2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
67 #define RCU_STATE_INITIALIZER(sname, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
71 .gpnum = 0UL - 300UL, \
72 .completed = 0UL - 300UL, \
73 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
77 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
81 struct rcu_state rcu_sched_state =
82 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
83 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
85 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
86 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
88 static struct rcu_state *rcu_state;
89 LIST_HEAD(rcu_struct_flavors);
91 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
92 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
93 module_param(rcu_fanout_leaf, int, 0444);
94 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
95 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
102 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
105 * The rcu_scheduler_active variable transitions from zero to one just
106 * before the first task is spawned. So when this variable is zero, RCU
107 * can assume that there is but one task, allowing RCU to (for example)
108 * optimized synchronize_sched() to a simple barrier(). When this variable
109 * is one, RCU must actually do all the hard work required to detect real
110 * grace periods. This variable is also used to suppress boot-time false
111 * positives from lockdep-RCU error checking.
113 int rcu_scheduler_active __read_mostly;
114 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
117 * The rcu_scheduler_fully_active variable transitions from zero to one
118 * during the early_initcall() processing, which is after the scheduler
119 * is capable of creating new tasks. So RCU processing (for example,
120 * creating tasks for RCU priority boosting) must be delayed until after
121 * rcu_scheduler_fully_active transitions from zero to one. We also
122 * currently delay invocation of any RCU callbacks until after this point.
124 * It might later prove better for people registering RCU callbacks during
125 * early boot to take responsibility for these callbacks, but one step at
128 static int rcu_scheduler_fully_active __read_mostly;
130 #ifdef CONFIG_RCU_BOOST
133 * Control variables for per-CPU and per-rcu_node kthreads. These
134 * handle all flavors of RCU.
136 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
139 DEFINE_PER_CPU(char, rcu_cpu_has_work);
141 #endif /* #ifdef CONFIG_RCU_BOOST */
143 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
144 static void invoke_rcu_core(void);
145 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
148 * Track the rcutorture test sequence number and the update version
149 * number within a given test. The rcutorture_testseq is incremented
150 * on every rcutorture module load and unload, so has an odd value
151 * when a test is running. The rcutorture_vernum is set to zero
152 * when rcutorture starts and is incremented on each rcutorture update.
153 * These variables enable correlating rcutorture output with the
154 * RCU tracing information.
156 unsigned long rcutorture_testseq;
157 unsigned long rcutorture_vernum;
160 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
161 * permit this function to be invoked without holding the root rcu_node
162 * structure's ->lock, but of course results can be subject to change.
164 static int rcu_gp_in_progress(struct rcu_state *rsp)
166 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
170 * Note a quiescent state. Because we do not need to know
171 * how many quiescent states passed, just if there was at least
172 * one since the start of the grace period, this just sets a flag.
173 * The caller must have disabled preemption.
175 void rcu_sched_qs(int cpu)
177 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
179 if (rdp->passed_quiesce == 0)
180 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
181 rdp->passed_quiesce = 1;
184 void rcu_bh_qs(int cpu)
186 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
188 if (rdp->passed_quiesce == 0)
189 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
190 rdp->passed_quiesce = 1;
194 * Note a context switch. This is a quiescent state for RCU-sched,
195 * and requires special handling for preemptible RCU.
196 * The caller must have disabled preemption.
198 void rcu_note_context_switch(int cpu)
200 trace_rcu_utilization("Start context switch");
202 rcu_preempt_note_context_switch(cpu);
203 trace_rcu_utilization("End context switch");
205 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
207 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
208 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
209 .dynticks = ATOMIC_INIT(1),
210 #if defined(CONFIG_RCU_USER_QS) && !defined(CONFIG_RCU_USER_QS_FORCE)
211 .ignore_user_qs = true,
215 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
216 static long qhimark = 10000; /* If this many pending, ignore blimit. */
217 static long qlowmark = 100; /* Once only this many pending, use blimit. */
219 module_param(blimit, long, 0444);
220 module_param(qhimark, long, 0444);
221 module_param(qlowmark, long, 0444);
223 static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
224 static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
226 module_param(jiffies_till_first_fqs, ulong, 0644);
227 module_param(jiffies_till_next_fqs, ulong, 0644);
229 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
230 static void force_quiescent_state(struct rcu_state *rsp);
231 static int rcu_pending(int cpu);
234 * Return the number of RCU-sched batches processed thus far for debug & stats.
236 long rcu_batches_completed_sched(void)
238 return rcu_sched_state.completed;
240 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
243 * Return the number of RCU BH batches processed thus far for debug & stats.
245 long rcu_batches_completed_bh(void)
247 return rcu_bh_state.completed;
249 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
252 * Force a quiescent state for RCU BH.
254 void rcu_bh_force_quiescent_state(void)
256 force_quiescent_state(&rcu_bh_state);
258 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
261 * Record the number of times rcutorture tests have been initiated and
262 * terminated. This information allows the debugfs tracing stats to be
263 * correlated to the rcutorture messages, even when the rcutorture module
264 * is being repeatedly loaded and unloaded. In other words, we cannot
265 * store this state in rcutorture itself.
267 void rcutorture_record_test_transition(void)
269 rcutorture_testseq++;
270 rcutorture_vernum = 0;
272 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
275 * Record the number of writer passes through the current rcutorture test.
276 * This is also used to correlate debugfs tracing stats with the rcutorture
279 void rcutorture_record_progress(unsigned long vernum)
283 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
286 * Force a quiescent state for RCU-sched.
288 void rcu_sched_force_quiescent_state(void)
290 force_quiescent_state(&rcu_sched_state);
292 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
295 * Does the CPU have callbacks ready to be invoked?
298 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
300 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
304 * Does the current CPU require a yet-as-unscheduled grace period?
307 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
309 return *rdp->nxttail[RCU_DONE_TAIL +
310 (ACCESS_ONCE(rsp->completed) != rdp->completed)] &&
311 !rcu_gp_in_progress(rsp);
315 * Return the root node of the specified rcu_state structure.
317 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
319 return &rsp->node[0];
323 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
325 * If the new value of the ->dynticks_nesting counter now is zero,
326 * we really have entered idle, and must do the appropriate accounting.
327 * The caller must have disabled interrupts.
329 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
332 trace_rcu_dyntick("Start", oldval, 0);
333 if (!user && !is_idle_task(current)) {
334 struct task_struct *idle = idle_task(smp_processor_id());
336 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
337 ftrace_dump(DUMP_ORIG);
338 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
339 current->pid, current->comm,
340 idle->pid, idle->comm); /* must be idle task! */
342 rcu_prepare_for_idle(smp_processor_id());
343 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
344 smp_mb__before_atomic_inc(); /* See above. */
345 atomic_inc(&rdtp->dynticks);
346 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
347 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
350 * It is illegal to enter an extended quiescent state while
351 * in an RCU read-side critical section.
353 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
354 "Illegal idle entry in RCU read-side critical section.");
355 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
356 "Illegal idle entry in RCU-bh read-side critical section.");
357 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
358 "Illegal idle entry in RCU-sched read-side critical section.");
362 * Enter an RCU extended quiescent state, which can be either the
363 * idle loop or adaptive-tickless usermode execution.
365 static void rcu_eqs_enter(bool user)
368 struct rcu_dynticks *rdtp;
370 rdtp = &__get_cpu_var(rcu_dynticks);
371 oldval = rdtp->dynticks_nesting;
372 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
373 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
374 rdtp->dynticks_nesting = 0;
376 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
377 rcu_eqs_enter_common(rdtp, oldval, user);
381 * rcu_idle_enter - inform RCU that current CPU is entering idle
383 * Enter idle mode, in other words, -leave- the mode in which RCU
384 * read-side critical sections can occur. (Though RCU read-side
385 * critical sections can occur in irq handlers in idle, a possibility
386 * handled by irq_enter() and irq_exit().)
388 * We crowbar the ->dynticks_nesting field to zero to allow for
389 * the possibility of usermode upcalls having messed up our count
390 * of interrupt nesting level during the prior busy period.
392 void rcu_idle_enter(void)
396 local_irq_save(flags);
397 rcu_eqs_enter(false);
398 local_irq_restore(flags);
400 EXPORT_SYMBOL_GPL(rcu_idle_enter);
402 #ifdef CONFIG_RCU_USER_QS
404 * rcu_user_enter - inform RCU that we are resuming userspace.
406 * Enter RCU idle mode right before resuming userspace. No use of RCU
407 * is permitted between this call and rcu_user_exit(). This way the
408 * CPU doesn't need to maintain the tick for RCU maintenance purposes
409 * when the CPU runs in userspace.
411 void rcu_user_enter(void)
414 struct rcu_dynticks *rdtp;
417 * Some contexts may involve an exception occuring in an irq,
418 * leading to that nesting:
419 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
420 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
421 * helpers are enough to protect RCU uses inside the exception. So
422 * just return immediately if we detect we are in an IRQ.
427 WARN_ON_ONCE(!current->mm);
429 local_irq_save(flags);
430 rdtp = &__get_cpu_var(rcu_dynticks);
431 if (!rdtp->ignore_user_qs && !rdtp->in_user) {
432 rdtp->in_user = true;
435 local_irq_restore(flags);
439 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
440 * after the current irq returns.
442 * This is similar to rcu_user_enter() but in the context of a non-nesting
443 * irq. After this call, RCU enters into idle mode when the interrupt
446 void rcu_user_enter_after_irq(void)
449 struct rcu_dynticks *rdtp;
451 local_irq_save(flags);
452 rdtp = &__get_cpu_var(rcu_dynticks);
453 /* Ensure this irq is interrupting a non-idle RCU state. */
454 WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
455 rdtp->dynticks_nesting = 1;
456 local_irq_restore(flags);
458 #endif /* CONFIG_RCU_USER_QS */
461 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
463 * Exit from an interrupt handler, which might possibly result in entering
464 * idle mode, in other words, leaving the mode in which read-side critical
465 * sections can occur.
467 * This code assumes that the idle loop never does anything that might
468 * result in unbalanced calls to irq_enter() and irq_exit(). If your
469 * architecture violates this assumption, RCU will give you what you
470 * deserve, good and hard. But very infrequently and irreproducibly.
472 * Use things like work queues to work around this limitation.
474 * You have been warned.
476 void rcu_irq_exit(void)
480 struct rcu_dynticks *rdtp;
482 local_irq_save(flags);
483 rdtp = &__get_cpu_var(rcu_dynticks);
484 oldval = rdtp->dynticks_nesting;
485 rdtp->dynticks_nesting--;
486 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
487 if (rdtp->dynticks_nesting)
488 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
490 rcu_eqs_enter_common(rdtp, oldval, true);
491 local_irq_restore(flags);
495 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
497 * If the new value of the ->dynticks_nesting counter was previously zero,
498 * we really have exited idle, and must do the appropriate accounting.
499 * The caller must have disabled interrupts.
501 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
504 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
505 atomic_inc(&rdtp->dynticks);
506 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
507 smp_mb__after_atomic_inc(); /* See above. */
508 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
509 rcu_cleanup_after_idle(smp_processor_id());
510 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
511 if (!user && !is_idle_task(current)) {
512 struct task_struct *idle = idle_task(smp_processor_id());
514 trace_rcu_dyntick("Error on exit: not idle task",
515 oldval, rdtp->dynticks_nesting);
516 ftrace_dump(DUMP_ORIG);
517 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
518 current->pid, current->comm,
519 idle->pid, idle->comm); /* must be idle task! */
524 * Exit an RCU extended quiescent state, which can be either the
525 * idle loop or adaptive-tickless usermode execution.
527 static void rcu_eqs_exit(bool user)
529 struct rcu_dynticks *rdtp;
532 rdtp = &__get_cpu_var(rcu_dynticks);
533 oldval = rdtp->dynticks_nesting;
534 WARN_ON_ONCE(oldval < 0);
535 if (oldval & DYNTICK_TASK_NEST_MASK)
536 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
538 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
539 rcu_eqs_exit_common(rdtp, oldval, user);
543 * rcu_idle_exit - inform RCU that current CPU is leaving idle
545 * Exit idle mode, in other words, -enter- the mode in which RCU
546 * read-side critical sections can occur.
548 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
549 * allow for the possibility of usermode upcalls messing up our count
550 * of interrupt nesting level during the busy period that is just
553 void rcu_idle_exit(void)
557 local_irq_save(flags);
559 local_irq_restore(flags);
561 EXPORT_SYMBOL_GPL(rcu_idle_exit);
563 #ifdef CONFIG_RCU_USER_QS
565 * rcu_user_exit - inform RCU that we are exiting userspace.
567 * Exit RCU idle mode while entering the kernel because it can
568 * run a RCU read side critical section anytime.
570 void rcu_user_exit(void)
573 struct rcu_dynticks *rdtp;
576 * Some contexts may involve an exception occuring in an irq,
577 * leading to that nesting:
578 * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
579 * This would mess up the dyntick_nesting count though. And rcu_irq_*()
580 * helpers are enough to protect RCU uses inside the exception. So
581 * just return immediately if we detect we are in an IRQ.
586 local_irq_save(flags);
587 rdtp = &__get_cpu_var(rcu_dynticks);
589 rdtp->in_user = false;
592 local_irq_restore(flags);
596 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
597 * idle mode after the current non-nesting irq returns.
599 * This is similar to rcu_user_exit() but in the context of an irq.
600 * This is called when the irq has interrupted a userspace RCU idle mode
601 * context. When the current non-nesting interrupt returns after this call,
602 * the CPU won't restore the RCU idle mode.
604 void rcu_user_exit_after_irq(void)
607 struct rcu_dynticks *rdtp;
609 local_irq_save(flags);
610 rdtp = &__get_cpu_var(rcu_dynticks);
611 /* Ensure we are interrupting an RCU idle mode. */
612 WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
613 rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
614 local_irq_restore(flags);
616 #endif /* CONFIG_RCU_USER_QS */
619 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
621 * Enter an interrupt handler, which might possibly result in exiting
622 * idle mode, in other words, entering the mode in which read-side critical
623 * sections can occur.
625 * Note that the Linux kernel is fully capable of entering an interrupt
626 * handler that it never exits, for example when doing upcalls to
627 * user mode! This code assumes that the idle loop never does upcalls to
628 * user mode. If your architecture does do upcalls from the idle loop (or
629 * does anything else that results in unbalanced calls to the irq_enter()
630 * and irq_exit() functions), RCU will give you what you deserve, good
631 * and hard. But very infrequently and irreproducibly.
633 * Use things like work queues to work around this limitation.
635 * You have been warned.
637 void rcu_irq_enter(void)
640 struct rcu_dynticks *rdtp;
643 local_irq_save(flags);
644 rdtp = &__get_cpu_var(rcu_dynticks);
645 oldval = rdtp->dynticks_nesting;
646 rdtp->dynticks_nesting++;
647 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
649 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
651 rcu_eqs_exit_common(rdtp, oldval, true);
652 local_irq_restore(flags);
656 * rcu_nmi_enter - inform RCU of entry to NMI context
658 * If the CPU was idle with dynamic ticks active, and there is no
659 * irq handler running, this updates rdtp->dynticks_nmi to let the
660 * RCU grace-period handling know that the CPU is active.
662 void rcu_nmi_enter(void)
664 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
666 if (rdtp->dynticks_nmi_nesting == 0 &&
667 (atomic_read(&rdtp->dynticks) & 0x1))
669 rdtp->dynticks_nmi_nesting++;
670 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
671 atomic_inc(&rdtp->dynticks);
672 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
673 smp_mb__after_atomic_inc(); /* See above. */
674 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
678 * rcu_nmi_exit - inform RCU of exit from NMI context
680 * If the CPU was idle with dynamic ticks active, and there is no
681 * irq handler running, this updates rdtp->dynticks_nmi to let the
682 * RCU grace-period handling know that the CPU is no longer active.
684 void rcu_nmi_exit(void)
686 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
688 if (rdtp->dynticks_nmi_nesting == 0 ||
689 --rdtp->dynticks_nmi_nesting != 0)
691 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
692 smp_mb__before_atomic_inc(); /* See above. */
693 atomic_inc(&rdtp->dynticks);
694 smp_mb__after_atomic_inc(); /* Force delay to next write. */
695 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
699 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
701 * If the current CPU is in its idle loop and is neither in an interrupt
702 * or NMI handler, return true.
704 int rcu_is_cpu_idle(void)
709 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
713 EXPORT_SYMBOL(rcu_is_cpu_idle);
715 #ifdef CONFIG_RCU_USER_QS
716 void rcu_user_hooks_switch(struct task_struct *prev,
717 struct task_struct *next)
719 struct rcu_dynticks *rdtp;
721 /* Interrupts are disabled in context switch */
722 rdtp = &__get_cpu_var(rcu_dynticks);
723 if (!rdtp->ignore_user_qs) {
724 clear_tsk_thread_flag(prev, TIF_NOHZ);
725 set_tsk_thread_flag(next, TIF_NOHZ);
728 #endif /* #ifdef CONFIG_RCU_USER_QS */
730 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
733 * Is the current CPU online? Disable preemption to avoid false positives
734 * that could otherwise happen due to the current CPU number being sampled,
735 * this task being preempted, its old CPU being taken offline, resuming
736 * on some other CPU, then determining that its old CPU is now offline.
737 * It is OK to use RCU on an offline processor during initial boot, hence
738 * the check for rcu_scheduler_fully_active. Note also that it is OK
739 * for a CPU coming online to use RCU for one jiffy prior to marking itself
740 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
741 * offline to continue to use RCU for one jiffy after marking itself
742 * offline in the cpu_online_mask. This leniency is necessary given the
743 * non-atomic nature of the online and offline processing, for example,
744 * the fact that a CPU enters the scheduler after completing the CPU_DYING
747 * This is also why RCU internally marks CPUs online during the
748 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
750 * Disable checking if in an NMI handler because we cannot safely report
751 * errors from NMI handlers anyway.
753 bool rcu_lockdep_current_cpu_online(void)
755 struct rcu_data *rdp;
756 struct rcu_node *rnp;
762 rdp = &__get_cpu_var(rcu_sched_data);
764 ret = (rdp->grpmask & rnp->qsmaskinit) ||
765 !rcu_scheduler_fully_active;
769 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
771 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
774 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
776 * If the current CPU is idle or running at a first-level (not nested)
777 * interrupt from idle, return true. The caller must have at least
778 * disabled preemption.
780 int rcu_is_cpu_rrupt_from_idle(void)
782 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
786 * Snapshot the specified CPU's dynticks counter so that we can later
787 * credit them with an implicit quiescent state. Return 1 if this CPU
788 * is in dynticks idle mode, which is an extended quiescent state.
790 static int dyntick_save_progress_counter(struct rcu_data *rdp)
792 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
793 return (rdp->dynticks_snap & 0x1) == 0;
797 * Return true if the specified CPU has passed through a quiescent
798 * state by virtue of being in or having passed through an dynticks
799 * idle state since the last call to dyntick_save_progress_counter()
800 * for this same CPU, or by virtue of having been offline.
802 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
807 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
808 snap = (unsigned int)rdp->dynticks_snap;
811 * If the CPU passed through or entered a dynticks idle phase with
812 * no active irq/NMI handlers, then we can safely pretend that the CPU
813 * already acknowledged the request to pass through a quiescent
814 * state. Either way, that CPU cannot possibly be in an RCU
815 * read-side critical section that started before the beginning
816 * of the current RCU grace period.
818 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
819 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
825 * Check for the CPU being offline, but only if the grace period
826 * is old enough. We don't need to worry about the CPU changing
827 * state: If we see it offline even once, it has been through a
830 * The reason for insisting that the grace period be at least
831 * one jiffy old is that CPUs that are not quite online and that
832 * have just gone offline can still execute RCU read-side critical
835 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
836 return 0; /* Grace period is not old enough. */
838 if (cpu_is_offline(rdp->cpu)) {
839 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
846 static void record_gp_stall_check_time(struct rcu_state *rsp)
848 rsp->gp_start = jiffies;
849 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
853 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
854 * for architectures that do not implement trigger_all_cpu_backtrace().
855 * The NMI-triggered stack traces are more accurate because they are
856 * printed by the target CPU.
858 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
862 struct rcu_node *rnp;
864 rcu_for_each_leaf_node(rsp, rnp) {
865 raw_spin_lock_irqsave(&rnp->lock, flags);
866 if (rnp->qsmask != 0) {
867 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
868 if (rnp->qsmask & (1UL << cpu))
869 dump_cpu_task(rnp->grplo + cpu);
871 raw_spin_unlock_irqrestore(&rnp->lock, flags);
875 static void print_other_cpu_stall(struct rcu_state *rsp)
881 struct rcu_node *rnp = rcu_get_root(rsp);
884 /* Only let one CPU complain about others per time interval. */
886 raw_spin_lock_irqsave(&rnp->lock, flags);
887 delta = jiffies - rsp->jiffies_stall;
888 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
889 raw_spin_unlock_irqrestore(&rnp->lock, flags);
892 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
893 raw_spin_unlock_irqrestore(&rnp->lock, flags);
896 * OK, time to rat on our buddy...
897 * See Documentation/RCU/stallwarn.txt for info on how to debug
898 * RCU CPU stall warnings.
900 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
902 print_cpu_stall_info_begin();
903 rcu_for_each_leaf_node(rsp, rnp) {
904 raw_spin_lock_irqsave(&rnp->lock, flags);
905 ndetected += rcu_print_task_stall(rnp);
906 if (rnp->qsmask != 0) {
907 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
908 if (rnp->qsmask & (1UL << cpu)) {
909 print_cpu_stall_info(rsp,
914 raw_spin_unlock_irqrestore(&rnp->lock, flags);
918 * Now rat on any tasks that got kicked up to the root rcu_node
919 * due to CPU offlining.
921 rnp = rcu_get_root(rsp);
922 raw_spin_lock_irqsave(&rnp->lock, flags);
923 ndetected += rcu_print_task_stall(rnp);
924 raw_spin_unlock_irqrestore(&rnp->lock, flags);
926 print_cpu_stall_info_end();
927 for_each_possible_cpu(cpu)
928 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
929 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
930 smp_processor_id(), (long)(jiffies - rsp->gp_start),
931 rsp->gpnum, rsp->completed, totqlen);
933 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
934 else if (!trigger_all_cpu_backtrace())
935 rcu_dump_cpu_stacks(rsp);
937 /* Complain about tasks blocking the grace period. */
939 rcu_print_detail_task_stall(rsp);
941 force_quiescent_state(rsp); /* Kick them all. */
944 static void print_cpu_stall(struct rcu_state *rsp)
948 struct rcu_node *rnp = rcu_get_root(rsp);
952 * OK, time to rat on ourselves...
953 * See Documentation/RCU/stallwarn.txt for info on how to debug
954 * RCU CPU stall warnings.
956 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
957 print_cpu_stall_info_begin();
958 print_cpu_stall_info(rsp, smp_processor_id());
959 print_cpu_stall_info_end();
960 for_each_possible_cpu(cpu)
961 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
962 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
963 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
964 if (!trigger_all_cpu_backtrace())
967 raw_spin_lock_irqsave(&rnp->lock, flags);
968 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
969 rsp->jiffies_stall = jiffies +
970 3 * rcu_jiffies_till_stall_check() + 3;
971 raw_spin_unlock_irqrestore(&rnp->lock, flags);
973 set_need_resched(); /* kick ourselves to get things going. */
976 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
980 struct rcu_node *rnp;
982 if (rcu_cpu_stall_suppress)
984 j = ACCESS_ONCE(jiffies);
985 js = ACCESS_ONCE(rsp->jiffies_stall);
987 if (rcu_gp_in_progress(rsp) &&
988 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
990 /* We haven't checked in, so go dump stack. */
991 print_cpu_stall(rsp);
993 } else if (rcu_gp_in_progress(rsp) &&
994 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
996 /* They had a few time units to dump stack, so complain. */
997 print_other_cpu_stall(rsp);
1002 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1004 * Set the stall-warning timeout way off into the future, thus preventing
1005 * any RCU CPU stall-warning messages from appearing in the current set of
1006 * RCU grace periods.
1008 * The caller must disable hard irqs.
1010 void rcu_cpu_stall_reset(void)
1012 struct rcu_state *rsp;
1014 for_each_rcu_flavor(rsp)
1015 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1019 * Update CPU-local rcu_data state to record the newly noticed grace period.
1020 * This is used both when we started the grace period and when we notice
1021 * that someone else started the grace period. The caller must hold the
1022 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
1023 * and must have irqs disabled.
1025 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1027 if (rdp->gpnum != rnp->gpnum) {
1029 * If the current grace period is waiting for this CPU,
1030 * set up to detect a quiescent state, otherwise don't
1031 * go looking for one.
1033 rdp->gpnum = rnp->gpnum;
1034 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
1035 rdp->passed_quiesce = 0;
1036 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1037 zero_cpu_stall_ticks(rdp);
1041 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1043 unsigned long flags;
1044 struct rcu_node *rnp;
1046 local_irq_save(flags);
1048 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1049 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1050 local_irq_restore(flags);
1053 __note_new_gpnum(rsp, rnp, rdp);
1054 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1058 * Did someone else start a new RCU grace period start since we last
1059 * checked? Update local state appropriately if so. Must be called
1060 * on the CPU corresponding to rdp.
1063 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1065 unsigned long flags;
1068 local_irq_save(flags);
1069 if (rdp->gpnum != rsp->gpnum) {
1070 note_new_gpnum(rsp, rdp);
1073 local_irq_restore(flags);
1078 * Initialize the specified rcu_data structure's callback list to empty.
1080 static void init_callback_list(struct rcu_data *rdp)
1084 rdp->nxtlist = NULL;
1085 for (i = 0; i < RCU_NEXT_SIZE; i++)
1086 rdp->nxttail[i] = &rdp->nxtlist;
1090 * Advance this CPU's callbacks, but only if the current grace period
1091 * has ended. This may be called only from the CPU to whom the rdp
1092 * belongs. In addition, the corresponding leaf rcu_node structure's
1093 * ->lock must be held by the caller, with irqs disabled.
1096 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1098 /* Did another grace period end? */
1099 if (rdp->completed != rnp->completed) {
1101 /* Advance callbacks. No harm if list empty. */
1102 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
1103 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
1104 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1106 /* Remember that we saw this grace-period completion. */
1107 rdp->completed = rnp->completed;
1108 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1111 * If we were in an extended quiescent state, we may have
1112 * missed some grace periods that others CPUs handled on
1113 * our behalf. Catch up with this state to avoid noting
1114 * spurious new grace periods. If another grace period
1115 * has started, then rnp->gpnum will have advanced, so
1116 * we will detect this later on. Of course, any quiescent
1117 * states we found for the old GP are now invalid.
1119 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1120 rdp->gpnum = rdp->completed;
1121 rdp->passed_quiesce = 0;
1125 * If RCU does not need a quiescent state from this CPU,
1126 * then make sure that this CPU doesn't go looking for one.
1128 if ((rnp->qsmask & rdp->grpmask) == 0)
1129 rdp->qs_pending = 0;
1134 * Advance this CPU's callbacks, but only if the current grace period
1135 * has ended. This may be called only from the CPU to whom the rdp
1139 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1141 unsigned long flags;
1142 struct rcu_node *rnp;
1144 local_irq_save(flags);
1146 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1147 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1148 local_irq_restore(flags);
1151 __rcu_process_gp_end(rsp, rnp, rdp);
1152 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1156 * Do per-CPU grace-period initialization for running CPU. The caller
1157 * must hold the lock of the leaf rcu_node structure corresponding to
1161 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1163 /* Prior grace period ended, so advance callbacks for current CPU. */
1164 __rcu_process_gp_end(rsp, rnp, rdp);
1166 /* Set state so that this CPU will detect the next quiescent state. */
1167 __note_new_gpnum(rsp, rnp, rdp);
1171 * Initialize a new grace period.
1173 static int rcu_gp_init(struct rcu_state *rsp)
1175 struct rcu_data *rdp;
1176 struct rcu_node *rnp = rcu_get_root(rsp);
1178 raw_spin_lock_irq(&rnp->lock);
1179 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1181 if (rcu_gp_in_progress(rsp)) {
1182 /* Grace period already in progress, don't start another. */
1183 raw_spin_unlock_irq(&rnp->lock);
1187 /* Advance to a new grace period and initialize state. */
1189 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1190 record_gp_stall_check_time(rsp);
1191 raw_spin_unlock_irq(&rnp->lock);
1193 /* Exclude any concurrent CPU-hotplug operations. */
1194 mutex_lock(&rsp->onoff_mutex);
1197 * Set the quiescent-state-needed bits in all the rcu_node
1198 * structures for all currently online CPUs in breadth-first order,
1199 * starting from the root rcu_node structure, relying on the layout
1200 * of the tree within the rsp->node[] array. Note that other CPUs
1201 * will access only the leaves of the hierarchy, thus seeing that no
1202 * grace period is in progress, at least until the corresponding
1203 * leaf node has been initialized. In addition, we have excluded
1204 * CPU-hotplug operations.
1206 * The grace period cannot complete until the initialization
1207 * process finishes, because this kthread handles both.
1209 rcu_for_each_node_breadth_first(rsp, rnp) {
1210 raw_spin_lock_irq(&rnp->lock);
1211 rdp = this_cpu_ptr(rsp->rda);
1212 rcu_preempt_check_blocked_tasks(rnp);
1213 rnp->qsmask = rnp->qsmaskinit;
1214 rnp->gpnum = rsp->gpnum;
1215 WARN_ON_ONCE(rnp->completed != rsp->completed);
1216 rnp->completed = rsp->completed;
1217 if (rnp == rdp->mynode)
1218 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1219 rcu_preempt_boost_start_gp(rnp);
1220 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1221 rnp->level, rnp->grplo,
1222 rnp->grphi, rnp->qsmask);
1223 raw_spin_unlock_irq(&rnp->lock);
1224 #ifdef CONFIG_PROVE_RCU_DELAY
1225 if ((random32() % (rcu_num_nodes * 8)) == 0)
1226 schedule_timeout_uninterruptible(2);
1227 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1231 mutex_unlock(&rsp->onoff_mutex);
1236 * Do one round of quiescent-state forcing.
1238 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1240 int fqs_state = fqs_state_in;
1241 struct rcu_node *rnp = rcu_get_root(rsp);
1244 if (fqs_state == RCU_SAVE_DYNTICK) {
1245 /* Collect dyntick-idle snapshots. */
1246 force_qs_rnp(rsp, dyntick_save_progress_counter);
1247 fqs_state = RCU_FORCE_QS;
1249 /* Handle dyntick-idle and offline CPUs. */
1250 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1252 /* Clear flag to prevent immediate re-entry. */
1253 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1254 raw_spin_lock_irq(&rnp->lock);
1255 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1256 raw_spin_unlock_irq(&rnp->lock);
1262 * Clean up after the old grace period.
1264 static void rcu_gp_cleanup(struct rcu_state *rsp)
1266 unsigned long gp_duration;
1267 struct rcu_data *rdp;
1268 struct rcu_node *rnp = rcu_get_root(rsp);
1270 raw_spin_lock_irq(&rnp->lock);
1271 gp_duration = jiffies - rsp->gp_start;
1272 if (gp_duration > rsp->gp_max)
1273 rsp->gp_max = gp_duration;
1276 * We know the grace period is complete, but to everyone else
1277 * it appears to still be ongoing. But it is also the case
1278 * that to everyone else it looks like there is nothing that
1279 * they can do to advance the grace period. It is therefore
1280 * safe for us to drop the lock in order to mark the grace
1281 * period as completed in all of the rcu_node structures.
1283 raw_spin_unlock_irq(&rnp->lock);
1286 * Propagate new ->completed value to rcu_node structures so
1287 * that other CPUs don't have to wait until the start of the next
1288 * grace period to process their callbacks. This also avoids
1289 * some nasty RCU grace-period initialization races by forcing
1290 * the end of the current grace period to be completely recorded in
1291 * all of the rcu_node structures before the beginning of the next
1292 * grace period is recorded in any of the rcu_node structures.
1294 rcu_for_each_node_breadth_first(rsp, rnp) {
1295 raw_spin_lock_irq(&rnp->lock);
1296 rnp->completed = rsp->gpnum;
1297 raw_spin_unlock_irq(&rnp->lock);
1300 rnp = rcu_get_root(rsp);
1301 raw_spin_lock_irq(&rnp->lock);
1303 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1304 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1305 rsp->fqs_state = RCU_GP_IDLE;
1306 rdp = this_cpu_ptr(rsp->rda);
1307 if (cpu_needs_another_gp(rsp, rdp))
1309 raw_spin_unlock_irq(&rnp->lock);
1313 * Body of kthread that handles grace periods.
1315 static int __noreturn rcu_gp_kthread(void *arg)
1320 struct rcu_state *rsp = arg;
1321 struct rcu_node *rnp = rcu_get_root(rsp);
1325 /* Handle grace-period start. */
1327 wait_event_interruptible(rsp->gp_wq,
1330 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1334 flush_signals(current);
1337 /* Handle quiescent-state forcing. */
1338 fqs_state = RCU_SAVE_DYNTICK;
1339 j = jiffies_till_first_fqs;
1342 jiffies_till_first_fqs = HZ;
1345 rsp->jiffies_force_qs = jiffies + j;
1346 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1347 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1348 (!ACCESS_ONCE(rnp->qsmask) &&
1349 !rcu_preempt_blocked_readers_cgp(rnp)),
1351 /* If grace period done, leave loop. */
1352 if (!ACCESS_ONCE(rnp->qsmask) &&
1353 !rcu_preempt_blocked_readers_cgp(rnp))
1355 /* If time for quiescent-state forcing, do it. */
1356 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1357 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1360 /* Deal with stray signal. */
1362 flush_signals(current);
1364 j = jiffies_till_next_fqs;
1367 jiffies_till_next_fqs = HZ;
1370 jiffies_till_next_fqs = 1;
1374 /* Handle grace-period end. */
1375 rcu_gp_cleanup(rsp);
1380 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1381 * in preparation for detecting the next grace period. The caller must hold
1382 * the root node's ->lock, which is released before return. Hard irqs must
1385 * Note that it is legal for a dying CPU (which is marked as offline) to
1386 * invoke this function. This can happen when the dying CPU reports its
1390 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1391 __releases(rcu_get_root(rsp)->lock)
1393 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1394 struct rcu_node *rnp = rcu_get_root(rsp);
1396 if (!rsp->gp_kthread ||
1397 !cpu_needs_another_gp(rsp, rdp)) {
1399 * Either we have not yet spawned the grace-period
1400 * task, this CPU does not need another grace period,
1401 * or a grace period is already in progress.
1402 * Either way, don't start a new grace period.
1404 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1409 * Because there is no grace period in progress right now,
1410 * any callbacks we have up to this point will be satisfied
1411 * by the next grace period. So promote all callbacks to be
1412 * handled after the end of the next grace period. If the
1413 * CPU is not yet aware of the end of the previous grace period,
1414 * we need to allow for the callback advancement that will
1415 * occur when it does become aware. Deadlock prevents us from
1416 * making it aware at this point: We cannot acquire a leaf
1417 * rcu_node ->lock while holding the root rcu_node ->lock.
1419 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1420 if (rdp->completed == rsp->completed)
1421 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1423 rsp->gp_flags = RCU_GP_FLAG_INIT;
1424 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
1426 /* Ensure that CPU is aware of completion of last grace period. */
1427 rcu_process_gp_end(rsp, rdp);
1428 local_irq_restore(flags);
1430 /* Wake up rcu_gp_kthread() to start the grace period. */
1431 wake_up(&rsp->gp_wq);
1435 * Report a full set of quiescent states to the specified rcu_state
1436 * data structure. This involves cleaning up after the prior grace
1437 * period and letting rcu_start_gp() start up the next grace period
1438 * if one is needed. Note that the caller must hold rnp->lock, as
1439 * required by rcu_start_gp(), which will release it.
1441 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1442 __releases(rcu_get_root(rsp)->lock)
1444 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1445 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1446 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1450 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1451 * Allows quiescent states for a group of CPUs to be reported at one go
1452 * to the specified rcu_node structure, though all the CPUs in the group
1453 * must be represented by the same rcu_node structure (which need not be
1454 * a leaf rcu_node structure, though it often will be). That structure's
1455 * lock must be held upon entry, and it is released before return.
1458 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1459 struct rcu_node *rnp, unsigned long flags)
1460 __releases(rnp->lock)
1462 struct rcu_node *rnp_c;
1464 /* Walk up the rcu_node hierarchy. */
1466 if (!(rnp->qsmask & mask)) {
1468 /* Our bit has already been cleared, so done. */
1469 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1472 rnp->qsmask &= ~mask;
1473 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1474 mask, rnp->qsmask, rnp->level,
1475 rnp->grplo, rnp->grphi,
1477 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1479 /* Other bits still set at this level, so done. */
1480 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1483 mask = rnp->grpmask;
1484 if (rnp->parent == NULL) {
1486 /* No more levels. Exit loop holding root lock. */
1490 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1493 raw_spin_lock_irqsave(&rnp->lock, flags);
1494 WARN_ON_ONCE(rnp_c->qsmask);
1498 * Get here if we are the last CPU to pass through a quiescent
1499 * state for this grace period. Invoke rcu_report_qs_rsp()
1500 * to clean up and start the next grace period if one is needed.
1502 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1506 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1507 * structure. This must be either called from the specified CPU, or
1508 * called when the specified CPU is known to be offline (and when it is
1509 * also known that no other CPU is concurrently trying to help the offline
1510 * CPU). The lastcomp argument is used to make sure we are still in the
1511 * grace period of interest. We don't want to end the current grace period
1512 * based on quiescent states detected in an earlier grace period!
1515 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1517 unsigned long flags;
1519 struct rcu_node *rnp;
1522 raw_spin_lock_irqsave(&rnp->lock, flags);
1523 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1524 rnp->completed == rnp->gpnum) {
1527 * The grace period in which this quiescent state was
1528 * recorded has ended, so don't report it upwards.
1529 * We will instead need a new quiescent state that lies
1530 * within the current grace period.
1532 rdp->passed_quiesce = 0; /* need qs for new gp. */
1533 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1536 mask = rdp->grpmask;
1537 if ((rnp->qsmask & mask) == 0) {
1538 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1540 rdp->qs_pending = 0;
1543 * This GP can't end until cpu checks in, so all of our
1544 * callbacks can be processed during the next GP.
1546 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1548 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1553 * Check to see if there is a new grace period of which this CPU
1554 * is not yet aware, and if so, set up local rcu_data state for it.
1555 * Otherwise, see if this CPU has just passed through its first
1556 * quiescent state for this grace period, and record that fact if so.
1559 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1561 /* If there is now a new grace period, record and return. */
1562 if (check_for_new_grace_period(rsp, rdp))
1566 * Does this CPU still need to do its part for current grace period?
1567 * If no, return and let the other CPUs do their part as well.
1569 if (!rdp->qs_pending)
1573 * Was there a quiescent state since the beginning of the grace
1574 * period? If no, then exit and wait for the next call.
1576 if (!rdp->passed_quiesce)
1580 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1583 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1586 #ifdef CONFIG_HOTPLUG_CPU
1589 * Send the specified CPU's RCU callbacks to the orphanage. The
1590 * specified CPU must be offline, and the caller must hold the
1594 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1595 struct rcu_node *rnp, struct rcu_data *rdp)
1598 * Orphan the callbacks. First adjust the counts. This is safe
1599 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1600 * cannot be running now. Thus no memory barrier is required.
1602 if (rdp->nxtlist != NULL) {
1603 rsp->qlen_lazy += rdp->qlen_lazy;
1604 rsp->qlen += rdp->qlen;
1605 rdp->n_cbs_orphaned += rdp->qlen;
1607 ACCESS_ONCE(rdp->qlen) = 0;
1611 * Next, move those callbacks still needing a grace period to
1612 * the orphanage, where some other CPU will pick them up.
1613 * Some of the callbacks might have gone partway through a grace
1614 * period, but that is too bad. They get to start over because we
1615 * cannot assume that grace periods are synchronized across CPUs.
1616 * We don't bother updating the ->nxttail[] array yet, instead
1617 * we just reset the whole thing later on.
1619 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1620 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1621 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1622 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1626 * Then move the ready-to-invoke callbacks to the orphanage,
1627 * where some other CPU will pick them up. These will not be
1628 * required to pass though another grace period: They are done.
1630 if (rdp->nxtlist != NULL) {
1631 *rsp->orphan_donetail = rdp->nxtlist;
1632 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1635 /* Finally, initialize the rcu_data structure's list to empty. */
1636 init_callback_list(rdp);
1640 * Adopt the RCU callbacks from the specified rcu_state structure's
1641 * orphanage. The caller must hold the ->orphan_lock.
1643 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1646 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1648 /* Do the accounting first. */
1649 rdp->qlen_lazy += rsp->qlen_lazy;
1650 rdp->qlen += rsp->qlen;
1651 rdp->n_cbs_adopted += rsp->qlen;
1652 if (rsp->qlen_lazy != rsp->qlen)
1653 rcu_idle_count_callbacks_posted();
1658 * We do not need a memory barrier here because the only way we
1659 * can get here if there is an rcu_barrier() in flight is if
1660 * we are the task doing the rcu_barrier().
1663 /* First adopt the ready-to-invoke callbacks. */
1664 if (rsp->orphan_donelist != NULL) {
1665 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1666 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1667 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1668 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1669 rdp->nxttail[i] = rsp->orphan_donetail;
1670 rsp->orphan_donelist = NULL;
1671 rsp->orphan_donetail = &rsp->orphan_donelist;
1674 /* And then adopt the callbacks that still need a grace period. */
1675 if (rsp->orphan_nxtlist != NULL) {
1676 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1677 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1678 rsp->orphan_nxtlist = NULL;
1679 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1684 * Trace the fact that this CPU is going offline.
1686 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1688 RCU_TRACE(unsigned long mask);
1689 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1690 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1692 RCU_TRACE(mask = rdp->grpmask);
1693 trace_rcu_grace_period(rsp->name,
1694 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1699 * The CPU has been completely removed, and some other CPU is reporting
1700 * this fact from process context. Do the remainder of the cleanup,
1701 * including orphaning the outgoing CPU's RCU callbacks, and also
1702 * adopting them. There can only be one CPU hotplug operation at a time,
1703 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1705 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1707 unsigned long flags;
1709 int need_report = 0;
1710 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1711 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1713 /* Adjust any no-longer-needed kthreads. */
1714 rcu_boost_kthread_setaffinity(rnp, -1);
1716 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1718 /* Exclude any attempts to start a new grace period. */
1719 mutex_lock(&rsp->onoff_mutex);
1720 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1722 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1723 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1724 rcu_adopt_orphan_cbs(rsp);
1726 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1727 mask = rdp->grpmask; /* rnp->grplo is constant. */
1729 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1730 rnp->qsmaskinit &= ~mask;
1731 if (rnp->qsmaskinit != 0) {
1732 if (rnp != rdp->mynode)
1733 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1736 if (rnp == rdp->mynode)
1737 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1739 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1740 mask = rnp->grpmask;
1742 } while (rnp != NULL);
1745 * We still hold the leaf rcu_node structure lock here, and
1746 * irqs are still disabled. The reason for this subterfuge is
1747 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1748 * held leads to deadlock.
1750 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1752 if (need_report & RCU_OFL_TASKS_NORM_GP)
1753 rcu_report_unblock_qs_rnp(rnp, flags);
1755 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1756 if (need_report & RCU_OFL_TASKS_EXP_GP)
1757 rcu_report_exp_rnp(rsp, rnp, true);
1758 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1759 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1760 cpu, rdp->qlen, rdp->nxtlist);
1761 init_callback_list(rdp);
1762 /* Disallow further callbacks on this CPU. */
1763 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1764 mutex_unlock(&rsp->onoff_mutex);
1767 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1769 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1773 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1777 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1780 * Invoke any RCU callbacks that have made it to the end of their grace
1781 * period. Thottle as specified by rdp->blimit.
1783 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1785 unsigned long flags;
1786 struct rcu_head *next, *list, **tail;
1787 long bl, count, count_lazy;
1790 /* If no callbacks are ready, just return.*/
1791 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1792 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1793 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1794 need_resched(), is_idle_task(current),
1795 rcu_is_callbacks_kthread());
1800 * Extract the list of ready callbacks, disabling to prevent
1801 * races with call_rcu() from interrupt handlers.
1803 local_irq_save(flags);
1804 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1806 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1807 list = rdp->nxtlist;
1808 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1809 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1810 tail = rdp->nxttail[RCU_DONE_TAIL];
1811 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1812 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1813 rdp->nxttail[i] = &rdp->nxtlist;
1814 local_irq_restore(flags);
1816 /* Invoke callbacks. */
1817 count = count_lazy = 0;
1821 debug_rcu_head_unqueue(list);
1822 if (__rcu_reclaim(rsp->name, list))
1825 /* Stop only if limit reached and CPU has something to do. */
1826 if (++count >= bl &&
1828 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1832 local_irq_save(flags);
1833 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1834 is_idle_task(current),
1835 rcu_is_callbacks_kthread());
1837 /* Update count, and requeue any remaining callbacks. */
1839 *tail = rdp->nxtlist;
1840 rdp->nxtlist = list;
1841 for (i = 0; i < RCU_NEXT_SIZE; i++)
1842 if (&rdp->nxtlist == rdp->nxttail[i])
1843 rdp->nxttail[i] = tail;
1847 smp_mb(); /* List handling before counting for rcu_barrier(). */
1848 rdp->qlen_lazy -= count_lazy;
1849 ACCESS_ONCE(rdp->qlen) -= count;
1850 rdp->n_cbs_invoked += count;
1852 /* Reinstate batch limit if we have worked down the excess. */
1853 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1854 rdp->blimit = blimit;
1856 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1857 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1858 rdp->qlen_last_fqs_check = 0;
1859 rdp->n_force_qs_snap = rsp->n_force_qs;
1860 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1861 rdp->qlen_last_fqs_check = rdp->qlen;
1862 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1864 local_irq_restore(flags);
1866 /* Re-invoke RCU core processing if there are callbacks remaining. */
1867 if (cpu_has_callbacks_ready_to_invoke(rdp))
1872 * Check to see if this CPU is in a non-context-switch quiescent state
1873 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1874 * Also schedule RCU core processing.
1876 * This function must be called from hardirq context. It is normally
1877 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1878 * false, there is no point in invoking rcu_check_callbacks().
1880 void rcu_check_callbacks(int cpu, int user)
1882 trace_rcu_utilization("Start scheduler-tick");
1883 increment_cpu_stall_ticks();
1884 if (user || rcu_is_cpu_rrupt_from_idle()) {
1887 * Get here if this CPU took its interrupt from user
1888 * mode or from the idle loop, and if this is not a
1889 * nested interrupt. In this case, the CPU is in
1890 * a quiescent state, so note it.
1892 * No memory barrier is required here because both
1893 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1894 * variables that other CPUs neither access nor modify,
1895 * at least not while the corresponding CPU is online.
1901 } else if (!in_softirq()) {
1904 * Get here if this CPU did not take its interrupt from
1905 * softirq, in other words, if it is not interrupting
1906 * a rcu_bh read-side critical section. This is an _bh
1907 * critical section, so note it.
1912 rcu_preempt_check_callbacks(cpu);
1913 if (rcu_pending(cpu))
1915 trace_rcu_utilization("End scheduler-tick");
1919 * Scan the leaf rcu_node structures, processing dyntick state for any that
1920 * have not yet encountered a quiescent state, using the function specified.
1921 * Also initiate boosting for any threads blocked on the root rcu_node.
1923 * The caller must have suppressed start of new grace periods.
1925 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1929 unsigned long flags;
1931 struct rcu_node *rnp;
1933 rcu_for_each_leaf_node(rsp, rnp) {
1936 raw_spin_lock_irqsave(&rnp->lock, flags);
1937 if (!rcu_gp_in_progress(rsp)) {
1938 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1941 if (rnp->qsmask == 0) {
1942 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1947 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1948 if ((rnp->qsmask & bit) != 0 &&
1949 f(per_cpu_ptr(rsp->rda, cpu)))
1954 /* rcu_report_qs_rnp() releases rnp->lock. */
1955 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1958 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1960 rnp = rcu_get_root(rsp);
1961 if (rnp->qsmask == 0) {
1962 raw_spin_lock_irqsave(&rnp->lock, flags);
1963 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1968 * Force quiescent states on reluctant CPUs, and also detect which
1969 * CPUs are in dyntick-idle mode.
1971 static void force_quiescent_state(struct rcu_state *rsp)
1973 unsigned long flags;
1975 struct rcu_node *rnp;
1976 struct rcu_node *rnp_old = NULL;
1978 /* Funnel through hierarchy to reduce memory contention. */
1979 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1980 for (; rnp != NULL; rnp = rnp->parent) {
1981 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1982 !raw_spin_trylock(&rnp->fqslock);
1983 if (rnp_old != NULL)
1984 raw_spin_unlock(&rnp_old->fqslock);
1986 rsp->n_force_qs_lh++;
1991 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1993 /* Reached the root of the rcu_node tree, acquire lock. */
1994 raw_spin_lock_irqsave(&rnp_old->lock, flags);
1995 raw_spin_unlock(&rnp_old->fqslock);
1996 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1997 rsp->n_force_qs_lh++;
1998 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1999 return; /* Someone beat us to it. */
2001 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2002 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2003 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2007 * This does the RCU core processing work for the specified rcu_state
2008 * and rcu_data structures. This may be called only from the CPU to
2009 * whom the rdp belongs.
2012 __rcu_process_callbacks(struct rcu_state *rsp)
2014 unsigned long flags;
2015 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2017 WARN_ON_ONCE(rdp->beenonline == 0);
2020 * Advance callbacks in response to end of earlier grace
2021 * period that some other CPU ended.
2023 rcu_process_gp_end(rsp, rdp);
2025 /* Update RCU state based on any recent quiescent states. */
2026 rcu_check_quiescent_state(rsp, rdp);
2028 /* Does this CPU require a not-yet-started grace period? */
2029 if (cpu_needs_another_gp(rsp, rdp)) {
2030 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
2031 rcu_start_gp(rsp, flags); /* releases above lock */
2034 /* If there are callbacks ready, invoke them. */
2035 if (cpu_has_callbacks_ready_to_invoke(rdp))
2036 invoke_rcu_callbacks(rsp, rdp);
2040 * Do RCU core processing for the current CPU.
2042 static void rcu_process_callbacks(struct softirq_action *unused)
2044 struct rcu_state *rsp;
2046 if (cpu_is_offline(smp_processor_id()))
2048 trace_rcu_utilization("Start RCU core");
2049 for_each_rcu_flavor(rsp)
2050 __rcu_process_callbacks(rsp);
2051 trace_rcu_utilization("End RCU core");
2055 * Schedule RCU callback invocation. If the specified type of RCU
2056 * does not support RCU priority boosting, just do a direct call,
2057 * otherwise wake up the per-CPU kernel kthread. Note that because we
2058 * are running on the current CPU with interrupts disabled, the
2059 * rcu_cpu_kthread_task cannot disappear out from under us.
2061 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2063 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2065 if (likely(!rsp->boost)) {
2066 rcu_do_batch(rsp, rdp);
2069 invoke_rcu_callbacks_kthread();
2072 static void invoke_rcu_core(void)
2074 raise_softirq(RCU_SOFTIRQ);
2078 * Handle any core-RCU processing required by a call_rcu() invocation.
2080 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2081 struct rcu_head *head, unsigned long flags)
2084 * If called from an extended quiescent state, invoke the RCU
2085 * core in order to force a re-evaluation of RCU's idleness.
2087 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2090 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2091 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2095 * Force the grace period if too many callbacks or too long waiting.
2096 * Enforce hysteresis, and don't invoke force_quiescent_state()
2097 * if some other CPU has recently done so. Also, don't bother
2098 * invoking force_quiescent_state() if the newly enqueued callback
2099 * is the only one waiting for a grace period to complete.
2101 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2103 /* Are we ignoring a completed grace period? */
2104 rcu_process_gp_end(rsp, rdp);
2105 check_for_new_grace_period(rsp, rdp);
2107 /* Start a new grace period if one not already started. */
2108 if (!rcu_gp_in_progress(rsp)) {
2109 unsigned long nestflag;
2110 struct rcu_node *rnp_root = rcu_get_root(rsp);
2112 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2113 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
2115 /* Give the grace period a kick. */
2116 rdp->blimit = LONG_MAX;
2117 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2118 *rdp->nxttail[RCU_DONE_TAIL] != head)
2119 force_quiescent_state(rsp);
2120 rdp->n_force_qs_snap = rsp->n_force_qs;
2121 rdp->qlen_last_fqs_check = rdp->qlen;
2127 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2128 struct rcu_state *rsp, bool lazy)
2130 unsigned long flags;
2131 struct rcu_data *rdp;
2133 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2134 debug_rcu_head_queue(head);
2139 * Opportunistically note grace-period endings and beginnings.
2140 * Note that we might see a beginning right after we see an
2141 * end, but never vice versa, since this CPU has to pass through
2142 * a quiescent state betweentimes.
2144 local_irq_save(flags);
2145 rdp = this_cpu_ptr(rsp->rda);
2147 /* Add the callback to our list. */
2148 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL)) {
2149 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2151 local_irq_restore(flags);
2154 ACCESS_ONCE(rdp->qlen)++;
2158 rcu_idle_count_callbacks_posted();
2159 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2160 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2161 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2163 if (__is_kfree_rcu_offset((unsigned long)func))
2164 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2165 rdp->qlen_lazy, rdp->qlen);
2167 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2169 /* Go handle any RCU core processing required. */
2170 __call_rcu_core(rsp, rdp, head, flags);
2171 local_irq_restore(flags);
2175 * Queue an RCU-sched callback for invocation after a grace period.
2177 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2179 __call_rcu(head, func, &rcu_sched_state, 0);
2181 EXPORT_SYMBOL_GPL(call_rcu_sched);
2184 * Queue an RCU callback for invocation after a quicker grace period.
2186 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2188 __call_rcu(head, func, &rcu_bh_state, 0);
2190 EXPORT_SYMBOL_GPL(call_rcu_bh);
2193 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2194 * any blocking grace-period wait automatically implies a grace period
2195 * if there is only one CPU online at any point time during execution
2196 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2197 * occasionally incorrectly indicate that there are multiple CPUs online
2198 * when there was in fact only one the whole time, as this just adds
2199 * some overhead: RCU still operates correctly.
2201 static inline int rcu_blocking_is_gp(void)
2205 might_sleep(); /* Check for RCU read-side critical section. */
2207 ret = num_online_cpus() <= 1;
2213 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2215 * Control will return to the caller some time after a full rcu-sched
2216 * grace period has elapsed, in other words after all currently executing
2217 * rcu-sched read-side critical sections have completed. These read-side
2218 * critical sections are delimited by rcu_read_lock_sched() and
2219 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2220 * local_irq_disable(), and so on may be used in place of
2221 * rcu_read_lock_sched().
2223 * This means that all preempt_disable code sequences, including NMI and
2224 * hardware-interrupt handlers, in progress on entry will have completed
2225 * before this primitive returns. However, this does not guarantee that
2226 * softirq handlers will have completed, since in some kernels, these
2227 * handlers can run in process context, and can block.
2229 * Note that this guarantee implies a further memory-ordering guarantee.
2230 * On systems with more than one CPU, when synchronize_sched() returns,
2231 * each CPU is guaranteed to have executed a full memory barrier since
2232 * the end of its last RCU-sched read-side critical section whose beginning
2233 * preceded the call to synchronize_sched(). Note that this guarantee
2234 * includes CPUs that are offline, idle, or executing in user mode, as
2235 * well as CPUs that are executing in the kernel. Furthermore, if CPU A
2236 * invoked synchronize_sched(), which returned to its caller on CPU B,
2237 * then both CPU A and CPU B are guaranteed to have executed a full memory
2238 * barrier during the execution of synchronize_sched() -- even if CPU A
2239 * and CPU B are the same CPU (but again only if the system has more than
2242 * This primitive provides the guarantees made by the (now removed)
2243 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2244 * guarantees that rcu_read_lock() sections will have completed.
2245 * In "classic RCU", these two guarantees happen to be one and
2246 * the same, but can differ in realtime RCU implementations.
2248 void synchronize_sched(void)
2250 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2251 !lock_is_held(&rcu_lock_map) &&
2252 !lock_is_held(&rcu_sched_lock_map),
2253 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2254 if (rcu_blocking_is_gp())
2257 synchronize_sched_expedited();
2259 wait_rcu_gp(call_rcu_sched);
2261 EXPORT_SYMBOL_GPL(synchronize_sched);
2264 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2266 * Control will return to the caller some time after a full rcu_bh grace
2267 * period has elapsed, in other words after all currently executing rcu_bh
2268 * read-side critical sections have completed. RCU read-side critical
2269 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2270 * and may be nested.
2272 * See the description of synchronize_sched() for more detailed information
2273 * on memory ordering guarantees.
2275 void synchronize_rcu_bh(void)
2277 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2278 !lock_is_held(&rcu_lock_map) &&
2279 !lock_is_held(&rcu_sched_lock_map),
2280 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2281 if (rcu_blocking_is_gp())
2284 synchronize_rcu_bh_expedited();
2286 wait_rcu_gp(call_rcu_bh);
2288 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2290 static int synchronize_sched_expedited_cpu_stop(void *data)
2293 * There must be a full memory barrier on each affected CPU
2294 * between the time that try_stop_cpus() is called and the
2295 * time that it returns.
2297 * In the current initial implementation of cpu_stop, the
2298 * above condition is already met when the control reaches
2299 * this point and the following smp_mb() is not strictly
2300 * necessary. Do smp_mb() anyway for documentation and
2301 * robustness against future implementation changes.
2303 smp_mb(); /* See above comment block. */
2308 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2310 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2311 * approach to force the grace period to end quickly. This consumes
2312 * significant time on all CPUs and is unfriendly to real-time workloads,
2313 * so is thus not recommended for any sort of common-case code. In fact,
2314 * if you are using synchronize_sched_expedited() in a loop, please
2315 * restructure your code to batch your updates, and then use a single
2316 * synchronize_sched() instead.
2318 * Note that it is illegal to call this function while holding any lock
2319 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2320 * to call this function from a CPU-hotplug notifier. Failing to observe
2321 * these restriction will result in deadlock.
2323 * This implementation can be thought of as an application of ticket
2324 * locking to RCU, with sync_sched_expedited_started and
2325 * sync_sched_expedited_done taking on the roles of the halves
2326 * of the ticket-lock word. Each task atomically increments
2327 * sync_sched_expedited_started upon entry, snapshotting the old value,
2328 * then attempts to stop all the CPUs. If this succeeds, then each
2329 * CPU will have executed a context switch, resulting in an RCU-sched
2330 * grace period. We are then done, so we use atomic_cmpxchg() to
2331 * update sync_sched_expedited_done to match our snapshot -- but
2332 * only if someone else has not already advanced past our snapshot.
2334 * On the other hand, if try_stop_cpus() fails, we check the value
2335 * of sync_sched_expedited_done. If it has advanced past our
2336 * initial snapshot, then someone else must have forced a grace period
2337 * some time after we took our snapshot. In this case, our work is
2338 * done for us, and we can simply return. Otherwise, we try again,
2339 * but keep our initial snapshot for purposes of checking for someone
2340 * doing our work for us.
2342 * If we fail too many times in a row, we fall back to synchronize_sched().
2344 void synchronize_sched_expedited(void)
2346 long firstsnap, s, snap;
2348 struct rcu_state *rsp = &rcu_sched_state;
2351 * If we are in danger of counter wrap, just do synchronize_sched().
2352 * By allowing sync_sched_expedited_started to advance no more than
2353 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2354 * that more than 3.5 billion CPUs would be required to force a
2355 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2356 * course be required on a 64-bit system.
2358 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2359 (ulong)atomic_long_read(&rsp->expedited_done) +
2361 synchronize_sched();
2362 atomic_long_inc(&rsp->expedited_wrap);
2367 * Take a ticket. Note that atomic_inc_return() implies a
2368 * full memory barrier.
2370 snap = atomic_long_inc_return(&rsp->expedited_start);
2373 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2376 * Each pass through the following loop attempts to force a
2377 * context switch on each CPU.
2379 while (try_stop_cpus(cpu_online_mask,
2380 synchronize_sched_expedited_cpu_stop,
2383 atomic_long_inc(&rsp->expedited_tryfail);
2385 /* Check to see if someone else did our work for us. */
2386 s = atomic_long_read(&rsp->expedited_done);
2387 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2388 /* ensure test happens before caller kfree */
2389 smp_mb__before_atomic_inc(); /* ^^^ */
2390 atomic_long_inc(&rsp->expedited_workdone1);
2394 /* No joy, try again later. Or just synchronize_sched(). */
2395 if (trycount++ < 10) {
2396 udelay(trycount * num_online_cpus());
2398 wait_rcu_gp(call_rcu_sched);
2399 atomic_long_inc(&rsp->expedited_normal);
2403 /* Recheck to see if someone else did our work for us. */
2404 s = atomic_long_read(&rsp->expedited_done);
2405 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2406 /* ensure test happens before caller kfree */
2407 smp_mb__before_atomic_inc(); /* ^^^ */
2408 atomic_long_inc(&rsp->expedited_workdone2);
2413 * Refetching sync_sched_expedited_started allows later
2414 * callers to piggyback on our grace period. We retry
2415 * after they started, so our grace period works for them,
2416 * and they started after our first try, so their grace
2417 * period works for us.
2420 snap = atomic_long_read(&rsp->expedited_start);
2421 smp_mb(); /* ensure read is before try_stop_cpus(). */
2423 atomic_long_inc(&rsp->expedited_stoppedcpus);
2426 * Everyone up to our most recent fetch is covered by our grace
2427 * period. Update the counter, but only if our work is still
2428 * relevant -- which it won't be if someone who started later
2429 * than we did already did their update.
2432 atomic_long_inc(&rsp->expedited_done_tries);
2433 s = atomic_long_read(&rsp->expedited_done);
2434 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2435 /* ensure test happens before caller kfree */
2436 smp_mb__before_atomic_inc(); /* ^^^ */
2437 atomic_long_inc(&rsp->expedited_done_lost);
2440 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2441 atomic_long_inc(&rsp->expedited_done_exit);
2445 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2448 * Check to see if there is any immediate RCU-related work to be done
2449 * by the current CPU, for the specified type of RCU, returning 1 if so.
2450 * The checks are in order of increasing expense: checks that can be
2451 * carried out against CPU-local state are performed first. However,
2452 * we must check for CPU stalls first, else we might not get a chance.
2454 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2456 struct rcu_node *rnp = rdp->mynode;
2458 rdp->n_rcu_pending++;
2460 /* Check for CPU stalls, if enabled. */
2461 check_cpu_stall(rsp, rdp);
2463 /* Is the RCU core waiting for a quiescent state from this CPU? */
2464 if (rcu_scheduler_fully_active &&
2465 rdp->qs_pending && !rdp->passed_quiesce) {
2466 rdp->n_rp_qs_pending++;
2467 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2468 rdp->n_rp_report_qs++;
2472 /* Does this CPU have callbacks ready to invoke? */
2473 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2474 rdp->n_rp_cb_ready++;
2478 /* Has RCU gone idle with this CPU needing another grace period? */
2479 if (cpu_needs_another_gp(rsp, rdp)) {
2480 rdp->n_rp_cpu_needs_gp++;
2484 /* Has another RCU grace period completed? */
2485 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2486 rdp->n_rp_gp_completed++;
2490 /* Has a new RCU grace period started? */
2491 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2492 rdp->n_rp_gp_started++;
2497 rdp->n_rp_need_nothing++;
2502 * Check to see if there is any immediate RCU-related work to be done
2503 * by the current CPU, returning 1 if so. This function is part of the
2504 * RCU implementation; it is -not- an exported member of the RCU API.
2506 static int rcu_pending(int cpu)
2508 struct rcu_state *rsp;
2510 for_each_rcu_flavor(rsp)
2511 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2517 * Check to see if any future RCU-related work will need to be done
2518 * by the current CPU, even if none need be done immediately, returning
2521 static int rcu_cpu_has_callbacks(int cpu)
2523 struct rcu_state *rsp;
2525 /* RCU callbacks either ready or pending? */
2526 for_each_rcu_flavor(rsp)
2527 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2533 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2534 * the compiler is expected to optimize this away.
2536 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2537 int cpu, unsigned long done)
2539 trace_rcu_barrier(rsp->name, s, cpu,
2540 atomic_read(&rsp->barrier_cpu_count), done);
2544 * RCU callback function for _rcu_barrier(). If we are last, wake
2545 * up the task executing _rcu_barrier().
2547 static void rcu_barrier_callback(struct rcu_head *rhp)
2549 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2550 struct rcu_state *rsp = rdp->rsp;
2552 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2553 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2554 complete(&rsp->barrier_completion);
2556 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2561 * Called with preemption disabled, and from cross-cpu IRQ context.
2563 static void rcu_barrier_func(void *type)
2565 struct rcu_state *rsp = type;
2566 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2568 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2569 atomic_inc(&rsp->barrier_cpu_count);
2570 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2574 * Orchestrate the specified type of RCU barrier, waiting for all
2575 * RCU callbacks of the specified type to complete.
2577 static void _rcu_barrier(struct rcu_state *rsp)
2580 struct rcu_data *rdp;
2581 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2582 unsigned long snap_done;
2584 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2586 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2587 mutex_lock(&rsp->barrier_mutex);
2590 * Ensure that all prior references, including to ->n_barrier_done,
2591 * are ordered before the _rcu_barrier() machinery.
2593 smp_mb(); /* See above block comment. */
2596 * Recheck ->n_barrier_done to see if others did our work for us.
2597 * This means checking ->n_barrier_done for an even-to-odd-to-even
2598 * transition. The "if" expression below therefore rounds the old
2599 * value up to the next even number and adds two before comparing.
2601 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2602 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2603 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2604 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2605 smp_mb(); /* caller's subsequent code after above check. */
2606 mutex_unlock(&rsp->barrier_mutex);
2611 * Increment ->n_barrier_done to avoid duplicate work. Use
2612 * ACCESS_ONCE() to prevent the compiler from speculating
2613 * the increment to precede the early-exit check.
2615 ACCESS_ONCE(rsp->n_barrier_done)++;
2616 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2617 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2618 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2621 * Initialize the count to one rather than to zero in order to
2622 * avoid a too-soon return to zero in case of a short grace period
2623 * (or preemption of this task). Exclude CPU-hotplug operations
2624 * to ensure that no offline CPU has callbacks queued.
2626 init_completion(&rsp->barrier_completion);
2627 atomic_set(&rsp->barrier_cpu_count, 1);
2631 * Force each CPU with callbacks to register a new callback.
2632 * When that callback is invoked, we will know that all of the
2633 * corresponding CPU's preceding callbacks have been invoked.
2635 for_each_online_cpu(cpu) {
2636 rdp = per_cpu_ptr(rsp->rda, cpu);
2637 if (ACCESS_ONCE(rdp->qlen)) {
2638 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2639 rsp->n_barrier_done);
2640 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2642 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2643 rsp->n_barrier_done);
2649 * Now that we have an rcu_barrier_callback() callback on each
2650 * CPU, and thus each counted, remove the initial count.
2652 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2653 complete(&rsp->barrier_completion);
2655 /* Increment ->n_barrier_done to prevent duplicate work. */
2656 smp_mb(); /* Keep increment after above mechanism. */
2657 ACCESS_ONCE(rsp->n_barrier_done)++;
2658 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2659 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2660 smp_mb(); /* Keep increment before caller's subsequent code. */
2662 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2663 wait_for_completion(&rsp->barrier_completion);
2665 /* Other rcu_barrier() invocations can now safely proceed. */
2666 mutex_unlock(&rsp->barrier_mutex);
2670 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2672 void rcu_barrier_bh(void)
2674 _rcu_barrier(&rcu_bh_state);
2676 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2679 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2681 void rcu_barrier_sched(void)
2683 _rcu_barrier(&rcu_sched_state);
2685 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2688 * Do boot-time initialization of a CPU's per-CPU RCU data.
2691 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2693 unsigned long flags;
2694 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2695 struct rcu_node *rnp = rcu_get_root(rsp);
2697 /* Set up local state, ensuring consistent view of global state. */
2698 raw_spin_lock_irqsave(&rnp->lock, flags);
2699 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2700 init_callback_list(rdp);
2702 ACCESS_ONCE(rdp->qlen) = 0;
2703 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2704 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2705 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2706 #ifdef CONFIG_RCU_USER_QS
2707 WARN_ON_ONCE(rdp->dynticks->in_user);
2711 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2715 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2716 * offline event can be happening at a given time. Note also that we
2717 * can accept some slop in the rsp->completed access due to the fact
2718 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2720 static void __cpuinit
2721 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2723 unsigned long flags;
2725 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2726 struct rcu_node *rnp = rcu_get_root(rsp);
2728 /* Exclude new grace periods. */
2729 mutex_lock(&rsp->onoff_mutex);
2731 /* Set up local state, ensuring consistent view of global state. */
2732 raw_spin_lock_irqsave(&rnp->lock, flags);
2733 rdp->beenonline = 1; /* We have now been online. */
2734 rdp->preemptible = preemptible;
2735 rdp->qlen_last_fqs_check = 0;
2736 rdp->n_force_qs_snap = rsp->n_force_qs;
2737 rdp->blimit = blimit;
2738 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2739 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2740 atomic_set(&rdp->dynticks->dynticks,
2741 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2742 rcu_prepare_for_idle_init(cpu);
2743 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2745 /* Add CPU to rcu_node bitmasks. */
2747 mask = rdp->grpmask;
2749 /* Exclude any attempts to start a new GP on small systems. */
2750 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2751 rnp->qsmaskinit |= mask;
2752 mask = rnp->grpmask;
2753 if (rnp == rdp->mynode) {
2755 * If there is a grace period in progress, we will
2756 * set up to wait for it next time we run the
2759 rdp->gpnum = rnp->completed;
2760 rdp->completed = rnp->completed;
2761 rdp->passed_quiesce = 0;
2762 rdp->qs_pending = 0;
2763 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2765 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2767 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2768 local_irq_restore(flags);
2770 mutex_unlock(&rsp->onoff_mutex);
2773 static void __cpuinit rcu_prepare_cpu(int cpu)
2775 struct rcu_state *rsp;
2777 for_each_rcu_flavor(rsp)
2778 rcu_init_percpu_data(cpu, rsp,
2779 strcmp(rsp->name, "rcu_preempt") == 0);
2783 * Handle CPU online/offline notification events.
2785 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2786 unsigned long action, void *hcpu)
2788 long cpu = (long)hcpu;
2789 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2790 struct rcu_node *rnp = rdp->mynode;
2791 struct rcu_state *rsp;
2793 trace_rcu_utilization("Start CPU hotplug");
2795 case CPU_UP_PREPARE:
2796 case CPU_UP_PREPARE_FROZEN:
2797 rcu_prepare_cpu(cpu);
2798 rcu_prepare_kthreads(cpu);
2801 case CPU_DOWN_FAILED:
2802 rcu_boost_kthread_setaffinity(rnp, -1);
2804 case CPU_DOWN_PREPARE:
2805 rcu_boost_kthread_setaffinity(rnp, cpu);
2808 case CPU_DYING_FROZEN:
2810 * The whole machine is "stopped" except this CPU, so we can
2811 * touch any data without introducing corruption. We send the
2812 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2814 for_each_rcu_flavor(rsp)
2815 rcu_cleanup_dying_cpu(rsp);
2816 rcu_cleanup_after_idle(cpu);
2819 case CPU_DEAD_FROZEN:
2820 case CPU_UP_CANCELED:
2821 case CPU_UP_CANCELED_FROZEN:
2822 for_each_rcu_flavor(rsp)
2823 rcu_cleanup_dead_cpu(cpu, rsp);
2828 trace_rcu_utilization("End CPU hotplug");
2833 * Spawn the kthread that handles this RCU flavor's grace periods.
2835 static int __init rcu_spawn_gp_kthread(void)
2837 unsigned long flags;
2838 struct rcu_node *rnp;
2839 struct rcu_state *rsp;
2840 struct task_struct *t;
2842 for_each_rcu_flavor(rsp) {
2843 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2845 rnp = rcu_get_root(rsp);
2846 raw_spin_lock_irqsave(&rnp->lock, flags);
2847 rsp->gp_kthread = t;
2848 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2852 early_initcall(rcu_spawn_gp_kthread);
2855 * This function is invoked towards the end of the scheduler's initialization
2856 * process. Before this is called, the idle task might contain
2857 * RCU read-side critical sections (during which time, this idle
2858 * task is booting the system). After this function is called, the
2859 * idle tasks are prohibited from containing RCU read-side critical
2860 * sections. This function also enables RCU lockdep checking.
2862 void rcu_scheduler_starting(void)
2864 WARN_ON(num_online_cpus() != 1);
2865 WARN_ON(nr_context_switches() > 0);
2866 rcu_scheduler_active = 1;
2870 * Compute the per-level fanout, either using the exact fanout specified
2871 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2873 #ifdef CONFIG_RCU_FANOUT_EXACT
2874 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2878 for (i = rcu_num_lvls - 1; i > 0; i--)
2879 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2880 rsp->levelspread[0] = rcu_fanout_leaf;
2882 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2883 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2890 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2891 ccur = rsp->levelcnt[i];
2892 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2896 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2899 * Helper function for rcu_init() that initializes one rcu_state structure.
2901 static void __init rcu_init_one(struct rcu_state *rsp,
2902 struct rcu_data __percpu *rda)
2904 static char *buf[] = { "rcu_node_0",
2907 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2908 static char *fqs[] = { "rcu_node_fqs_0",
2911 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2915 struct rcu_node *rnp;
2917 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2919 /* Initialize the level-tracking arrays. */
2921 for (i = 0; i < rcu_num_lvls; i++)
2922 rsp->levelcnt[i] = num_rcu_lvl[i];
2923 for (i = 1; i < rcu_num_lvls; i++)
2924 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2925 rcu_init_levelspread(rsp);
2927 /* Initialize the elements themselves, starting from the leaves. */
2929 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2930 cpustride *= rsp->levelspread[i];
2931 rnp = rsp->level[i];
2932 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2933 raw_spin_lock_init(&rnp->lock);
2934 lockdep_set_class_and_name(&rnp->lock,
2935 &rcu_node_class[i], buf[i]);
2936 raw_spin_lock_init(&rnp->fqslock);
2937 lockdep_set_class_and_name(&rnp->fqslock,
2938 &rcu_fqs_class[i], fqs[i]);
2939 rnp->gpnum = rsp->gpnum;
2940 rnp->completed = rsp->completed;
2942 rnp->qsmaskinit = 0;
2943 rnp->grplo = j * cpustride;
2944 rnp->grphi = (j + 1) * cpustride - 1;
2945 if (rnp->grphi >= NR_CPUS)
2946 rnp->grphi = NR_CPUS - 1;
2952 rnp->grpnum = j % rsp->levelspread[i - 1];
2953 rnp->grpmask = 1UL << rnp->grpnum;
2954 rnp->parent = rsp->level[i - 1] +
2955 j / rsp->levelspread[i - 1];
2958 INIT_LIST_HEAD(&rnp->blkd_tasks);
2963 init_waitqueue_head(&rsp->gp_wq);
2964 rnp = rsp->level[rcu_num_lvls - 1];
2965 for_each_possible_cpu(i) {
2966 while (i > rnp->grphi)
2968 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2969 rcu_boot_init_percpu_data(i, rsp);
2971 list_add(&rsp->flavors, &rcu_struct_flavors);
2975 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2976 * replace the definitions in rcutree.h because those are needed to size
2977 * the ->node array in the rcu_state structure.
2979 static void __init rcu_init_geometry(void)
2984 int rcu_capacity[MAX_RCU_LVLS + 1];
2986 /* If the compile-time values are accurate, just leave. */
2987 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
2988 nr_cpu_ids == NR_CPUS)
2992 * Compute number of nodes that can be handled an rcu_node tree
2993 * with the given number of levels. Setting rcu_capacity[0] makes
2994 * some of the arithmetic easier.
2996 rcu_capacity[0] = 1;
2997 rcu_capacity[1] = rcu_fanout_leaf;
2998 for (i = 2; i <= MAX_RCU_LVLS; i++)
2999 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3002 * The boot-time rcu_fanout_leaf parameter is only permitted
3003 * to increase the leaf-level fanout, not decrease it. Of course,
3004 * the leaf-level fanout cannot exceed the number of bits in
3005 * the rcu_node masks. Finally, the tree must be able to accommodate
3006 * the configured number of CPUs. Complain and fall back to the
3007 * compile-time values if these limits are exceeded.
3009 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3010 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3011 n > rcu_capacity[MAX_RCU_LVLS]) {
3016 /* Calculate the number of rcu_nodes at each level of the tree. */
3017 for (i = 1; i <= MAX_RCU_LVLS; i++)
3018 if (n <= rcu_capacity[i]) {
3019 for (j = 0; j <= i; j++)
3021 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3023 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3028 /* Calculate the total number of rcu_node structures. */
3030 for (i = 0; i <= MAX_RCU_LVLS; i++)
3031 rcu_num_nodes += num_rcu_lvl[i];
3035 void __init rcu_init(void)
3039 rcu_bootup_announce();
3040 rcu_init_geometry();
3041 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3042 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3043 __rcu_init_preempt();
3044 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3047 * We don't need protection against CPU-hotplug here because
3048 * this is called early in boot, before either interrupts
3049 * or the scheduler are operational.
3051 cpu_notifier(rcu_cpu_notify, 0);
3052 for_each_online_cpu(cpu)
3053 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3056 #include "rcutree_plugin.h"