2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
57 #include <trace/events/rcu.h>
61 /* Data structures. */
63 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 #define RCU_STATE_INITIALIZER(sname, cr) { \
66 .level = { &sname##_state.node[0] }, \
68 .fqs_state = RCU_GP_IDLE, \
71 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
72 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
73 .orphan_donetail = &sname##_state.orphan_donelist, \
74 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \
76 .n_force_qs_ngp = 0, \
80 struct rcu_state rcu_sched_state =
81 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
82 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
84 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
85 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
87 static struct rcu_state *rcu_state;
89 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
90 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
91 module_param(rcu_fanout_leaf, int, 0);
92 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
93 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
100 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
103 * The rcu_scheduler_active variable transitions from zero to one just
104 * before the first task is spawned. So when this variable is zero, RCU
105 * can assume that there is but one task, allowing RCU to (for example)
106 * optimized synchronize_sched() to a simple barrier(). When this variable
107 * is one, RCU must actually do all the hard work required to detect real
108 * grace periods. This variable is also used to suppress boot-time false
109 * positives from lockdep-RCU error checking.
111 int rcu_scheduler_active __read_mostly;
112 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
115 * The rcu_scheduler_fully_active variable transitions from zero to one
116 * during the early_initcall() processing, which is after the scheduler
117 * is capable of creating new tasks. So RCU processing (for example,
118 * creating tasks for RCU priority boosting) must be delayed until after
119 * rcu_scheduler_fully_active transitions from zero to one. We also
120 * currently delay invocation of any RCU callbacks until after this point.
122 * It might later prove better for people registering RCU callbacks during
123 * early boot to take responsibility for these callbacks, but one step at
126 static int rcu_scheduler_fully_active __read_mostly;
128 #ifdef CONFIG_RCU_BOOST
131 * Control variables for per-CPU and per-rcu_node kthreads. These
132 * handle all flavors of RCU.
134 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
135 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
136 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
138 DEFINE_PER_CPU(char, rcu_cpu_has_work);
140 #endif /* #ifdef CONFIG_RCU_BOOST */
142 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
143 static void invoke_rcu_core(void);
144 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
147 * Track the rcutorture test sequence number and the update version
148 * number within a given test. The rcutorture_testseq is incremented
149 * on every rcutorture module load and unload, so has an odd value
150 * when a test is running. The rcutorture_vernum is set to zero
151 * when rcutorture starts and is incremented on each rcutorture update.
152 * These variables enable correlating rcutorture output with the
153 * RCU tracing information.
155 unsigned long rcutorture_testseq;
156 unsigned long rcutorture_vernum;
158 /* State information for rcu_barrier() and friends. */
160 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
161 static atomic_t rcu_barrier_cpu_count;
162 static DEFINE_MUTEX(rcu_barrier_mutex);
163 static struct completion rcu_barrier_completion;
166 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
167 * permit this function to be invoked without holding the root rcu_node
168 * structure's ->lock, but of course results can be subject to change.
170 static int rcu_gp_in_progress(struct rcu_state *rsp)
172 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
176 * Note a quiescent state. Because we do not need to know
177 * how many quiescent states passed, just if there was at least
178 * one since the start of the grace period, this just sets a flag.
179 * The caller must have disabled preemption.
181 void rcu_sched_qs(int cpu)
183 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
185 rdp->passed_quiesce_gpnum = rdp->gpnum;
187 if (rdp->passed_quiesce == 0)
188 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
189 rdp->passed_quiesce = 1;
192 void rcu_bh_qs(int cpu)
194 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
196 rdp->passed_quiesce_gpnum = rdp->gpnum;
198 if (rdp->passed_quiesce == 0)
199 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
200 rdp->passed_quiesce = 1;
204 * Note a context switch. This is a quiescent state for RCU-sched,
205 * and requires special handling for preemptible RCU.
206 * The caller must have disabled preemption.
208 void rcu_note_context_switch(int cpu)
210 trace_rcu_utilization("Start context switch");
212 rcu_preempt_note_context_switch(cpu);
213 trace_rcu_utilization("End context switch");
215 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
217 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
218 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
219 .dynticks = ATOMIC_INIT(1),
222 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
223 static int qhimark = 10000; /* If this many pending, ignore blimit. */
224 static int qlowmark = 100; /* Once only this many pending, use blimit. */
226 module_param(blimit, int, 0);
227 module_param(qhimark, int, 0);
228 module_param(qlowmark, int, 0);
230 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
231 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
233 module_param(rcu_cpu_stall_suppress, int, 0644);
234 module_param(rcu_cpu_stall_timeout, int, 0644);
236 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
237 static int rcu_pending(int cpu);
240 * Return the number of RCU-sched batches processed thus far for debug & stats.
242 long rcu_batches_completed_sched(void)
244 return rcu_sched_state.completed;
246 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
249 * Return the number of RCU BH batches processed thus far for debug & stats.
251 long rcu_batches_completed_bh(void)
253 return rcu_bh_state.completed;
255 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
258 * Force a quiescent state for RCU BH.
260 void rcu_bh_force_quiescent_state(void)
262 force_quiescent_state(&rcu_bh_state, 0);
264 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
267 * Record the number of times rcutorture tests have been initiated and
268 * terminated. This information allows the debugfs tracing stats to be
269 * correlated to the rcutorture messages, even when the rcutorture module
270 * is being repeatedly loaded and unloaded. In other words, we cannot
271 * store this state in rcutorture itself.
273 void rcutorture_record_test_transition(void)
275 rcutorture_testseq++;
276 rcutorture_vernum = 0;
278 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
281 * Record the number of writer passes through the current rcutorture test.
282 * This is also used to correlate debugfs tracing stats with the rcutorture
285 void rcutorture_record_progress(unsigned long vernum)
289 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
292 * Force a quiescent state for RCU-sched.
294 void rcu_sched_force_quiescent_state(void)
296 force_quiescent_state(&rcu_sched_state, 0);
298 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
301 * Does the CPU have callbacks ready to be invoked?
304 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
306 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
310 * Does the current CPU require a yet-as-unscheduled grace period?
313 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
315 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
319 * Return the root node of the specified rcu_state structure.
321 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
323 return &rsp->node[0];
327 * If the specified CPU is offline, tell the caller that it is in
328 * a quiescent state. Otherwise, whack it with a reschedule IPI.
329 * Grace periods can end up waiting on an offline CPU when that
330 * CPU is in the process of coming online -- it will be added to the
331 * rcu_node bitmasks before it actually makes it online. The same thing
332 * can happen while a CPU is in the process of coming online. Because this
333 * race is quite rare, we check for it after detecting that the grace
334 * period has been delayed rather than checking each and every CPU
335 * each and every time we start a new grace period.
337 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
340 * If the CPU is offline for more than a jiffy, it is in a quiescent
341 * state. We can trust its state not to change because interrupts
342 * are disabled. The reason for the jiffy's worth of slack is to
343 * handle CPUs initializing on the way up and finding their way
344 * to the idle loop on the way down.
346 if (cpu_is_offline(rdp->cpu) &&
347 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
348 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
356 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
358 * If the new value of the ->dynticks_nesting counter now is zero,
359 * we really have entered idle, and must do the appropriate accounting.
360 * The caller must have disabled interrupts.
362 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
364 trace_rcu_dyntick("Start", oldval, 0);
365 if (!is_idle_task(current)) {
366 struct task_struct *idle = idle_task(smp_processor_id());
368 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
369 ftrace_dump(DUMP_ALL);
370 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
371 current->pid, current->comm,
372 idle->pid, idle->comm); /* must be idle task! */
374 rcu_prepare_for_idle(smp_processor_id());
375 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
376 smp_mb__before_atomic_inc(); /* See above. */
377 atomic_inc(&rdtp->dynticks);
378 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
379 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
382 * The idle task is not permitted to enter the idle loop while
383 * in an RCU read-side critical section.
385 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
386 "Illegal idle entry in RCU read-side critical section.");
387 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
388 "Illegal idle entry in RCU-bh read-side critical section.");
389 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
390 "Illegal idle entry in RCU-sched read-side critical section.");
394 * rcu_idle_enter - inform RCU that current CPU is entering idle
396 * Enter idle mode, in other words, -leave- the mode in which RCU
397 * read-side critical sections can occur. (Though RCU read-side
398 * critical sections can occur in irq handlers in idle, a possibility
399 * handled by irq_enter() and irq_exit().)
401 * We crowbar the ->dynticks_nesting field to zero to allow for
402 * the possibility of usermode upcalls having messed up our count
403 * of interrupt nesting level during the prior busy period.
405 void rcu_idle_enter(void)
409 struct rcu_dynticks *rdtp;
411 local_irq_save(flags);
412 rdtp = &__get_cpu_var(rcu_dynticks);
413 oldval = rdtp->dynticks_nesting;
414 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
415 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
416 rdtp->dynticks_nesting = 0;
418 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
419 rcu_idle_enter_common(rdtp, oldval);
420 local_irq_restore(flags);
422 EXPORT_SYMBOL_GPL(rcu_idle_enter);
425 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
427 * Exit from an interrupt handler, which might possibly result in entering
428 * idle mode, in other words, leaving the mode in which read-side critical
429 * sections can occur.
431 * This code assumes that the idle loop never does anything that might
432 * result in unbalanced calls to irq_enter() and irq_exit(). If your
433 * architecture violates this assumption, RCU will give you what you
434 * deserve, good and hard. But very infrequently and irreproducibly.
436 * Use things like work queues to work around this limitation.
438 * You have been warned.
440 void rcu_irq_exit(void)
444 struct rcu_dynticks *rdtp;
446 local_irq_save(flags);
447 rdtp = &__get_cpu_var(rcu_dynticks);
448 oldval = rdtp->dynticks_nesting;
449 rdtp->dynticks_nesting--;
450 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
451 if (rdtp->dynticks_nesting)
452 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
454 rcu_idle_enter_common(rdtp, oldval);
455 local_irq_restore(flags);
459 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
461 * If the new value of the ->dynticks_nesting counter was previously zero,
462 * we really have exited idle, and must do the appropriate accounting.
463 * The caller must have disabled interrupts.
465 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
467 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
468 atomic_inc(&rdtp->dynticks);
469 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
470 smp_mb__after_atomic_inc(); /* See above. */
471 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
472 rcu_cleanup_after_idle(smp_processor_id());
473 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
474 if (!is_idle_task(current)) {
475 struct task_struct *idle = idle_task(smp_processor_id());
477 trace_rcu_dyntick("Error on exit: not idle task",
478 oldval, rdtp->dynticks_nesting);
479 ftrace_dump(DUMP_ALL);
480 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
481 current->pid, current->comm,
482 idle->pid, idle->comm); /* must be idle task! */
487 * rcu_idle_exit - inform RCU that current CPU is leaving idle
489 * Exit idle mode, in other words, -enter- the mode in which RCU
490 * read-side critical sections can occur.
492 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
493 * allow for the possibility of usermode upcalls messing up our count
494 * of interrupt nesting level during the busy period that is just
497 void rcu_idle_exit(void)
500 struct rcu_dynticks *rdtp;
503 local_irq_save(flags);
504 rdtp = &__get_cpu_var(rcu_dynticks);
505 oldval = rdtp->dynticks_nesting;
506 WARN_ON_ONCE(oldval < 0);
507 if (oldval & DYNTICK_TASK_NEST_MASK)
508 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
510 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
511 rcu_idle_exit_common(rdtp, oldval);
512 local_irq_restore(flags);
514 EXPORT_SYMBOL_GPL(rcu_idle_exit);
517 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
519 * Enter an interrupt handler, which might possibly result in exiting
520 * idle mode, in other words, entering the mode in which read-side critical
521 * sections can occur.
523 * Note that the Linux kernel is fully capable of entering an interrupt
524 * handler that it never exits, for example when doing upcalls to
525 * user mode! This code assumes that the idle loop never does upcalls to
526 * user mode. If your architecture does do upcalls from the idle loop (or
527 * does anything else that results in unbalanced calls to the irq_enter()
528 * and irq_exit() functions), RCU will give you what you deserve, good
529 * and hard. But very infrequently and irreproducibly.
531 * Use things like work queues to work around this limitation.
533 * You have been warned.
535 void rcu_irq_enter(void)
538 struct rcu_dynticks *rdtp;
541 local_irq_save(flags);
542 rdtp = &__get_cpu_var(rcu_dynticks);
543 oldval = rdtp->dynticks_nesting;
544 rdtp->dynticks_nesting++;
545 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
547 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
549 rcu_idle_exit_common(rdtp, oldval);
550 local_irq_restore(flags);
554 * rcu_nmi_enter - inform RCU of entry to NMI context
556 * If the CPU was idle with dynamic ticks active, and there is no
557 * irq handler running, this updates rdtp->dynticks_nmi to let the
558 * RCU grace-period handling know that the CPU is active.
560 void rcu_nmi_enter(void)
562 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
564 if (rdtp->dynticks_nmi_nesting == 0 &&
565 (atomic_read(&rdtp->dynticks) & 0x1))
567 rdtp->dynticks_nmi_nesting++;
568 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
569 atomic_inc(&rdtp->dynticks);
570 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
571 smp_mb__after_atomic_inc(); /* See above. */
572 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
576 * rcu_nmi_exit - inform RCU of exit from NMI context
578 * If the CPU was idle with dynamic ticks active, and there is no
579 * irq handler running, this updates rdtp->dynticks_nmi to let the
580 * RCU grace-period handling know that the CPU is no longer active.
582 void rcu_nmi_exit(void)
584 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
586 if (rdtp->dynticks_nmi_nesting == 0 ||
587 --rdtp->dynticks_nmi_nesting != 0)
589 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
590 smp_mb__before_atomic_inc(); /* See above. */
591 atomic_inc(&rdtp->dynticks);
592 smp_mb__after_atomic_inc(); /* Force delay to next write. */
593 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
596 #ifdef CONFIG_PROVE_RCU
599 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
601 * If the current CPU is in its idle loop and is neither in an interrupt
602 * or NMI handler, return true.
604 int rcu_is_cpu_idle(void)
609 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
613 EXPORT_SYMBOL(rcu_is_cpu_idle);
615 #ifdef CONFIG_HOTPLUG_CPU
618 * Is the current CPU online? Disable preemption to avoid false positives
619 * that could otherwise happen due to the current CPU number being sampled,
620 * this task being preempted, its old CPU being taken offline, resuming
621 * on some other CPU, then determining that its old CPU is now offline.
622 * It is OK to use RCU on an offline processor during initial boot, hence
623 * the check for rcu_scheduler_fully_active. Note also that it is OK
624 * for a CPU coming online to use RCU for one jiffy prior to marking itself
625 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
626 * offline to continue to use RCU for one jiffy after marking itself
627 * offline in the cpu_online_mask. This leniency is necessary given the
628 * non-atomic nature of the online and offline processing, for example,
629 * the fact that a CPU enters the scheduler after completing the CPU_DYING
632 * This is also why RCU internally marks CPUs online during the
633 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
635 * Disable checking if in an NMI handler because we cannot safely report
636 * errors from NMI handlers anyway.
638 bool rcu_lockdep_current_cpu_online(void)
640 struct rcu_data *rdp;
641 struct rcu_node *rnp;
647 rdp = &__get_cpu_var(rcu_sched_data);
649 ret = (rdp->grpmask & rnp->qsmaskinit) ||
650 !rcu_scheduler_fully_active;
654 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
656 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
658 #endif /* #ifdef CONFIG_PROVE_RCU */
661 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
663 * If the current CPU is idle or running at a first-level (not nested)
664 * interrupt from idle, return true. The caller must have at least
665 * disabled preemption.
667 int rcu_is_cpu_rrupt_from_idle(void)
669 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
673 * Snapshot the specified CPU's dynticks counter so that we can later
674 * credit them with an implicit quiescent state. Return 1 if this CPU
675 * is in dynticks idle mode, which is an extended quiescent state.
677 static int dyntick_save_progress_counter(struct rcu_data *rdp)
679 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
680 return (rdp->dynticks_snap & 0x1) == 0;
684 * Return true if the specified CPU has passed through a quiescent
685 * state by virtue of being in or having passed through an dynticks
686 * idle state since the last call to dyntick_save_progress_counter()
689 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
694 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
695 snap = (unsigned int)rdp->dynticks_snap;
698 * If the CPU passed through or entered a dynticks idle phase with
699 * no active irq/NMI handlers, then we can safely pretend that the CPU
700 * already acknowledged the request to pass through a quiescent
701 * state. Either way, that CPU cannot possibly be in an RCU
702 * read-side critical section that started before the beginning
703 * of the current RCU grace period.
705 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
706 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
711 /* Go check for the CPU being offline. */
712 return rcu_implicit_offline_qs(rdp);
715 static int jiffies_till_stall_check(void)
717 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
720 * Limit check must be consistent with the Kconfig limits
721 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
723 if (till_stall_check < 3) {
724 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
725 till_stall_check = 3;
726 } else if (till_stall_check > 300) {
727 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
728 till_stall_check = 300;
730 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
733 static void record_gp_stall_check_time(struct rcu_state *rsp)
735 rsp->gp_start = jiffies;
736 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
739 static void print_other_cpu_stall(struct rcu_state *rsp)
745 struct rcu_node *rnp = rcu_get_root(rsp);
747 /* Only let one CPU complain about others per time interval. */
749 raw_spin_lock_irqsave(&rnp->lock, flags);
750 delta = jiffies - rsp->jiffies_stall;
751 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
752 raw_spin_unlock_irqrestore(&rnp->lock, flags);
755 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
756 raw_spin_unlock_irqrestore(&rnp->lock, flags);
759 * OK, time to rat on our buddy...
760 * See Documentation/RCU/stallwarn.txt for info on how to debug
761 * RCU CPU stall warnings.
763 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
765 print_cpu_stall_info_begin();
766 rcu_for_each_leaf_node(rsp, rnp) {
767 raw_spin_lock_irqsave(&rnp->lock, flags);
768 ndetected += rcu_print_task_stall(rnp);
769 raw_spin_unlock_irqrestore(&rnp->lock, flags);
770 if (rnp->qsmask == 0)
772 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
773 if (rnp->qsmask & (1UL << cpu)) {
774 print_cpu_stall_info(rsp, rnp->grplo + cpu);
780 * Now rat on any tasks that got kicked up to the root rcu_node
781 * due to CPU offlining.
783 rnp = rcu_get_root(rsp);
784 raw_spin_lock_irqsave(&rnp->lock, flags);
785 ndetected = rcu_print_task_stall(rnp);
786 raw_spin_unlock_irqrestore(&rnp->lock, flags);
788 print_cpu_stall_info_end();
789 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
790 smp_processor_id(), (long)(jiffies - rsp->gp_start));
792 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
793 else if (!trigger_all_cpu_backtrace())
796 /* If so configured, complain about tasks blocking the grace period. */
798 rcu_print_detail_task_stall(rsp);
800 force_quiescent_state(rsp, 0); /* Kick them all. */
803 static void print_cpu_stall(struct rcu_state *rsp)
806 struct rcu_node *rnp = rcu_get_root(rsp);
809 * OK, time to rat on ourselves...
810 * See Documentation/RCU/stallwarn.txt for info on how to debug
811 * RCU CPU stall warnings.
813 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
814 print_cpu_stall_info_begin();
815 print_cpu_stall_info(rsp, smp_processor_id());
816 print_cpu_stall_info_end();
817 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
818 if (!trigger_all_cpu_backtrace())
821 raw_spin_lock_irqsave(&rnp->lock, flags);
822 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
823 rsp->jiffies_stall = jiffies +
824 3 * jiffies_till_stall_check() + 3;
825 raw_spin_unlock_irqrestore(&rnp->lock, flags);
827 set_need_resched(); /* kick ourselves to get things going. */
830 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
834 struct rcu_node *rnp;
836 if (rcu_cpu_stall_suppress)
838 j = ACCESS_ONCE(jiffies);
839 js = ACCESS_ONCE(rsp->jiffies_stall);
841 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
843 /* We haven't checked in, so go dump stack. */
844 print_cpu_stall(rsp);
846 } else if (rcu_gp_in_progress(rsp) &&
847 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
849 /* They had a few time units to dump stack, so complain. */
850 print_other_cpu_stall(rsp);
854 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
856 rcu_cpu_stall_suppress = 1;
861 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
863 * Set the stall-warning timeout way off into the future, thus preventing
864 * any RCU CPU stall-warning messages from appearing in the current set of
867 * The caller must disable hard irqs.
869 void rcu_cpu_stall_reset(void)
871 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
872 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
873 rcu_preempt_stall_reset();
876 static struct notifier_block rcu_panic_block = {
877 .notifier_call = rcu_panic,
880 static void __init check_cpu_stall_init(void)
882 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
886 * Update CPU-local rcu_data state to record the newly noticed grace period.
887 * This is used both when we started the grace period and when we notice
888 * that someone else started the grace period. The caller must hold the
889 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
890 * and must have irqs disabled.
892 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
894 if (rdp->gpnum != rnp->gpnum) {
896 * If the current grace period is waiting for this CPU,
897 * set up to detect a quiescent state, otherwise don't
898 * go looking for one.
900 rdp->gpnum = rnp->gpnum;
901 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
902 if (rnp->qsmask & rdp->grpmask) {
904 rdp->passed_quiesce = 0;
907 zero_cpu_stall_ticks(rdp);
911 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
914 struct rcu_node *rnp;
916 local_irq_save(flags);
918 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
919 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
920 local_irq_restore(flags);
923 __note_new_gpnum(rsp, rnp, rdp);
924 raw_spin_unlock_irqrestore(&rnp->lock, flags);
928 * Did someone else start a new RCU grace period start since we last
929 * checked? Update local state appropriately if so. Must be called
930 * on the CPU corresponding to rdp.
933 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
938 local_irq_save(flags);
939 if (rdp->gpnum != rsp->gpnum) {
940 note_new_gpnum(rsp, rdp);
943 local_irq_restore(flags);
948 * Advance this CPU's callbacks, but only if the current grace period
949 * has ended. This may be called only from the CPU to whom the rdp
950 * belongs. In addition, the corresponding leaf rcu_node structure's
951 * ->lock must be held by the caller, with irqs disabled.
954 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
956 /* Did another grace period end? */
957 if (rdp->completed != rnp->completed) {
959 /* Advance callbacks. No harm if list empty. */
960 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
961 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
962 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
964 /* Remember that we saw this grace-period completion. */
965 rdp->completed = rnp->completed;
966 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
969 * If we were in an extended quiescent state, we may have
970 * missed some grace periods that others CPUs handled on
971 * our behalf. Catch up with this state to avoid noting
972 * spurious new grace periods. If another grace period
973 * has started, then rnp->gpnum will have advanced, so
974 * we will detect this later on.
976 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
977 rdp->gpnum = rdp->completed;
980 * If RCU does not need a quiescent state from this CPU,
981 * then make sure that this CPU doesn't go looking for one.
983 if ((rnp->qsmask & rdp->grpmask) == 0)
989 * Advance this CPU's callbacks, but only if the current grace period
990 * has ended. This may be called only from the CPU to whom the rdp
994 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
997 struct rcu_node *rnp;
999 local_irq_save(flags);
1001 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1002 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1003 local_irq_restore(flags);
1006 __rcu_process_gp_end(rsp, rnp, rdp);
1007 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1011 * Do per-CPU grace-period initialization for running CPU. The caller
1012 * must hold the lock of the leaf rcu_node structure corresponding to
1016 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1018 /* Prior grace period ended, so advance callbacks for current CPU. */
1019 __rcu_process_gp_end(rsp, rnp, rdp);
1022 * Because this CPU just now started the new grace period, we know
1023 * that all of its callbacks will be covered by this upcoming grace
1024 * period, even the ones that were registered arbitrarily recently.
1025 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1027 * Other CPUs cannot be sure exactly when the grace period started.
1028 * Therefore, their recently registered callbacks must pass through
1029 * an additional RCU_NEXT_READY stage, so that they will be handled
1030 * by the next RCU grace period.
1032 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1033 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1035 /* Set state so that this CPU will detect the next quiescent state. */
1036 __note_new_gpnum(rsp, rnp, rdp);
1040 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1041 * in preparation for detecting the next grace period. The caller must hold
1042 * the root node's ->lock, which is released before return. Hard irqs must
1045 * Note that it is legal for a dying CPU (which is marked as offline) to
1046 * invoke this function. This can happen when the dying CPU reports its
1050 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1051 __releases(rcu_get_root(rsp)->lock)
1053 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1054 struct rcu_node *rnp = rcu_get_root(rsp);
1056 if (!rcu_scheduler_fully_active ||
1057 !cpu_needs_another_gp(rsp, rdp)) {
1059 * Either the scheduler hasn't yet spawned the first
1060 * non-idle task or this CPU does not need another
1061 * grace period. Either way, don't start a new grace
1064 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1068 if (rsp->fqs_active) {
1070 * This CPU needs a grace period, but force_quiescent_state()
1071 * is running. Tell it to start one on this CPU's behalf.
1073 rsp->fqs_need_gp = 1;
1074 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1078 /* Advance to a new grace period and initialize state. */
1080 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1081 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1082 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1083 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1084 record_gp_stall_check_time(rsp);
1085 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1087 /* Exclude any concurrent CPU-hotplug operations. */
1088 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1091 * Set the quiescent-state-needed bits in all the rcu_node
1092 * structures for all currently online CPUs in breadth-first
1093 * order, starting from the root rcu_node structure. This
1094 * operation relies on the layout of the hierarchy within the
1095 * rsp->node[] array. Note that other CPUs will access only
1096 * the leaves of the hierarchy, which still indicate that no
1097 * grace period is in progress, at least until the corresponding
1098 * leaf node has been initialized. In addition, we have excluded
1099 * CPU-hotplug operations.
1101 * Note that the grace period cannot complete until we finish
1102 * the initialization process, as there will be at least one
1103 * qsmask bit set in the root node until that time, namely the
1104 * one corresponding to this CPU, due to the fact that we have
1107 rcu_for_each_node_breadth_first(rsp, rnp) {
1108 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1109 rcu_preempt_check_blocked_tasks(rnp);
1110 rnp->qsmask = rnp->qsmaskinit;
1111 rnp->gpnum = rsp->gpnum;
1112 rnp->completed = rsp->completed;
1113 if (rnp == rdp->mynode)
1114 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1115 rcu_preempt_boost_start_gp(rnp);
1116 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1117 rnp->level, rnp->grplo,
1118 rnp->grphi, rnp->qsmask);
1119 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1122 rnp = rcu_get_root(rsp);
1123 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1124 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1125 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1126 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1130 * Report a full set of quiescent states to the specified rcu_state
1131 * data structure. This involves cleaning up after the prior grace
1132 * period and letting rcu_start_gp() start up the next grace period
1133 * if one is needed. Note that the caller must hold rnp->lock, as
1134 * required by rcu_start_gp(), which will release it.
1136 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1137 __releases(rcu_get_root(rsp)->lock)
1139 unsigned long gp_duration;
1140 struct rcu_node *rnp = rcu_get_root(rsp);
1141 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1143 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1146 * Ensure that all grace-period and pre-grace-period activity
1147 * is seen before the assignment to rsp->completed.
1149 smp_mb(); /* See above block comment. */
1150 gp_duration = jiffies - rsp->gp_start;
1151 if (gp_duration > rsp->gp_max)
1152 rsp->gp_max = gp_duration;
1155 * We know the grace period is complete, but to everyone else
1156 * it appears to still be ongoing. But it is also the case
1157 * that to everyone else it looks like there is nothing that
1158 * they can do to advance the grace period. It is therefore
1159 * safe for us to drop the lock in order to mark the grace
1160 * period as completed in all of the rcu_node structures.
1162 * But if this CPU needs another grace period, it will take
1163 * care of this while initializing the next grace period.
1164 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1165 * because the callbacks have not yet been advanced: Those
1166 * callbacks are waiting on the grace period that just now
1169 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1170 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1173 * Propagate new ->completed value to rcu_node structures
1174 * so that other CPUs don't have to wait until the start
1175 * of the next grace period to process their callbacks.
1177 rcu_for_each_node_breadth_first(rsp, rnp) {
1178 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1179 rnp->completed = rsp->gpnum;
1180 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1182 rnp = rcu_get_root(rsp);
1183 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1186 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1187 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1188 rsp->fqs_state = RCU_GP_IDLE;
1189 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1193 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1194 * Allows quiescent states for a group of CPUs to be reported at one go
1195 * to the specified rcu_node structure, though all the CPUs in the group
1196 * must be represented by the same rcu_node structure (which need not be
1197 * a leaf rcu_node structure, though it often will be). That structure's
1198 * lock must be held upon entry, and it is released before return.
1201 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1202 struct rcu_node *rnp, unsigned long flags)
1203 __releases(rnp->lock)
1205 struct rcu_node *rnp_c;
1207 /* Walk up the rcu_node hierarchy. */
1209 if (!(rnp->qsmask & mask)) {
1211 /* Our bit has already been cleared, so done. */
1212 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1215 rnp->qsmask &= ~mask;
1216 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1217 mask, rnp->qsmask, rnp->level,
1218 rnp->grplo, rnp->grphi,
1220 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1222 /* Other bits still set at this level, so done. */
1223 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1226 mask = rnp->grpmask;
1227 if (rnp->parent == NULL) {
1229 /* No more levels. Exit loop holding root lock. */
1233 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1236 raw_spin_lock_irqsave(&rnp->lock, flags);
1237 WARN_ON_ONCE(rnp_c->qsmask);
1241 * Get here if we are the last CPU to pass through a quiescent
1242 * state for this grace period. Invoke rcu_report_qs_rsp()
1243 * to clean up and start the next grace period if one is needed.
1245 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1249 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1250 * structure. This must be either called from the specified CPU, or
1251 * called when the specified CPU is known to be offline (and when it is
1252 * also known that no other CPU is concurrently trying to help the offline
1253 * CPU). The lastcomp argument is used to make sure we are still in the
1254 * grace period of interest. We don't want to end the current grace period
1255 * based on quiescent states detected in an earlier grace period!
1258 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1260 unsigned long flags;
1262 struct rcu_node *rnp;
1265 raw_spin_lock_irqsave(&rnp->lock, flags);
1266 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1269 * The grace period in which this quiescent state was
1270 * recorded has ended, so don't report it upwards.
1271 * We will instead need a new quiescent state that lies
1272 * within the current grace period.
1274 rdp->passed_quiesce = 0; /* need qs for new gp. */
1275 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1278 mask = rdp->grpmask;
1279 if ((rnp->qsmask & mask) == 0) {
1280 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1282 rdp->qs_pending = 0;
1285 * This GP can't end until cpu checks in, so all of our
1286 * callbacks can be processed during the next GP.
1288 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1290 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1295 * Check to see if there is a new grace period of which this CPU
1296 * is not yet aware, and if so, set up local rcu_data state for it.
1297 * Otherwise, see if this CPU has just passed through its first
1298 * quiescent state for this grace period, and record that fact if so.
1301 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1303 /* If there is now a new grace period, record and return. */
1304 if (check_for_new_grace_period(rsp, rdp))
1308 * Does this CPU still need to do its part for current grace period?
1309 * If no, return and let the other CPUs do their part as well.
1311 if (!rdp->qs_pending)
1315 * Was there a quiescent state since the beginning of the grace
1316 * period? If no, then exit and wait for the next call.
1318 if (!rdp->passed_quiesce)
1322 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1325 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1328 #ifdef CONFIG_HOTPLUG_CPU
1331 * Send the specified CPU's RCU callbacks to the orphanage. The
1332 * specified CPU must be offline, and the caller must hold the
1336 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1337 struct rcu_node *rnp, struct rcu_data *rdp)
1342 * Orphan the callbacks. First adjust the counts. This is safe
1343 * because ->onofflock excludes _rcu_barrier()'s adoption of
1344 * the callbacks, thus no memory barrier is required.
1346 if (rdp->nxtlist != NULL) {
1347 rsp->qlen_lazy += rdp->qlen_lazy;
1348 rsp->qlen += rdp->qlen;
1349 rdp->n_cbs_orphaned += rdp->qlen;
1355 * Next, move those callbacks still needing a grace period to
1356 * the orphanage, where some other CPU will pick them up.
1357 * Some of the callbacks might have gone partway through a grace
1358 * period, but that is too bad. They get to start over because we
1359 * cannot assume that grace periods are synchronized across CPUs.
1360 * We don't bother updating the ->nxttail[] array yet, instead
1361 * we just reset the whole thing later on.
1363 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1364 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1365 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1366 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1370 * Then move the ready-to-invoke callbacks to the orphanage,
1371 * where some other CPU will pick them up. These will not be
1372 * required to pass though another grace period: They are done.
1374 if (rdp->nxtlist != NULL) {
1375 *rsp->orphan_donetail = rdp->nxtlist;
1376 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1379 /* Finally, initialize the rcu_data structure's list to empty. */
1380 rdp->nxtlist = NULL;
1381 for (i = 0; i < RCU_NEXT_SIZE; i++)
1382 rdp->nxttail[i] = &rdp->nxtlist;
1386 * Adopt the RCU callbacks from the specified rcu_state structure's
1387 * orphanage. The caller must hold the ->onofflock.
1389 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1392 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1395 * If there is an rcu_barrier() operation in progress, then
1396 * only the task doing that operation is permitted to adopt
1397 * callbacks. To do otherwise breaks rcu_barrier() and friends
1398 * by causing them to fail to wait for the callbacks in the
1401 if (rsp->rcu_barrier_in_progress &&
1402 rsp->rcu_barrier_in_progress != current)
1405 /* Do the accounting first. */
1406 rdp->qlen_lazy += rsp->qlen_lazy;
1407 rdp->qlen += rsp->qlen;
1408 rdp->n_cbs_adopted += rsp->qlen;
1409 if (rsp->qlen_lazy != rsp->qlen)
1410 rcu_idle_count_callbacks_posted();
1415 * We do not need a memory barrier here because the only way we
1416 * can get here if there is an rcu_barrier() in flight is if
1417 * we are the task doing the rcu_barrier().
1420 /* First adopt the ready-to-invoke callbacks. */
1421 if (rsp->orphan_donelist != NULL) {
1422 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1423 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1424 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1425 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1426 rdp->nxttail[i] = rsp->orphan_donetail;
1427 rsp->orphan_donelist = NULL;
1428 rsp->orphan_donetail = &rsp->orphan_donelist;
1431 /* And then adopt the callbacks that still need a grace period. */
1432 if (rsp->orphan_nxtlist != NULL) {
1433 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1434 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1435 rsp->orphan_nxtlist = NULL;
1436 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1441 * Trace the fact that this CPU is going offline.
1443 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1445 RCU_TRACE(unsigned long mask);
1446 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1447 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1449 RCU_TRACE(mask = rdp->grpmask);
1450 trace_rcu_grace_period(rsp->name,
1451 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1456 * The CPU has been completely removed, and some other CPU is reporting
1457 * this fact from process context. Do the remainder of the cleanup,
1458 * including orphaning the outgoing CPU's RCU callbacks, and also
1459 * adopting them, if there is no _rcu_barrier() instance running.
1460 * There can only be one CPU hotplug operation at a time, so no other
1461 * CPU can be attempting to update rcu_cpu_kthread_task.
1463 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1465 unsigned long flags;
1467 int need_report = 0;
1468 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1469 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1471 /* Adjust any no-longer-needed kthreads. */
1472 rcu_stop_cpu_kthread(cpu);
1473 rcu_node_kthread_setaffinity(rnp, -1);
1475 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1477 /* Exclude any attempts to start a new grace period. */
1478 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1480 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1481 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1482 rcu_adopt_orphan_cbs(rsp);
1484 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1485 mask = rdp->grpmask; /* rnp->grplo is constant. */
1487 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1488 rnp->qsmaskinit &= ~mask;
1489 if (rnp->qsmaskinit != 0) {
1490 if (rnp != rdp->mynode)
1491 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1494 if (rnp == rdp->mynode)
1495 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1497 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1498 mask = rnp->grpmask;
1500 } while (rnp != NULL);
1503 * We still hold the leaf rcu_node structure lock here, and
1504 * irqs are still disabled. The reason for this subterfuge is
1505 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1506 * held leads to deadlock.
1508 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1510 if (need_report & RCU_OFL_TASKS_NORM_GP)
1511 rcu_report_unblock_qs_rnp(rnp, flags);
1513 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1514 if (need_report & RCU_OFL_TASKS_EXP_GP)
1515 rcu_report_exp_rnp(rsp, rnp, true);
1518 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1520 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1524 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1528 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1532 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1535 * Invoke any RCU callbacks that have made it to the end of their grace
1536 * period. Thottle as specified by rdp->blimit.
1538 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1540 unsigned long flags;
1541 struct rcu_head *next, *list, **tail;
1542 int bl, count, count_lazy, i;
1544 /* If no callbacks are ready, just return.*/
1545 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1546 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1547 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1548 need_resched(), is_idle_task(current),
1549 rcu_is_callbacks_kthread());
1554 * Extract the list of ready callbacks, disabling to prevent
1555 * races with call_rcu() from interrupt handlers.
1557 local_irq_save(flags);
1558 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1560 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1561 list = rdp->nxtlist;
1562 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1563 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1564 tail = rdp->nxttail[RCU_DONE_TAIL];
1565 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1566 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1567 rdp->nxttail[i] = &rdp->nxtlist;
1568 local_irq_restore(flags);
1570 /* Invoke callbacks. */
1571 count = count_lazy = 0;
1575 debug_rcu_head_unqueue(list);
1576 if (__rcu_reclaim(rsp->name, list))
1579 /* Stop only if limit reached and CPU has something to do. */
1580 if (++count >= bl &&
1582 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1586 local_irq_save(flags);
1587 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1588 is_idle_task(current),
1589 rcu_is_callbacks_kthread());
1591 /* Update count, and requeue any remaining callbacks. */
1593 *tail = rdp->nxtlist;
1594 rdp->nxtlist = list;
1595 for (i = 0; i < RCU_NEXT_SIZE; i++)
1596 if (&rdp->nxtlist == rdp->nxttail[i])
1597 rdp->nxttail[i] = tail;
1601 smp_mb(); /* List handling before counting for rcu_barrier(). */
1602 rdp->qlen_lazy -= count_lazy;
1604 rdp->n_cbs_invoked += count;
1606 /* Reinstate batch limit if we have worked down the excess. */
1607 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1608 rdp->blimit = blimit;
1610 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1611 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1612 rdp->qlen_last_fqs_check = 0;
1613 rdp->n_force_qs_snap = rsp->n_force_qs;
1614 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1615 rdp->qlen_last_fqs_check = rdp->qlen;
1617 local_irq_restore(flags);
1619 /* Re-invoke RCU core processing if there are callbacks remaining. */
1620 if (cpu_has_callbacks_ready_to_invoke(rdp))
1625 * Check to see if this CPU is in a non-context-switch quiescent state
1626 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1627 * Also schedule RCU core processing.
1629 * This function must be called from hardirq context. It is normally
1630 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1631 * false, there is no point in invoking rcu_check_callbacks().
1633 void rcu_check_callbacks(int cpu, int user)
1635 trace_rcu_utilization("Start scheduler-tick");
1636 increment_cpu_stall_ticks();
1637 if (user || rcu_is_cpu_rrupt_from_idle()) {
1640 * Get here if this CPU took its interrupt from user
1641 * mode or from the idle loop, and if this is not a
1642 * nested interrupt. In this case, the CPU is in
1643 * a quiescent state, so note it.
1645 * No memory barrier is required here because both
1646 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1647 * variables that other CPUs neither access nor modify,
1648 * at least not while the corresponding CPU is online.
1654 } else if (!in_softirq()) {
1657 * Get here if this CPU did not take its interrupt from
1658 * softirq, in other words, if it is not interrupting
1659 * a rcu_bh read-side critical section. This is an _bh
1660 * critical section, so note it.
1665 rcu_preempt_check_callbacks(cpu);
1666 if (rcu_pending(cpu))
1668 trace_rcu_utilization("End scheduler-tick");
1672 * Scan the leaf rcu_node structures, processing dyntick state for any that
1673 * have not yet encountered a quiescent state, using the function specified.
1674 * Also initiate boosting for any threads blocked on the root rcu_node.
1676 * The caller must have suppressed start of new grace periods.
1678 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1682 unsigned long flags;
1684 struct rcu_node *rnp;
1686 rcu_for_each_leaf_node(rsp, rnp) {
1688 raw_spin_lock_irqsave(&rnp->lock, flags);
1689 if (!rcu_gp_in_progress(rsp)) {
1690 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1693 if (rnp->qsmask == 0) {
1694 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1699 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1700 if ((rnp->qsmask & bit) != 0 &&
1701 f(per_cpu_ptr(rsp->rda, cpu)))
1706 /* rcu_report_qs_rnp() releases rnp->lock. */
1707 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1710 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1712 rnp = rcu_get_root(rsp);
1713 if (rnp->qsmask == 0) {
1714 raw_spin_lock_irqsave(&rnp->lock, flags);
1715 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1720 * Force quiescent states on reluctant CPUs, and also detect which
1721 * CPUs are in dyntick-idle mode.
1723 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1725 unsigned long flags;
1726 struct rcu_node *rnp = rcu_get_root(rsp);
1728 trace_rcu_utilization("Start fqs");
1729 if (!rcu_gp_in_progress(rsp)) {
1730 trace_rcu_utilization("End fqs");
1731 return; /* No grace period in progress, nothing to force. */
1733 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1734 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1735 trace_rcu_utilization("End fqs");
1736 return; /* Someone else is already on the job. */
1738 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1739 goto unlock_fqs_ret; /* no emergency and done recently. */
1741 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1742 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1743 if(!rcu_gp_in_progress(rsp)) {
1744 rsp->n_force_qs_ngp++;
1745 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1746 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1748 rsp->fqs_active = 1;
1749 switch (rsp->fqs_state) {
1753 break; /* grace period idle or initializing, ignore. */
1755 case RCU_SAVE_DYNTICK:
1756 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1757 break; /* So gcc recognizes the dead code. */
1759 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1761 /* Record dyntick-idle state. */
1762 force_qs_rnp(rsp, dyntick_save_progress_counter);
1763 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1764 if (rcu_gp_in_progress(rsp))
1765 rsp->fqs_state = RCU_FORCE_QS;
1770 /* Check dyntick-idle state, send IPI to laggarts. */
1771 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1772 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1774 /* Leave state in case more forcing is required. */
1776 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1779 rsp->fqs_active = 0;
1780 if (rsp->fqs_need_gp) {
1781 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1782 rsp->fqs_need_gp = 0;
1783 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1784 trace_rcu_utilization("End fqs");
1787 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1789 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1790 trace_rcu_utilization("End fqs");
1794 * This does the RCU core processing work for the specified rcu_state
1795 * and rcu_data structures. This may be called only from the CPU to
1796 * whom the rdp belongs.
1799 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1801 unsigned long flags;
1803 WARN_ON_ONCE(rdp->beenonline == 0);
1806 * If an RCU GP has gone long enough, go check for dyntick
1807 * idle CPUs and, if needed, send resched IPIs.
1809 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1810 force_quiescent_state(rsp, 1);
1813 * Advance callbacks in response to end of earlier grace
1814 * period that some other CPU ended.
1816 rcu_process_gp_end(rsp, rdp);
1818 /* Update RCU state based on any recent quiescent states. */
1819 rcu_check_quiescent_state(rsp, rdp);
1821 /* Does this CPU require a not-yet-started grace period? */
1822 if (cpu_needs_another_gp(rsp, rdp)) {
1823 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1824 rcu_start_gp(rsp, flags); /* releases above lock */
1827 /* If there are callbacks ready, invoke them. */
1828 if (cpu_has_callbacks_ready_to_invoke(rdp))
1829 invoke_rcu_callbacks(rsp, rdp);
1833 * Do RCU core processing for the current CPU.
1835 static void rcu_process_callbacks(struct softirq_action *unused)
1837 trace_rcu_utilization("Start RCU core");
1838 __rcu_process_callbacks(&rcu_sched_state,
1839 &__get_cpu_var(rcu_sched_data));
1840 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1841 rcu_preempt_process_callbacks();
1842 trace_rcu_utilization("End RCU core");
1846 * Schedule RCU callback invocation. If the specified type of RCU
1847 * does not support RCU priority boosting, just do a direct call,
1848 * otherwise wake up the per-CPU kernel kthread. Note that because we
1849 * are running on the current CPU with interrupts disabled, the
1850 * rcu_cpu_kthread_task cannot disappear out from under us.
1852 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1854 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1856 if (likely(!rsp->boost)) {
1857 rcu_do_batch(rsp, rdp);
1860 invoke_rcu_callbacks_kthread();
1863 static void invoke_rcu_core(void)
1865 raise_softirq(RCU_SOFTIRQ);
1869 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1870 struct rcu_state *rsp, bool lazy)
1872 unsigned long flags;
1873 struct rcu_data *rdp;
1875 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1876 debug_rcu_head_queue(head);
1880 smp_mb(); /* Ensure RCU update seen before callback registry. */
1883 * Opportunistically note grace-period endings and beginnings.
1884 * Note that we might see a beginning right after we see an
1885 * end, but never vice versa, since this CPU has to pass through
1886 * a quiescent state betweentimes.
1888 local_irq_save(flags);
1889 rdp = this_cpu_ptr(rsp->rda);
1891 /* Add the callback to our list. */
1896 rcu_idle_count_callbacks_posted();
1897 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1898 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1899 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1901 if (__is_kfree_rcu_offset((unsigned long)func))
1902 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1903 rdp->qlen_lazy, rdp->qlen);
1905 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1907 /* If interrupts were disabled, don't dive into RCU core. */
1908 if (irqs_disabled_flags(flags)) {
1909 local_irq_restore(flags);
1914 * Force the grace period if too many callbacks or too long waiting.
1915 * Enforce hysteresis, and don't invoke force_quiescent_state()
1916 * if some other CPU has recently done so. Also, don't bother
1917 * invoking force_quiescent_state() if the newly enqueued callback
1918 * is the only one waiting for a grace period to complete.
1920 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1922 /* Are we ignoring a completed grace period? */
1923 rcu_process_gp_end(rsp, rdp);
1924 check_for_new_grace_period(rsp, rdp);
1926 /* Start a new grace period if one not already started. */
1927 if (!rcu_gp_in_progress(rsp)) {
1928 unsigned long nestflag;
1929 struct rcu_node *rnp_root = rcu_get_root(rsp);
1931 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1932 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1934 /* Give the grace period a kick. */
1935 rdp->blimit = LONG_MAX;
1936 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1937 *rdp->nxttail[RCU_DONE_TAIL] != head)
1938 force_quiescent_state(rsp, 0);
1939 rdp->n_force_qs_snap = rsp->n_force_qs;
1940 rdp->qlen_last_fqs_check = rdp->qlen;
1942 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1943 force_quiescent_state(rsp, 1);
1944 local_irq_restore(flags);
1948 * Queue an RCU-sched callback for invocation after a grace period.
1950 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1952 __call_rcu(head, func, &rcu_sched_state, 0);
1954 EXPORT_SYMBOL_GPL(call_rcu_sched);
1957 * Queue an RCU callback for invocation after a quicker grace period.
1959 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1961 __call_rcu(head, func, &rcu_bh_state, 0);
1963 EXPORT_SYMBOL_GPL(call_rcu_bh);
1966 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1967 * any blocking grace-period wait automatically implies a grace period
1968 * if there is only one CPU online at any point time during execution
1969 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
1970 * occasionally incorrectly indicate that there are multiple CPUs online
1971 * when there was in fact only one the whole time, as this just adds
1972 * some overhead: RCU still operates correctly.
1974 * Of course, sampling num_online_cpus() with preemption enabled can
1975 * give erroneous results if there are concurrent CPU-hotplug operations.
1976 * For example, given a demonic sequence of preemptions in num_online_cpus()
1977 * and CPU-hotplug operations, there could be two or more CPUs online at
1978 * all times, but num_online_cpus() might well return one (or even zero).
1980 * However, all such demonic sequences require at least one CPU-offline
1981 * operation. Furthermore, rcu_blocking_is_gp() giving the wrong answer
1982 * is only a problem if there is an RCU read-side critical section executing
1983 * throughout. But RCU-sched and RCU-bh read-side critical sections
1984 * disable either preemption or bh, which prevents a CPU from going offline.
1985 * Therefore, the only way that rcu_blocking_is_gp() can incorrectly return
1986 * that there is only one CPU when in fact there was more than one throughout
1987 * is when there were no RCU readers in the system. If there are no
1988 * RCU readers, the grace period by definition can be of zero length,
1989 * regardless of the number of online CPUs.
1991 static inline int rcu_blocking_is_gp(void)
1993 might_sleep(); /* Check for RCU read-side critical section. */
1994 return num_online_cpus() <= 1;
1998 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2000 * Control will return to the caller some time after a full rcu-sched
2001 * grace period has elapsed, in other words after all currently executing
2002 * rcu-sched read-side critical sections have completed. These read-side
2003 * critical sections are delimited by rcu_read_lock_sched() and
2004 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2005 * local_irq_disable(), and so on may be used in place of
2006 * rcu_read_lock_sched().
2008 * This means that all preempt_disable code sequences, including NMI and
2009 * hardware-interrupt handlers, in progress on entry will have completed
2010 * before this primitive returns. However, this does not guarantee that
2011 * softirq handlers will have completed, since in some kernels, these
2012 * handlers can run in process context, and can block.
2014 * This primitive provides the guarantees made by the (now removed)
2015 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2016 * guarantees that rcu_read_lock() sections will have completed.
2017 * In "classic RCU", these two guarantees happen to be one and
2018 * the same, but can differ in realtime RCU implementations.
2020 void synchronize_sched(void)
2022 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2023 !lock_is_held(&rcu_lock_map) &&
2024 !lock_is_held(&rcu_sched_lock_map),
2025 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2026 if (rcu_blocking_is_gp())
2028 wait_rcu_gp(call_rcu_sched);
2030 EXPORT_SYMBOL_GPL(synchronize_sched);
2033 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2035 * Control will return to the caller some time after a full rcu_bh grace
2036 * period has elapsed, in other words after all currently executing rcu_bh
2037 * read-side critical sections have completed. RCU read-side critical
2038 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2039 * and may be nested.
2041 void synchronize_rcu_bh(void)
2043 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2044 !lock_is_held(&rcu_lock_map) &&
2045 !lock_is_held(&rcu_sched_lock_map),
2046 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2047 if (rcu_blocking_is_gp())
2049 wait_rcu_gp(call_rcu_bh);
2051 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2053 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2054 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2056 static int synchronize_sched_expedited_cpu_stop(void *data)
2059 * There must be a full memory barrier on each affected CPU
2060 * between the time that try_stop_cpus() is called and the
2061 * time that it returns.
2063 * In the current initial implementation of cpu_stop, the
2064 * above condition is already met when the control reaches
2065 * this point and the following smp_mb() is not strictly
2066 * necessary. Do smp_mb() anyway for documentation and
2067 * robustness against future implementation changes.
2069 smp_mb(); /* See above comment block. */
2074 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2076 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2077 * approach to force the grace period to end quickly. This consumes
2078 * significant time on all CPUs and is unfriendly to real-time workloads,
2079 * so is thus not recommended for any sort of common-case code. In fact,
2080 * if you are using synchronize_sched_expedited() in a loop, please
2081 * restructure your code to batch your updates, and then use a single
2082 * synchronize_sched() instead.
2084 * Note that it is illegal to call this function while holding any lock
2085 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2086 * to call this function from a CPU-hotplug notifier. Failing to observe
2087 * these restriction will result in deadlock.
2089 * This implementation can be thought of as an application of ticket
2090 * locking to RCU, with sync_sched_expedited_started and
2091 * sync_sched_expedited_done taking on the roles of the halves
2092 * of the ticket-lock word. Each task atomically increments
2093 * sync_sched_expedited_started upon entry, snapshotting the old value,
2094 * then attempts to stop all the CPUs. If this succeeds, then each
2095 * CPU will have executed a context switch, resulting in an RCU-sched
2096 * grace period. We are then done, so we use atomic_cmpxchg() to
2097 * update sync_sched_expedited_done to match our snapshot -- but
2098 * only if someone else has not already advanced past our snapshot.
2100 * On the other hand, if try_stop_cpus() fails, we check the value
2101 * of sync_sched_expedited_done. If it has advanced past our
2102 * initial snapshot, then someone else must have forced a grace period
2103 * some time after we took our snapshot. In this case, our work is
2104 * done for us, and we can simply return. Otherwise, we try again,
2105 * but keep our initial snapshot for purposes of checking for someone
2106 * doing our work for us.
2108 * If we fail too many times in a row, we fall back to synchronize_sched().
2110 void synchronize_sched_expedited(void)
2112 int firstsnap, s, snap, trycount = 0;
2114 /* Note that atomic_inc_return() implies full memory barrier. */
2115 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2117 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2120 * Each pass through the following loop attempts to force a
2121 * context switch on each CPU.
2123 while (try_stop_cpus(cpu_online_mask,
2124 synchronize_sched_expedited_cpu_stop,
2128 /* No joy, try again later. Or just synchronize_sched(). */
2129 if (trycount++ < 10)
2130 udelay(trycount * num_online_cpus());
2132 synchronize_sched();
2136 /* Check to see if someone else did our work for us. */
2137 s = atomic_read(&sync_sched_expedited_done);
2138 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2139 smp_mb(); /* ensure test happens before caller kfree */
2144 * Refetching sync_sched_expedited_started allows later
2145 * callers to piggyback on our grace period. We subtract
2146 * 1 to get the same token that the last incrementer got.
2147 * We retry after they started, so our grace period works
2148 * for them, and they started after our first try, so their
2149 * grace period works for us.
2152 snap = atomic_read(&sync_sched_expedited_started);
2153 smp_mb(); /* ensure read is before try_stop_cpus(). */
2157 * Everyone up to our most recent fetch is covered by our grace
2158 * period. Update the counter, but only if our work is still
2159 * relevant -- which it won't be if someone who started later
2160 * than we did beat us to the punch.
2163 s = atomic_read(&sync_sched_expedited_done);
2164 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2165 smp_mb(); /* ensure test happens before caller kfree */
2168 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2172 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2175 * Check to see if there is any immediate RCU-related work to be done
2176 * by the current CPU, for the specified type of RCU, returning 1 if so.
2177 * The checks are in order of increasing expense: checks that can be
2178 * carried out against CPU-local state are performed first. However,
2179 * we must check for CPU stalls first, else we might not get a chance.
2181 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2183 struct rcu_node *rnp = rdp->mynode;
2185 rdp->n_rcu_pending++;
2187 /* Check for CPU stalls, if enabled. */
2188 check_cpu_stall(rsp, rdp);
2190 /* Is the RCU core waiting for a quiescent state from this CPU? */
2191 if (rcu_scheduler_fully_active &&
2192 rdp->qs_pending && !rdp->passed_quiesce) {
2195 * If force_quiescent_state() coming soon and this CPU
2196 * needs a quiescent state, and this is either RCU-sched
2197 * or RCU-bh, force a local reschedule.
2199 rdp->n_rp_qs_pending++;
2200 if (!rdp->preemptible &&
2201 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2204 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2205 rdp->n_rp_report_qs++;
2209 /* Does this CPU have callbacks ready to invoke? */
2210 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2211 rdp->n_rp_cb_ready++;
2215 /* Has RCU gone idle with this CPU needing another grace period? */
2216 if (cpu_needs_another_gp(rsp, rdp)) {
2217 rdp->n_rp_cpu_needs_gp++;
2221 /* Has another RCU grace period completed? */
2222 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2223 rdp->n_rp_gp_completed++;
2227 /* Has a new RCU grace period started? */
2228 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2229 rdp->n_rp_gp_started++;
2233 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2234 if (rcu_gp_in_progress(rsp) &&
2235 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2236 rdp->n_rp_need_fqs++;
2241 rdp->n_rp_need_nothing++;
2246 * Check to see if there is any immediate RCU-related work to be done
2247 * by the current CPU, returning 1 if so. This function is part of the
2248 * RCU implementation; it is -not- an exported member of the RCU API.
2250 static int rcu_pending(int cpu)
2252 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
2253 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2254 rcu_preempt_pending(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 /* RCU callbacks either ready or pending? */
2265 return per_cpu(rcu_sched_data, cpu).nxtlist ||
2266 per_cpu(rcu_bh_data, cpu).nxtlist ||
2267 rcu_preempt_cpu_has_callbacks(cpu);
2271 * RCU callback function for _rcu_barrier(). If we are last, wake
2272 * up the task executing _rcu_barrier().
2274 static void rcu_barrier_callback(struct rcu_head *notused)
2276 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2277 complete(&rcu_barrier_completion);
2281 * Called with preemption disabled, and from cross-cpu IRQ context.
2283 static void rcu_barrier_func(void *type)
2285 int cpu = smp_processor_id();
2286 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
2287 struct rcu_state *rsp = type;
2289 atomic_inc(&rcu_barrier_cpu_count);
2290 rsp->call(head, rcu_barrier_callback);
2294 * Orchestrate the specified type of RCU barrier, waiting for all
2295 * RCU callbacks of the specified type to complete.
2297 static void _rcu_barrier(struct rcu_state *rsp)
2300 unsigned long flags;
2301 struct rcu_data *rdp;
2304 init_rcu_head_on_stack(&rh);
2306 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2307 mutex_lock(&rcu_barrier_mutex);
2309 smp_mb(); /* Prevent any prior operations from leaking in. */
2312 * Initialize the count to one rather than to zero in order to
2313 * avoid a too-soon return to zero in case of a short grace period
2314 * (or preemption of this task). Also flag this task as doing
2315 * an rcu_barrier(). This will prevent anyone else from adopting
2316 * orphaned callbacks, which could cause otherwise failure if a
2317 * CPU went offline and quickly came back online. To see this,
2318 * consider the following sequence of events:
2320 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2321 * 2. CPU 1 goes offline, orphaning its callbacks.
2322 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2323 * 4. CPU 1 comes back online.
2324 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2325 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2326 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2328 init_completion(&rcu_barrier_completion);
2329 atomic_set(&rcu_barrier_cpu_count, 1);
2330 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2331 rsp->rcu_barrier_in_progress = current;
2332 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2335 * Force every CPU with callbacks to register a new callback
2336 * that will tell us when all the preceding callbacks have
2337 * been invoked. If an offline CPU has callbacks, wait for
2338 * it to either come back online or to finish orphaning those
2341 for_each_possible_cpu(cpu) {
2343 rdp = per_cpu_ptr(rsp->rda, cpu);
2344 if (cpu_is_offline(cpu)) {
2346 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2347 schedule_timeout_interruptible(1);
2348 } else if (ACCESS_ONCE(rdp->qlen)) {
2349 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2357 * Now that all online CPUs have rcu_barrier_callback() callbacks
2358 * posted, we can adopt all of the orphaned callbacks and place
2359 * an rcu_barrier_callback() callback after them. When that is done,
2360 * we are guaranteed to have an rcu_barrier_callback() callback
2361 * following every callback that could possibly have been
2362 * registered before _rcu_barrier() was called.
2364 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2365 rcu_adopt_orphan_cbs(rsp);
2366 rsp->rcu_barrier_in_progress = NULL;
2367 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2368 atomic_inc(&rcu_barrier_cpu_count);
2369 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2370 rsp->call(&rh, rcu_barrier_callback);
2373 * Now that we have an rcu_barrier_callback() callback on each
2374 * CPU, and thus each counted, remove the initial count.
2376 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2377 complete(&rcu_barrier_completion);
2379 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2380 wait_for_completion(&rcu_barrier_completion);
2382 /* Other rcu_barrier() invocations can now safely proceed. */
2383 mutex_unlock(&rcu_barrier_mutex);
2385 destroy_rcu_head_on_stack(&rh);
2389 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2391 void rcu_barrier_bh(void)
2393 _rcu_barrier(&rcu_bh_state);
2395 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2398 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2400 void rcu_barrier_sched(void)
2402 _rcu_barrier(&rcu_sched_state);
2404 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2407 * Do boot-time initialization of a CPU's per-CPU RCU data.
2410 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2412 unsigned long flags;
2414 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2415 struct rcu_node *rnp = rcu_get_root(rsp);
2417 /* Set up local state, ensuring consistent view of global state. */
2418 raw_spin_lock_irqsave(&rnp->lock, flags);
2419 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2420 rdp->nxtlist = NULL;
2421 for (i = 0; i < RCU_NEXT_SIZE; i++)
2422 rdp->nxttail[i] = &rdp->nxtlist;
2425 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2426 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2427 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2430 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2434 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2435 * offline event can be happening at a given time. Note also that we
2436 * can accept some slop in the rsp->completed access due to the fact
2437 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2439 static void __cpuinit
2440 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2442 unsigned long flags;
2444 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2445 struct rcu_node *rnp = rcu_get_root(rsp);
2447 /* Set up local state, ensuring consistent view of global state. */
2448 raw_spin_lock_irqsave(&rnp->lock, flags);
2449 rdp->beenonline = 1; /* We have now been online. */
2450 rdp->preemptible = preemptible;
2451 rdp->qlen_last_fqs_check = 0;
2452 rdp->n_force_qs_snap = rsp->n_force_qs;
2453 rdp->blimit = blimit;
2454 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2455 atomic_set(&rdp->dynticks->dynticks,
2456 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2457 rcu_prepare_for_idle_init(cpu);
2458 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2461 * A new grace period might start here. If so, we won't be part
2462 * of it, but that is OK, as we are currently in a quiescent state.
2465 /* Exclude any attempts to start a new GP on large systems. */
2466 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2468 /* Add CPU to rcu_node bitmasks. */
2470 mask = rdp->grpmask;
2472 /* Exclude any attempts to start a new GP on small systems. */
2473 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2474 rnp->qsmaskinit |= mask;
2475 mask = rnp->grpmask;
2476 if (rnp == rdp->mynode) {
2478 * If there is a grace period in progress, we will
2479 * set up to wait for it next time we run the
2482 rdp->gpnum = rnp->completed;
2483 rdp->completed = rnp->completed;
2484 rdp->passed_quiesce = 0;
2485 rdp->qs_pending = 0;
2486 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2487 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2489 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2491 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2493 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2496 static void __cpuinit rcu_prepare_cpu(int cpu)
2498 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2499 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2500 rcu_preempt_init_percpu_data(cpu);
2504 * Handle CPU online/offline notification events.
2506 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2507 unsigned long action, void *hcpu)
2509 long cpu = (long)hcpu;
2510 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2511 struct rcu_node *rnp = rdp->mynode;
2513 trace_rcu_utilization("Start CPU hotplug");
2515 case CPU_UP_PREPARE:
2516 case CPU_UP_PREPARE_FROZEN:
2517 rcu_prepare_cpu(cpu);
2518 rcu_prepare_kthreads(cpu);
2521 case CPU_DOWN_FAILED:
2522 rcu_node_kthread_setaffinity(rnp, -1);
2523 rcu_cpu_kthread_setrt(cpu, 1);
2525 case CPU_DOWN_PREPARE:
2526 rcu_node_kthread_setaffinity(rnp, cpu);
2527 rcu_cpu_kthread_setrt(cpu, 0);
2530 case CPU_DYING_FROZEN:
2532 * The whole machine is "stopped" except this CPU, so we can
2533 * touch any data without introducing corruption. We send the
2534 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2536 rcu_cleanup_dying_cpu(&rcu_bh_state);
2537 rcu_cleanup_dying_cpu(&rcu_sched_state);
2538 rcu_preempt_cleanup_dying_cpu();
2539 rcu_cleanup_after_idle(cpu);
2542 case CPU_DEAD_FROZEN:
2543 case CPU_UP_CANCELED:
2544 case CPU_UP_CANCELED_FROZEN:
2545 rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2546 rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2547 rcu_preempt_cleanup_dead_cpu(cpu);
2552 trace_rcu_utilization("End CPU hotplug");
2557 * This function is invoked towards the end of the scheduler's initialization
2558 * process. Before this is called, the idle task might contain
2559 * RCU read-side critical sections (during which time, this idle
2560 * task is booting the system). After this function is called, the
2561 * idle tasks are prohibited from containing RCU read-side critical
2562 * sections. This function also enables RCU lockdep checking.
2564 void rcu_scheduler_starting(void)
2566 WARN_ON(num_online_cpus() != 1);
2567 WARN_ON(nr_context_switches() > 0);
2568 rcu_scheduler_active = 1;
2572 * Compute the per-level fanout, either using the exact fanout specified
2573 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2575 #ifdef CONFIG_RCU_FANOUT_EXACT
2576 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2580 for (i = rcu_num_lvls - 1; i > 0; i--)
2581 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2582 rsp->levelspread[0] = rcu_fanout_leaf;
2584 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2585 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2592 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2593 ccur = rsp->levelcnt[i];
2594 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2598 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2601 * Helper function for rcu_init() that initializes one rcu_state structure.
2603 static void __init rcu_init_one(struct rcu_state *rsp,
2604 struct rcu_data __percpu *rda)
2606 static char *buf[] = { "rcu_node_level_0",
2609 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2613 struct rcu_node *rnp;
2615 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2617 /* Initialize the level-tracking arrays. */
2619 for (i = 0; i < rcu_num_lvls; i++)
2620 rsp->levelcnt[i] = num_rcu_lvl[i];
2621 for (i = 1; i < rcu_num_lvls; i++)
2622 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2623 rcu_init_levelspread(rsp);
2625 /* Initialize the elements themselves, starting from the leaves. */
2627 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2628 cpustride *= rsp->levelspread[i];
2629 rnp = rsp->level[i];
2630 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2631 raw_spin_lock_init(&rnp->lock);
2632 lockdep_set_class_and_name(&rnp->lock,
2633 &rcu_node_class[i], buf[i]);
2636 rnp->qsmaskinit = 0;
2637 rnp->grplo = j * cpustride;
2638 rnp->grphi = (j + 1) * cpustride - 1;
2639 if (rnp->grphi >= NR_CPUS)
2640 rnp->grphi = NR_CPUS - 1;
2646 rnp->grpnum = j % rsp->levelspread[i - 1];
2647 rnp->grpmask = 1UL << rnp->grpnum;
2648 rnp->parent = rsp->level[i - 1] +
2649 j / rsp->levelspread[i - 1];
2652 INIT_LIST_HEAD(&rnp->blkd_tasks);
2657 rnp = rsp->level[rcu_num_lvls - 1];
2658 for_each_possible_cpu(i) {
2659 while (i > rnp->grphi)
2661 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2662 rcu_boot_init_percpu_data(i, rsp);
2667 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2668 * replace the definitions in rcutree.h because those are needed to size
2669 * the ->node array in the rcu_state structure.
2671 static void __init rcu_init_geometry(void)
2676 int rcu_capacity[MAX_RCU_LVLS + 1];
2678 /* If the compile-time values are accurate, just leave. */
2679 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2683 * Compute number of nodes that can be handled an rcu_node tree
2684 * with the given number of levels. Setting rcu_capacity[0] makes
2685 * some of the arithmetic easier.
2687 rcu_capacity[0] = 1;
2688 rcu_capacity[1] = rcu_fanout_leaf;
2689 for (i = 2; i <= MAX_RCU_LVLS; i++)
2690 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2693 * The boot-time rcu_fanout_leaf parameter is only permitted
2694 * to increase the leaf-level fanout, not decrease it. Of course,
2695 * the leaf-level fanout cannot exceed the number of bits in
2696 * the rcu_node masks. Finally, the tree must be able to accommodate
2697 * the configured number of CPUs. Complain and fall back to the
2698 * compile-time values if these limits are exceeded.
2700 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2701 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2702 n > rcu_capacity[MAX_RCU_LVLS]) {
2707 /* Calculate the number of rcu_nodes at each level of the tree. */
2708 for (i = 1; i <= MAX_RCU_LVLS; i++)
2709 if (n <= rcu_capacity[i]) {
2710 for (j = 0; j <= i; j++)
2712 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2714 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2719 /* Calculate the total number of rcu_node structures. */
2721 for (i = 0; i <= MAX_RCU_LVLS; i++)
2722 rcu_num_nodes += num_rcu_lvl[i];
2726 void __init rcu_init(void)
2730 rcu_bootup_announce();
2731 rcu_init_geometry();
2732 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2733 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2734 __rcu_init_preempt();
2735 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2738 * We don't need protection against CPU-hotplug here because
2739 * this is called early in boot, before either interrupts
2740 * or the scheduler are operational.
2742 cpu_notifier(rcu_cpu_notify, 0);
2743 for_each_online_cpu(cpu)
2744 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2745 check_cpu_stall_init();
2748 #include "rcutree_plugin.h"