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 * If the specified CPU is offline, tell the caller that it is in
323 * a quiescent state. Otherwise, whack it with a reschedule IPI.
324 * Grace periods can end up waiting on an offline CPU when that
325 * CPU is in the process of coming online -- it will be added to the
326 * rcu_node bitmasks before it actually makes it online. The same thing
327 * can happen while a CPU is in the process of coming online. Because this
328 * race is quite rare, we check for it after detecting that the grace
329 * period has been delayed rather than checking each and every CPU
330 * each and every time we start a new grace period.
332 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
335 * If the CPU is offline for more than a jiffy, it is in a quiescent
336 * state. We can trust its state not to change because interrupts
337 * are disabled. The reason for the jiffy's worth of slack is to
338 * handle CPUs initializing on the way up and finding their way
339 * to the idle loop on the way down.
341 if (cpu_is_offline(rdp->cpu) &&
342 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
343 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
351 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
353 * If the new value of the ->dynticks_nesting counter now is zero,
354 * we really have entered idle, and must do the appropriate accounting.
355 * The caller must have disabled interrupts.
357 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
359 trace_rcu_dyntick("Start", oldval, 0);
360 if (!is_idle_task(current)) {
361 struct task_struct *idle = idle_task(smp_processor_id());
363 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
364 ftrace_dump(DUMP_ORIG);
365 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
366 current->pid, current->comm,
367 idle->pid, idle->comm); /* must be idle task! */
369 rcu_prepare_for_idle(smp_processor_id());
370 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
371 smp_mb__before_atomic_inc(); /* See above. */
372 atomic_inc(&rdtp->dynticks);
373 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
374 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
377 * The idle task is not permitted to enter the idle loop while
378 * in an RCU read-side critical section.
380 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
381 "Illegal idle entry in RCU read-side critical section.");
382 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
383 "Illegal idle entry in RCU-bh read-side critical section.");
384 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
385 "Illegal idle entry in RCU-sched read-side critical section.");
389 * rcu_idle_enter - inform RCU that current CPU is entering idle
391 * Enter idle mode, in other words, -leave- the mode in which RCU
392 * read-side critical sections can occur. (Though RCU read-side
393 * critical sections can occur in irq handlers in idle, a possibility
394 * handled by irq_enter() and irq_exit().)
396 * We crowbar the ->dynticks_nesting field to zero to allow for
397 * the possibility of usermode upcalls having messed up our count
398 * of interrupt nesting level during the prior busy period.
400 void rcu_idle_enter(void)
404 struct rcu_dynticks *rdtp;
406 local_irq_save(flags);
407 rdtp = &__get_cpu_var(rcu_dynticks);
408 oldval = rdtp->dynticks_nesting;
409 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
410 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
411 rdtp->dynticks_nesting = 0;
413 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
414 rcu_idle_enter_common(rdtp, oldval);
415 local_irq_restore(flags);
417 EXPORT_SYMBOL_GPL(rcu_idle_enter);
420 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
422 * Exit from an interrupt handler, which might possibly result in entering
423 * idle mode, in other words, leaving the mode in which read-side critical
424 * sections can occur.
426 * This code assumes that the idle loop never does anything that might
427 * result in unbalanced calls to irq_enter() and irq_exit(). If your
428 * architecture violates this assumption, RCU will give you what you
429 * deserve, good and hard. But very infrequently and irreproducibly.
431 * Use things like work queues to work around this limitation.
433 * You have been warned.
435 void rcu_irq_exit(void)
439 struct rcu_dynticks *rdtp;
441 local_irq_save(flags);
442 rdtp = &__get_cpu_var(rcu_dynticks);
443 oldval = rdtp->dynticks_nesting;
444 rdtp->dynticks_nesting--;
445 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
446 if (rdtp->dynticks_nesting)
447 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
449 rcu_idle_enter_common(rdtp, oldval);
450 local_irq_restore(flags);
454 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
456 * If the new value of the ->dynticks_nesting counter was previously zero,
457 * we really have exited idle, and must do the appropriate accounting.
458 * The caller must have disabled interrupts.
460 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
462 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
463 atomic_inc(&rdtp->dynticks);
464 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
465 smp_mb__after_atomic_inc(); /* See above. */
466 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
467 rcu_cleanup_after_idle(smp_processor_id());
468 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
469 if (!is_idle_task(current)) {
470 struct task_struct *idle = idle_task(smp_processor_id());
472 trace_rcu_dyntick("Error on exit: not idle task",
473 oldval, rdtp->dynticks_nesting);
474 ftrace_dump(DUMP_ORIG);
475 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
476 current->pid, current->comm,
477 idle->pid, idle->comm); /* must be idle task! */
482 * rcu_idle_exit - inform RCU that current CPU is leaving idle
484 * Exit idle mode, in other words, -enter- the mode in which RCU
485 * read-side critical sections can occur.
487 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
488 * allow for the possibility of usermode upcalls messing up our count
489 * of interrupt nesting level during the busy period that is just
492 void rcu_idle_exit(void)
495 struct rcu_dynticks *rdtp;
498 local_irq_save(flags);
499 rdtp = &__get_cpu_var(rcu_dynticks);
500 oldval = rdtp->dynticks_nesting;
501 WARN_ON_ONCE(oldval < 0);
502 if (oldval & DYNTICK_TASK_NEST_MASK)
503 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
505 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
506 rcu_idle_exit_common(rdtp, oldval);
507 local_irq_restore(flags);
509 EXPORT_SYMBOL_GPL(rcu_idle_exit);
512 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
514 * Enter an interrupt handler, which might possibly result in exiting
515 * idle mode, in other words, entering the mode in which read-side critical
516 * sections can occur.
518 * Note that the Linux kernel is fully capable of entering an interrupt
519 * handler that it never exits, for example when doing upcalls to
520 * user mode! This code assumes that the idle loop never does upcalls to
521 * user mode. If your architecture does do upcalls from the idle loop (or
522 * does anything else that results in unbalanced calls to the irq_enter()
523 * and irq_exit() functions), RCU will give you what you deserve, good
524 * and hard. But very infrequently and irreproducibly.
526 * Use things like work queues to work around this limitation.
528 * You have been warned.
530 void rcu_irq_enter(void)
533 struct rcu_dynticks *rdtp;
536 local_irq_save(flags);
537 rdtp = &__get_cpu_var(rcu_dynticks);
538 oldval = rdtp->dynticks_nesting;
539 rdtp->dynticks_nesting++;
540 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
542 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
544 rcu_idle_exit_common(rdtp, oldval);
545 local_irq_restore(flags);
549 * rcu_nmi_enter - inform RCU of entry to NMI context
551 * If the CPU was idle with dynamic ticks active, and there is no
552 * irq handler running, this updates rdtp->dynticks_nmi to let the
553 * RCU grace-period handling know that the CPU is active.
555 void rcu_nmi_enter(void)
557 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
559 if (rdtp->dynticks_nmi_nesting == 0 &&
560 (atomic_read(&rdtp->dynticks) & 0x1))
562 rdtp->dynticks_nmi_nesting++;
563 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
564 atomic_inc(&rdtp->dynticks);
565 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
566 smp_mb__after_atomic_inc(); /* See above. */
567 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
571 * rcu_nmi_exit - inform RCU of exit from NMI context
573 * If the CPU was idle with dynamic ticks active, and there is no
574 * irq handler running, this updates rdtp->dynticks_nmi to let the
575 * RCU grace-period handling know that the CPU is no longer active.
577 void rcu_nmi_exit(void)
579 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
581 if (rdtp->dynticks_nmi_nesting == 0 ||
582 --rdtp->dynticks_nmi_nesting != 0)
584 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
585 smp_mb__before_atomic_inc(); /* See above. */
586 atomic_inc(&rdtp->dynticks);
587 smp_mb__after_atomic_inc(); /* Force delay to next write. */
588 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
592 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
594 * If the current CPU is in its idle loop and is neither in an interrupt
595 * or NMI handler, return true.
597 int rcu_is_cpu_idle(void)
602 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
606 EXPORT_SYMBOL(rcu_is_cpu_idle);
608 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
611 * Is the current CPU online? Disable preemption to avoid false positives
612 * that could otherwise happen due to the current CPU number being sampled,
613 * this task being preempted, its old CPU being taken offline, resuming
614 * on some other CPU, then determining that its old CPU is now offline.
615 * It is OK to use RCU on an offline processor during initial boot, hence
616 * the check for rcu_scheduler_fully_active. Note also that it is OK
617 * for a CPU coming online to use RCU for one jiffy prior to marking itself
618 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
619 * offline to continue to use RCU for one jiffy after marking itself
620 * offline in the cpu_online_mask. This leniency is necessary given the
621 * non-atomic nature of the online and offline processing, for example,
622 * the fact that a CPU enters the scheduler after completing the CPU_DYING
625 * This is also why RCU internally marks CPUs online during the
626 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
628 * Disable checking if in an NMI handler because we cannot safely report
629 * errors from NMI handlers anyway.
631 bool rcu_lockdep_current_cpu_online(void)
633 struct rcu_data *rdp;
634 struct rcu_node *rnp;
640 rdp = &__get_cpu_var(rcu_sched_data);
642 ret = (rdp->grpmask & rnp->qsmaskinit) ||
643 !rcu_scheduler_fully_active;
647 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
649 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
652 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
654 * If the current CPU is idle or running at a first-level (not nested)
655 * interrupt from idle, return true. The caller must have at least
656 * disabled preemption.
658 int rcu_is_cpu_rrupt_from_idle(void)
660 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
664 * Snapshot the specified CPU's dynticks counter so that we can later
665 * credit them with an implicit quiescent state. Return 1 if this CPU
666 * is in dynticks idle mode, which is an extended quiescent state.
668 static int dyntick_save_progress_counter(struct rcu_data *rdp)
670 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
671 return (rdp->dynticks_snap & 0x1) == 0;
675 * Return true if the specified CPU has passed through a quiescent
676 * state by virtue of being in or having passed through an dynticks
677 * idle state since the last call to dyntick_save_progress_counter()
680 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
685 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
686 snap = (unsigned int)rdp->dynticks_snap;
689 * If the CPU passed through or entered a dynticks idle phase with
690 * no active irq/NMI handlers, then we can safely pretend that the CPU
691 * already acknowledged the request to pass through a quiescent
692 * state. Either way, that CPU cannot possibly be in an RCU
693 * read-side critical section that started before the beginning
694 * of the current RCU grace period.
696 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
697 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
702 /* Go check for the CPU being offline. */
703 return rcu_implicit_offline_qs(rdp);
706 static int jiffies_till_stall_check(void)
708 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
711 * Limit check must be consistent with the Kconfig limits
712 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
714 if (till_stall_check < 3) {
715 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
716 till_stall_check = 3;
717 } else if (till_stall_check > 300) {
718 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
719 till_stall_check = 300;
721 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
724 static void record_gp_stall_check_time(struct rcu_state *rsp)
726 rsp->gp_start = jiffies;
727 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
730 static void print_other_cpu_stall(struct rcu_state *rsp)
736 struct rcu_node *rnp = rcu_get_root(rsp);
738 /* Only let one CPU complain about others per time interval. */
740 raw_spin_lock_irqsave(&rnp->lock, flags);
741 delta = jiffies - rsp->jiffies_stall;
742 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
743 raw_spin_unlock_irqrestore(&rnp->lock, flags);
746 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
747 raw_spin_unlock_irqrestore(&rnp->lock, flags);
750 * OK, time to rat on our buddy...
751 * See Documentation/RCU/stallwarn.txt for info on how to debug
752 * RCU CPU stall warnings.
754 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
756 print_cpu_stall_info_begin();
757 rcu_for_each_leaf_node(rsp, rnp) {
758 raw_spin_lock_irqsave(&rnp->lock, flags);
759 ndetected += rcu_print_task_stall(rnp);
760 raw_spin_unlock_irqrestore(&rnp->lock, flags);
761 if (rnp->qsmask == 0)
763 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
764 if (rnp->qsmask & (1UL << cpu)) {
765 print_cpu_stall_info(rsp, rnp->grplo + cpu);
771 * Now rat on any tasks that got kicked up to the root rcu_node
772 * due to CPU offlining.
774 rnp = rcu_get_root(rsp);
775 raw_spin_lock_irqsave(&rnp->lock, flags);
776 ndetected += rcu_print_task_stall(rnp);
777 raw_spin_unlock_irqrestore(&rnp->lock, flags);
779 print_cpu_stall_info_end();
780 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
781 smp_processor_id(), (long)(jiffies - rsp->gp_start));
783 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
784 else if (!trigger_all_cpu_backtrace())
787 /* If so configured, complain about tasks blocking the grace period. */
789 rcu_print_detail_task_stall(rsp);
791 force_quiescent_state(rsp, 0); /* Kick them all. */
794 static void print_cpu_stall(struct rcu_state *rsp)
797 struct rcu_node *rnp = rcu_get_root(rsp);
800 * OK, time to rat on ourselves...
801 * See Documentation/RCU/stallwarn.txt for info on how to debug
802 * RCU CPU stall warnings.
804 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
805 print_cpu_stall_info_begin();
806 print_cpu_stall_info(rsp, smp_processor_id());
807 print_cpu_stall_info_end();
808 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
809 if (!trigger_all_cpu_backtrace())
812 raw_spin_lock_irqsave(&rnp->lock, flags);
813 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
814 rsp->jiffies_stall = jiffies +
815 3 * jiffies_till_stall_check() + 3;
816 raw_spin_unlock_irqrestore(&rnp->lock, flags);
818 set_need_resched(); /* kick ourselves to get things going. */
821 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
825 struct rcu_node *rnp;
827 if (rcu_cpu_stall_suppress)
829 j = ACCESS_ONCE(jiffies);
830 js = ACCESS_ONCE(rsp->jiffies_stall);
832 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
834 /* We haven't checked in, so go dump stack. */
835 print_cpu_stall(rsp);
837 } else if (rcu_gp_in_progress(rsp) &&
838 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
840 /* They had a few time units to dump stack, so complain. */
841 print_other_cpu_stall(rsp);
845 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
847 rcu_cpu_stall_suppress = 1;
852 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
854 * Set the stall-warning timeout way off into the future, thus preventing
855 * any RCU CPU stall-warning messages from appearing in the current set of
858 * The caller must disable hard irqs.
860 void rcu_cpu_stall_reset(void)
862 struct rcu_state *rsp;
864 for_each_rcu_flavor(rsp)
865 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
868 static struct notifier_block rcu_panic_block = {
869 .notifier_call = rcu_panic,
872 static void __init check_cpu_stall_init(void)
874 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
878 * Update CPU-local rcu_data state to record the newly noticed grace period.
879 * This is used both when we started the grace period and when we notice
880 * that someone else started the grace period. The caller must hold the
881 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
882 * and must have irqs disabled.
884 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
886 if (rdp->gpnum != rnp->gpnum) {
888 * If the current grace period is waiting for this CPU,
889 * set up to detect a quiescent state, otherwise don't
890 * go looking for one.
892 rdp->gpnum = rnp->gpnum;
893 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
894 if (rnp->qsmask & rdp->grpmask) {
896 rdp->passed_quiesce = 0;
900 zero_cpu_stall_ticks(rdp);
904 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
907 struct rcu_node *rnp;
909 local_irq_save(flags);
911 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
912 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
913 local_irq_restore(flags);
916 __note_new_gpnum(rsp, rnp, rdp);
917 raw_spin_unlock_irqrestore(&rnp->lock, flags);
921 * Did someone else start a new RCU grace period start since we last
922 * checked? Update local state appropriately if so. Must be called
923 * on the CPU corresponding to rdp.
926 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
931 local_irq_save(flags);
932 if (rdp->gpnum != rsp->gpnum) {
933 note_new_gpnum(rsp, rdp);
936 local_irq_restore(flags);
941 * Initialize the specified rcu_data structure's callback list to empty.
943 static void init_callback_list(struct rcu_data *rdp)
948 for (i = 0; i < RCU_NEXT_SIZE; i++)
949 rdp->nxttail[i] = &rdp->nxtlist;
953 * Advance this CPU's callbacks, but only if the current grace period
954 * has ended. This may be called only from the CPU to whom the rdp
955 * belongs. In addition, the corresponding leaf rcu_node structure's
956 * ->lock must be held by the caller, with irqs disabled.
959 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
961 /* Did another grace period end? */
962 if (rdp->completed != rnp->completed) {
964 /* Advance callbacks. No harm if list empty. */
965 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
966 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
967 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
969 /* Remember that we saw this grace-period completion. */
970 rdp->completed = rnp->completed;
971 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
974 * If we were in an extended quiescent state, we may have
975 * missed some grace periods that others CPUs handled on
976 * our behalf. Catch up with this state to avoid noting
977 * spurious new grace periods. If another grace period
978 * has started, then rnp->gpnum will have advanced, so
979 * we will detect this later on.
981 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
982 rdp->gpnum = rdp->completed;
985 * If RCU does not need a quiescent state from this CPU,
986 * then make sure that this CPU doesn't go looking for one.
988 if ((rnp->qsmask & rdp->grpmask) == 0)
994 * Advance this CPU's callbacks, but only if the current grace period
995 * has ended. This may be called only from the CPU to whom the rdp
999 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1001 unsigned long flags;
1002 struct rcu_node *rnp;
1004 local_irq_save(flags);
1006 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1007 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1008 local_irq_restore(flags);
1011 __rcu_process_gp_end(rsp, rnp, rdp);
1012 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1016 * Do per-CPU grace-period initialization for running CPU. The caller
1017 * must hold the lock of the leaf rcu_node structure corresponding to
1021 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1023 /* Prior grace period ended, so advance callbacks for current CPU. */
1024 __rcu_process_gp_end(rsp, rnp, rdp);
1026 /* Set state so that this CPU will detect the next quiescent state. */
1027 __note_new_gpnum(rsp, rnp, rdp);
1031 * Initialize a new grace period.
1033 static int rcu_gp_init(struct rcu_state *rsp)
1035 struct rcu_data *rdp;
1036 struct rcu_node *rnp = rcu_get_root(rsp);
1038 raw_spin_lock_irq(&rnp->lock);
1041 if (rcu_gp_in_progress(rsp)) {
1042 /* Grace period already in progress, don't start another. */
1043 raw_spin_unlock_irq(&rnp->lock);
1047 if (rsp->fqs_active) {
1049 * We need a grace period, but force_quiescent_state()
1050 * is running. Tell it to start one on our behalf.
1052 rsp->fqs_need_gp = 1;
1053 raw_spin_unlock_irq(&rnp->lock);
1057 /* Advance to a new grace period and initialize state. */
1059 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1060 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1061 rsp->fqs_state = RCU_GP_INIT; /* Stop force_quiescent_state. */
1062 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1063 record_gp_stall_check_time(rsp);
1064 raw_spin_unlock_irq(&rnp->lock);
1066 /* Exclude any concurrent CPU-hotplug operations. */
1070 * Set the quiescent-state-needed bits in all the rcu_node
1071 * structures for all currently online CPUs in breadth-first order,
1072 * starting from the root rcu_node structure, relying on the layout
1073 * of the tree within the rsp->node[] array. Note that other CPUs
1074 * will access only the leaves of the hierarchy, thus seeing that no
1075 * grace period is in progress, at least until the corresponding
1076 * leaf node has been initialized. In addition, we have excluded
1077 * CPU-hotplug operations.
1079 * The grace period cannot complete until the initialization
1080 * process finishes, because this kthread handles both.
1082 rcu_for_each_node_breadth_first(rsp, rnp) {
1083 raw_spin_lock_irq(&rnp->lock);
1084 rdp = this_cpu_ptr(rsp->rda);
1085 rcu_preempt_check_blocked_tasks(rnp);
1086 rnp->qsmask = rnp->qsmaskinit;
1087 rnp->gpnum = rsp->gpnum;
1088 rnp->completed = rsp->completed;
1089 if (rnp == rdp->mynode)
1090 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1091 rcu_preempt_boost_start_gp(rnp);
1092 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1093 rnp->level, rnp->grplo,
1094 rnp->grphi, rnp->qsmask);
1095 raw_spin_unlock_irq(&rnp->lock);
1099 rnp = rcu_get_root(rsp);
1100 raw_spin_lock_irq(&rnp->lock);
1101 /* force_quiescent_state() now OK. */
1102 rsp->fqs_state = RCU_SIGNAL_INIT;
1103 raw_spin_unlock_irq(&rnp->lock);
1109 * Clean up after the old grace period.
1111 static int rcu_gp_cleanup(struct rcu_state *rsp)
1113 unsigned long gp_duration;
1114 struct rcu_data *rdp;
1115 struct rcu_node *rnp = rcu_get_root(rsp);
1117 raw_spin_lock_irq(&rnp->lock);
1118 gp_duration = jiffies - rsp->gp_start;
1119 if (gp_duration > rsp->gp_max)
1120 rsp->gp_max = gp_duration;
1123 * We know the grace period is complete, but to everyone else
1124 * it appears to still be ongoing. But it is also the case
1125 * that to everyone else it looks like there is nothing that
1126 * they can do to advance the grace period. It is therefore
1127 * safe for us to drop the lock in order to mark the grace
1128 * period as completed in all of the rcu_node structures.
1130 * But if this CPU needs another grace period, it will take
1131 * care of this while initializing the next grace period.
1132 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1133 * because the callbacks have not yet been advanced: Those
1134 * callbacks are waiting on the grace period that just now
1137 rdp = this_cpu_ptr(rsp->rda);
1138 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1139 raw_spin_unlock_irq(&rnp->lock);
1142 * Propagate new ->completed value to rcu_node
1143 * structures so that other CPUs don't have to
1144 * wait until the start of the next grace period
1145 * to process their callbacks.
1147 rcu_for_each_node_breadth_first(rsp, rnp) {
1148 raw_spin_lock_irq(&rnp->lock);
1149 rnp->completed = rsp->gpnum;
1150 raw_spin_unlock_irq(&rnp->lock);
1153 rnp = rcu_get_root(rsp);
1154 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 if (cpu_needs_another_gp(rsp, rdp))
1162 raw_spin_unlock_irq(&rnp->lock);
1167 * Body of kthread that handles grace periods.
1169 static int __noreturn rcu_gp_kthread(void *arg)
1171 struct rcu_state *rsp = arg;
1172 struct rcu_node *rnp = rcu_get_root(rsp);
1176 /* Handle grace-period start. */
1178 wait_event_interruptible(rsp->gp_wq, rsp->gp_flags);
1179 if (rsp->gp_flags && rcu_gp_init(rsp))
1182 flush_signals(current);
1185 /* Handle grace-period end. */
1186 rnp = rcu_get_root(rsp);
1188 wait_event_interruptible(rsp->gp_wq,
1189 !ACCESS_ONCE(rnp->qsmask) &&
1190 !rcu_preempt_blocked_readers_cgp(rnp));
1191 if (!ACCESS_ONCE(rnp->qsmask) &&
1192 !rcu_preempt_blocked_readers_cgp(rnp) &&
1193 rcu_gp_cleanup(rsp))
1196 flush_signals(current);
1202 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1203 * in preparation for detecting the next grace period. The caller must hold
1204 * the root node's ->lock, which is released before return. Hard irqs must
1207 * Note that it is legal for a dying CPU (which is marked as offline) to
1208 * invoke this function. This can happen when the dying CPU reports its
1212 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1213 __releases(rcu_get_root(rsp)->lock)
1215 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1216 struct rcu_node *rnp = rcu_get_root(rsp);
1218 if (!rsp->gp_kthread ||
1219 !cpu_needs_another_gp(rsp, rdp)) {
1221 * Either we have not yet spawned the grace-period
1222 * task or this CPU does not need another grace period.
1223 * Either way, don't start a new grace period.
1225 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1230 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1231 wake_up(&rsp->gp_wq);
1235 * Report a full set of quiescent states to the specified rcu_state
1236 * data structure. This involves cleaning up after the prior grace
1237 * period and letting rcu_start_gp() start up the next grace period
1238 * if one is needed. Note that the caller must hold rnp->lock, as
1239 * required by rcu_start_gp(), which will release it.
1241 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1242 __releases(rcu_get_root(rsp)->lock)
1244 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1245 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1246 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1250 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1251 * Allows quiescent states for a group of CPUs to be reported at one go
1252 * to the specified rcu_node structure, though all the CPUs in the group
1253 * must be represented by the same rcu_node structure (which need not be
1254 * a leaf rcu_node structure, though it often will be). That structure's
1255 * lock must be held upon entry, and it is released before return.
1258 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1259 struct rcu_node *rnp, unsigned long flags)
1260 __releases(rnp->lock)
1262 struct rcu_node *rnp_c;
1264 /* Walk up the rcu_node hierarchy. */
1266 if (!(rnp->qsmask & mask)) {
1268 /* Our bit has already been cleared, so done. */
1269 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1272 rnp->qsmask &= ~mask;
1273 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1274 mask, rnp->qsmask, rnp->level,
1275 rnp->grplo, rnp->grphi,
1277 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1279 /* Other bits still set at this level, so done. */
1280 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1283 mask = rnp->grpmask;
1284 if (rnp->parent == NULL) {
1286 /* No more levels. Exit loop holding root lock. */
1290 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1293 raw_spin_lock_irqsave(&rnp->lock, flags);
1294 WARN_ON_ONCE(rnp_c->qsmask);
1298 * Get here if we are the last CPU to pass through a quiescent
1299 * state for this grace period. Invoke rcu_report_qs_rsp()
1300 * to clean up and start the next grace period if one is needed.
1302 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1306 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1307 * structure. This must be either called from the specified CPU, or
1308 * called when the specified CPU is known to be offline (and when it is
1309 * also known that no other CPU is concurrently trying to help the offline
1310 * CPU). The lastcomp argument is used to make sure we are still in the
1311 * grace period of interest. We don't want to end the current grace period
1312 * based on quiescent states detected in an earlier grace period!
1315 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1317 unsigned long flags;
1319 struct rcu_node *rnp;
1322 raw_spin_lock_irqsave(&rnp->lock, flags);
1323 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1326 * The grace period in which this quiescent state was
1327 * recorded has ended, so don't report it upwards.
1328 * We will instead need a new quiescent state that lies
1329 * within the current grace period.
1331 rdp->passed_quiesce = 0; /* need qs for new gp. */
1332 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1335 mask = rdp->grpmask;
1336 if ((rnp->qsmask & mask) == 0) {
1337 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1339 rdp->qs_pending = 0;
1342 * This GP can't end until cpu checks in, so all of our
1343 * callbacks can be processed during the next GP.
1345 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1347 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1352 * Check to see if there is a new grace period of which this CPU
1353 * is not yet aware, and if so, set up local rcu_data state for it.
1354 * Otherwise, see if this CPU has just passed through its first
1355 * quiescent state for this grace period, and record that fact if so.
1358 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1360 /* If there is now a new grace period, record and return. */
1361 if (check_for_new_grace_period(rsp, rdp))
1365 * Does this CPU still need to do its part for current grace period?
1366 * If no, return and let the other CPUs do their part as well.
1368 if (!rdp->qs_pending)
1372 * Was there a quiescent state since the beginning of the grace
1373 * period? If no, then exit and wait for the next call.
1375 if (!rdp->passed_quiesce)
1379 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1382 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1385 #ifdef CONFIG_HOTPLUG_CPU
1388 * Send the specified CPU's RCU callbacks to the orphanage. The
1389 * specified CPU must be offline, and the caller must hold the
1393 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1394 struct rcu_node *rnp, struct rcu_data *rdp)
1397 * Orphan the callbacks. First adjust the counts. This is safe
1398 * because ->onofflock excludes _rcu_barrier()'s adoption of
1399 * the callbacks, thus no memory barrier is required.
1401 if (rdp->nxtlist != NULL) {
1402 rsp->qlen_lazy += rdp->qlen_lazy;
1403 rsp->qlen += rdp->qlen;
1404 rdp->n_cbs_orphaned += rdp->qlen;
1406 ACCESS_ONCE(rdp->qlen) = 0;
1410 * Next, move those callbacks still needing a grace period to
1411 * the orphanage, where some other CPU will pick them up.
1412 * Some of the callbacks might have gone partway through a grace
1413 * period, but that is too bad. They get to start over because we
1414 * cannot assume that grace periods are synchronized across CPUs.
1415 * We don't bother updating the ->nxttail[] array yet, instead
1416 * we just reset the whole thing later on.
1418 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1419 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1420 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1421 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1425 * Then move the ready-to-invoke callbacks to the orphanage,
1426 * where some other CPU will pick them up. These will not be
1427 * required to pass though another grace period: They are done.
1429 if (rdp->nxtlist != NULL) {
1430 *rsp->orphan_donetail = rdp->nxtlist;
1431 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1434 /* Finally, initialize the rcu_data structure's list to empty. */
1435 init_callback_list(rdp);
1439 * Adopt the RCU callbacks from the specified rcu_state structure's
1440 * orphanage. The caller must hold the ->onofflock.
1442 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1445 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1448 * If there is an rcu_barrier() operation in progress, then
1449 * only the task doing that operation is permitted to adopt
1450 * callbacks. To do otherwise breaks rcu_barrier() and friends
1451 * by causing them to fail to wait for the callbacks in the
1454 if (rsp->rcu_barrier_in_progress &&
1455 rsp->rcu_barrier_in_progress != current)
1458 /* Do the accounting first. */
1459 rdp->qlen_lazy += rsp->qlen_lazy;
1460 rdp->qlen += rsp->qlen;
1461 rdp->n_cbs_adopted += rsp->qlen;
1462 if (rsp->qlen_lazy != rsp->qlen)
1463 rcu_idle_count_callbacks_posted();
1468 * We do not need a memory barrier here because the only way we
1469 * can get here if there is an rcu_barrier() in flight is if
1470 * we are the task doing the rcu_barrier().
1473 /* First adopt the ready-to-invoke callbacks. */
1474 if (rsp->orphan_donelist != NULL) {
1475 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1476 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1477 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1478 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1479 rdp->nxttail[i] = rsp->orphan_donetail;
1480 rsp->orphan_donelist = NULL;
1481 rsp->orphan_donetail = &rsp->orphan_donelist;
1484 /* And then adopt the callbacks that still need a grace period. */
1485 if (rsp->orphan_nxtlist != NULL) {
1486 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1487 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1488 rsp->orphan_nxtlist = NULL;
1489 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1494 * Trace the fact that this CPU is going offline.
1496 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1498 RCU_TRACE(unsigned long mask);
1499 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1500 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1502 RCU_TRACE(mask = rdp->grpmask);
1503 trace_rcu_grace_period(rsp->name,
1504 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1509 * The CPU has been completely removed, and some other CPU is reporting
1510 * this fact from process context. Do the remainder of the cleanup,
1511 * including orphaning the outgoing CPU's RCU callbacks, and also
1512 * adopting them, if there is no _rcu_barrier() instance running.
1513 * There can only be one CPU hotplug operation at a time, so no other
1514 * CPU can be attempting to update rcu_cpu_kthread_task.
1516 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1518 unsigned long flags;
1520 int need_report = 0;
1521 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1522 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1524 /* Adjust any no-longer-needed kthreads. */
1525 rcu_stop_cpu_kthread(cpu);
1526 rcu_node_kthread_setaffinity(rnp, -1);
1528 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1530 /* Exclude any attempts to start a new grace period. */
1531 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1533 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1534 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1535 rcu_adopt_orphan_cbs(rsp);
1537 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1538 mask = rdp->grpmask; /* rnp->grplo is constant. */
1540 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1541 rnp->qsmaskinit &= ~mask;
1542 if (rnp->qsmaskinit != 0) {
1543 if (rnp != rdp->mynode)
1544 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1547 if (rnp == rdp->mynode)
1548 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1550 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1551 mask = rnp->grpmask;
1553 } while (rnp != NULL);
1556 * We still hold the leaf rcu_node structure lock here, and
1557 * irqs are still disabled. The reason for this subterfuge is
1558 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1559 * held leads to deadlock.
1561 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1563 if (need_report & RCU_OFL_TASKS_NORM_GP)
1564 rcu_report_unblock_qs_rnp(rnp, flags);
1566 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1567 if (need_report & RCU_OFL_TASKS_EXP_GP)
1568 rcu_report_exp_rnp(rsp, rnp, true);
1569 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1570 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1571 cpu, rdp->qlen, rdp->nxtlist);
1574 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1576 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1580 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1584 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1588 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1591 * Invoke any RCU callbacks that have made it to the end of their grace
1592 * period. Thottle as specified by rdp->blimit.
1594 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1596 unsigned long flags;
1597 struct rcu_head *next, *list, **tail;
1598 int bl, count, count_lazy, i;
1600 /* If no callbacks are ready, just return.*/
1601 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1602 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1603 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1604 need_resched(), is_idle_task(current),
1605 rcu_is_callbacks_kthread());
1610 * Extract the list of ready callbacks, disabling to prevent
1611 * races with call_rcu() from interrupt handlers.
1613 local_irq_save(flags);
1614 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1616 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1617 list = rdp->nxtlist;
1618 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1619 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1620 tail = rdp->nxttail[RCU_DONE_TAIL];
1621 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1622 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1623 rdp->nxttail[i] = &rdp->nxtlist;
1624 local_irq_restore(flags);
1626 /* Invoke callbacks. */
1627 count = count_lazy = 0;
1631 debug_rcu_head_unqueue(list);
1632 if (__rcu_reclaim(rsp->name, list))
1635 /* Stop only if limit reached and CPU has something to do. */
1636 if (++count >= bl &&
1638 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1642 local_irq_save(flags);
1643 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1644 is_idle_task(current),
1645 rcu_is_callbacks_kthread());
1647 /* Update count, and requeue any remaining callbacks. */
1649 *tail = rdp->nxtlist;
1650 rdp->nxtlist = list;
1651 for (i = 0; i < RCU_NEXT_SIZE; i++)
1652 if (&rdp->nxtlist == rdp->nxttail[i])
1653 rdp->nxttail[i] = tail;
1657 smp_mb(); /* List handling before counting for rcu_barrier(). */
1658 rdp->qlen_lazy -= count_lazy;
1659 ACCESS_ONCE(rdp->qlen) -= count;
1660 rdp->n_cbs_invoked += count;
1662 /* Reinstate batch limit if we have worked down the excess. */
1663 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1664 rdp->blimit = blimit;
1666 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1667 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1668 rdp->qlen_last_fqs_check = 0;
1669 rdp->n_force_qs_snap = rsp->n_force_qs;
1670 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1671 rdp->qlen_last_fqs_check = rdp->qlen;
1672 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1674 local_irq_restore(flags);
1676 /* Re-invoke RCU core processing if there are callbacks remaining. */
1677 if (cpu_has_callbacks_ready_to_invoke(rdp))
1682 * Check to see if this CPU is in a non-context-switch quiescent state
1683 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1684 * Also schedule RCU core processing.
1686 * This function must be called from hardirq context. It is normally
1687 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1688 * false, there is no point in invoking rcu_check_callbacks().
1690 void rcu_check_callbacks(int cpu, int user)
1692 trace_rcu_utilization("Start scheduler-tick");
1693 increment_cpu_stall_ticks();
1694 if (user || rcu_is_cpu_rrupt_from_idle()) {
1697 * Get here if this CPU took its interrupt from user
1698 * mode or from the idle loop, and if this is not a
1699 * nested interrupt. In this case, the CPU is in
1700 * a quiescent state, so note it.
1702 * No memory barrier is required here because both
1703 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1704 * variables that other CPUs neither access nor modify,
1705 * at least not while the corresponding CPU is online.
1711 } else if (!in_softirq()) {
1714 * Get here if this CPU did not take its interrupt from
1715 * softirq, in other words, if it is not interrupting
1716 * a rcu_bh read-side critical section. This is an _bh
1717 * critical section, so note it.
1722 rcu_preempt_check_callbacks(cpu);
1723 if (rcu_pending(cpu))
1725 trace_rcu_utilization("End scheduler-tick");
1729 * Scan the leaf rcu_node structures, processing dyntick state for any that
1730 * have not yet encountered a quiescent state, using the function specified.
1731 * Also initiate boosting for any threads blocked on the root rcu_node.
1733 * The caller must have suppressed start of new grace periods.
1735 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1739 unsigned long flags;
1741 struct rcu_node *rnp;
1743 rcu_for_each_leaf_node(rsp, rnp) {
1745 raw_spin_lock_irqsave(&rnp->lock, flags);
1746 if (!rcu_gp_in_progress(rsp)) {
1747 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1750 if (rnp->qsmask == 0) {
1751 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1756 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1757 if ((rnp->qsmask & bit) != 0 &&
1758 f(per_cpu_ptr(rsp->rda, cpu)))
1763 /* rcu_report_qs_rnp() releases rnp->lock. */
1764 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1767 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1769 rnp = rcu_get_root(rsp);
1770 if (rnp->qsmask == 0) {
1771 raw_spin_lock_irqsave(&rnp->lock, flags);
1772 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1777 * Force quiescent states on reluctant CPUs, and also detect which
1778 * CPUs are in dyntick-idle mode.
1780 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1782 unsigned long flags;
1783 struct rcu_node *rnp = rcu_get_root(rsp);
1785 trace_rcu_utilization("Start fqs");
1786 if (!rcu_gp_in_progress(rsp)) {
1787 trace_rcu_utilization("End fqs");
1788 return; /* No grace period in progress, nothing to force. */
1790 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1791 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1792 trace_rcu_utilization("End fqs");
1793 return; /* Someone else is already on the job. */
1795 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1796 goto unlock_fqs_ret; /* no emergency and done recently. */
1798 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1799 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1800 if(!rcu_gp_in_progress(rsp)) {
1801 rsp->n_force_qs_ngp++;
1802 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1803 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1805 rsp->fqs_active = 1;
1806 switch (rsp->fqs_state) {
1810 break; /* grace period idle or initializing, ignore. */
1812 case RCU_SAVE_DYNTICK:
1814 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1816 /* Record dyntick-idle state. */
1817 force_qs_rnp(rsp, dyntick_save_progress_counter);
1818 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1819 if (rcu_gp_in_progress(rsp))
1820 rsp->fqs_state = RCU_FORCE_QS;
1825 /* Check dyntick-idle state, send IPI to laggarts. */
1826 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1827 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1829 /* Leave state in case more forcing is required. */
1831 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1834 rsp->fqs_active = 0;
1835 if (rsp->fqs_need_gp) {
1836 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1837 rsp->fqs_need_gp = 0;
1838 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1839 trace_rcu_utilization("End fqs");
1842 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1844 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1845 trace_rcu_utilization("End fqs");
1849 * This does the RCU core processing work for the specified rcu_state
1850 * and rcu_data structures. This may be called only from the CPU to
1851 * whom the rdp belongs.
1854 __rcu_process_callbacks(struct rcu_state *rsp)
1856 unsigned long flags;
1857 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1859 WARN_ON_ONCE(rdp->beenonline == 0);
1862 * If an RCU GP has gone long enough, go check for dyntick
1863 * idle CPUs and, if needed, send resched IPIs.
1865 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1866 force_quiescent_state(rsp, 1);
1869 * Advance callbacks in response to end of earlier grace
1870 * period that some other CPU ended.
1872 rcu_process_gp_end(rsp, rdp);
1874 /* Update RCU state based on any recent quiescent states. */
1875 rcu_check_quiescent_state(rsp, rdp);
1877 /* Does this CPU require a not-yet-started grace period? */
1878 if (cpu_needs_another_gp(rsp, rdp)) {
1879 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1880 rcu_start_gp(rsp, flags); /* releases above lock */
1883 /* If there are callbacks ready, invoke them. */
1884 if (cpu_has_callbacks_ready_to_invoke(rdp))
1885 invoke_rcu_callbacks(rsp, rdp);
1889 * Do RCU core processing for the current CPU.
1891 static void rcu_process_callbacks(struct softirq_action *unused)
1893 struct rcu_state *rsp;
1895 trace_rcu_utilization("Start RCU core");
1896 for_each_rcu_flavor(rsp)
1897 __rcu_process_callbacks(rsp);
1898 trace_rcu_utilization("End RCU core");
1902 * Schedule RCU callback invocation. If the specified type of RCU
1903 * does not support RCU priority boosting, just do a direct call,
1904 * otherwise wake up the per-CPU kernel kthread. Note that because we
1905 * are running on the current CPU with interrupts disabled, the
1906 * rcu_cpu_kthread_task cannot disappear out from under us.
1908 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1910 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1912 if (likely(!rsp->boost)) {
1913 rcu_do_batch(rsp, rdp);
1916 invoke_rcu_callbacks_kthread();
1919 static void invoke_rcu_core(void)
1921 raise_softirq(RCU_SOFTIRQ);
1925 * Handle any core-RCU processing required by a call_rcu() invocation.
1927 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1928 struct rcu_head *head, unsigned long flags)
1931 * If called from an extended quiescent state, invoke the RCU
1932 * core in order to force a re-evaluation of RCU's idleness.
1934 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1937 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1938 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1942 * Force the grace period if too many callbacks or too long waiting.
1943 * Enforce hysteresis, and don't invoke force_quiescent_state()
1944 * if some other CPU has recently done so. Also, don't bother
1945 * invoking force_quiescent_state() if the newly enqueued callback
1946 * is the only one waiting for a grace period to complete.
1948 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1950 /* Are we ignoring a completed grace period? */
1951 rcu_process_gp_end(rsp, rdp);
1952 check_for_new_grace_period(rsp, rdp);
1954 /* Start a new grace period if one not already started. */
1955 if (!rcu_gp_in_progress(rsp)) {
1956 unsigned long nestflag;
1957 struct rcu_node *rnp_root = rcu_get_root(rsp);
1959 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1960 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1962 /* Give the grace period a kick. */
1963 rdp->blimit = LONG_MAX;
1964 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1965 *rdp->nxttail[RCU_DONE_TAIL] != head)
1966 force_quiescent_state(rsp, 0);
1967 rdp->n_force_qs_snap = rsp->n_force_qs;
1968 rdp->qlen_last_fqs_check = rdp->qlen;
1970 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1971 force_quiescent_state(rsp, 1);
1975 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1976 struct rcu_state *rsp, bool lazy)
1978 unsigned long flags;
1979 struct rcu_data *rdp;
1981 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1982 debug_rcu_head_queue(head);
1986 smp_mb(); /* Ensure RCU update seen before callback registry. */
1989 * Opportunistically note grace-period endings and beginnings.
1990 * Note that we might see a beginning right after we see an
1991 * end, but never vice versa, since this CPU has to pass through
1992 * a quiescent state betweentimes.
1994 local_irq_save(flags);
1995 rdp = this_cpu_ptr(rsp->rda);
1997 /* Add the callback to our list. */
1998 ACCESS_ONCE(rdp->qlen)++;
2002 rcu_idle_count_callbacks_posted();
2003 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2004 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2005 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2007 if (__is_kfree_rcu_offset((unsigned long)func))
2008 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2009 rdp->qlen_lazy, rdp->qlen);
2011 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2013 /* Go handle any RCU core processing required. */
2014 __call_rcu_core(rsp, rdp, head, flags);
2015 local_irq_restore(flags);
2019 * Queue an RCU-sched callback for invocation after a grace period.
2021 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2023 __call_rcu(head, func, &rcu_sched_state, 0);
2025 EXPORT_SYMBOL_GPL(call_rcu_sched);
2028 * Queue an RCU callback for invocation after a quicker grace period.
2030 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2032 __call_rcu(head, func, &rcu_bh_state, 0);
2034 EXPORT_SYMBOL_GPL(call_rcu_bh);
2037 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2038 * any blocking grace-period wait automatically implies a grace period
2039 * if there is only one CPU online at any point time during execution
2040 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2041 * occasionally incorrectly indicate that there are multiple CPUs online
2042 * when there was in fact only one the whole time, as this just adds
2043 * some overhead: RCU still operates correctly.
2045 static inline int rcu_blocking_is_gp(void)
2049 might_sleep(); /* Check for RCU read-side critical section. */
2051 ret = num_online_cpus() <= 1;
2057 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2059 * Control will return to the caller some time after a full rcu-sched
2060 * grace period has elapsed, in other words after all currently executing
2061 * rcu-sched read-side critical sections have completed. These read-side
2062 * critical sections are delimited by rcu_read_lock_sched() and
2063 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2064 * local_irq_disable(), and so on may be used in place of
2065 * rcu_read_lock_sched().
2067 * This means that all preempt_disable code sequences, including NMI and
2068 * hardware-interrupt handlers, in progress on entry will have completed
2069 * before this primitive returns. However, this does not guarantee that
2070 * softirq handlers will have completed, since in some kernels, these
2071 * handlers can run in process context, and can block.
2073 * This primitive provides the guarantees made by the (now removed)
2074 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2075 * guarantees that rcu_read_lock() sections will have completed.
2076 * In "classic RCU", these two guarantees happen to be one and
2077 * the same, but can differ in realtime RCU implementations.
2079 void synchronize_sched(void)
2081 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2082 !lock_is_held(&rcu_lock_map) &&
2083 !lock_is_held(&rcu_sched_lock_map),
2084 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2085 if (rcu_blocking_is_gp())
2087 wait_rcu_gp(call_rcu_sched);
2089 EXPORT_SYMBOL_GPL(synchronize_sched);
2092 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2094 * Control will return to the caller some time after a full rcu_bh grace
2095 * period has elapsed, in other words after all currently executing rcu_bh
2096 * read-side critical sections have completed. RCU read-side critical
2097 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2098 * and may be nested.
2100 void synchronize_rcu_bh(void)
2102 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2103 !lock_is_held(&rcu_lock_map) &&
2104 !lock_is_held(&rcu_sched_lock_map),
2105 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2106 if (rcu_blocking_is_gp())
2108 wait_rcu_gp(call_rcu_bh);
2110 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2112 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2113 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2115 static int synchronize_sched_expedited_cpu_stop(void *data)
2118 * There must be a full memory barrier on each affected CPU
2119 * between the time that try_stop_cpus() is called and the
2120 * time that it returns.
2122 * In the current initial implementation of cpu_stop, the
2123 * above condition is already met when the control reaches
2124 * this point and the following smp_mb() is not strictly
2125 * necessary. Do smp_mb() anyway for documentation and
2126 * robustness against future implementation changes.
2128 smp_mb(); /* See above comment block. */
2133 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2135 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2136 * approach to force the grace period to end quickly. This consumes
2137 * significant time on all CPUs and is unfriendly to real-time workloads,
2138 * so is thus not recommended for any sort of common-case code. In fact,
2139 * if you are using synchronize_sched_expedited() in a loop, please
2140 * restructure your code to batch your updates, and then use a single
2141 * synchronize_sched() instead.
2143 * Note that it is illegal to call this function while holding any lock
2144 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2145 * to call this function from a CPU-hotplug notifier. Failing to observe
2146 * these restriction will result in deadlock.
2148 * This implementation can be thought of as an application of ticket
2149 * locking to RCU, with sync_sched_expedited_started and
2150 * sync_sched_expedited_done taking on the roles of the halves
2151 * of the ticket-lock word. Each task atomically increments
2152 * sync_sched_expedited_started upon entry, snapshotting the old value,
2153 * then attempts to stop all the CPUs. If this succeeds, then each
2154 * CPU will have executed a context switch, resulting in an RCU-sched
2155 * grace period. We are then done, so we use atomic_cmpxchg() to
2156 * update sync_sched_expedited_done to match our snapshot -- but
2157 * only if someone else has not already advanced past our snapshot.
2159 * On the other hand, if try_stop_cpus() fails, we check the value
2160 * of sync_sched_expedited_done. If it has advanced past our
2161 * initial snapshot, then someone else must have forced a grace period
2162 * some time after we took our snapshot. In this case, our work is
2163 * done for us, and we can simply return. Otherwise, we try again,
2164 * but keep our initial snapshot for purposes of checking for someone
2165 * doing our work for us.
2167 * If we fail too many times in a row, we fall back to synchronize_sched().
2169 void synchronize_sched_expedited(void)
2171 int firstsnap, s, snap, trycount = 0;
2173 /* Note that atomic_inc_return() implies full memory barrier. */
2174 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2176 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2179 * Each pass through the following loop attempts to force a
2180 * context switch on each CPU.
2182 while (try_stop_cpus(cpu_online_mask,
2183 synchronize_sched_expedited_cpu_stop,
2187 /* No joy, try again later. Or just synchronize_sched(). */
2188 if (trycount++ < 10) {
2189 udelay(trycount * num_online_cpus());
2191 synchronize_sched();
2195 /* Check to see if someone else did our work for us. */
2196 s = atomic_read(&sync_sched_expedited_done);
2197 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2198 smp_mb(); /* ensure test happens before caller kfree */
2203 * Refetching sync_sched_expedited_started allows later
2204 * callers to piggyback on our grace period. We subtract
2205 * 1 to get the same token that the last incrementer got.
2206 * We retry after they started, so our grace period works
2207 * for them, and they started after our first try, so their
2208 * grace period works for us.
2211 snap = atomic_read(&sync_sched_expedited_started);
2212 smp_mb(); /* ensure read is before try_stop_cpus(). */
2216 * Everyone up to our most recent fetch is covered by our grace
2217 * period. Update the counter, but only if our work is still
2218 * relevant -- which it won't be if someone who started later
2219 * than we did beat us to the punch.
2222 s = atomic_read(&sync_sched_expedited_done);
2223 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2224 smp_mb(); /* ensure test happens before caller kfree */
2227 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2231 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2234 * Check to see if there is any immediate RCU-related work to be done
2235 * by the current CPU, for the specified type of RCU, returning 1 if so.
2236 * The checks are in order of increasing expense: checks that can be
2237 * carried out against CPU-local state are performed first. However,
2238 * we must check for CPU stalls first, else we might not get a chance.
2240 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2242 struct rcu_node *rnp = rdp->mynode;
2244 rdp->n_rcu_pending++;
2246 /* Check for CPU stalls, if enabled. */
2247 check_cpu_stall(rsp, rdp);
2249 /* Is the RCU core waiting for a quiescent state from this CPU? */
2250 if (rcu_scheduler_fully_active &&
2251 rdp->qs_pending && !rdp->passed_quiesce) {
2254 * If force_quiescent_state() coming soon and this CPU
2255 * needs a quiescent state, and this is either RCU-sched
2256 * or RCU-bh, force a local reschedule.
2258 rdp->n_rp_qs_pending++;
2259 if (!rdp->preemptible &&
2260 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2263 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2264 rdp->n_rp_report_qs++;
2268 /* Does this CPU have callbacks ready to invoke? */
2269 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2270 rdp->n_rp_cb_ready++;
2274 /* Has RCU gone idle with this CPU needing another grace period? */
2275 if (cpu_needs_another_gp(rsp, rdp)) {
2276 rdp->n_rp_cpu_needs_gp++;
2280 /* Has another RCU grace period completed? */
2281 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2282 rdp->n_rp_gp_completed++;
2286 /* Has a new RCU grace period started? */
2287 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2288 rdp->n_rp_gp_started++;
2292 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2293 if (rcu_gp_in_progress(rsp) &&
2294 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2295 rdp->n_rp_need_fqs++;
2300 rdp->n_rp_need_nothing++;
2305 * Check to see if there is any immediate RCU-related work to be done
2306 * by the current CPU, returning 1 if so. This function is part of the
2307 * RCU implementation; it is -not- an exported member of the RCU API.
2309 static int rcu_pending(int cpu)
2311 struct rcu_state *rsp;
2313 for_each_rcu_flavor(rsp)
2314 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2320 * Check to see if any future RCU-related work will need to be done
2321 * by the current CPU, even if none need be done immediately, returning
2324 static int rcu_cpu_has_callbacks(int cpu)
2326 struct rcu_state *rsp;
2328 /* RCU callbacks either ready or pending? */
2329 for_each_rcu_flavor(rsp)
2330 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2336 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2337 * the compiler is expected to optimize this away.
2339 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2340 int cpu, unsigned long done)
2342 trace_rcu_barrier(rsp->name, s, cpu,
2343 atomic_read(&rsp->barrier_cpu_count), done);
2347 * RCU callback function for _rcu_barrier(). If we are last, wake
2348 * up the task executing _rcu_barrier().
2350 static void rcu_barrier_callback(struct rcu_head *rhp)
2352 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2353 struct rcu_state *rsp = rdp->rsp;
2355 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2356 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2357 complete(&rsp->barrier_completion);
2359 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2364 * Called with preemption disabled, and from cross-cpu IRQ context.
2366 static void rcu_barrier_func(void *type)
2368 struct rcu_state *rsp = type;
2369 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2371 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2372 atomic_inc(&rsp->barrier_cpu_count);
2373 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2377 * Orchestrate the specified type of RCU barrier, waiting for all
2378 * RCU callbacks of the specified type to complete.
2380 static void _rcu_barrier(struct rcu_state *rsp)
2383 unsigned long flags;
2384 struct rcu_data *rdp;
2386 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2387 unsigned long snap_done;
2389 init_rcu_head_on_stack(&rd.barrier_head);
2390 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2392 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2393 mutex_lock(&rsp->barrier_mutex);
2396 * Ensure that all prior references, including to ->n_barrier_done,
2397 * are ordered before the _rcu_barrier() machinery.
2399 smp_mb(); /* See above block comment. */
2402 * Recheck ->n_barrier_done to see if others did our work for us.
2403 * This means checking ->n_barrier_done for an even-to-odd-to-even
2404 * transition. The "if" expression below therefore rounds the old
2405 * value up to the next even number and adds two before comparing.
2407 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2408 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2409 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2410 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2411 smp_mb(); /* caller's subsequent code after above check. */
2412 mutex_unlock(&rsp->barrier_mutex);
2417 * Increment ->n_barrier_done to avoid duplicate work. Use
2418 * ACCESS_ONCE() to prevent the compiler from speculating
2419 * the increment to precede the early-exit check.
2421 ACCESS_ONCE(rsp->n_barrier_done)++;
2422 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2423 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2424 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2427 * Initialize the count to one rather than to zero in order to
2428 * avoid a too-soon return to zero in case of a short grace period
2429 * (or preemption of this task). Also flag this task as doing
2430 * an rcu_barrier(). This will prevent anyone else from adopting
2431 * orphaned callbacks, which could cause otherwise failure if a
2432 * CPU went offline and quickly came back online. To see this,
2433 * consider the following sequence of events:
2435 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2436 * 2. CPU 1 goes offline, orphaning its callbacks.
2437 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2438 * 4. CPU 1 comes back online.
2439 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2440 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2441 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2443 init_completion(&rsp->barrier_completion);
2444 atomic_set(&rsp->barrier_cpu_count, 1);
2445 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2446 rsp->rcu_barrier_in_progress = current;
2447 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2450 * Force every CPU with callbacks to register a new callback
2451 * that will tell us when all the preceding callbacks have
2452 * been invoked. If an offline CPU has callbacks, wait for
2453 * it to either come back online or to finish orphaning those
2456 for_each_possible_cpu(cpu) {
2458 rdp = per_cpu_ptr(rsp->rda, cpu);
2459 if (cpu_is_offline(cpu)) {
2460 _rcu_barrier_trace(rsp, "Offline", cpu,
2461 rsp->n_barrier_done);
2463 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2464 schedule_timeout_interruptible(1);
2465 } else if (ACCESS_ONCE(rdp->qlen)) {
2466 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2467 rsp->n_barrier_done);
2468 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2471 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2472 rsp->n_barrier_done);
2478 * Now that all online CPUs have rcu_barrier_callback() callbacks
2479 * posted, we can adopt all of the orphaned callbacks and place
2480 * an rcu_barrier_callback() callback after them. When that is done,
2481 * we are guaranteed to have an rcu_barrier_callback() callback
2482 * following every callback that could possibly have been
2483 * registered before _rcu_barrier() was called.
2485 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2486 rcu_adopt_orphan_cbs(rsp);
2487 rsp->rcu_barrier_in_progress = NULL;
2488 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2489 atomic_inc(&rsp->barrier_cpu_count);
2490 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2492 rsp->call(&rd.barrier_head, rcu_barrier_callback);
2495 * Now that we have an rcu_barrier_callback() callback on each
2496 * CPU, and thus each counted, remove the initial count.
2498 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2499 complete(&rsp->barrier_completion);
2501 /* Increment ->n_barrier_done to prevent duplicate work. */
2502 smp_mb(); /* Keep increment after above mechanism. */
2503 ACCESS_ONCE(rsp->n_barrier_done)++;
2504 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2505 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2506 smp_mb(); /* Keep increment before caller's subsequent code. */
2508 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2509 wait_for_completion(&rsp->barrier_completion);
2511 /* Other rcu_barrier() invocations can now safely proceed. */
2512 mutex_unlock(&rsp->barrier_mutex);
2514 destroy_rcu_head_on_stack(&rd.barrier_head);
2518 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2520 void rcu_barrier_bh(void)
2522 _rcu_barrier(&rcu_bh_state);
2524 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2527 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2529 void rcu_barrier_sched(void)
2531 _rcu_barrier(&rcu_sched_state);
2533 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2536 * Do boot-time initialization of a CPU's per-CPU RCU data.
2539 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2541 unsigned long flags;
2542 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2543 struct rcu_node *rnp = rcu_get_root(rsp);
2545 /* Set up local state, ensuring consistent view of global state. */
2546 raw_spin_lock_irqsave(&rnp->lock, flags);
2547 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2548 init_callback_list(rdp);
2550 ACCESS_ONCE(rdp->qlen) = 0;
2551 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2552 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2553 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2556 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2560 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2561 * offline event can be happening at a given time. Note also that we
2562 * can accept some slop in the rsp->completed access due to the fact
2563 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2565 static void __cpuinit
2566 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2568 unsigned long flags;
2570 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2571 struct rcu_node *rnp = rcu_get_root(rsp);
2573 /* Set up local state, ensuring consistent view of global state. */
2574 raw_spin_lock_irqsave(&rnp->lock, flags);
2575 rdp->beenonline = 1; /* We have now been online. */
2576 rdp->preemptible = preemptible;
2577 rdp->qlen_last_fqs_check = 0;
2578 rdp->n_force_qs_snap = rsp->n_force_qs;
2579 rdp->blimit = blimit;
2580 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2581 atomic_set(&rdp->dynticks->dynticks,
2582 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2583 rcu_prepare_for_idle_init(cpu);
2584 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2587 * A new grace period might start here. If so, we won't be part
2588 * of it, but that is OK, as we are currently in a quiescent state.
2591 /* Exclude any attempts to start a new GP on large systems. */
2592 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2594 /* Add CPU to rcu_node bitmasks. */
2596 mask = rdp->grpmask;
2598 /* Exclude any attempts to start a new GP on small systems. */
2599 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2600 rnp->qsmaskinit |= mask;
2601 mask = rnp->grpmask;
2602 if (rnp == rdp->mynode) {
2604 * If there is a grace period in progress, we will
2605 * set up to wait for it next time we run the
2608 rdp->gpnum = rnp->completed;
2609 rdp->completed = rnp->completed;
2610 rdp->passed_quiesce = 0;
2611 rdp->qs_pending = 0;
2612 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2613 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2615 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2617 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2619 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2622 static void __cpuinit rcu_prepare_cpu(int cpu)
2624 struct rcu_state *rsp;
2626 for_each_rcu_flavor(rsp)
2627 rcu_init_percpu_data(cpu, rsp,
2628 strcmp(rsp->name, "rcu_preempt") == 0);
2632 * Handle CPU online/offline notification events.
2634 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2635 unsigned long action, void *hcpu)
2637 long cpu = (long)hcpu;
2638 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2639 struct rcu_node *rnp = rdp->mynode;
2640 struct rcu_state *rsp;
2642 trace_rcu_utilization("Start CPU hotplug");
2644 case CPU_UP_PREPARE:
2645 case CPU_UP_PREPARE_FROZEN:
2646 rcu_prepare_cpu(cpu);
2647 rcu_prepare_kthreads(cpu);
2650 case CPU_DOWN_FAILED:
2651 rcu_node_kthread_setaffinity(rnp, -1);
2652 rcu_cpu_kthread_setrt(cpu, 1);
2654 case CPU_DOWN_PREPARE:
2655 rcu_node_kthread_setaffinity(rnp, cpu);
2656 rcu_cpu_kthread_setrt(cpu, 0);
2659 case CPU_DYING_FROZEN:
2661 * The whole machine is "stopped" except this CPU, so we can
2662 * touch any data without introducing corruption. We send the
2663 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2665 for_each_rcu_flavor(rsp)
2666 rcu_cleanup_dying_cpu(rsp);
2667 rcu_cleanup_after_idle(cpu);
2670 case CPU_DEAD_FROZEN:
2671 case CPU_UP_CANCELED:
2672 case CPU_UP_CANCELED_FROZEN:
2673 for_each_rcu_flavor(rsp)
2674 rcu_cleanup_dead_cpu(cpu, rsp);
2679 trace_rcu_utilization("End CPU hotplug");
2684 * Spawn the kthread that handles this RCU flavor's grace periods.
2686 static int __init rcu_spawn_gp_kthread(void)
2688 unsigned long flags;
2689 struct rcu_node *rnp;
2690 struct rcu_state *rsp;
2691 struct task_struct *t;
2693 for_each_rcu_flavor(rsp) {
2694 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2696 rnp = rcu_get_root(rsp);
2697 raw_spin_lock_irqsave(&rnp->lock, flags);
2698 rsp->gp_kthread = t;
2699 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2703 early_initcall(rcu_spawn_gp_kthread);
2706 * This function is invoked towards the end of the scheduler's initialization
2707 * process. Before this is called, the idle task might contain
2708 * RCU read-side critical sections (during which time, this idle
2709 * task is booting the system). After this function is called, the
2710 * idle tasks are prohibited from containing RCU read-side critical
2711 * sections. This function also enables RCU lockdep checking.
2713 void rcu_scheduler_starting(void)
2715 WARN_ON(num_online_cpus() != 1);
2716 WARN_ON(nr_context_switches() > 0);
2717 rcu_scheduler_active = 1;
2721 * Compute the per-level fanout, either using the exact fanout specified
2722 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2724 #ifdef CONFIG_RCU_FANOUT_EXACT
2725 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2729 for (i = rcu_num_lvls - 1; i > 0; i--)
2730 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2731 rsp->levelspread[0] = rcu_fanout_leaf;
2733 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2734 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2741 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2742 ccur = rsp->levelcnt[i];
2743 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2747 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2750 * Helper function for rcu_init() that initializes one rcu_state structure.
2752 static void __init rcu_init_one(struct rcu_state *rsp,
2753 struct rcu_data __percpu *rda)
2755 static char *buf[] = { "rcu_node_level_0",
2758 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2762 struct rcu_node *rnp;
2764 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2766 /* Initialize the level-tracking arrays. */
2768 for (i = 0; i < rcu_num_lvls; i++)
2769 rsp->levelcnt[i] = num_rcu_lvl[i];
2770 for (i = 1; i < rcu_num_lvls; i++)
2771 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2772 rcu_init_levelspread(rsp);
2774 /* Initialize the elements themselves, starting from the leaves. */
2776 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2777 cpustride *= rsp->levelspread[i];
2778 rnp = rsp->level[i];
2779 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2780 raw_spin_lock_init(&rnp->lock);
2781 lockdep_set_class_and_name(&rnp->lock,
2782 &rcu_node_class[i], buf[i]);
2785 rnp->qsmaskinit = 0;
2786 rnp->grplo = j * cpustride;
2787 rnp->grphi = (j + 1) * cpustride - 1;
2788 if (rnp->grphi >= NR_CPUS)
2789 rnp->grphi = NR_CPUS - 1;
2795 rnp->grpnum = j % rsp->levelspread[i - 1];
2796 rnp->grpmask = 1UL << rnp->grpnum;
2797 rnp->parent = rsp->level[i - 1] +
2798 j / rsp->levelspread[i - 1];
2801 INIT_LIST_HEAD(&rnp->blkd_tasks);
2806 init_waitqueue_head(&rsp->gp_wq);
2807 rnp = rsp->level[rcu_num_lvls - 1];
2808 for_each_possible_cpu(i) {
2809 while (i > rnp->grphi)
2811 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2812 rcu_boot_init_percpu_data(i, rsp);
2814 list_add(&rsp->flavors, &rcu_struct_flavors);
2818 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2819 * replace the definitions in rcutree.h because those are needed to size
2820 * the ->node array in the rcu_state structure.
2822 static void __init rcu_init_geometry(void)
2827 int rcu_capacity[MAX_RCU_LVLS + 1];
2829 /* If the compile-time values are accurate, just leave. */
2830 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2834 * Compute number of nodes that can be handled an rcu_node tree
2835 * with the given number of levels. Setting rcu_capacity[0] makes
2836 * some of the arithmetic easier.
2838 rcu_capacity[0] = 1;
2839 rcu_capacity[1] = rcu_fanout_leaf;
2840 for (i = 2; i <= MAX_RCU_LVLS; i++)
2841 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2844 * The boot-time rcu_fanout_leaf parameter is only permitted
2845 * to increase the leaf-level fanout, not decrease it. Of course,
2846 * the leaf-level fanout cannot exceed the number of bits in
2847 * the rcu_node masks. Finally, the tree must be able to accommodate
2848 * the configured number of CPUs. Complain and fall back to the
2849 * compile-time values if these limits are exceeded.
2851 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2852 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2853 n > rcu_capacity[MAX_RCU_LVLS]) {
2858 /* Calculate the number of rcu_nodes at each level of the tree. */
2859 for (i = 1; i <= MAX_RCU_LVLS; i++)
2860 if (n <= rcu_capacity[i]) {
2861 for (j = 0; j <= i; j++)
2863 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2865 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2870 /* Calculate the total number of rcu_node structures. */
2872 for (i = 0; i <= MAX_RCU_LVLS; i++)
2873 rcu_num_nodes += num_rcu_lvl[i];
2877 void __init rcu_init(void)
2881 rcu_bootup_announce();
2882 rcu_init_geometry();
2883 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2884 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2885 __rcu_init_preempt();
2886 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2889 * We don't need protection against CPU-hotplug here because
2890 * this is called early in boot, before either interrupts
2891 * or the scheduler are operational.
2893 cpu_notifier(rcu_cpu_notify, 0);
2894 for_each_online_cpu(cpu)
2895 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2896 check_cpu_stall_init();
2899 #include "rcutree_plugin.h"