_root(rsp);
1708 if (rnp->qsmask == 0) {
1709 raw_spin_lock_irqsave(&rnp->lock, flags);
1710 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1715 * Force quiescent states on reluctant CPUs, and also detect which
1716 * CPUs are in dyntick-idle mode.
1718 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1720 unsigned long flags;
1721 struct rcu_node *rnp = rcu_get_root(rsp);
1723 trace_rcu_utilization("Start fqs");
1724 if (!rcu_gp_in_progress(rsp)) {
1725 trace_rcu_utilization("End fqs");
1726 return; /* No grace period in progress, nothing to force. */
1728 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1729 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1730 trace_rcu_utilization("End fqs");
1731 return; /* Someone else is already on the job. */
1733 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1734 goto unlock_fqs_ret; /* no emergency and done recently. */
1736 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1737 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1738 if(!rcu_gp_in_progress(rsp)) {
1739 rsp->n_force_qs_ngp++;
1740 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1741 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1743 rsp->fqs_active = 1;
1744 switch (rsp->fqs_state) {
1748 break; /* grace period idle or initializing, ignore. */
1750 case RCU_SAVE_DYNTICK:
1752 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1754 /* Record dyntick-idle state. */
1755 force_qs_rnp(rsp, dyntick_save_progress_counter);
1756 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1757 if (rcu_gp_in_progress(rsp))
1758 rsp->fqs_state = RCU_FORCE_QS;
1763 /* Check dyntick-idle state, send IPI to laggarts. */
1764 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1765 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1767 /* Leave state in case more forcing is required. */
1769 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1772 rsp->fqs_active = 0;
1773 if (rsp->fqs_need_gp) {
1774 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1775 rsp->fqs_need_gp = 0;
1776 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1777 trace_rcu_utilization("End fqs");
1780 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1782 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1783 trace_rcu_utilization("End fqs");
1787 * This does the RCU core processing work for the specified rcu_state
1788 * and rcu_data structures. This may be called only from the CPU to
1789 * whom the rdp belongs.
1792 __rcu_process_callbacks(struct rcu_state *rsp)
1794 unsigned long flags;
1795 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1797 WARN_ON_ONCE(rdp->beenonline == 0);
1800 * If an RCU GP has gone long enough, go check for dyntick
1801 * idle CPUs and, if needed, send resched IPIs.
1803 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1804 force_quiescent_state(rsp, 1);
1807 * Advance callbacks in response to end of earlier grace
1808 * period that some other CPU ended.
1810 rcu_process_gp_end(rsp, rdp);
1812 /* Update RCU state based on any recent quiescent states. */
1813 rcu_check_quiescent_state(rsp, rdp);
1815 /* Does this CPU require a not-yet-started grace period? */
1816 if (cpu_needs_another_gp(rsp, rdp)) {
1817 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1818 rcu_start_gp(rsp, flags); /* releases above lock */
1821 /* If there are callbacks ready, invoke them. */
1822 if (cpu_has_callbacks_ready_to_invoke(rdp))
1823 invoke_rcu_callbacks(rsp, rdp);
1827 * Do RCU core processing for the current CPU.
1829 static void rcu_process_callbacks(struct softirq_action *unused)
1831 struct rcu_state *rsp;
1833 trace_rcu_utilization("Start RCU core");
1834 for_each_rcu_flavor(rsp)
1835 __rcu_process_callbacks(rsp);
1836 trace_rcu_utilization("End RCU core");
1840 * Schedule RCU callback invocation. If the specified type of RCU
1841 * does not support RCU priority boosting, just do a direct call,
1842 * otherwise wake up the per-CPU kernel kthread. Note that because we
1843 * are running on the current CPU with interrupts disabled, the
1844 * rcu_cpu_kthread_task cannot disappear out from under us.
1846 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1848 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1850 if (likely(!rsp->boost)) {
1851 rcu_do_batch(rsp, rdp);
1854 invoke_rcu_callbacks_kthread();
1857 static void invoke_rcu_core(void)
1859 raise_softirq(RCU_SOFTIRQ);
1863 * Handle any core-RCU processing required by a call_rcu() invocation.
1865 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1866 struct rcu_head *head, unsigned long flags)
1869 * If called from an extended quiescent state, invoke the RCU
1870 * core in order to force a re-evaluation of RCU's idleness.
1872 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1875 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1876 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1880 * Force the grace period if too many callbacks or too long waiting.
1881 * Enforce hysteresis, and don't invoke force_quiescent_state()
1882 * if some other CPU has recently done so. Also, don't bother
1883 * invoking force_quiescent_state() if the newly enqueued callback
1884 * is the only one waiting for a grace period to complete.
1886 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1888 /* Are we ignoring a completed grace period? */
1889 rcu_process_gp_end(rsp, rdp);
1890 check_for_new_grace_period(rsp, rdp);
1892 /* Start a new grace period if one not already started. */
1893 if (!rcu_gp_in_progress(rsp)) {
1894 unsigned long nestflag;
1895 struct rcu_node *rnp_root = rcu_get_root(rsp);
1897 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1898 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1900 /* Give the grace period a kick. */
1901 rdp->blimit = LONG_MAX;
1902 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1903 *rdp->nxttail[RCU_DONE_TAIL] != head)
1904 force_quiescent_state(rsp, 0);
1905 rdp->n_force_qs_snap = rsp->n_force_qs;
1906 rdp->qlen_last_fqs_check = rdp->qlen;
1908 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1909 force_quiescent_state(rsp, 1);
1913 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1914 struct rcu_state *rsp, bool lazy)
1916 unsigned long flags;
1917 struct rcu_data *rdp;
1919 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1920 debug_rcu_head_queue(head);
1924 smp_mb(); /* Ensure RCU update seen before callback registry. */
1927 * Opportunistically note grace-period endings and beginnings.
1928 * Note that we might see a beginning right after we see an
1929 * end, but never vice versa, since this CPU has to pass through
1930 * a quiescent state betweentimes.
1932 local_irq_save(flags);
1933 rdp = this_cpu_ptr(rsp->rda);
1935 /* Add the callback to our list. */
1936 ACCESS_ONCE(rdp->qlen)++;
1940 rcu_idle_count_callbacks_posted();
1941 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1942 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1943 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1945 if (__is_kfree_rcu_offset((unsigned long)func))
1946 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1947 rdp->qlen_lazy, rdp->qlen);
1949 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1951 /* Go handle any RCU core processing required. */
1952 __call_rcu_core(rsp, rdp, head, flags);
1953 local_irq_restore(flags);
1957 * Queue an RCU-sched callback for invocation after a grace period.
1959 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1961 __call_rcu(head, func, &rcu_sched_state, 0);
1963 EXPORT_SYMBOL_GPL(call_rcu_sched);
1966 * Queue an RCU callback for invocation after a quicker grace period.
1968 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1970 __call_rcu(head, func, &rcu_bh_state, 0);
1972 EXPORT_SYMBOL_GPL(call_rcu_bh);
1975 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1976 * any blocking grace-period wait automatically implies a grace period
1977 * if there is only one CPU online at any point time during execution
1978 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
1979 * occasionally incorrectly indicate that there are multiple CPUs online
1980 * when there was in fact only one the whole time, as this just adds
1981 * some overhead: RCU still operates correctly.
1983 static inline int rcu_blocking_is_gp(void)
1987 might_sleep(); /* Check for RCU read-side critical section. */
1989 ret = num_online_cpus() <= 1;
1995 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1997 * Control will return to the caller some time after a full rcu-sched
1998 * grace period has elapsed, in other words after all currently executing
1999 * rcu-sched read-side critical sections have completed. These read-side
2000 * critical sections are delimited by rcu_read_lock_sched() and
2001 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2002 * local_irq_disable(), and so on may be used in place of
2003 * rcu_read_lock_sched().
2005 * This means that all preempt_disable code sequences, including NMI and
2006 * hardware-interrupt handlers, in progress on entry will have completed
2007 * before this primitive returns. However, this does not guarantee that
2008 * softirq handlers will have completed, since in some kernels, these
2009 * handlers can run in process context, and can block.
2011 * This primitive provides the guarantees made by the (now removed)
2012 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2013 * guarantees that rcu_read_lock() sections will have completed.
2014 * In "classic RCU", these two guarantees happen to be one and
2015 * the same, but can differ in realtime RCU implementations.
2017 void synchronize_sched(void)
2019 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2020 !lock_is_held(&rcu_lock_map) &&
2021 !lock_is_held(&rcu_sched_lock_map),
2022 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2023 if (rcu_blocking_is_gp())
2025 wait_rcu_gp(call_rcu_sched);
2027 EXPORT_SYMBOL_GPL(synchronize_sched);
2030 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2032 * Control will return to the caller some time after a full rcu_bh grace
2033 * period has elapsed, in other words after all currently executing rcu_bh
2034 * read-side critical sections have completed. RCU read-side critical
2035 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2036 * and may be nested.
2038 void synchronize_rcu_bh(void)
2040 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2041 !lock_is_held(&rcu_lock_map) &&
2042 !lock_is_held(&rcu_sched_lock_map),
2043 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2044 if (rcu_blocking_is_gp())
2046 wait_rcu_gp(call_rcu_bh);
2048 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2050 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2051 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2053 static int synchronize_sched_expedited_cpu_stop(void *data)
2056 * There must be a full memory barrier on each affected CPU
2057 * between the time that try_stop_cpus() is called and the
2058 * time that it returns.
2060 * In the current initial implementation of cpu_stop, the
2061 * above condition is already met when the control reaches
2062 * this point and the following smp_mb() is not strictly
2063 * necessary. Do smp_mb() anyway for documentation and
2064 * robustness against future implementation changes.
2066 smp_mb(); /* See above comment block. */
2071 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2073 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2074 * approach to force the grace period to end quickly. This consumes
2075 * significant time on all CPUs and is unfriendly to real-time workloads,
2076 * so is thus not recommended for any sort of common-case code. In fact,
2077 * if you are using synchronize_sched_expedited() in a loop, please
2078 * restructure your code to batch your updates, and then use a single
2079 * synchronize_sched() instead.
2081 * Note that it is illegal to call this function while holding any lock
2082 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2083 * to call this function from a CPU-hotplug notifier. Failing to observe
2084 * these restriction will result in deadlock.
2086 * This implementation can be thought of as an application of ticket
2087 * locking to RCU, with sync_sched_expedited_started and
2088 * sync_sched_expedited_done taking on the roles of the halves
2089 * of the ticket-lock word. Each task atomically increments
2090 * sync_sched_expedited_started upon entry, snapshotting the old value,
2091 * then attempts to stop all the CPUs. If this succeeds, then each
2092 * CPU will have executed a context switch, resulting in an RCU-sched
2093 * grace period. We are then done, so we use atomic_cmpxchg() to
2094 * update sync_sched_expedited_done to match our snapshot -- but
2095 * only if someone else has not already advanced past our snapshot.
2097 * On the other hand, if try_stop_cpus() fails, we check the value
2098 * of sync_sched_expedited_done. If it has advanced past our
2099 * initial snapshot, then someone else must have forced a grace period
2100 * some time after we took our snapshot. In this case, our work is
2101 * done for us, and we can simply return. Otherwise, we try again,
2102 * but keep our initial snapshot for purposes of checking for someone
2103 * doing our work for us.
2105 * If we fail too many times in a row, we fall back to synchronize_sched().
2107 void synchronize_sched_expedited(void)
2109 int firstsnap, s, snap, trycount = 0;
2111 /* Note that atomic_inc_return() implies full memory barrier. */
2112 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2114 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2117 * Each pass through the following loop attempts to force a
2118 * context switch on each CPU.
2120 while (try_stop_cpus(cpu_online_mask,
2121 synchronize_sched_expedited_cpu_stop,
2125 /* No joy, try again later. Or just synchronize_sched(). */
2126 if (trycount++ < 10) {
2127 udelay(trycount * num_online_cpus());
2129 synchronize_sched();
2133 /* Check to see if someone else did our work for us. */
2134 s = atomic_read(&sync_sched_expedited_done);
2135 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2136 smp_mb(); /* ensure test happens before caller kfree */
2141 * Refetching sync_sched_expedited_started allows later
2142 * callers to piggyback on our grace period. We subtract
2143 * 1 to get the same token that the last incrementer got.
2144 * We retry after they started, so our grace period works
2145 * for them, and they started after our first try, so their
2146 * grace period works for us.
2149 snap = atomic_read(&sync_sched_expedited_started);
2150 smp_mb(); /* ensure read is before try_stop_cpus(). */
2154 * Everyone up to our most recent fetch is covered by our grace
2155 * period. Update the counter, but only if our work is still
2156 * relevant -- which it won't be if someone who started later
2157 * than we did beat us to the punch.
2160 s = atomic_read(&sync_sched_expedited_done);
2161 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2162 smp_mb(); /* ensure test happens before caller kfree */
2165 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2169 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2172 * Check to see if there is any immediate RCU-related work to be done
2173 * by the current CPU, for the specified type of RCU, returning 1 if so.
2174 * The checks are in order of increasing expense: checks that can be
2175 * carried out against CPU-local state are performed first. However,
2176 * we must check for CPU stalls first, else we might not get a chance.
2178 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2180 struct rcu_node *rnp = rdp->mynode;
2182 rdp->n_rcu_pending++;
2184 /* Check for CPU stalls, if enabled. */
2185 check_cpu_stall(rsp, rdp);
2187 /* Is the RCU core waiting for a quiescent state from this CPU? */
2188 if (rcu_scheduler_fully_active &&
2189 rdp->qs_pending && !rdp->passed_quiesce) {
2192 * If force_quiescent_state() coming soon and this CPU
2193 * needs a quiescent state, and this is either RCU-sched
2194 * or RCU-bh, force a local reschedule.
2196 rdp->n_rp_qs_pending++;
2197 if (!rdp->preemptible &&
2198 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2201 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2202 rdp->n_rp_report_qs++;
2206 /* Does this CPU have callbacks ready to invoke? */
2207 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2208 rdp->n_rp_cb_ready++;
2212 /* Has RCU gone idle with this CPU needing another grace period? */
2213 if (cpu_needs_another_gp(rsp, rdp)) {
2214 rdp->n_rp_cpu_needs_gp++;
2218 /* Has another RCU grace period completed? */
2219 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2220 rdp->n_rp_gp_completed++;
2224 /* Has a new RCU grace period started? */
2225 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2226 rdp->n_rp_gp_started++;
2230 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2231 if (rcu_gp_in_progress(rsp) &&
2232 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2233 rdp->n_rp_need_fqs++;
2238 rdp->n_rp_need_nothing++;
2243 * Check to see if there is any immediate RCU-related work to be done
2244 * by the current CPU, returning 1 if so. This function is part of the
2245 * RCU implementation; it is -not- an exported member of the RCU API.
2247 static int rcu_pending(int cpu)
2249 struct rcu_state *rsp;
2251 for_each_rcu_flavor(rsp)
2252 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2258 * Check to see if any future RCU-related work will need to be done
2259 * by the current CPU, even if none need be done immediately, returning
2262 static int rcu_cpu_has_callbacks(int cpu)
2264 struct rcu_state *rsp;
2266 /* RCU callbacks either ready or pending? */
2267 for_each_rcu_flavor(rsp)
2268 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2274 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2275 * the compiler is expected to optimize this away.
2277 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2278 int cpu, unsigned long done)
2280 trace_rcu_barrier(rsp->name, s, cpu,
2281 atomic_read(&rsp->barrier_cpu_count), done);
2285 * RCU callback function for _rcu_barrier(). If we are last, wake
2286 * up the task executing _rcu_barrier().
2288 static void rcu_barrier_callback(struct rcu_head *rhp)
2290 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2291 struct rcu_state *rsp = rdp->rsp;
2293 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2294 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2295 complete(&rsp->barrier_completion);
2297 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2302 * Called with preemption disabled, and from cross-cpu IRQ context.
2304 static void rcu_barrier_func(void *type)
2306 struct rcu_state *rsp = type;
2307 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2309 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2310 atomic_inc(&rsp->barrier_cpu_count);
2311 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2315 * Orchestrate the specified type of RCU barrier, waiting for all
2316 * RCU callbacks of the specified type to complete.
2318 static void _rcu_barrier(struct rcu_state *rsp)
2321 unsigned long flags;
2322 struct rcu_data *rdp;
2324 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2325 unsigned long snap_done;
2327 init_rcu_head_on_stack(&rd.barrier_head);
2328 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2330 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2331 mutex_lock(&rsp->barrier_mutex);
2334 * Ensure that all prior references, including to ->n_barrier_done,
2335 * are ordered before the _rcu_barrier() machinery.
2337 smp_mb(); /* See above block comment. */
2340 * Recheck ->n_barrier_done to see if others did our work for us.
2341 * This means checking ->n_barrier_done for an even-to-odd-to-even
2342 * transition. The "if" expression below therefore rounds the old
2343 * value up to the next even number and adds two before comparing.
2345 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2346 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2347 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2348 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2349 smp_mb(); /* caller's subsequent code after above check. */
2350 mutex_unlock(&rsp->barrier_mutex);
2355 * Increment ->n_barrier_done to avoid duplicate work. Use
2356 * ACCESS_ONCE() to prevent the compiler from speculating
2357 * the increment to precede the early-exit check.
2359 ACCESS_ONCE(rsp->n_barrier_done)++;
2360 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2361 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2362 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2365 * Initialize the count to one rather than to zero in order to
2366 * avoid a too-soon return to zero in case of a short grace period
2367 * (or preemption of this task). Also flag this task as doing
2368 * an rcu_barrier(). This will prevent anyone else from adopting
2369 * orphaned callbacks, which could cause otherwise failure if a
2370 * CPU went offline and quickly came back online. To see this,
2371 * consider the following sequence of events:
2373 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2374 * 2. CPU 1 goes offline, orphaning its callbacks.
2375 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2376 * 4. CPU 1 comes back online.
2377 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2378 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2379 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2381 init_completion(&rsp->barrier_completion);
2382 atomic_set(&rsp->barrier_cpu_count, 1);
2383 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2384 rsp->rcu_barrier_in_progress = current;
2385 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2388 * Force every CPU with callbacks to register a new callback
2389 * that will tell us when all the preceding callbacks have
2390 * been invoked. If an offline CPU has callbacks, wait for
2391 * it to either come back online or to finish orphaning those
2394 for_each_possible_cpu(cpu) {
2396 rdp = per_cpu_ptr(rsp->rda, cpu);
2397 if (cpu_is_offline(cpu)) {
2398 _rcu_barrier_trace(rsp, "Offline", cpu,
2399 rsp->n_barrier_done);
2401 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2402 schedule_timeout_interruptible(1);
2403 } else if (ACCESS_ONCE(rdp->qlen)) {
2404 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2405 rsp->n_barrier_done);
2406 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2409 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2410 rsp->n_barrier_done);
2416 * Now that all online CPUs have rcu_barrier_callback() callbacks
2417 * posted, we can adopt all of the orphaned callbacks and place
2418 * an rcu_barrier_callback() callback after them. When that is done,
2419 * we are guaranteed to have an rcu_barrier_callback() callback
2420 * following every callback that could possibly have been
2421 * registered before _rcu_barrier() was called.
2423 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2424 rcu_adopt_orphan_cbs(rsp);
2425 rsp->rcu_barrier_in_progress = NULL;
2426 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2427 atomic_inc(&rsp->barrier_cpu_count);
2428 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2430 rsp->call(&rd.barrier_head, rcu_barrier_callback);
2433 * Now that we have an rcu_barrier_callback() callback on each
2434 * CPU, and thus each counted, remove the initial count.
2436 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2437 complete(&rsp->barrier_completion);
2439 /* Increment ->n_barrier_done to prevent duplicate work. */
2440 smp_mb(); /* Keep increment after above mechanism. */
2441 ACCESS_ONCE(rsp->n_barrier_done)++;
2442 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2443 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2444 smp_mb(); /* Keep increment before caller's subsequent code. */
2446 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2447 wait_for_completion(&rsp->barrier_completion);
2449 /* Other rcu_barrier() invocations can now safely proceed. */
2450 mutex_unlock(&rsp->barrier_mutex);
2452 destroy_rcu_head_on_stack(&rd.barrier_head);
2456 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2458 void rcu_barrier_bh(void)
2460 _rcu_barrier(&rcu_bh_state);
2462 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2465 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2467 void rcu_barrier_sched(void)
2469 _rcu_barrier(&rcu_sched_state);
2471 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2474 * Do boot-time initialization of a CPU's per-CPU RCU data.
2477 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2479 unsigned long flags;
2480 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2481 struct rcu_node *rnp = rcu_get_root(rsp);
2483 /* Set up local state, ensuring consistent view of global state. */
2484 raw_spin_lock_irqsave(&rnp->lock, flags);
2485 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2486 init_callback_list(rdp);
2488 ACCESS_ONCE(rdp->qlen) = 0;
2489 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2490 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2491 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2494 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2498 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2499 * offline event can be happening at a given time. Note also that we
2500 * can accept some slop in the rsp->completed access due to the fact
2501 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2503 static void __cpuinit
2504 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2506 unsigned long flags;
2508 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2509 struct rcu_node *rnp = rcu_get_root(rsp);
2511 /* Set up local state, ensuring consistent view of global state. */
2512 raw_spin_lock_irqsave(&rnp->lock, flags);
2513 rdp->beenonline = 1; /* We have now been online. */
2514 rdp->preemptible = preemptible;
2515 rdp->qlen_last_fqs_check = 0;
2516 rdp->n_force_qs_snap = rsp->n_force_qs;
2517 rdp->blimit = blimit;
2518 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2519 atomic_set(&rdp->dynticks->dynticks,
2520 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2521 rcu_prepare_for_idle_init(cpu);
2522 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2525 * A new grace period might start here. If so, we won't be part
2526 * of it, but that is OK, as we are currently in a quiescent state.
2529 /* Exclude any attempts to start a new GP on large systems. */
2530 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2532 /* Add CPU to rcu_node bitmasks. */
2534 mask = rdp->grpmask;
2536 /* Exclude any attempts to start a new GP on small systems. */
2537 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2538 rnp->qsmaskinit |= mask;
2539 mask = rnp->grpmask;
2540 if (rnp == rdp->mynode) {
2542 * If there is a grace period in progress, we will
2543 * set up to wait for it next time we run the
2546 rdp->gpnum = rnp->completed;
2547 rdp->completed = rnp->completed;
2548 rdp->passed_quiesce = 0;
2549 rdp->qs_pending = 0;
2550 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2551 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2553 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2555 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2557 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2560 static void __cpuinit rcu_prepare_cpu(int cpu)
2562 struct rcu_state *rsp;
2564 for_each_rcu_flavor(rsp)
2565 rcu_init_percpu_data(cpu, rsp,
2566 strcmp(rsp->name, "rcu_preempt") == 0);
2570 * Handle CPU online/offline notification events.
2572 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2573 unsigned long action, void *hcpu)
2575 long cpu = (long)hcpu;
2576 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2577 struct rcu_node *rnp = rdp->mynode;
2578 struct rcu_state *rsp;
2580 trace_rcu_utilization("Start CPU hotplug");
2582 case CPU_UP_PREPARE:
2583 case CPU_UP_PREPARE_FROZEN:
2584 rcu_prepare_cpu(cpu);
2585 rcu_prepare_kthreads(cpu);
2588 case CPU_DOWN_FAILED:
2589 rcu_node_kthread_setaffinity(rnp, -1);
2590 rcu_cpu_kthread_setrt(cpu, 1);
2592 case CPU_DOWN_PREPARE:
2593 rcu_node_kthread_setaffinity(rnp, cpu);
2594 rcu_cpu_kthread_setrt(cpu, 0);
2597 case CPU_DYING_FROZEN:
2599 * The whole machine is "stopped" except this CPU, so we can
2600 * touch any data without introducing corruption. We send the
2601 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2603 for_each_rcu_flavor(rsp)
2604 rcu_cleanup_dying_cpu(rsp);
2605 rcu_cleanup_after_idle(cpu);
2608 case CPU_DEAD_FROZEN:
2609 case CPU_UP_CANCELED:
2610 case CPU_UP_CANCELED_FROZEN:
2611 for_each_rcu_flavor(rsp)
2612 rcu_cleanup_dead_cpu(cpu, rsp);
2617 trace_rcu_utilization("End CPU hotplug");
2622 * This function is invoked towards the end of the scheduler's initialization
2623 * process. Before this is called, the idle task might contain
2624 * RCU read-side critical sections (during which time, this idle
2625 * task is booting the system). After this function is called, the
2626 * idle tasks are prohibited from containing RCU read-side critical
2627 * sections. This function also enables RCU lockdep checking.
2629 void rcu_scheduler_starting(void)
2631 WARN_ON(num_online_cpus() != 1);
2632 WARN_ON(nr_context_switches() > 0);
2633 rcu_scheduler_active = 1;
2637 * Compute the per-level fanout, either using the exact fanout specified
2638 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2640 #ifdef CONFIG_RCU_FANOUT_EXACT
2641 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2645 for (i = rcu_num_lvls - 1; i > 0; i--)
2646 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2647 rsp->levelspread[0] = rcu_fanout_leaf;
2649 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2650 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2657 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2658 ccur = rsp->levelcnt[i];
2659 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2663 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2666 * Helper function for rcu_init() that initializes one rcu_state structure.
2668 static void __init rcu_init_one(struct rcu_state *rsp,
2669 struct rcu_data __percpu *rda)
2671 static char *buf[] = { "rcu_node_level_0",
2674 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2678 struct rcu_node *rnp;
2680 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2682 /* Initialize the level-tracking arrays. */
2684 for (i = 0; i < rcu_num_lvls; i++)
2685 rsp->levelcnt[i] = num_rcu_lvl[i];
2686 for (i = 1; i < rcu_num_lvls; i++)
2687 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2688 rcu_init_levelspread(rsp);
2690 /* Initialize the elements themselves, starting from the leaves. */
2692 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2693 cpustride *= rsp->levelspread[i];
2694 rnp = rsp->level[i];
2695 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2696 raw_spin_lock_init(&rnp->lock);
2697 lockdep_set_class_and_name(&rnp->lock,
2698 &rcu_node_class[i], buf[i]);
2701 rnp->qsmaskinit = 0;
2702 rnp->grplo = j * cpustride;
2703 rnp->grphi = (j + 1) * cpustride - 1;
2704 if (rnp->grphi >= NR_CPUS)
2705 rnp->grphi = NR_CPUS - 1;
2711 rnp->grpnum = j % rsp->levelspread[i - 1];
2712 rnp->grpmask = 1UL << rnp->grpnum;
2713 rnp->parent = rsp->level[i - 1] +
2714 j / rsp->levelspread[i - 1];
2717 INIT_LIST_HEAD(&rnp->blkd_tasks);
2722 rnp = rsp->level[rcu_num_lvls - 1];
2723 for_each_possible_cpu(i) {
2724 while (i > rnp->grphi)
2726 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2727 rcu_boot_init_percpu_data(i, rsp);
2729 list_add(&rsp->flavors, &rcu_struct_flavors);
2733 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2734 * replace the definitions in rcutree.h because those are needed to size
2735 * the ->node array in the rcu_state structure.
2737 static void __init rcu_init_geometry(void)
2742 int rcu_capacity[MAX_RCU_LVLS + 1];
2744 /* If the compile-time values are accurate, just leave. */
2745 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2749 * Compute number of nodes that can be handled an rcu_node tree
2750 * with the given number of levels. Setting rcu_capacity[0] makes
2751 * some of the arithmetic easier.
2753 rcu_capacity[0] = 1;
2754 rcu_capacity[1] = rcu_fanout_leaf;
2755 for (i = 2; i <= MAX_RCU_LVLS; i++)
2756 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2759 * The boot-time rcu_fanout_leaf parameter is only permitted
2760 * to increase the leaf-level fanout, not decrease it. Of course,
2761 * the leaf-level fanout cannot exceed the number of bits in
2762 * the rcu_node masks. Finally, the tree must be able to accommodate
2763 * the configured number of CPUs. Complain and fall back to the
2764 * compile-time values if these limits are exceeded.
2766 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2767 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2768 n > rcu_capacity[MAX_RCU_LVLS]) {
2773 /* Calculate the number of rcu_nodes at each level of the tree. */
2774 for (i = 1; i <= MAX_RCU_LVLS; i++)
2775 if (n <= rcu_capacity[i]) {
2776 for (j = 0; j <= i; j++)
2778 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2780 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2785 /* Calculate the total number of rcu_node structures. */
2787 for (i = 0; i <= MAX_RCU_LVLS; i++)
2788 rcu_num_nodes += num_rcu_lvl[i];
2792 void __init rcu_init(void)
2796 rcu_bootup_announce();
2797 rcu_init_geometry();
2798 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2799 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2800 __rcu_init_preempt();
2801 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2804 * We don't need protection against CPU-hotplug here because
2805 * this is called early in boot, before either interrupts
2806 * or the scheduler are operational.
2808 cpu_notifier(rcu_cpu_notify, 0);
2809 for_each_online_cpu(cpu)
2810 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2811 check_cpu_stall_init();
2814 #include "rcutree_plugin.h"