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>
55 #include <trace/events/rcu.h>
59 /* Data structures. */
61 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
63 #define RCU_STATE_INITIALIZER(structname) { \
64 .level = { &structname##_state.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .fqs_state = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched);
83 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
85 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh);
86 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
88 static struct rcu_state *rcu_state;
91 * The rcu_scheduler_active variable transitions from zero to one just
92 * before the first task is spawned. So when this variable is zero, RCU
93 * can assume that there is but one task, allowing RCU to (for example)
94 * optimized synchronize_sched() to a simple barrier(). When this variable
95 * is one, RCU must actually do all the hard work required to detect real
96 * grace periods. This variable is also used to suppress boot-time false
97 * positives from lockdep-RCU error checking.
99 int rcu_scheduler_active __read_mostly;
100 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
103 * The rcu_scheduler_fully_active variable transitions from zero to one
104 * during the early_initcall() processing, which is after the scheduler
105 * is capable of creating new tasks. So RCU processing (for example,
106 * creating tasks for RCU priority boosting) must be delayed until after
107 * rcu_scheduler_fully_active transitions from zero to one. We also
108 * currently delay invocation of any RCU callbacks until after this point.
110 * It might later prove better for people registering RCU callbacks during
111 * early boot to take responsibility for these callbacks, but one step at
114 static int rcu_scheduler_fully_active __read_mostly;
116 #ifdef CONFIG_RCU_BOOST
119 * Control variables for per-CPU and per-rcu_node kthreads. These
120 * handle all flavors of RCU.
122 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
123 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
124 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
126 DEFINE_PER_CPU(char, rcu_cpu_has_work);
128 #endif /* #ifdef CONFIG_RCU_BOOST */
130 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
131 static void invoke_rcu_core(void);
132 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
135 * Track the rcutorture test sequence number and the update version
136 * number within a given test. The rcutorture_testseq is incremented
137 * on every rcutorture module load and unload, so has an odd value
138 * when a test is running. The rcutorture_vernum is set to zero
139 * when rcutorture starts and is incremented on each rcutorture update.
140 * These variables enable correlating rcutorture output with the
141 * RCU tracing information.
143 unsigned long rcutorture_testseq;
144 unsigned long rcutorture_vernum;
147 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
148 * permit this function to be invoked without holding the root rcu_node
149 * structure's ->lock, but of course results can be subject to change.
151 static int rcu_gp_in_progress(struct rcu_state *rsp)
153 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
157 * Note a quiescent state. Because we do not need to know
158 * how many quiescent states passed, just if there was at least
159 * one since the start of the grace period, this just sets a flag.
160 * The caller must have disabled preemption.
162 void rcu_sched_qs(int cpu)
164 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
166 rdp->passed_quiesce_gpnum = rdp->gpnum;
168 if (rdp->passed_quiesce == 0)
169 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
170 rdp->passed_quiesce = 1;
173 void rcu_bh_qs(int cpu)
175 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
177 rdp->passed_quiesce_gpnum = rdp->gpnum;
179 if (rdp->passed_quiesce == 0)
180 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
181 rdp->passed_quiesce = 1;
185 * Note a context switch. This is a quiescent state for RCU-sched,
186 * and requires special handling for preemptible RCU.
187 * The caller must have disabled preemption.
189 void rcu_note_context_switch(int cpu)
191 trace_rcu_utilization("Start context switch");
193 rcu_preempt_note_context_switch(cpu);
194 trace_rcu_utilization("End context switch");
196 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
198 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
199 .dynticks_nesting = DYNTICK_TASK_NESTING,
200 .dynticks = ATOMIC_INIT(1),
203 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
204 static int qhimark = 10000; /* If this many pending, ignore blimit. */
205 static int qlowmark = 100; /* Once only this many pending, use blimit. */
207 module_param(blimit, int, 0);
208 module_param(qhimark, int, 0);
209 module_param(qlowmark, int, 0);
211 int rcu_cpu_stall_suppress __read_mostly;
212 module_param(rcu_cpu_stall_suppress, int, 0644);
214 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
215 static int rcu_pending(int cpu);
218 * Return the number of RCU-sched batches processed thus far for debug & stats.
220 long rcu_batches_completed_sched(void)
222 return rcu_sched_state.completed;
224 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
227 * Return the number of RCU BH batches processed thus far for debug & stats.
229 long rcu_batches_completed_bh(void)
231 return rcu_bh_state.completed;
233 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
236 * Force a quiescent state for RCU BH.
238 void rcu_bh_force_quiescent_state(void)
240 force_quiescent_state(&rcu_bh_state, 0);
242 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
245 * Record the number of times rcutorture tests have been initiated and
246 * terminated. This information allows the debugfs tracing stats to be
247 * correlated to the rcutorture messages, even when the rcutorture module
248 * is being repeatedly loaded and unloaded. In other words, we cannot
249 * store this state in rcutorture itself.
251 void rcutorture_record_test_transition(void)
253 rcutorture_testseq++;
254 rcutorture_vernum = 0;
256 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
259 * Record the number of writer passes through the current rcutorture test.
260 * This is also used to correlate debugfs tracing stats with the rcutorture
263 void rcutorture_record_progress(unsigned long vernum)
267 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
270 * Force a quiescent state for RCU-sched.
272 void rcu_sched_force_quiescent_state(void)
274 force_quiescent_state(&rcu_sched_state, 0);
276 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
279 * Does the CPU have callbacks ready to be invoked?
282 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
284 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
288 * Does the current CPU require a yet-as-unscheduled grace period?
291 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
293 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
297 * Return the root node of the specified rcu_state structure.
299 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
301 return &rsp->node[0];
307 * If the specified CPU is offline, tell the caller that it is in
308 * a quiescent state. Otherwise, whack it with a reschedule IPI.
309 * Grace periods can end up waiting on an offline CPU when that
310 * CPU is in the process of coming online -- it will be added to the
311 * rcu_node bitmasks before it actually makes it online. The same thing
312 * can happen while a CPU is in the process of coming online. Because this
313 * race is quite rare, we check for it after detecting that the grace
314 * period has been delayed rather than checking each and every CPU
315 * each and every time we start a new grace period.
317 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
320 * If the CPU is offline, it is in a quiescent state. We can
321 * trust its state not to change because interrupts are disabled.
323 if (cpu_is_offline(rdp->cpu)) {
324 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
330 * The CPU is online, so send it a reschedule IPI. This forces
331 * it through the scheduler, and (inefficiently) also handles cases
332 * where idle loops fail to inform RCU about the CPU being idle.
334 if (rdp->cpu != smp_processor_id())
335 smp_send_reschedule(rdp->cpu);
342 #endif /* #ifdef CONFIG_SMP */
345 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
347 * If the new value of the ->dynticks_nesting counter now is zero,
348 * we really have entered idle, and must do the appropriate accounting.
349 * The caller must have disabled interrupts.
351 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
353 if (rdtp->dynticks_nesting) {
354 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
357 trace_rcu_dyntick("Start", oldval, rdtp->dynticks_nesting);
358 if (!is_idle_task(current)) {
359 struct task_struct *idle = idle_task(smp_processor_id());
361 trace_rcu_dyntick("Error on entry: not idle task",
362 oldval, rdtp->dynticks_nesting);
363 ftrace_dump(DUMP_ALL);
364 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
365 current->pid, current->comm,
366 idle->pid, idle->comm); /* must be idle task! */
368 rcu_prepare_for_idle(smp_processor_id());
369 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
370 smp_mb__before_atomic_inc(); /* See above. */
371 atomic_inc(&rdtp->dynticks);
372 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
373 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
377 * rcu_idle_enter - inform RCU that current CPU is entering idle
379 * Enter idle mode, in other words, -leave- the mode in which RCU
380 * read-side critical sections can occur. (Though RCU read-side
381 * critical sections can occur in irq handlers in idle, a possibility
382 * handled by irq_enter() and irq_exit().)
384 * We crowbar the ->dynticks_nesting field to zero to allow for
385 * the possibility of usermode upcalls having messed up our count
386 * of interrupt nesting level during the prior busy period.
388 void rcu_idle_enter(void)
392 struct rcu_dynticks *rdtp;
394 local_irq_save(flags);
395 rdtp = &__get_cpu_var(rcu_dynticks);
396 oldval = rdtp->dynticks_nesting;
397 rdtp->dynticks_nesting = 0;
398 rcu_idle_enter_common(rdtp, oldval);
399 local_irq_restore(flags);
403 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
405 * Exit from an interrupt handler, which might possibly result in entering
406 * idle mode, in other words, leaving the mode in which read-side critical
407 * sections can occur.
409 * This code assumes that the idle loop never does anything that might
410 * result in unbalanced calls to irq_enter() and irq_exit(). If your
411 * architecture violates this assumption, RCU will give you what you
412 * deserve, good and hard. But very infrequently and irreproducibly.
414 * Use things like work queues to work around this limitation.
416 * You have been warned.
418 void rcu_irq_exit(void)
422 struct rcu_dynticks *rdtp;
424 local_irq_save(flags);
425 rdtp = &__get_cpu_var(rcu_dynticks);
426 oldval = rdtp->dynticks_nesting;
427 rdtp->dynticks_nesting--;
428 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
429 rcu_idle_enter_common(rdtp, oldval);
430 local_irq_restore(flags);
434 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
436 * If the new value of the ->dynticks_nesting counter was previously zero,
437 * we really have exited idle, and must do the appropriate accounting.
438 * The caller must have disabled interrupts.
440 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
443 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
446 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
447 atomic_inc(&rdtp->dynticks);
448 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
449 smp_mb__after_atomic_inc(); /* See above. */
450 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
451 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
452 if (!is_idle_task(current)) {
453 struct task_struct *idle = idle_task(smp_processor_id());
455 trace_rcu_dyntick("Error on exit: not idle task",
456 oldval, rdtp->dynticks_nesting);
457 ftrace_dump(DUMP_ALL);
458 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
459 current->pid, current->comm,
460 idle->pid, idle->comm); /* must be idle task! */
465 * rcu_idle_exit - inform RCU that current CPU is leaving idle
467 * Exit idle mode, in other words, -enter- the mode in which RCU
468 * read-side critical sections can occur.
470 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NESTING to
471 * allow for the possibility of usermode upcalls messing up our count
472 * of interrupt nesting level during the busy period that is just
475 void rcu_idle_exit(void)
478 struct rcu_dynticks *rdtp;
481 local_irq_save(flags);
482 rdtp = &__get_cpu_var(rcu_dynticks);
483 oldval = rdtp->dynticks_nesting;
484 WARN_ON_ONCE(oldval != 0);
485 rdtp->dynticks_nesting = DYNTICK_TASK_NESTING;
486 rcu_idle_exit_common(rdtp, oldval);
487 local_irq_restore(flags);
491 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
493 * Enter an interrupt handler, which might possibly result in exiting
494 * idle mode, in other words, entering the mode in which read-side critical
495 * sections can occur.
497 * Note that the Linux kernel is fully capable of entering an interrupt
498 * handler that it never exits, for example when doing upcalls to
499 * user mode! This code assumes that the idle loop never does upcalls to
500 * user mode. If your architecture does do upcalls from the idle loop (or
501 * does anything else that results in unbalanced calls to the irq_enter()
502 * and irq_exit() functions), RCU will give you what you deserve, good
503 * and hard. But very infrequently and irreproducibly.
505 * Use things like work queues to work around this limitation.
507 * You have been warned.
509 void rcu_irq_enter(void)
512 struct rcu_dynticks *rdtp;
515 local_irq_save(flags);
516 rdtp = &__get_cpu_var(rcu_dynticks);
517 oldval = rdtp->dynticks_nesting;
518 rdtp->dynticks_nesting++;
519 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
520 rcu_idle_exit_common(rdtp, oldval);
521 local_irq_restore(flags);
525 * rcu_nmi_enter - inform RCU of entry to NMI context
527 * If the CPU was idle with dynamic ticks active, and there is no
528 * irq handler running, this updates rdtp->dynticks_nmi to let the
529 * RCU grace-period handling know that the CPU is active.
531 void rcu_nmi_enter(void)
533 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
535 if (rdtp->dynticks_nmi_nesting == 0 &&
536 (atomic_read(&rdtp->dynticks) & 0x1))
538 rdtp->dynticks_nmi_nesting++;
539 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
540 atomic_inc(&rdtp->dynticks);
541 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
542 smp_mb__after_atomic_inc(); /* See above. */
543 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
547 * rcu_nmi_exit - inform RCU of exit from NMI context
549 * If the CPU was idle with dynamic ticks active, and there is no
550 * irq handler running, this updates rdtp->dynticks_nmi to let the
551 * RCU grace-period handling know that the CPU is no longer active.
553 void rcu_nmi_exit(void)
555 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
557 if (rdtp->dynticks_nmi_nesting == 0 ||
558 --rdtp->dynticks_nmi_nesting != 0)
560 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
561 smp_mb__before_atomic_inc(); /* See above. */
562 atomic_inc(&rdtp->dynticks);
563 smp_mb__after_atomic_inc(); /* Force delay to next write. */
564 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
567 #ifdef CONFIG_PROVE_RCU
570 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
572 * If the current CPU is in its idle loop and is neither in an interrupt
573 * or NMI handler, return true.
575 int rcu_is_cpu_idle(void)
580 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
584 EXPORT_SYMBOL(rcu_is_cpu_idle);
586 #endif /* #ifdef CONFIG_PROVE_RCU */
589 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
591 * If the current CPU is idle or running at a first-level (not nested)
592 * interrupt from idle, return true. The caller must have at least
593 * disabled preemption.
595 int rcu_is_cpu_rrupt_from_idle(void)
597 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
603 * Snapshot the specified CPU's dynticks counter so that we can later
604 * credit them with an implicit quiescent state. Return 1 if this CPU
605 * is in dynticks idle mode, which is an extended quiescent state.
607 static int dyntick_save_progress_counter(struct rcu_data *rdp)
609 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
614 * Return true if the specified CPU has passed through a quiescent
615 * state by virtue of being in or having passed through an dynticks
616 * idle state since the last call to dyntick_save_progress_counter()
619 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
624 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
625 snap = (unsigned int)rdp->dynticks_snap;
628 * If the CPU passed through or entered a dynticks idle phase with
629 * no active irq/NMI handlers, then we can safely pretend that the CPU
630 * already acknowledged the request to pass through a quiescent
631 * state. Either way, that CPU cannot possibly be in an RCU
632 * read-side critical section that started before the beginning
633 * of the current RCU grace period.
635 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
636 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
641 /* Go check for the CPU being offline. */
642 return rcu_implicit_offline_qs(rdp);
645 #endif /* #ifdef CONFIG_SMP */
647 int rcu_cpu_stall_suppress __read_mostly;
649 static void record_gp_stall_check_time(struct rcu_state *rsp)
651 rsp->gp_start = jiffies;
652 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
655 static void print_other_cpu_stall(struct rcu_state *rsp)
661 struct rcu_node *rnp = rcu_get_root(rsp);
663 /* Only let one CPU complain about others per time interval. */
665 raw_spin_lock_irqsave(&rnp->lock, flags);
666 delta = jiffies - rsp->jiffies_stall;
667 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
668 raw_spin_unlock_irqrestore(&rnp->lock, flags);
671 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
674 * Now rat on any tasks that got kicked up to the root rcu_node
675 * due to CPU offlining.
677 ndetected = rcu_print_task_stall(rnp);
678 raw_spin_unlock_irqrestore(&rnp->lock, flags);
681 * OK, time to rat on our buddy...
682 * See Documentation/RCU/stallwarn.txt for info on how to debug
683 * RCU CPU stall warnings.
685 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
687 rcu_for_each_leaf_node(rsp, rnp) {
688 raw_spin_lock_irqsave(&rnp->lock, flags);
689 ndetected += rcu_print_task_stall(rnp);
690 raw_spin_unlock_irqrestore(&rnp->lock, flags);
691 if (rnp->qsmask == 0)
693 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
694 if (rnp->qsmask & (1UL << cpu)) {
695 printk(" %d", rnp->grplo + cpu);
699 printk("} (detected by %d, t=%ld jiffies)\n",
700 smp_processor_id(), (long)(jiffies - rsp->gp_start));
702 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
703 else if (!trigger_all_cpu_backtrace())
706 /* If so configured, complain about tasks blocking the grace period. */
708 rcu_print_detail_task_stall(rsp);
710 force_quiescent_state(rsp, 0); /* Kick them all. */
713 static void print_cpu_stall(struct rcu_state *rsp)
716 struct rcu_node *rnp = rcu_get_root(rsp);
719 * OK, time to rat on ourselves...
720 * See Documentation/RCU/stallwarn.txt for info on how to debug
721 * RCU CPU stall warnings.
723 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
724 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
725 if (!trigger_all_cpu_backtrace())
728 raw_spin_lock_irqsave(&rnp->lock, flags);
729 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
731 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
732 raw_spin_unlock_irqrestore(&rnp->lock, flags);
734 set_need_resched(); /* kick ourselves to get things going. */
737 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
741 struct rcu_node *rnp;
743 if (rcu_cpu_stall_suppress)
745 j = ACCESS_ONCE(jiffies);
746 js = ACCESS_ONCE(rsp->jiffies_stall);
748 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
750 /* We haven't checked in, so go dump stack. */
751 print_cpu_stall(rsp);
753 } else if (rcu_gp_in_progress(rsp) &&
754 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
756 /* They had a few time units to dump stack, so complain. */
757 print_other_cpu_stall(rsp);
761 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
763 rcu_cpu_stall_suppress = 1;
768 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
770 * Set the stall-warning timeout way off into the future, thus preventing
771 * any RCU CPU stall-warning messages from appearing in the current set of
774 * The caller must disable hard irqs.
776 void rcu_cpu_stall_reset(void)
778 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
779 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
780 rcu_preempt_stall_reset();
783 static struct notifier_block rcu_panic_block = {
784 .notifier_call = rcu_panic,
787 static void __init check_cpu_stall_init(void)
789 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
793 * Update CPU-local rcu_data state to record the newly noticed grace period.
794 * This is used both when we started the grace period and when we notice
795 * that someone else started the grace period. The caller must hold the
796 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
797 * and must have irqs disabled.
799 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
801 if (rdp->gpnum != rnp->gpnum) {
803 * If the current grace period is waiting for this CPU,
804 * set up to detect a quiescent state, otherwise don't
805 * go looking for one.
807 rdp->gpnum = rnp->gpnum;
808 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
809 if (rnp->qsmask & rdp->grpmask) {
811 rdp->passed_quiesce = 0;
817 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
820 struct rcu_node *rnp;
822 local_irq_save(flags);
824 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
825 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
826 local_irq_restore(flags);
829 __note_new_gpnum(rsp, rnp, rdp);
830 raw_spin_unlock_irqrestore(&rnp->lock, flags);
834 * Did someone else start a new RCU grace period start since we last
835 * checked? Update local state appropriately if so. Must be called
836 * on the CPU corresponding to rdp.
839 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
844 local_irq_save(flags);
845 if (rdp->gpnum != rsp->gpnum) {
846 note_new_gpnum(rsp, rdp);
849 local_irq_restore(flags);
854 * Advance this CPU's callbacks, but only if the current grace period
855 * has ended. This may be called only from the CPU to whom the rdp
856 * belongs. In addition, the corresponding leaf rcu_node structure's
857 * ->lock must be held by the caller, with irqs disabled.
860 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
862 /* Did another grace period end? */
863 if (rdp->completed != rnp->completed) {
865 /* Advance callbacks. No harm if list empty. */
866 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
867 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
868 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
870 /* Remember that we saw this grace-period completion. */
871 rdp->completed = rnp->completed;
872 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
875 * If we were in an extended quiescent state, we may have
876 * missed some grace periods that others CPUs handled on
877 * our behalf. Catch up with this state to avoid noting
878 * spurious new grace periods. If another grace period
879 * has started, then rnp->gpnum will have advanced, so
880 * we will detect this later on.
882 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
883 rdp->gpnum = rdp->completed;
886 * If RCU does not need a quiescent state from this CPU,
887 * then make sure that this CPU doesn't go looking for one.
889 if ((rnp->qsmask & rdp->grpmask) == 0)
895 * Advance this CPU's callbacks, but only if the current grace period
896 * has ended. This may be called only from the CPU to whom the rdp
900 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
903 struct rcu_node *rnp;
905 local_irq_save(flags);
907 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
908 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
909 local_irq_restore(flags);
912 __rcu_process_gp_end(rsp, rnp, rdp);
913 raw_spin_unlock_irqrestore(&rnp->lock, flags);
917 * Do per-CPU grace-period initialization for running CPU. The caller
918 * must hold the lock of the leaf rcu_node structure corresponding to
922 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
924 /* Prior grace period ended, so advance callbacks for current CPU. */
925 __rcu_process_gp_end(rsp, rnp, rdp);
928 * Because this CPU just now started the new grace period, we know
929 * that all of its callbacks will be covered by this upcoming grace
930 * period, even the ones that were registered arbitrarily recently.
931 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
933 * Other CPUs cannot be sure exactly when the grace period started.
934 * Therefore, their recently registered callbacks must pass through
935 * an additional RCU_NEXT_READY stage, so that they will be handled
936 * by the next RCU grace period.
938 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
939 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
941 /* Set state so that this CPU will detect the next quiescent state. */
942 __note_new_gpnum(rsp, rnp, rdp);
946 * Start a new RCU grace period if warranted, re-initializing the hierarchy
947 * in preparation for detecting the next grace period. The caller must hold
948 * the root node's ->lock, which is released before return. Hard irqs must
952 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
953 __releases(rcu_get_root(rsp)->lock)
955 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
956 struct rcu_node *rnp = rcu_get_root(rsp);
958 if (!rcu_scheduler_fully_active ||
959 !cpu_needs_another_gp(rsp, rdp)) {
961 * Either the scheduler hasn't yet spawned the first
962 * non-idle task or this CPU does not need another
963 * grace period. Either way, don't start a new grace
966 raw_spin_unlock_irqrestore(&rnp->lock, flags);
970 if (rsp->fqs_active) {
972 * This CPU needs a grace period, but force_quiescent_state()
973 * is running. Tell it to start one on this CPU's behalf.
975 rsp->fqs_need_gp = 1;
976 raw_spin_unlock_irqrestore(&rnp->lock, flags);
980 /* Advance to a new grace period and initialize state. */
982 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
983 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
984 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
985 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
986 record_gp_stall_check_time(rsp);
988 /* Special-case the common single-level case. */
989 if (NUM_RCU_NODES == 1) {
990 rcu_preempt_check_blocked_tasks(rnp);
991 rnp->qsmask = rnp->qsmaskinit;
992 rnp->gpnum = rsp->gpnum;
993 rnp->completed = rsp->completed;
994 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state OK */
995 rcu_start_gp_per_cpu(rsp, rnp, rdp);
996 rcu_preempt_boost_start_gp(rnp);
997 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
998 rnp->level, rnp->grplo,
999 rnp->grphi, rnp->qsmask);
1000 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1004 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1007 /* Exclude any concurrent CPU-hotplug operations. */
1008 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1011 * Set the quiescent-state-needed bits in all the rcu_node
1012 * structures for all currently online CPUs in breadth-first
1013 * order, starting from the root rcu_node structure. This
1014 * operation relies on the layout of the hierarchy within the
1015 * rsp->node[] array. Note that other CPUs will access only
1016 * the leaves of the hierarchy, which still indicate that no
1017 * grace period is in progress, at least until the corresponding
1018 * leaf node has been initialized. In addition, we have excluded
1019 * CPU-hotplug operations.
1021 * Note that the grace period cannot complete until we finish
1022 * the initialization process, as there will be at least one
1023 * qsmask bit set in the root node until that time, namely the
1024 * one corresponding to this CPU, due to the fact that we have
1027 rcu_for_each_node_breadth_first(rsp, rnp) {
1028 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1029 rcu_preempt_check_blocked_tasks(rnp);
1030 rnp->qsmask = rnp->qsmaskinit;
1031 rnp->gpnum = rsp->gpnum;
1032 rnp->completed = rsp->completed;
1033 if (rnp == rdp->mynode)
1034 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1035 rcu_preempt_boost_start_gp(rnp);
1036 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1037 rnp->level, rnp->grplo,
1038 rnp->grphi, rnp->qsmask);
1039 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1042 rnp = rcu_get_root(rsp);
1043 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1044 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1045 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1046 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1050 * Report a full set of quiescent states to the specified rcu_state
1051 * data structure. This involves cleaning up after the prior grace
1052 * period and letting rcu_start_gp() start up the next grace period
1053 * if one is needed. Note that the caller must hold rnp->lock, as
1054 * required by rcu_start_gp(), which will release it.
1056 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1057 __releases(rcu_get_root(rsp)->lock)
1059 unsigned long gp_duration;
1060 struct rcu_node *rnp = rcu_get_root(rsp);
1061 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1063 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1066 * Ensure that all grace-period and pre-grace-period activity
1067 * is seen before the assignment to rsp->completed.
1069 smp_mb(); /* See above block comment. */
1070 gp_duration = jiffies - rsp->gp_start;
1071 if (gp_duration > rsp->gp_max)
1072 rsp->gp_max = gp_duration;
1075 * We know the grace period is complete, but to everyone else
1076 * it appears to still be ongoing. But it is also the case
1077 * that to everyone else it looks like there is nothing that
1078 * they can do to advance the grace period. It is therefore
1079 * safe for us to drop the lock in order to mark the grace
1080 * period as completed in all of the rcu_node structures.
1082 * But if this CPU needs another grace period, it will take
1083 * care of this while initializing the next grace period.
1084 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1085 * because the callbacks have not yet been advanced: Those
1086 * callbacks are waiting on the grace period that just now
1089 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1090 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1093 * Propagate new ->completed value to rcu_node structures
1094 * so that other CPUs don't have to wait until the start
1095 * of the next grace period to process their callbacks.
1097 rcu_for_each_node_breadth_first(rsp, rnp) {
1098 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1099 rnp->completed = rsp->gpnum;
1100 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1102 rnp = rcu_get_root(rsp);
1103 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1106 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1107 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1108 rsp->fqs_state = RCU_GP_IDLE;
1109 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1113 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1114 * Allows quiescent states for a group of CPUs to be reported at one go
1115 * to the specified rcu_node structure, though all the CPUs in the group
1116 * must be represented by the same rcu_node structure (which need not be
1117 * a leaf rcu_node structure, though it often will be). That structure's
1118 * lock must be held upon entry, and it is released before return.
1121 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1122 struct rcu_node *rnp, unsigned long flags)
1123 __releases(rnp->lock)
1125 struct rcu_node *rnp_c;
1127 /* Walk up the rcu_node hierarchy. */
1129 if (!(rnp->qsmask & mask)) {
1131 /* Our bit has already been cleared, so done. */
1132 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1135 rnp->qsmask &= ~mask;
1136 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1137 mask, rnp->qsmask, rnp->level,
1138 rnp->grplo, rnp->grphi,
1140 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1142 /* Other bits still set at this level, so done. */
1143 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1146 mask = rnp->grpmask;
1147 if (rnp->parent == NULL) {
1149 /* No more levels. Exit loop holding root lock. */
1153 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1156 raw_spin_lock_irqsave(&rnp->lock, flags);
1157 WARN_ON_ONCE(rnp_c->qsmask);
1161 * Get here if we are the last CPU to pass through a quiescent
1162 * state for this grace period. Invoke rcu_report_qs_rsp()
1163 * to clean up and start the next grace period if one is needed.
1165 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1169 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1170 * structure. This must be either called from the specified CPU, or
1171 * called when the specified CPU is known to be offline (and when it is
1172 * also known that no other CPU is concurrently trying to help the offline
1173 * CPU). The lastcomp argument is used to make sure we are still in the
1174 * grace period of interest. We don't want to end the current grace period
1175 * based on quiescent states detected in an earlier grace period!
1178 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1180 unsigned long flags;
1182 struct rcu_node *rnp;
1185 raw_spin_lock_irqsave(&rnp->lock, flags);
1186 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1189 * The grace period in which this quiescent state was
1190 * recorded has ended, so don't report it upwards.
1191 * We will instead need a new quiescent state that lies
1192 * within the current grace period.
1194 rdp->passed_quiesce = 0; /* need qs for new gp. */
1195 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1198 mask = rdp->grpmask;
1199 if ((rnp->qsmask & mask) == 0) {
1200 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1202 rdp->qs_pending = 0;
1205 * This GP can't end until cpu checks in, so all of our
1206 * callbacks can be processed during the next GP.
1208 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1210 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1215 * Check to see if there is a new grace period of which this CPU
1216 * is not yet aware, and if so, set up local rcu_data state for it.
1217 * Otherwise, see if this CPU has just passed through its first
1218 * quiescent state for this grace period, and record that fact if so.
1221 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1223 /* If there is now a new grace period, record and return. */
1224 if (check_for_new_grace_period(rsp, rdp))
1228 * Does this CPU still need to do its part for current grace period?
1229 * If no, return and let the other CPUs do their part as well.
1231 if (!rdp->qs_pending)
1235 * Was there a quiescent state since the beginning of the grace
1236 * period? If no, then exit and wait for the next call.
1238 if (!rdp->passed_quiesce)
1242 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1245 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1248 #ifdef CONFIG_HOTPLUG_CPU
1251 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1252 * Synchronization is not required because this function executes
1253 * in stop_machine() context.
1255 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1258 /* current DYING CPU is cleared in the cpu_online_mask */
1259 int receive_cpu = cpumask_any(cpu_online_mask);
1260 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1261 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1263 if (rdp->nxtlist == NULL)
1264 return; /* irqs disabled, so comparison is stable. */
1266 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1267 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1268 receive_rdp->qlen += rdp->qlen;
1269 receive_rdp->n_cbs_adopted += rdp->qlen;
1270 rdp->n_cbs_orphaned += rdp->qlen;
1272 rdp->nxtlist = NULL;
1273 for (i = 0; i < RCU_NEXT_SIZE; i++)
1274 rdp->nxttail[i] = &rdp->nxtlist;
1279 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1280 * and move all callbacks from the outgoing CPU to the current one.
1281 * There can only be one CPU hotplug operation at a time, so no other
1282 * CPU can be attempting to update rcu_cpu_kthread_task.
1284 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1286 unsigned long flags;
1288 int need_report = 0;
1289 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1290 struct rcu_node *rnp;
1292 rcu_stop_cpu_kthread(cpu);
1294 /* Exclude any attempts to start a new grace period. */
1295 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1297 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1298 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1299 mask = rdp->grpmask; /* rnp->grplo is constant. */
1301 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1302 rnp->qsmaskinit &= ~mask;
1303 if (rnp->qsmaskinit != 0) {
1304 if (rnp != rdp->mynode)
1305 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1307 trace_rcu_grace_period(rsp->name,
1309 !!(rnp->qsmask & mask),
1313 if (rnp == rdp->mynode) {
1314 trace_rcu_grace_period(rsp->name,
1316 !!(rnp->qsmask & mask),
1318 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1320 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1321 mask = rnp->grpmask;
1323 } while (rnp != NULL);
1326 * We still hold the leaf rcu_node structure lock here, and
1327 * irqs are still disabled. The reason for this subterfuge is
1328 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1329 * held leads to deadlock.
1331 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1333 if (need_report & RCU_OFL_TASKS_NORM_GP)
1334 rcu_report_unblock_qs_rnp(rnp, flags);
1336 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1337 if (need_report & RCU_OFL_TASKS_EXP_GP)
1338 rcu_report_exp_rnp(rsp, rnp, true);
1339 rcu_node_kthread_setaffinity(rnp, -1);
1343 * Remove the specified CPU from the RCU hierarchy and move any pending
1344 * callbacks that it might have to the current CPU. This code assumes
1345 * that at least one CPU in the system will remain running at all times.
1346 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1348 static void rcu_offline_cpu(int cpu)
1350 __rcu_offline_cpu(cpu, &rcu_sched_state);
1351 __rcu_offline_cpu(cpu, &rcu_bh_state);
1352 rcu_preempt_offline_cpu(cpu);
1355 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1357 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1361 static void rcu_offline_cpu(int cpu)
1365 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1368 * Invoke any RCU callbacks that have made it to the end of their grace
1369 * period. Thottle as specified by rdp->blimit.
1371 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1373 unsigned long flags;
1374 struct rcu_head *next, *list, **tail;
1377 /* If no callbacks are ready, just return.*/
1378 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1379 trace_rcu_batch_start(rsp->name, 0, 0);
1380 trace_rcu_batch_end(rsp->name, 0);
1385 * Extract the list of ready callbacks, disabling to prevent
1386 * races with call_rcu() from interrupt handlers.
1388 local_irq_save(flags);
1390 trace_rcu_batch_start(rsp->name, rdp->qlen, bl);
1391 list = rdp->nxtlist;
1392 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1393 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1394 tail = rdp->nxttail[RCU_DONE_TAIL];
1395 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1396 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1397 rdp->nxttail[count] = &rdp->nxtlist;
1398 local_irq_restore(flags);
1400 /* Invoke callbacks. */
1405 debug_rcu_head_unqueue(list);
1406 __rcu_reclaim(rsp->name, list);
1412 local_irq_save(flags);
1413 trace_rcu_batch_end(rsp->name, count);
1415 /* Update count, and requeue any remaining callbacks. */
1417 rdp->n_cbs_invoked += count;
1419 *tail = rdp->nxtlist;
1420 rdp->nxtlist = list;
1421 for (count = 0; count < RCU_NEXT_SIZE; count++)
1422 if (&rdp->nxtlist == rdp->nxttail[count])
1423 rdp->nxttail[count] = tail;
1428 /* Reinstate batch limit if we have worked down the excess. */
1429 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1430 rdp->blimit = blimit;
1432 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1433 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1434 rdp->qlen_last_fqs_check = 0;
1435 rdp->n_force_qs_snap = rsp->n_force_qs;
1436 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1437 rdp->qlen_last_fqs_check = rdp->qlen;
1439 local_irq_restore(flags);
1441 /* Re-invoke RCU core processing if there are callbacks remaining. */
1442 if (cpu_has_callbacks_ready_to_invoke(rdp))
1447 * Check to see if this CPU is in a non-context-switch quiescent state
1448 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1449 * Also schedule RCU core processing.
1451 * This function must be called from hardirq context. It is normally
1452 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1453 * false, there is no point in invoking rcu_check_callbacks().
1455 void rcu_check_callbacks(int cpu, int user)
1457 trace_rcu_utilization("Start scheduler-tick");
1458 if (user || rcu_is_cpu_rrupt_from_idle()) {
1461 * Get here if this CPU took its interrupt from user
1462 * mode or from the idle loop, and if this is not a
1463 * nested interrupt. In this case, the CPU is in
1464 * a quiescent state, so note it.
1466 * No memory barrier is required here because both
1467 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1468 * variables that other CPUs neither access nor modify,
1469 * at least not while the corresponding CPU is online.
1475 } else if (!in_softirq()) {
1478 * Get here if this CPU did not take its interrupt from
1479 * softirq, in other words, if it is not interrupting
1480 * a rcu_bh read-side critical section. This is an _bh
1481 * critical section, so note it.
1486 rcu_preempt_check_callbacks(cpu);
1487 if (rcu_pending(cpu))
1489 trace_rcu_utilization("End scheduler-tick");
1495 * Scan the leaf rcu_node structures, processing dyntick state for any that
1496 * have not yet encountered a quiescent state, using the function specified.
1497 * Also initiate boosting for any threads blocked on the root rcu_node.
1499 * The caller must have suppressed start of new grace periods.
1501 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1505 unsigned long flags;
1507 struct rcu_node *rnp;
1509 rcu_for_each_leaf_node(rsp, rnp) {
1511 raw_spin_lock_irqsave(&rnp->lock, flags);
1512 if (!rcu_gp_in_progress(rsp)) {
1513 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1516 if (rnp->qsmask == 0) {
1517 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1522 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1523 if ((rnp->qsmask & bit) != 0 &&
1524 f(per_cpu_ptr(rsp->rda, cpu)))
1529 /* rcu_report_qs_rnp() releases rnp->lock. */
1530 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1533 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1535 rnp = rcu_get_root(rsp);
1536 if (rnp->qsmask == 0) {
1537 raw_spin_lock_irqsave(&rnp->lock, flags);
1538 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1543 * Force quiescent states on reluctant CPUs, and also detect which
1544 * CPUs are in dyntick-idle mode.
1546 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1548 unsigned long flags;
1549 struct rcu_node *rnp = rcu_get_root(rsp);
1551 trace_rcu_utilization("Start fqs");
1552 if (!rcu_gp_in_progress(rsp)) {
1553 trace_rcu_utilization("End fqs");
1554 return; /* No grace period in progress, nothing to force. */
1556 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1557 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1558 trace_rcu_utilization("End fqs");
1559 return; /* Someone else is already on the job. */
1561 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1562 goto unlock_fqs_ret; /* no emergency and done recently. */
1564 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1565 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1566 if(!rcu_gp_in_progress(rsp)) {
1567 rsp->n_force_qs_ngp++;
1568 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1569 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1571 rsp->fqs_active = 1;
1572 switch (rsp->fqs_state) {
1576 break; /* grace period idle or initializing, ignore. */
1578 case RCU_SAVE_DYNTICK:
1579 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1580 break; /* So gcc recognizes the dead code. */
1582 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1584 /* Record dyntick-idle state. */
1585 force_qs_rnp(rsp, dyntick_save_progress_counter);
1586 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1587 if (rcu_gp_in_progress(rsp))
1588 rsp->fqs_state = RCU_FORCE_QS;
1593 /* Check dyntick-idle state, send IPI to laggarts. */
1594 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1595 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1597 /* Leave state in case more forcing is required. */
1599 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1602 rsp->fqs_active = 0;
1603 if (rsp->fqs_need_gp) {
1604 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1605 rsp->fqs_need_gp = 0;
1606 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1607 trace_rcu_utilization("End fqs");
1610 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1612 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1613 trace_rcu_utilization("End fqs");
1616 #else /* #ifdef CONFIG_SMP */
1618 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1623 #endif /* #else #ifdef CONFIG_SMP */
1626 * This does the RCU core processing work for the specified rcu_state
1627 * and rcu_data structures. This may be called only from the CPU to
1628 * whom the rdp belongs.
1631 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1633 unsigned long flags;
1635 WARN_ON_ONCE(rdp->beenonline == 0);
1638 * If an RCU GP has gone long enough, go check for dyntick
1639 * idle CPUs and, if needed, send resched IPIs.
1641 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1642 force_quiescent_state(rsp, 1);
1645 * Advance callbacks in response to end of earlier grace
1646 * period that some other CPU ended.
1648 rcu_process_gp_end(rsp, rdp);
1650 /* Update RCU state based on any recent quiescent states. */
1651 rcu_check_quiescent_state(rsp, rdp);
1653 /* Does this CPU require a not-yet-started grace period? */
1654 if (cpu_needs_another_gp(rsp, rdp)) {
1655 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1656 rcu_start_gp(rsp, flags); /* releases above lock */
1659 /* If there are callbacks ready, invoke them. */
1660 if (cpu_has_callbacks_ready_to_invoke(rdp))
1661 invoke_rcu_callbacks(rsp, rdp);
1665 * Do RCU core processing for the current CPU.
1667 static void rcu_process_callbacks(struct softirq_action *unused)
1669 trace_rcu_utilization("Start RCU core");
1670 __rcu_process_callbacks(&rcu_sched_state,
1671 &__get_cpu_var(rcu_sched_data));
1672 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1673 rcu_preempt_process_callbacks();
1674 trace_rcu_utilization("End RCU core");
1678 * Schedule RCU callback invocation. If the specified type of RCU
1679 * does not support RCU priority boosting, just do a direct call,
1680 * otherwise wake up the per-CPU kernel kthread. Note that because we
1681 * are running on the current CPU with interrupts disabled, the
1682 * rcu_cpu_kthread_task cannot disappear out from under us.
1684 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1686 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1688 if (likely(!rsp->boost)) {
1689 rcu_do_batch(rsp, rdp);
1692 invoke_rcu_callbacks_kthread();
1695 static void invoke_rcu_core(void)
1697 raise_softirq(RCU_SOFTIRQ);
1701 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1702 struct rcu_state *rsp)
1704 unsigned long flags;
1705 struct rcu_data *rdp;
1707 debug_rcu_head_queue(head);
1711 smp_mb(); /* Ensure RCU update seen before callback registry. */
1714 * Opportunistically note grace-period endings and beginnings.
1715 * Note that we might see a beginning right after we see an
1716 * end, but never vice versa, since this CPU has to pass through
1717 * a quiescent state betweentimes.
1719 local_irq_save(flags);
1720 rdp = this_cpu_ptr(rsp->rda);
1722 /* Add the callback to our list. */
1723 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1724 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1727 if (__is_kfree_rcu_offset((unsigned long)func))
1728 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1731 trace_rcu_callback(rsp->name, head, rdp->qlen);
1733 /* If interrupts were disabled, don't dive into RCU core. */
1734 if (irqs_disabled_flags(flags)) {
1735 local_irq_restore(flags);
1740 * Force the grace period if too many callbacks or too long waiting.
1741 * Enforce hysteresis, and don't invoke force_quiescent_state()
1742 * if some other CPU has recently done so. Also, don't bother
1743 * invoking force_quiescent_state() if the newly enqueued callback
1744 * is the only one waiting for a grace period to complete.
1746 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1748 /* Are we ignoring a completed grace period? */
1749 rcu_process_gp_end(rsp, rdp);
1750 check_for_new_grace_period(rsp, rdp);
1752 /* Start a new grace period if one not already started. */
1753 if (!rcu_gp_in_progress(rsp)) {
1754 unsigned long nestflag;
1755 struct rcu_node *rnp_root = rcu_get_root(rsp);
1757 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1758 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1760 /* Give the grace period a kick. */
1761 rdp->blimit = LONG_MAX;
1762 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1763 *rdp->nxttail[RCU_DONE_TAIL] != head)
1764 force_quiescent_state(rsp, 0);
1765 rdp->n_force_qs_snap = rsp->n_force_qs;
1766 rdp->qlen_last_fqs_check = rdp->qlen;
1768 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1769 force_quiescent_state(rsp, 1);
1770 local_irq_restore(flags);
1774 * Queue an RCU-sched callback for invocation after a grace period.
1776 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1778 __call_rcu(head, func, &rcu_sched_state);
1780 EXPORT_SYMBOL_GPL(call_rcu_sched);
1783 * Queue an RCU for invocation after a quicker grace period.
1785 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1787 __call_rcu(head, func, &rcu_bh_state);
1789 EXPORT_SYMBOL_GPL(call_rcu_bh);
1792 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1794 * Control will return to the caller some time after a full rcu-sched
1795 * grace period has elapsed, in other words after all currently executing
1796 * rcu-sched read-side critical sections have completed. These read-side
1797 * critical sections are delimited by rcu_read_lock_sched() and
1798 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1799 * local_irq_disable(), and so on may be used in place of
1800 * rcu_read_lock_sched().
1802 * This means that all preempt_disable code sequences, including NMI and
1803 * hardware-interrupt handlers, in progress on entry will have completed
1804 * before this primitive returns. However, this does not guarantee that
1805 * softirq handlers will have completed, since in some kernels, these
1806 * handlers can run in process context, and can block.
1808 * This primitive provides the guarantees made by the (now removed)
1809 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1810 * guarantees that rcu_read_lock() sections will have completed.
1811 * In "classic RCU", these two guarantees happen to be one and
1812 * the same, but can differ in realtime RCU implementations.
1814 void synchronize_sched(void)
1816 if (rcu_blocking_is_gp())
1818 wait_rcu_gp(call_rcu_sched);
1820 EXPORT_SYMBOL_GPL(synchronize_sched);
1823 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1825 * Control will return to the caller some time after a full rcu_bh grace
1826 * period has elapsed, in other words after all currently executing rcu_bh
1827 * read-side critical sections have completed. RCU read-side critical
1828 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1829 * and may be nested.
1831 void synchronize_rcu_bh(void)
1833 if (rcu_blocking_is_gp())
1835 wait_rcu_gp(call_rcu_bh);
1837 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1840 * Check to see if there is any immediate RCU-related work to be done
1841 * by the current CPU, for the specified type of RCU, returning 1 if so.
1842 * The checks are in order of increasing expense: checks that can be
1843 * carried out against CPU-local state are performed first. However,
1844 * we must check for CPU stalls first, else we might not get a chance.
1846 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1848 struct rcu_node *rnp = rdp->mynode;
1850 rdp->n_rcu_pending++;
1852 /* Check for CPU stalls, if enabled. */
1853 check_cpu_stall(rsp, rdp);
1855 /* Is the RCU core waiting for a quiescent state from this CPU? */
1856 if (rcu_scheduler_fully_active &&
1857 rdp->qs_pending && !rdp->passed_quiesce) {
1860 * If force_quiescent_state() coming soon and this CPU
1861 * needs a quiescent state, and this is either RCU-sched
1862 * or RCU-bh, force a local reschedule.
1864 rdp->n_rp_qs_pending++;
1865 if (!rdp->preemptible &&
1866 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1869 } else if (rdp->qs_pending && rdp->passed_quiesce) {
1870 rdp->n_rp_report_qs++;
1874 /* Does this CPU have callbacks ready to invoke? */
1875 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1876 rdp->n_rp_cb_ready++;
1880 /* Has RCU gone idle with this CPU needing another grace period? */
1881 if (cpu_needs_another_gp(rsp, rdp)) {
1882 rdp->n_rp_cpu_needs_gp++;
1886 /* Has another RCU grace period completed? */
1887 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1888 rdp->n_rp_gp_completed++;
1892 /* Has a new RCU grace period started? */
1893 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1894 rdp->n_rp_gp_started++;
1898 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1899 if (rcu_gp_in_progress(rsp) &&
1900 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1901 rdp->n_rp_need_fqs++;
1906 rdp->n_rp_need_nothing++;
1911 * Check to see if there is any immediate RCU-related work to be done
1912 * by the current CPU, returning 1 if so. This function is part of the
1913 * RCU implementation; it is -not- an exported member of the RCU API.
1915 static int rcu_pending(int cpu)
1917 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1918 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1919 rcu_preempt_pending(cpu);
1923 * Check to see if any future RCU-related work will need to be done
1924 * by the current CPU, even if none need be done immediately, returning
1927 static int rcu_cpu_has_callbacks(int cpu)
1929 /* RCU callbacks either ready or pending? */
1930 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1931 per_cpu(rcu_bh_data, cpu).nxtlist ||
1932 rcu_preempt_needs_cpu(cpu);
1935 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1936 static atomic_t rcu_barrier_cpu_count;
1937 static DEFINE_MUTEX(rcu_barrier_mutex);
1938 static struct completion rcu_barrier_completion;
1940 static void rcu_barrier_callback(struct rcu_head *notused)
1942 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1943 complete(&rcu_barrier_completion);
1947 * Called with preemption disabled, and from cross-cpu IRQ context.
1949 static void rcu_barrier_func(void *type)
1951 int cpu = smp_processor_id();
1952 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1953 void (*call_rcu_func)(struct rcu_head *head,
1954 void (*func)(struct rcu_head *head));
1956 atomic_inc(&rcu_barrier_cpu_count);
1957 call_rcu_func = type;
1958 call_rcu_func(head, rcu_barrier_callback);
1962 * Orchestrate the specified type of RCU barrier, waiting for all
1963 * RCU callbacks of the specified type to complete.
1965 static void _rcu_barrier(struct rcu_state *rsp,
1966 void (*call_rcu_func)(struct rcu_head *head,
1967 void (*func)(struct rcu_head *head)))
1969 BUG_ON(in_interrupt());
1970 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1971 mutex_lock(&rcu_barrier_mutex);
1972 init_completion(&rcu_barrier_completion);
1974 * Initialize rcu_barrier_cpu_count to 1, then invoke
1975 * rcu_barrier_func() on each CPU, so that each CPU also has
1976 * incremented rcu_barrier_cpu_count. Only then is it safe to
1977 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1978 * might complete its grace period before all of the other CPUs
1979 * did their increment, causing this function to return too
1980 * early. Note that on_each_cpu() disables irqs, which prevents
1981 * any CPUs from coming online or going offline until each online
1982 * CPU has queued its RCU-barrier callback.
1984 atomic_set(&rcu_barrier_cpu_count, 1);
1985 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1986 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1987 complete(&rcu_barrier_completion);
1988 wait_for_completion(&rcu_barrier_completion);
1989 mutex_unlock(&rcu_barrier_mutex);
1993 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1995 void rcu_barrier_bh(void)
1997 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1999 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2002 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2004 void rcu_barrier_sched(void)
2006 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
2008 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2011 * Do boot-time initialization of a CPU's per-CPU RCU data.
2014 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2016 unsigned long flags;
2018 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2019 struct rcu_node *rnp = rcu_get_root(rsp);
2021 /* Set up local state, ensuring consistent view of global state. */
2022 raw_spin_lock_irqsave(&rnp->lock, flags);
2023 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2024 rdp->nxtlist = NULL;
2025 for (i = 0; i < RCU_NEXT_SIZE; i++)
2026 rdp->nxttail[i] = &rdp->nxtlist;
2028 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2029 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_NESTING);
2030 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2033 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2037 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2038 * offline event can be happening at a given time. Note also that we
2039 * can accept some slop in the rsp->completed access due to the fact
2040 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2042 static void __cpuinit
2043 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2045 unsigned long flags;
2047 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2048 struct rcu_node *rnp = rcu_get_root(rsp);
2050 /* Set up local state, ensuring consistent view of global state. */
2051 raw_spin_lock_irqsave(&rnp->lock, flags);
2052 rdp->beenonline = 1; /* We have now been online. */
2053 rdp->preemptible = preemptible;
2054 rdp->qlen_last_fqs_check = 0;
2055 rdp->n_force_qs_snap = rsp->n_force_qs;
2056 rdp->blimit = blimit;
2057 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_NESTING;
2058 atomic_set(&rdp->dynticks->dynticks,
2059 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2060 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2063 * A new grace period might start here. If so, we won't be part
2064 * of it, but that is OK, as we are currently in a quiescent state.
2067 /* Exclude any attempts to start a new GP on large systems. */
2068 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2070 /* Add CPU to rcu_node bitmasks. */
2072 mask = rdp->grpmask;
2074 /* Exclude any attempts to start a new GP on small systems. */
2075 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2076 rnp->qsmaskinit |= mask;
2077 mask = rnp->grpmask;
2078 if (rnp == rdp->mynode) {
2080 * If there is a grace period in progress, we will
2081 * set up to wait for it next time we run the
2084 rdp->gpnum = rnp->completed;
2085 rdp->completed = rnp->completed;
2086 rdp->passed_quiesce = 0;
2087 rdp->qs_pending = 0;
2088 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2089 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2091 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2093 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2095 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2098 static void __cpuinit rcu_prepare_cpu(int cpu)
2100 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2101 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2102 rcu_preempt_init_percpu_data(cpu);
2106 * Handle CPU online/offline notification events.
2108 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2109 unsigned long action, void *hcpu)
2111 long cpu = (long)hcpu;
2112 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2113 struct rcu_node *rnp = rdp->mynode;
2115 trace_rcu_utilization("Start CPU hotplug");
2117 case CPU_UP_PREPARE:
2118 case CPU_UP_PREPARE_FROZEN:
2119 rcu_prepare_cpu(cpu);
2120 rcu_prepare_kthreads(cpu);
2123 case CPU_DOWN_FAILED:
2124 rcu_node_kthread_setaffinity(rnp, -1);
2125 rcu_cpu_kthread_setrt(cpu, 1);
2127 case CPU_DOWN_PREPARE:
2128 rcu_node_kthread_setaffinity(rnp, cpu);
2129 rcu_cpu_kthread_setrt(cpu, 0);
2132 case CPU_DYING_FROZEN:
2134 * The whole machine is "stopped" except this CPU, so we can
2135 * touch any data without introducing corruption. We send the
2136 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2138 rcu_send_cbs_to_online(&rcu_bh_state);
2139 rcu_send_cbs_to_online(&rcu_sched_state);
2140 rcu_preempt_send_cbs_to_online();
2143 case CPU_DEAD_FROZEN:
2144 case CPU_UP_CANCELED:
2145 case CPU_UP_CANCELED_FROZEN:
2146 rcu_offline_cpu(cpu);
2151 trace_rcu_utilization("End CPU hotplug");
2156 * This function is invoked towards the end of the scheduler's initialization
2157 * process. Before this is called, the idle task might contain
2158 * RCU read-side critical sections (during which time, this idle
2159 * task is booting the system). After this function is called, the
2160 * idle tasks are prohibited from containing RCU read-side critical
2161 * sections. This function also enables RCU lockdep checking.
2163 void rcu_scheduler_starting(void)
2165 WARN_ON(num_online_cpus() != 1);
2166 WARN_ON(nr_context_switches() > 0);
2167 rcu_scheduler_active = 1;
2171 * Compute the per-level fanout, either using the exact fanout specified
2172 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2174 #ifdef CONFIG_RCU_FANOUT_EXACT
2175 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2179 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2180 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2181 rsp->levelspread[0] = RCU_FANOUT_LEAF;
2183 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2184 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2191 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2192 ccur = rsp->levelcnt[i];
2193 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2197 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2200 * Helper function for rcu_init() that initializes one rcu_state structure.
2202 static void __init rcu_init_one(struct rcu_state *rsp,
2203 struct rcu_data __percpu *rda)
2205 static char *buf[] = { "rcu_node_level_0",
2208 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2212 struct rcu_node *rnp;
2214 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2216 /* Initialize the level-tracking arrays. */
2218 for (i = 1; i < NUM_RCU_LVLS; i++)
2219 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2220 rcu_init_levelspread(rsp);
2222 /* Initialize the elements themselves, starting from the leaves. */
2224 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2225 cpustride *= rsp->levelspread[i];
2226 rnp = rsp->level[i];
2227 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2228 raw_spin_lock_init(&rnp->lock);
2229 lockdep_set_class_and_name(&rnp->lock,
2230 &rcu_node_class[i], buf[i]);
2233 rnp->qsmaskinit = 0;
2234 rnp->grplo = j * cpustride;
2235 rnp->grphi = (j + 1) * cpustride - 1;
2236 if (rnp->grphi >= NR_CPUS)
2237 rnp->grphi = NR_CPUS - 1;
2243 rnp->grpnum = j % rsp->levelspread[i - 1];
2244 rnp->grpmask = 1UL << rnp->grpnum;
2245 rnp->parent = rsp->level[i - 1] +
2246 j / rsp->levelspread[i - 1];
2249 INIT_LIST_HEAD(&rnp->blkd_tasks);
2254 rnp = rsp->level[NUM_RCU_LVLS - 1];
2255 for_each_possible_cpu(i) {
2256 while (i > rnp->grphi)
2258 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2259 rcu_boot_init_percpu_data(i, rsp);
2263 void __init rcu_init(void)
2267 rcu_bootup_announce();
2268 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2269 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2270 __rcu_init_preempt();
2271 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2274 * We don't need protection against CPU-hotplug here because
2275 * this is called early in boot, before either interrupts
2276 * or the scheduler are operational.
2278 cpu_notifier(rcu_cpu_notify, 0);
2279 for_each_online_cpu(cpu)
2280 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2281 check_cpu_stall_init();
2284 #include "rcutree_plugin.h"