2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
57 #include <trace/events/rcu.h>
61 /* Data structures. */
63 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
65 #define RCU_STATE_INITIALIZER(structname) { \
66 .level = { &structname##_state.node[0] }, \
68 NUM_RCU_LVL_0, /* root of hierarchy. */ \
72 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
74 .fqs_state = RCU_GP_IDLE, \
77 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
78 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
80 .n_force_qs_ngp = 0, \
81 .name = #structname, \
84 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched);
85 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
87 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh);
88 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
90 static struct rcu_state *rcu_state;
93 * The rcu_scheduler_active variable transitions from zero to one just
94 * before the first task is spawned. So when this variable is zero, RCU
95 * can assume that there is but one task, allowing RCU to (for example)
96 * optimized synchronize_sched() to a simple barrier(). When this variable
97 * is one, RCU must actually do all the hard work required to detect real
98 * grace periods. This variable is also used to suppress boot-time false
99 * positives from lockdep-RCU error checking.
101 int rcu_scheduler_active __read_mostly;
102 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
105 * The rcu_scheduler_fully_active variable transitions from zero to one
106 * during the early_initcall() processing, which is after the scheduler
107 * is capable of creating new tasks. So RCU processing (for example,
108 * creating tasks for RCU priority boosting) must be delayed until after
109 * rcu_scheduler_fully_active transitions from zero to one. We also
110 * currently delay invocation of any RCU callbacks until after this point.
112 * It might later prove better for people registering RCU callbacks during
113 * early boot to take responsibility for these callbacks, but one step at
116 static int rcu_scheduler_fully_active __read_mostly;
118 #ifdef CONFIG_RCU_BOOST
121 * Control variables for per-CPU and per-rcu_node kthreads. These
122 * handle all flavors of RCU.
124 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
126 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
127 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
128 DEFINE_PER_CPU(char, rcu_cpu_has_work);
130 #endif /* #ifdef CONFIG_RCU_BOOST */
132 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
133 static void invoke_rcu_core(void);
134 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
137 * Track the rcutorture test sequence number and the update version
138 * number within a given test. The rcutorture_testseq is incremented
139 * on every rcutorture module load and unload, so has an odd value
140 * when a test is running. The rcutorture_vernum is set to zero
141 * when rcutorture starts and is incremented on each rcutorture update.
142 * These variables enable correlating rcutorture output with the
143 * RCU tracing information.
145 unsigned long rcutorture_testseq;
146 unsigned long rcutorture_vernum;
149 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
150 * permit this function to be invoked without holding the root rcu_node
151 * structure's ->lock, but of course results can be subject to change.
153 static int rcu_gp_in_progress(struct rcu_state *rsp)
155 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
159 * Note a quiescent state. Because we do not need to know
160 * how many quiescent states passed, just if there was at least
161 * one since the start of the grace period, this just sets a flag.
162 * The caller must have disabled preemption.
164 void rcu_sched_qs(int cpu)
166 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
168 rdp->passed_quiesce_gpnum = rdp->gpnum;
170 if (rdp->passed_quiesce == 0)
171 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
172 rdp->passed_quiesce = 1;
175 void rcu_bh_qs(int cpu)
177 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
179 rdp->passed_quiesce_gpnum = rdp->gpnum;
181 if (rdp->passed_quiesce == 0)
182 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
183 rdp->passed_quiesce = 1;
187 * Note a context switch. This is a quiescent state for RCU-sched,
188 * and requires special handling for preemptible RCU.
189 * The caller must have disabled preemption.
191 void rcu_note_context_switch(int cpu)
193 trace_rcu_utilization("Start context switch");
195 rcu_preempt_note_context_switch(cpu);
196 trace_rcu_utilization("End context switch");
198 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
200 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
201 .dynticks_nesting = DYNTICK_TASK_NESTING,
202 .dynticks = ATOMIC_INIT(1),
205 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
206 static int qhimark = 10000; /* If this many pending, ignore blimit. */
207 static int qlowmark = 100; /* Once only this many pending, use blimit. */
209 module_param(blimit, int, 0);
210 module_param(qhimark, int, 0);
211 module_param(qlowmark, int, 0);
213 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
214 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
216 module_param(rcu_cpu_stall_suppress, int, 0644);
217 module_param(rcu_cpu_stall_timeout, int, 0644);
219 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
220 static int rcu_pending(int cpu);
223 * Return the number of RCU-sched batches processed thus far for debug & stats.
225 long rcu_batches_completed_sched(void)
227 return rcu_sched_state.completed;
229 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
232 * Return the number of RCU BH batches processed thus far for debug & stats.
234 long rcu_batches_completed_bh(void)
236 return rcu_bh_state.completed;
238 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
241 * Force a quiescent state for RCU BH.
243 void rcu_bh_force_quiescent_state(void)
245 force_quiescent_state(&rcu_bh_state, 0);
247 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
250 * Record the number of times rcutorture tests have been initiated and
251 * terminated. This information allows the debugfs tracing stats to be
252 * correlated to the rcutorture messages, even when the rcutorture module
253 * is being repeatedly loaded and unloaded. In other words, we cannot
254 * store this state in rcutorture itself.
256 void rcutorture_record_test_transition(void)
258 rcutorture_testseq++;
259 rcutorture_vernum = 0;
261 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
264 * Record the number of writer passes through the current rcutorture test.
265 * This is also used to correlate debugfs tracing stats with the rcutorture
268 void rcutorture_record_progress(unsigned long vernum)
272 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
275 * Force a quiescent state for RCU-sched.
277 void rcu_sched_force_quiescent_state(void)
279 force_quiescent_state(&rcu_sched_state, 0);
281 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
284 * Does the CPU have callbacks ready to be invoked?
287 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
289 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
293 * Does the current CPU require a yet-as-unscheduled grace period?
296 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
298 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
302 * Return the root node of the specified rcu_state structure.
304 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
306 return &rsp->node[0];
310 * If the specified CPU is offline, tell the caller that it is in
311 * a quiescent state. Otherwise, whack it with a reschedule IPI.
312 * Grace periods can end up waiting on an offline CPU when that
313 * CPU is in the process of coming online -- it will be added to the
314 * rcu_node bitmasks before it actually makes it online. The same thing
315 * can happen while a CPU is in the process of coming online. Because this
316 * race is quite rare, we check for it after detecting that the grace
317 * period has been delayed rather than checking each and every CPU
318 * each and every time we start a new grace period.
320 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
323 * If the CPU is offline, it is in a quiescent state. We can
324 * trust its state not to change because interrupts are disabled.
326 if (cpu_is_offline(rdp->cpu)) {
327 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
333 * The CPU is online, so send it a reschedule IPI. This forces
334 * it through the scheduler, and (inefficiently) also handles cases
335 * where idle loops fail to inform RCU about the CPU being idle.
337 if (rdp->cpu != smp_processor_id())
338 smp_send_reschedule(rdp->cpu);
346 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
348 * If the new value of the ->dynticks_nesting counter now is zero,
349 * we really have entered idle, and must do the appropriate accounting.
350 * The caller must have disabled interrupts.
352 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
354 trace_rcu_dyntick("Start", oldval, 0);
355 if (!is_idle_task(current)) {
356 struct task_struct *idle = idle_task(smp_processor_id());
358 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
359 ftrace_dump(DUMP_ALL);
360 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
361 current->pid, current->comm,
362 idle->pid, idle->comm); /* must be idle task! */
364 rcu_prepare_for_idle(smp_processor_id());
365 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
366 smp_mb__before_atomic_inc(); /* See above. */
367 atomic_inc(&rdtp->dynticks);
368 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
369 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
372 * The idle task is not permitted to enter the idle loop while
373 * in an RCU read-side critical section.
375 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
376 "Illegal idle entry in RCU read-side critical section.");
377 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
378 "Illegal idle entry in RCU-bh read-side critical section.");
379 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
380 "Illegal idle entry in RCU-sched read-side critical section.");
384 * rcu_idle_enter - inform RCU that current CPU is entering idle
386 * Enter idle mode, in other words, -leave- the mode in which RCU
387 * read-side critical sections can occur. (Though RCU read-side
388 * critical sections can occur in irq handlers in idle, a possibility
389 * handled by irq_enter() and irq_exit().)
391 * We crowbar the ->dynticks_nesting field to zero to allow for
392 * the possibility of usermode upcalls having messed up our count
393 * of interrupt nesting level during the prior busy period.
395 void rcu_idle_enter(void)
399 struct rcu_dynticks *rdtp;
401 local_irq_save(flags);
402 rdtp = &__get_cpu_var(rcu_dynticks);
403 oldval = rdtp->dynticks_nesting;
404 rdtp->dynticks_nesting = 0;
405 rcu_idle_enter_common(rdtp, oldval);
406 local_irq_restore(flags);
410 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
412 * Exit from an interrupt handler, which might possibly result in entering
413 * idle mode, in other words, leaving the mode in which read-side critical
414 * sections can occur.
416 * This code assumes that the idle loop never does anything that might
417 * result in unbalanced calls to irq_enter() and irq_exit(). If your
418 * architecture violates this assumption, RCU will give you what you
419 * deserve, good and hard. But very infrequently and irreproducibly.
421 * Use things like work queues to work around this limitation.
423 * You have been warned.
425 void rcu_irq_exit(void)
429 struct rcu_dynticks *rdtp;
431 local_irq_save(flags);
432 rdtp = &__get_cpu_var(rcu_dynticks);
433 oldval = rdtp->dynticks_nesting;
434 rdtp->dynticks_nesting--;
435 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
436 if (rdtp->dynticks_nesting)
437 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
439 rcu_idle_enter_common(rdtp, oldval);
440 local_irq_restore(flags);
444 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
446 * If the new value of the ->dynticks_nesting counter was previously zero,
447 * we really have exited idle, and must do the appropriate accounting.
448 * The caller must have disabled interrupts.
450 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
452 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
453 atomic_inc(&rdtp->dynticks);
454 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
455 smp_mb__after_atomic_inc(); /* See above. */
456 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
457 rcu_cleanup_after_idle(smp_processor_id());
458 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
459 if (!is_idle_task(current)) {
460 struct task_struct *idle = idle_task(smp_processor_id());
462 trace_rcu_dyntick("Error on exit: not idle task",
463 oldval, rdtp->dynticks_nesting);
464 ftrace_dump(DUMP_ALL);
465 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
466 current->pid, current->comm,
467 idle->pid, idle->comm); /* must be idle task! */
472 * rcu_idle_exit - inform RCU that current CPU is leaving idle
474 * Exit idle mode, in other words, -enter- the mode in which RCU
475 * read-side critical sections can occur.
477 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NESTING to
478 * allow for the possibility of usermode upcalls messing up our count
479 * of interrupt nesting level during the busy period that is just
482 void rcu_idle_exit(void)
485 struct rcu_dynticks *rdtp;
488 local_irq_save(flags);
489 rdtp = &__get_cpu_var(rcu_dynticks);
490 oldval = rdtp->dynticks_nesting;
491 WARN_ON_ONCE(oldval != 0);
492 rdtp->dynticks_nesting = DYNTICK_TASK_NESTING;
493 rcu_idle_exit_common(rdtp, oldval);
494 local_irq_restore(flags);
498 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
500 * Enter an interrupt handler, which might possibly result in exiting
501 * idle mode, in other words, entering the mode in which read-side critical
502 * sections can occur.
504 * Note that the Linux kernel is fully capable of entering an interrupt
505 * handler that it never exits, for example when doing upcalls to
506 * user mode! This code assumes that the idle loop never does upcalls to
507 * user mode. If your architecture does do upcalls from the idle loop (or
508 * does anything else that results in unbalanced calls to the irq_enter()
509 * and irq_exit() functions), RCU will give you what you deserve, good
510 * and hard. But very infrequently and irreproducibly.
512 * Use things like work queues to work around this limitation.
514 * You have been warned.
516 void rcu_irq_enter(void)
519 struct rcu_dynticks *rdtp;
522 local_irq_save(flags);
523 rdtp = &__get_cpu_var(rcu_dynticks);
524 oldval = rdtp->dynticks_nesting;
525 rdtp->dynticks_nesting++;
526 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
528 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
530 rcu_idle_exit_common(rdtp, oldval);
531 local_irq_restore(flags);
535 * rcu_nmi_enter - inform RCU of entry to NMI context
537 * If the CPU was idle with dynamic ticks active, and there is no
538 * irq handler running, this updates rdtp->dynticks_nmi to let the
539 * RCU grace-period handling know that the CPU is active.
541 void rcu_nmi_enter(void)
543 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
545 if (rdtp->dynticks_nmi_nesting == 0 &&
546 (atomic_read(&rdtp->dynticks) & 0x1))
548 rdtp->dynticks_nmi_nesting++;
549 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
550 atomic_inc(&rdtp->dynticks);
551 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
552 smp_mb__after_atomic_inc(); /* See above. */
553 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
557 * rcu_nmi_exit - inform RCU of exit from NMI context
559 * If the CPU was idle with dynamic ticks active, and there is no
560 * irq handler running, this updates rdtp->dynticks_nmi to let the
561 * RCU grace-period handling know that the CPU is no longer active.
563 void rcu_nmi_exit(void)
565 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
567 if (rdtp->dynticks_nmi_nesting == 0 ||
568 --rdtp->dynticks_nmi_nesting != 0)
570 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
571 smp_mb__before_atomic_inc(); /* See above. */
572 atomic_inc(&rdtp->dynticks);
573 smp_mb__after_atomic_inc(); /* Force delay to next write. */
574 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
577 #ifdef CONFIG_PROVE_RCU
580 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
582 * If the current CPU is in its idle loop and is neither in an interrupt
583 * or NMI handler, return true.
585 int rcu_is_cpu_idle(void)
590 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
594 EXPORT_SYMBOL(rcu_is_cpu_idle);
596 #ifdef CONFIG_HOTPLUG_CPU
599 * Is the current CPU online? Disable preemption to avoid false positives
600 * that could otherwise happen due to the current CPU number being sampled,
601 * this task being preempted, its old CPU being taken offline, resuming
602 * on some other CPU, then determining that its old CPU is now offline.
603 * It is OK to use RCU on an offline processor during initial boot, hence
604 * the check for rcu_scheduler_fully_active.
606 * Disable checking if in an NMI handler because we cannot safely report
607 * errors from NMI handlers anyway.
609 bool rcu_lockdep_current_cpu_online(void)
616 ret = cpu_online(smp_processor_id()) ||
617 !rcu_scheduler_fully_active;
621 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
623 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
625 #endif /* #ifdef CONFIG_PROVE_RCU */
628 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
630 * If the current CPU is idle or running at a first-level (not nested)
631 * interrupt from idle, return true. The caller must have at least
632 * disabled preemption.
634 int rcu_is_cpu_rrupt_from_idle(void)
636 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
640 * Snapshot the specified CPU's dynticks counter so that we can later
641 * credit them with an implicit quiescent state. Return 1 if this CPU
642 * is in dynticks idle mode, which is an extended quiescent state.
644 static int dyntick_save_progress_counter(struct rcu_data *rdp)
646 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
647 return (rdp->dynticks_snap & 0x1) == 0;
651 * Return true if the specified CPU has passed through a quiescent
652 * state by virtue of being in or having passed through an dynticks
653 * idle state since the last call to dyntick_save_progress_counter()
656 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
661 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
662 snap = (unsigned int)rdp->dynticks_snap;
665 * If the CPU passed through or entered a dynticks idle phase with
666 * no active irq/NMI handlers, then we can safely pretend that the CPU
667 * already acknowledged the request to pass through a quiescent
668 * state. Either way, that CPU cannot possibly be in an RCU
669 * read-side critical section that started before the beginning
670 * of the current RCU grace period.
672 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
673 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
678 /* Go check for the CPU being offline. */
679 return rcu_implicit_offline_qs(rdp);
682 static int jiffies_till_stall_check(void)
684 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
687 * Limit check must be consistent with the Kconfig limits
688 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
690 if (till_stall_check < 3) {
691 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
692 till_stall_check = 3;
693 } else if (till_stall_check > 300) {
694 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
695 till_stall_check = 300;
697 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
700 static void record_gp_stall_check_time(struct rcu_state *rsp)
702 rsp->gp_start = jiffies;
703 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
706 static void print_other_cpu_stall(struct rcu_state *rsp)
712 struct rcu_node *rnp = rcu_get_root(rsp);
714 /* Only let one CPU complain about others per time interval. */
716 raw_spin_lock_irqsave(&rnp->lock, flags);
717 delta = jiffies - rsp->jiffies_stall;
718 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
719 raw_spin_unlock_irqrestore(&rnp->lock, flags);
722 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
723 raw_spin_unlock_irqrestore(&rnp->lock, flags);
726 * OK, time to rat on our buddy...
727 * See Documentation/RCU/stallwarn.txt for info on how to debug
728 * RCU CPU stall warnings.
730 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
732 print_cpu_stall_info_begin();
733 rcu_for_each_leaf_node(rsp, rnp) {
734 raw_spin_lock_irqsave(&rnp->lock, flags);
735 ndetected += rcu_print_task_stall(rnp);
736 raw_spin_unlock_irqrestore(&rnp->lock, flags);
737 if (rnp->qsmask == 0)
739 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
740 if (rnp->qsmask & (1UL << cpu)) {
741 print_cpu_stall_info(rsp, rnp->grplo + cpu);
747 * Now rat on any tasks that got kicked up to the root rcu_node
748 * due to CPU offlining.
750 rnp = rcu_get_root(rsp);
751 raw_spin_lock_irqsave(&rnp->lock, flags);
752 ndetected = rcu_print_task_stall(rnp);
753 raw_spin_unlock_irqrestore(&rnp->lock, flags);
755 print_cpu_stall_info_end();
756 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
757 smp_processor_id(), (long)(jiffies - rsp->gp_start));
759 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
760 else if (!trigger_all_cpu_backtrace())
763 /* If so configured, complain about tasks blocking the grace period. */
765 rcu_print_detail_task_stall(rsp);
767 force_quiescent_state(rsp, 0); /* Kick them all. */
770 static void print_cpu_stall(struct rcu_state *rsp)
773 struct rcu_node *rnp = rcu_get_root(rsp);
776 * OK, time to rat on ourselves...
777 * See Documentation/RCU/stallwarn.txt for info on how to debug
778 * RCU CPU stall warnings.
780 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
781 print_cpu_stall_info_begin();
782 print_cpu_stall_info(rsp, smp_processor_id());
783 print_cpu_stall_info_end();
784 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
785 if (!trigger_all_cpu_backtrace())
788 raw_spin_lock_irqsave(&rnp->lock, flags);
789 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
790 rsp->jiffies_stall = jiffies +
791 3 * jiffies_till_stall_check() + 3;
792 raw_spin_unlock_irqrestore(&rnp->lock, flags);
794 set_need_resched(); /* kick ourselves to get things going. */
797 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
801 struct rcu_node *rnp;
803 if (rcu_cpu_stall_suppress)
805 j = ACCESS_ONCE(jiffies);
806 js = ACCESS_ONCE(rsp->jiffies_stall);
808 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
810 /* We haven't checked in, so go dump stack. */
811 print_cpu_stall(rsp);
813 } else if (rcu_gp_in_progress(rsp) &&
814 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
816 /* They had a few time units to dump stack, so complain. */
817 print_other_cpu_stall(rsp);
821 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
823 rcu_cpu_stall_suppress = 1;
828 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
830 * Set the stall-warning timeout way off into the future, thus preventing
831 * any RCU CPU stall-warning messages from appearing in the current set of
834 * The caller must disable hard irqs.
836 void rcu_cpu_stall_reset(void)
838 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
839 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
840 rcu_preempt_stall_reset();
843 static struct notifier_block rcu_panic_block = {
844 .notifier_call = rcu_panic,
847 static void __init check_cpu_stall_init(void)
849 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
853 * Update CPU-local rcu_data state to record the newly noticed grace period.
854 * This is used both when we started the grace period and when we notice
855 * that someone else started the grace period. The caller must hold the
856 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
857 * and must have irqs disabled.
859 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
861 if (rdp->gpnum != rnp->gpnum) {
863 * If the current grace period is waiting for this CPU,
864 * set up to detect a quiescent state, otherwise don't
865 * go looking for one.
867 rdp->gpnum = rnp->gpnum;
868 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
869 if (rnp->qsmask & rdp->grpmask) {
871 rdp->passed_quiesce = 0;
874 zero_cpu_stall_ticks(rdp);
878 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
881 struct rcu_node *rnp;
883 local_irq_save(flags);
885 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
886 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
887 local_irq_restore(flags);
890 __note_new_gpnum(rsp, rnp, rdp);
891 raw_spin_unlock_irqrestore(&rnp->lock, flags);
895 * Did someone else start a new RCU grace period start since we last
896 * checked? Update local state appropriately if so. Must be called
897 * on the CPU corresponding to rdp.
900 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
905 local_irq_save(flags);
906 if (rdp->gpnum != rsp->gpnum) {
907 note_new_gpnum(rsp, rdp);
910 local_irq_restore(flags);
915 * Advance this CPU's callbacks, but only if the current grace period
916 * has ended. This may be called only from the CPU to whom the rdp
917 * belongs. In addition, the corresponding leaf rcu_node structure's
918 * ->lock must be held by the caller, with irqs disabled.
921 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
923 /* Did another grace period end? */
924 if (rdp->completed != rnp->completed) {
926 /* Advance callbacks. No harm if list empty. */
927 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
928 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
929 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
931 /* Remember that we saw this grace-period completion. */
932 rdp->completed = rnp->completed;
933 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
936 * If we were in an extended quiescent state, we may have
937 * missed some grace periods that others CPUs handled on
938 * our behalf. Catch up with this state to avoid noting
939 * spurious new grace periods. If another grace period
940 * has started, then rnp->gpnum will have advanced, so
941 * we will detect this later on.
943 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
944 rdp->gpnum = rdp->completed;
947 * If RCU does not need a quiescent state from this CPU,
948 * then make sure that this CPU doesn't go looking for one.
950 if ((rnp->qsmask & rdp->grpmask) == 0)
956 * Advance this CPU's callbacks, but only if the current grace period
957 * has ended. This may be called only from the CPU to whom the rdp
961 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
964 struct rcu_node *rnp;
966 local_irq_save(flags);
968 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
969 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
970 local_irq_restore(flags);
973 __rcu_process_gp_end(rsp, rnp, rdp);
974 raw_spin_unlock_irqrestore(&rnp->lock, flags);
978 * Do per-CPU grace-period initialization for running CPU. The caller
979 * must hold the lock of the leaf rcu_node structure corresponding to
983 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
985 /* Prior grace period ended, so advance callbacks for current CPU. */
986 __rcu_process_gp_end(rsp, rnp, rdp);
989 * Because this CPU just now started the new grace period, we know
990 * that all of its callbacks will be covered by this upcoming grace
991 * period, even the ones that were registered arbitrarily recently.
992 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
994 * Other CPUs cannot be sure exactly when the grace period started.
995 * Therefore, their recently registered callbacks must pass through
996 * an additional RCU_NEXT_READY stage, so that they will be handled
997 * by the next RCU grace period.
999 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1000 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1002 /* Set state so that this CPU will detect the next quiescent state. */
1003 __note_new_gpnum(rsp, rnp, rdp);
1007 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1008 * in preparation for detecting the next grace period. The caller must hold
1009 * the root node's ->lock, which is released before return. Hard irqs must
1012 * Note that it is legal for a dying CPU (which is marked as offline) to
1013 * invoke this function. This can happen when the dying CPU reports its
1017 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1018 __releases(rcu_get_root(rsp)->lock)
1020 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1021 struct rcu_node *rnp = rcu_get_root(rsp);
1023 if (!rcu_scheduler_fully_active ||
1024 !cpu_needs_another_gp(rsp, rdp)) {
1026 * Either the scheduler hasn't yet spawned the first
1027 * non-idle task or this CPU does not need another
1028 * grace period. Either way, don't start a new grace
1031 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1035 if (rsp->fqs_active) {
1037 * This CPU needs a grace period, but force_quiescent_state()
1038 * is running. Tell it to start one on this CPU's behalf.
1040 rsp->fqs_need_gp = 1;
1041 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1045 /* Advance to a new grace period and initialize state. */
1047 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1048 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1049 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1050 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1051 record_gp_stall_check_time(rsp);
1052 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1054 /* Exclude any concurrent CPU-hotplug operations. */
1055 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1058 * Set the quiescent-state-needed bits in all the rcu_node
1059 * structures for all currently online CPUs in breadth-first
1060 * order, starting from the root rcu_node structure. This
1061 * operation relies on the layout of the hierarchy within the
1062 * rsp->node[] array. Note that other CPUs will access only
1063 * the leaves of the hierarchy, which still indicate that no
1064 * grace period is in progress, at least until the corresponding
1065 * leaf node has been initialized. In addition, we have excluded
1066 * CPU-hotplug operations.
1068 * Note that the grace period cannot complete until we finish
1069 * the initialization process, as there will be at least one
1070 * qsmask bit set in the root node until that time, namely the
1071 * one corresponding to this CPU, due to the fact that we have
1074 rcu_for_each_node_breadth_first(rsp, rnp) {
1075 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1076 rcu_preempt_check_blocked_tasks(rnp);
1077 rnp->qsmask = rnp->qsmaskinit;
1078 rnp->gpnum = rsp->gpnum;
1079 rnp->completed = rsp->completed;
1080 if (rnp == rdp->mynode)
1081 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1082 rcu_preempt_boost_start_gp(rnp);
1083 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1084 rnp->level, rnp->grplo,
1085 rnp->grphi, rnp->qsmask);
1086 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1089 rnp = rcu_get_root(rsp);
1090 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1091 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1092 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1093 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1097 * Report a full set of quiescent states to the specified rcu_state
1098 * data structure. This involves cleaning up after the prior grace
1099 * period and letting rcu_start_gp() start up the next grace period
1100 * if one is needed. Note that the caller must hold rnp->lock, as
1101 * required by rcu_start_gp(), which will release it.
1103 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1104 __releases(rcu_get_root(rsp)->lock)
1106 unsigned long gp_duration;
1107 struct rcu_node *rnp = rcu_get_root(rsp);
1108 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1110 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1113 * Ensure that all grace-period and pre-grace-period activity
1114 * is seen before the assignment to rsp->completed.
1116 smp_mb(); /* See above block comment. */
1117 gp_duration = jiffies - rsp->gp_start;
1118 if (gp_duration > rsp->gp_max)
1119 rsp->gp_max = gp_duration;
1122 * We know the grace period is complete, but to everyone else
1123 * it appears to still be ongoing. But it is also the case
1124 * that to everyone else it looks like there is nothing that
1125 * they can do to advance the grace period. It is therefore
1126 * safe for us to drop the lock in order to mark the grace
1127 * period as completed in all of the rcu_node structures.
1129 * But if this CPU needs another grace period, it will take
1130 * care of this while initializing the next grace period.
1131 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1132 * because the callbacks have not yet been advanced: Those
1133 * callbacks are waiting on the grace period that just now
1136 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1137 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1140 * Propagate new ->completed value to rcu_node structures
1141 * so that other CPUs don't have to wait until the start
1142 * of the next grace period to process their callbacks.
1144 rcu_for_each_node_breadth_first(rsp, rnp) {
1145 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1146 rnp->completed = rsp->gpnum;
1147 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1149 rnp = rcu_get_root(rsp);
1150 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1153 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1154 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1155 rsp->fqs_state = RCU_GP_IDLE;
1156 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1160 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1161 * Allows quiescent states for a group of CPUs to be reported at one go
1162 * to the specified rcu_node structure, though all the CPUs in the group
1163 * must be represented by the same rcu_node structure (which need not be
1164 * a leaf rcu_node structure, though it often will be). That structure's
1165 * lock must be held upon entry, and it is released before return.
1168 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1169 struct rcu_node *rnp, unsigned long flags)
1170 __releases(rnp->lock)
1172 struct rcu_node *rnp_c;
1174 /* Walk up the rcu_node hierarchy. */
1176 if (!(rnp->qsmask & mask)) {
1178 /* Our bit has already been cleared, so done. */
1179 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1182 rnp->qsmask &= ~mask;
1183 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1184 mask, rnp->qsmask, rnp->level,
1185 rnp->grplo, rnp->grphi,
1187 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1189 /* Other bits still set at this level, so done. */
1190 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1193 mask = rnp->grpmask;
1194 if (rnp->parent == NULL) {
1196 /* No more levels. Exit loop holding root lock. */
1200 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1203 raw_spin_lock_irqsave(&rnp->lock, flags);
1204 WARN_ON_ONCE(rnp_c->qsmask);
1208 * Get here if we are the last CPU to pass through a quiescent
1209 * state for this grace period. Invoke rcu_report_qs_rsp()
1210 * to clean up and start the next grace period if one is needed.
1212 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1216 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1217 * structure. This must be either called from the specified CPU, or
1218 * called when the specified CPU is known to be offline (and when it is
1219 * also known that no other CPU is concurrently trying to help the offline
1220 * CPU). The lastcomp argument is used to make sure we are still in the
1221 * grace period of interest. We don't want to end the current grace period
1222 * based on quiescent states detected in an earlier grace period!
1225 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1227 unsigned long flags;
1229 struct rcu_node *rnp;
1232 raw_spin_lock_irqsave(&rnp->lock, flags);
1233 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1236 * The grace period in which this quiescent state was
1237 * recorded has ended, so don't report it upwards.
1238 * We will instead need a new quiescent state that lies
1239 * within the current grace period.
1241 rdp->passed_quiesce = 0; /* need qs for new gp. */
1242 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1245 mask = rdp->grpmask;
1246 if ((rnp->qsmask & mask) == 0) {
1247 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1249 rdp->qs_pending = 0;
1252 * This GP can't end until cpu checks in, so all of our
1253 * callbacks can be processed during the next GP.
1255 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1257 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1262 * Check to see if there is a new grace period of which this CPU
1263 * is not yet aware, and if so, set up local rcu_data state for it.
1264 * Otherwise, see if this CPU has just passed through its first
1265 * quiescent state for this grace period, and record that fact if so.
1268 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1270 /* If there is now a new grace period, record and return. */
1271 if (check_for_new_grace_period(rsp, rdp))
1275 * Does this CPU still need to do its part for current grace period?
1276 * If no, return and let the other CPUs do their part as well.
1278 if (!rdp->qs_pending)
1282 * Was there a quiescent state since the beginning of the grace
1283 * period? If no, then exit and wait for the next call.
1285 if (!rdp->passed_quiesce)
1289 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1292 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1295 #ifdef CONFIG_HOTPLUG_CPU
1298 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1299 * Also record a quiescent state for this CPU for the current grace period.
1300 * Synchronization and interrupt disabling are not required because
1301 * this function executes in stop_machine() context. Therefore, cleanup
1302 * operations that might block must be done later from the CPU_DEAD
1305 * Note that the outgoing CPU's bit has already been cleared in the
1306 * cpu_online_mask. This allows us to randomly pick a callback
1307 * destination from the bits set in that mask.
1309 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1311 unsigned long flags;
1315 int receive_cpu = cpumask_any(cpu_online_mask);
1316 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1317 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1318 struct rcu_node *rnp = rdp->mynode; /* For dying CPU. */
1320 /* First, adjust the counts. */
1321 if (rdp->nxtlist != NULL) {
1322 receive_rdp->qlen_lazy += rdp->qlen_lazy;
1323 receive_rdp->qlen += rdp->qlen;
1329 * Next, move ready-to-invoke callbacks to be invoked on some
1330 * other CPU. These will not be required to pass through another
1331 * grace period: They are done, regardless of CPU.
1333 if (rdp->nxtlist != NULL &&
1334 rdp->nxttail[RCU_DONE_TAIL] != &rdp->nxtlist) {
1335 struct rcu_head *oldhead;
1336 struct rcu_head **oldtail;
1337 struct rcu_head **newtail;
1339 oldhead = rdp->nxtlist;
1340 oldtail = receive_rdp->nxttail[RCU_DONE_TAIL];
1341 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1342 *rdp->nxttail[RCU_DONE_TAIL] = *oldtail;
1343 *receive_rdp->nxttail[RCU_DONE_TAIL] = oldhead;
1344 newtail = rdp->nxttail[RCU_DONE_TAIL];
1345 for (i = RCU_DONE_TAIL; i < RCU_NEXT_SIZE; i++) {
1346 if (receive_rdp->nxttail[i] == oldtail)
1347 receive_rdp->nxttail[i] = newtail;
1348 if (rdp->nxttail[i] == newtail)
1349 rdp->nxttail[i] = &rdp->nxtlist;
1354 * Finally, put the rest of the callbacks at the end of the list.
1355 * The ones that made it partway through get to start over: We
1356 * cannot assume that grace periods are synchronized across CPUs.
1357 * (We could splice RCU_WAIT_TAIL into RCU_NEXT_READY_TAIL, but
1358 * this does not seem compelling. Not yet, anyway.)
1360 if (rdp->nxtlist != NULL) {
1361 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1362 receive_rdp->nxttail[RCU_NEXT_TAIL] =
1363 rdp->nxttail[RCU_NEXT_TAIL];
1364 receive_rdp->n_cbs_adopted += rdp->qlen;
1365 rdp->n_cbs_orphaned += rdp->qlen;
1367 rdp->nxtlist = NULL;
1368 for (i = 0; i < RCU_NEXT_SIZE; i++)
1369 rdp->nxttail[i] = &rdp->nxtlist;
1373 * Record a quiescent state for the dying CPU. This is safe
1374 * only because we have already cleared out the callbacks.
1375 * (Otherwise, the RCU core might try to schedule the invocation
1376 * of callbacks on this now-offline CPU, which would be bad.)
1378 mask = rdp->grpmask; /* rnp->grplo is constant. */
1379 trace_rcu_grace_period(rsp->name,
1380 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1382 rcu_report_qs_rdp(smp_processor_id(), rsp, rdp, rsp->gpnum);
1383 /* Note that rcu_report_qs_rdp() might call trace_rcu_grace_period(). */
1386 * Remove the dying CPU from the bitmasks in the rcu_node
1387 * hierarchy. Because we are in stop_machine() context, we
1388 * automatically exclude ->onofflock critical sections.
1391 raw_spin_lock_irqsave(&rnp->lock, flags);
1392 rnp->qsmaskinit &= ~mask;
1393 if (rnp->qsmaskinit != 0) {
1394 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1397 if (rnp == rdp->mynode) {
1398 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1399 if (need_report & RCU_OFL_TASKS_NORM_GP)
1400 rcu_report_unblock_qs_rnp(rnp, flags);
1402 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1403 if (need_report & RCU_OFL_TASKS_EXP_GP)
1404 rcu_report_exp_rnp(rsp, rnp, true);
1406 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1407 mask = rnp->grpmask;
1409 } while (rnp != NULL);
1413 * The CPU has been completely removed, and some other CPU is reporting
1414 * this fact from process context. Do the remainder of the cleanup.
1415 * There can only be one CPU hotplug operation at a time, so no other
1416 * CPU can be attempting to update rcu_cpu_kthread_task.
1418 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1420 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1421 struct rcu_node *rnp = rdp->mynode;
1423 rcu_stop_cpu_kthread(cpu);
1424 rcu_node_kthread_setaffinity(rnp, -1);
1427 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1429 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1433 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1437 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1440 * Invoke any RCU callbacks that have made it to the end of their grace
1441 * period. Thottle as specified by rdp->blimit.
1443 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1445 unsigned long flags;
1446 struct rcu_head *next, *list, **tail;
1447 int bl, count, count_lazy;
1449 /* If no callbacks are ready, just return.*/
1450 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1451 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1452 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1453 need_resched(), is_idle_task(current),
1454 rcu_is_callbacks_kthread());
1459 * Extract the list of ready callbacks, disabling to prevent
1460 * races with call_rcu() from interrupt handlers.
1462 local_irq_save(flags);
1463 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1465 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1466 list = rdp->nxtlist;
1467 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1468 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1469 tail = rdp->nxttail[RCU_DONE_TAIL];
1470 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1471 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1472 rdp->nxttail[count] = &rdp->nxtlist;
1473 local_irq_restore(flags);
1475 /* Invoke callbacks. */
1476 count = count_lazy = 0;
1480 debug_rcu_head_unqueue(list);
1481 if (__rcu_reclaim(rsp->name, list))
1484 /* Stop only if limit reached and CPU has something to do. */
1485 if (++count >= bl &&
1487 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1491 local_irq_save(flags);
1492 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1493 is_idle_task(current),
1494 rcu_is_callbacks_kthread());
1496 /* Update count, and requeue any remaining callbacks. */
1497 rdp->qlen_lazy -= count_lazy;
1499 rdp->n_cbs_invoked += count;
1501 *tail = rdp->nxtlist;
1502 rdp->nxtlist = list;
1503 for (count = 0; count < RCU_NEXT_SIZE; count++)
1504 if (&rdp->nxtlist == rdp->nxttail[count])
1505 rdp->nxttail[count] = tail;
1510 /* Reinstate batch limit if we have worked down the excess. */
1511 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1512 rdp->blimit = blimit;
1514 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1515 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1516 rdp->qlen_last_fqs_check = 0;
1517 rdp->n_force_qs_snap = rsp->n_force_qs;
1518 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1519 rdp->qlen_last_fqs_check = rdp->qlen;
1521 local_irq_restore(flags);
1523 /* Re-invoke RCU core processing if there are callbacks remaining. */
1524 if (cpu_has_callbacks_ready_to_invoke(rdp))
1529 * Check to see if this CPU is in a non-context-switch quiescent state
1530 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1531 * Also schedule RCU core processing.
1533 * This function must be called from hardirq context. It is normally
1534 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1535 * false, there is no point in invoking rcu_check_callbacks().
1537 void rcu_check_callbacks(int cpu, int user)
1539 trace_rcu_utilization("Start scheduler-tick");
1540 increment_cpu_stall_ticks();
1541 if (user || rcu_is_cpu_rrupt_from_idle()) {
1544 * Get here if this CPU took its interrupt from user
1545 * mode or from the idle loop, and if this is not a
1546 * nested interrupt. In this case, the CPU is in
1547 * a quiescent state, so note it.
1549 * No memory barrier is required here because both
1550 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1551 * variables that other CPUs neither access nor modify,
1552 * at least not while the corresponding CPU is online.
1558 } else if (!in_softirq()) {
1561 * Get here if this CPU did not take its interrupt from
1562 * softirq, in other words, if it is not interrupting
1563 * a rcu_bh read-side critical section. This is an _bh
1564 * critical section, so note it.
1569 rcu_preempt_check_callbacks(cpu);
1570 if (rcu_pending(cpu))
1572 trace_rcu_utilization("End scheduler-tick");
1576 * Scan the leaf rcu_node structures, processing dyntick state for any that
1577 * have not yet encountered a quiescent state, using the function specified.
1578 * Also initiate boosting for any threads blocked on the root rcu_node.
1580 * The caller must have suppressed start of new grace periods.
1582 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1586 unsigned long flags;
1588 struct rcu_node *rnp;
1590 rcu_for_each_leaf_node(rsp, rnp) {
1592 raw_spin_lock_irqsave(&rnp->lock, flags);
1593 if (!rcu_gp_in_progress(rsp)) {
1594 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1597 if (rnp->qsmask == 0) {
1598 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1603 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1604 if ((rnp->qsmask & bit) != 0 &&
1605 f(per_cpu_ptr(rsp->rda, cpu)))
1610 /* rcu_report_qs_rnp() releases rnp->lock. */
1611 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1614 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1616 rnp = rcu_get_root(rsp);
1617 if (rnp->qsmask == 0) {
1618 raw_spin_lock_irqsave(&rnp->lock, flags);
1619 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1624 * Force quiescent states on reluctant CPUs, and also detect which
1625 * CPUs are in dyntick-idle mode.
1627 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1629 unsigned long flags;
1630 struct rcu_node *rnp = rcu_get_root(rsp);
1632 trace_rcu_utilization("Start fqs");
1633 if (!rcu_gp_in_progress(rsp)) {
1634 trace_rcu_utilization("End fqs");
1635 return; /* No grace period in progress, nothing to force. */
1637 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1638 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1639 trace_rcu_utilization("End fqs");
1640 return; /* Someone else is already on the job. */
1642 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1643 goto unlock_fqs_ret; /* no emergency and done recently. */
1645 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1646 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1647 if(!rcu_gp_in_progress(rsp)) {
1648 rsp->n_force_qs_ngp++;
1649 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1650 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1652 rsp->fqs_active = 1;
1653 switch (rsp->fqs_state) {
1657 break; /* grace period idle or initializing, ignore. */
1659 case RCU_SAVE_DYNTICK:
1660 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1661 break; /* So gcc recognizes the dead code. */
1663 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1665 /* Record dyntick-idle state. */
1666 force_qs_rnp(rsp, dyntick_save_progress_counter);
1667 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1668 if (rcu_gp_in_progress(rsp))
1669 rsp->fqs_state = RCU_FORCE_QS;
1674 /* Check dyntick-idle state, send IPI to laggarts. */
1675 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1676 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1678 /* Leave state in case more forcing is required. */
1680 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1683 rsp->fqs_active = 0;
1684 if (rsp->fqs_need_gp) {
1685 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1686 rsp->fqs_need_gp = 0;
1687 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1688 trace_rcu_utilization("End fqs");
1691 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1693 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1694 trace_rcu_utilization("End fqs");
1698 * This does the RCU core processing work for the specified rcu_state
1699 * and rcu_data structures. This may be called only from the CPU to
1700 * whom the rdp belongs.
1703 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1705 unsigned long flags;
1707 WARN_ON_ONCE(rdp->beenonline == 0);
1710 * If an RCU GP has gone long enough, go check for dyntick
1711 * idle CPUs and, if needed, send resched IPIs.
1713 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1714 force_quiescent_state(rsp, 1);
1717 * Advance callbacks in response to end of earlier grace
1718 * period that some other CPU ended.
1720 rcu_process_gp_end(rsp, rdp);
1722 /* Update RCU state based on any recent quiescent states. */
1723 rcu_check_quiescent_state(rsp, rdp);
1725 /* Does this CPU require a not-yet-started grace period? */
1726 if (cpu_needs_another_gp(rsp, rdp)) {
1727 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1728 rcu_start_gp(rsp, flags); /* releases above lock */
1731 /* If there are callbacks ready, invoke them. */
1732 if (cpu_has_callbacks_ready_to_invoke(rdp))
1733 invoke_rcu_callbacks(rsp, rdp);
1737 * Do RCU core processing for the current CPU.
1739 static void rcu_process_callbacks(struct softirq_action *unused)
1741 trace_rcu_utilization("Start RCU core");
1742 __rcu_process_callbacks(&rcu_sched_state,
1743 &__get_cpu_var(rcu_sched_data));
1744 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1745 rcu_preempt_process_callbacks();
1746 trace_rcu_utilization("End RCU core");
1750 * Schedule RCU callback invocation. If the specified type of RCU
1751 * does not support RCU priority boosting, just do a direct call,
1752 * otherwise wake up the per-CPU kernel kthread. Note that because we
1753 * are running on the current CPU with interrupts disabled, the
1754 * rcu_cpu_kthread_task cannot disappear out from under us.
1756 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1758 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1760 if (likely(!rsp->boost)) {
1761 rcu_do_batch(rsp, rdp);
1764 invoke_rcu_callbacks_kthread();
1767 static void invoke_rcu_core(void)
1769 raise_softirq(RCU_SOFTIRQ);
1773 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1774 struct rcu_state *rsp, bool lazy)
1776 unsigned long flags;
1777 struct rcu_data *rdp;
1779 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1780 debug_rcu_head_queue(head);
1784 smp_mb(); /* Ensure RCU update seen before callback registry. */
1787 * Opportunistically note grace-period endings and beginnings.
1788 * Note that we might see a beginning right after we see an
1789 * end, but never vice versa, since this CPU has to pass through
1790 * a quiescent state betweentimes.
1792 local_irq_save(flags);
1793 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1794 rdp = this_cpu_ptr(rsp->rda);
1796 /* Add the callback to our list. */
1797 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1798 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1803 if (__is_kfree_rcu_offset((unsigned long)func))
1804 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1805 rdp->qlen_lazy, rdp->qlen);
1807 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1809 /* If interrupts were disabled, don't dive into RCU core. */
1810 if (irqs_disabled_flags(flags)) {
1811 local_irq_restore(flags);
1816 * Force the grace period if too many callbacks or too long waiting.
1817 * Enforce hysteresis, and don't invoke force_quiescent_state()
1818 * if some other CPU has recently done so. Also, don't bother
1819 * invoking force_quiescent_state() if the newly enqueued callback
1820 * is the only one waiting for a grace period to complete.
1822 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1824 /* Are we ignoring a completed grace period? */
1825 rcu_process_gp_end(rsp, rdp);
1826 check_for_new_grace_period(rsp, rdp);
1828 /* Start a new grace period if one not already started. */
1829 if (!rcu_gp_in_progress(rsp)) {
1830 unsigned long nestflag;
1831 struct rcu_node *rnp_root = rcu_get_root(rsp);
1833 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1834 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1836 /* Give the grace period a kick. */
1837 rdp->blimit = LONG_MAX;
1838 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1839 *rdp->nxttail[RCU_DONE_TAIL] != head)
1840 force_quiescent_state(rsp, 0);
1841 rdp->n_force_qs_snap = rsp->n_force_qs;
1842 rdp->qlen_last_fqs_check = rdp->qlen;
1844 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1845 force_quiescent_state(rsp, 1);
1846 local_irq_restore(flags);
1850 * Queue an RCU-sched callback for invocation after a grace period.
1852 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1854 __call_rcu(head, func, &rcu_sched_state, 0);
1856 EXPORT_SYMBOL_GPL(call_rcu_sched);
1859 * Queue an RCU callback for invocation after a quicker grace period.
1861 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1863 __call_rcu(head, func, &rcu_bh_state, 0);
1865 EXPORT_SYMBOL_GPL(call_rcu_bh);
1868 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1870 * Control will return to the caller some time after a full rcu-sched
1871 * grace period has elapsed, in other words after all currently executing
1872 * rcu-sched read-side critical sections have completed. These read-side
1873 * critical sections are delimited by rcu_read_lock_sched() and
1874 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1875 * local_irq_disable(), and so on may be used in place of
1876 * rcu_read_lock_sched().
1878 * This means that all preempt_disable code sequences, including NMI and
1879 * hardware-interrupt handlers, in progress on entry will have completed
1880 * before this primitive returns. However, this does not guarantee that
1881 * softirq handlers will have completed, since in some kernels, these
1882 * handlers can run in process context, and can block.
1884 * This primitive provides the guarantees made by the (now removed)
1885 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1886 * guarantees that rcu_read_lock() sections will have completed.
1887 * In "classic RCU", these two guarantees happen to be one and
1888 * the same, but can differ in realtime RCU implementations.
1890 void synchronize_sched(void)
1892 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1893 !lock_is_held(&rcu_lock_map) &&
1894 !lock_is_held(&rcu_sched_lock_map),
1895 "Illegal synchronize_sched() in RCU-sched read-side critical section");
1896 if (rcu_blocking_is_gp())
1898 wait_rcu_gp(call_rcu_sched);
1900 EXPORT_SYMBOL_GPL(synchronize_sched);
1903 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1905 * Control will return to the caller some time after a full rcu_bh grace
1906 * period has elapsed, in other words after all currently executing rcu_bh
1907 * read-side critical sections have completed. RCU read-side critical
1908 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1909 * and may be nested.
1911 void synchronize_rcu_bh(void)
1913 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1914 !lock_is_held(&rcu_lock_map) &&
1915 !lock_is_held(&rcu_sched_lock_map),
1916 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
1917 if (rcu_blocking_is_gp())
1919 wait_rcu_gp(call_rcu_bh);
1921 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1923 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
1924 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
1926 static int synchronize_sched_expedited_cpu_stop(void *data)
1929 * There must be a full memory barrier on each affected CPU
1930 * between the time that try_stop_cpus() is called and the
1931 * time that it returns.
1933 * In the current initial implementation of cpu_stop, the
1934 * above condition is already met when the control reaches
1935 * this point and the following smp_mb() is not strictly
1936 * necessary. Do smp_mb() anyway for documentation and
1937 * robustness against future implementation changes.
1939 smp_mb(); /* See above comment block. */
1944 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
1945 * approach to force grace period to end quickly. This consumes
1946 * significant time on all CPUs, and is thus not recommended for
1947 * any sort of common-case code.
1949 * Note that it is illegal to call this function while holding any
1950 * lock that is acquired by a CPU-hotplug notifier. Failing to
1951 * observe this restriction will result in deadlock.
1953 * This implementation can be thought of as an application of ticket
1954 * locking to RCU, with sync_sched_expedited_started and
1955 * sync_sched_expedited_done taking on the roles of the halves
1956 * of the ticket-lock word. Each task atomically increments
1957 * sync_sched_expedited_started upon entry, snapshotting the old value,
1958 * then attempts to stop all the CPUs. If this succeeds, then each
1959 * CPU will have executed a context switch, resulting in an RCU-sched
1960 * grace period. We are then done, so we use atomic_cmpxchg() to
1961 * update sync_sched_expedited_done to match our snapshot -- but
1962 * only if someone else has not already advanced past our snapshot.
1964 * On the other hand, if try_stop_cpus() fails, we check the value
1965 * of sync_sched_expedited_done. If it has advanced past our
1966 * initial snapshot, then someone else must have forced a grace period
1967 * some time after we took our snapshot. In this case, our work is
1968 * done for us, and we can simply return. Otherwise, we try again,
1969 * but keep our initial snapshot for purposes of checking for someone
1970 * doing our work for us.
1972 * If we fail too many times in a row, we fall back to synchronize_sched().
1974 void synchronize_sched_expedited(void)
1976 int firstsnap, s, snap, trycount = 0;
1978 /* Note that atomic_inc_return() implies full memory barrier. */
1979 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
1981 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1984 * Each pass through the following loop attempts to force a
1985 * context switch on each CPU.
1987 while (try_stop_cpus(cpu_online_mask,
1988 synchronize_sched_expedited_cpu_stop,
1992 /* No joy, try again later. Or just synchronize_sched(). */
1993 if (trycount++ < 10)
1994 udelay(trycount * num_online_cpus());
1996 synchronize_sched();
2000 /* Check to see if someone else did our work for us. */
2001 s = atomic_read(&sync_sched_expedited_done);
2002 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2003 smp_mb(); /* ensure test happens before caller kfree */
2008 * Refetching sync_sched_expedited_started allows later
2009 * callers to piggyback on our grace period. We subtract
2010 * 1 to get the same token that the last incrementer got.
2011 * We retry after they started, so our grace period works
2012 * for them, and they started after our first try, so their
2013 * grace period works for us.
2016 snap = atomic_read(&sync_sched_expedited_started);
2017 smp_mb(); /* ensure read is before try_stop_cpus(). */
2021 * Everyone up to our most recent fetch is covered by our grace
2022 * period. Update the counter, but only if our work is still
2023 * relevant -- which it won't be if someone who started later
2024 * than we did beat us to the punch.
2027 s = atomic_read(&sync_sched_expedited_done);
2028 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2029 smp_mb(); /* ensure test happens before caller kfree */
2032 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2036 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2039 * Check to see if there is any immediate RCU-related work to be done
2040 * by the current CPU, for the specified type of RCU, returning 1 if so.
2041 * The checks are in order of increasing expense: checks that can be
2042 * carried out against CPU-local state are performed first. However,
2043 * we must check for CPU stalls first, else we might not get a chance.
2045 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2047 struct rcu_node *rnp = rdp->mynode;
2049 rdp->n_rcu_pending++;
2051 /* Check for CPU stalls, if enabled. */
2052 check_cpu_stall(rsp, rdp);
2054 /* Is the RCU core waiting for a quiescent state from this CPU? */
2055 if (rcu_scheduler_fully_active &&
2056 rdp->qs_pending && !rdp->passed_quiesce) {
2059 * If force_quiescent_state() coming soon and this CPU
2060 * needs a quiescent state, and this is either RCU-sched
2061 * or RCU-bh, force a local reschedule.
2063 rdp->n_rp_qs_pending++;
2064 if (!rdp->preemptible &&
2065 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2068 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2069 rdp->n_rp_report_qs++;
2073 /* Does this CPU have callbacks ready to invoke? */
2074 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2075 rdp->n_rp_cb_ready++;
2079 /* Has RCU gone idle with this CPU needing another grace period? */
2080 if (cpu_needs_another_gp(rsp, rdp)) {
2081 rdp->n_rp_cpu_needs_gp++;
2085 /* Has another RCU grace period completed? */
2086 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2087 rdp->n_rp_gp_completed++;
2091 /* Has a new RCU grace period started? */
2092 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2093 rdp->n_rp_gp_started++;
2097 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2098 if (rcu_gp_in_progress(rsp) &&
2099 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2100 rdp->n_rp_need_fqs++;
2105 rdp->n_rp_need_nothing++;
2110 * Check to see if there is any immediate RCU-related work to be done
2111 * by the current CPU, returning 1 if so. This function is part of the
2112 * RCU implementation; it is -not- an exported member of the RCU API.
2114 static int rcu_pending(int cpu)
2116 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
2117 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2118 rcu_preempt_pending(cpu);
2122 * Check to see if any future RCU-related work will need to be done
2123 * by the current CPU, even if none need be done immediately, returning
2126 static int rcu_cpu_has_callbacks(int cpu)
2128 /* RCU callbacks either ready or pending? */
2129 return per_cpu(rcu_sched_data, cpu).nxtlist ||
2130 per_cpu(rcu_bh_data, cpu).nxtlist ||
2131 rcu_preempt_cpu_has_callbacks(cpu);
2134 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
2135 static atomic_t rcu_barrier_cpu_count;
2136 static DEFINE_MUTEX(rcu_barrier_mutex);
2137 static struct completion rcu_barrier_completion;
2139 static void rcu_barrier_callback(struct rcu_head *notused)
2141 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2142 complete(&rcu_barrier_completion);
2146 * Called with preemption disabled, and from cross-cpu IRQ context.
2148 static void rcu_barrier_func(void *type)
2150 int cpu = smp_processor_id();
2151 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
2152 void (*call_rcu_func)(struct rcu_head *head,
2153 void (*func)(struct rcu_head *head));
2155 atomic_inc(&rcu_barrier_cpu_count);
2156 call_rcu_func = type;
2157 call_rcu_func(head, rcu_barrier_callback);
2161 * Orchestrate the specified type of RCU barrier, waiting for all
2162 * RCU callbacks of the specified type to complete.
2164 static void _rcu_barrier(struct rcu_state *rsp,
2165 void (*call_rcu_func)(struct rcu_head *head,
2166 void (*func)(struct rcu_head *head)))
2168 BUG_ON(in_interrupt());
2169 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2170 mutex_lock(&rcu_barrier_mutex);
2171 init_completion(&rcu_barrier_completion);
2173 * Initialize rcu_barrier_cpu_count to 1, then invoke
2174 * rcu_barrier_func() on each CPU, so that each CPU also has
2175 * incremented rcu_barrier_cpu_count. Only then is it safe to
2176 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2177 * might complete its grace period before all of the other CPUs
2178 * did their increment, causing this function to return too
2179 * early. Note that on_each_cpu() disables irqs, which prevents
2180 * any CPUs from coming online or going offline until each online
2181 * CPU has queued its RCU-barrier callback.
2183 atomic_set(&rcu_barrier_cpu_count, 1);
2184 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
2185 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2186 complete(&rcu_barrier_completion);
2187 wait_for_completion(&rcu_barrier_completion);
2188 mutex_unlock(&rcu_barrier_mutex);
2192 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2194 void rcu_barrier_bh(void)
2196 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
2198 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2201 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2203 void rcu_barrier_sched(void)
2205 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
2207 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2210 * Do boot-time initialization of a CPU's per-CPU RCU data.
2213 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2215 unsigned long flags;
2217 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2218 struct rcu_node *rnp = rcu_get_root(rsp);
2220 /* Set up local state, ensuring consistent view of global state. */
2221 raw_spin_lock_irqsave(&rnp->lock, flags);
2222 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2223 rdp->nxtlist = NULL;
2224 for (i = 0; i < RCU_NEXT_SIZE; i++)
2225 rdp->nxttail[i] = &rdp->nxtlist;
2228 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2229 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_NESTING);
2230 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2233 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2237 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2238 * offline event can be happening at a given time. Note also that we
2239 * can accept some slop in the rsp->completed access due to the fact
2240 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2242 static void __cpuinit
2243 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2245 unsigned long flags;
2247 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2248 struct rcu_node *rnp = rcu_get_root(rsp);
2250 /* Set up local state, ensuring consistent view of global state. */
2251 raw_spin_lock_irqsave(&rnp->lock, flags);
2252 rdp->beenonline = 1; /* We have now been online. */
2253 rdp->preemptible = preemptible;
2254 rdp->qlen_last_fqs_check = 0;
2255 rdp->n_force_qs_snap = rsp->n_force_qs;
2256 rdp->blimit = blimit;
2257 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_NESTING;
2258 atomic_set(&rdp->dynticks->dynticks,
2259 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2260 rcu_prepare_for_idle_init(cpu);
2261 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2264 * A new grace period might start here. If so, we won't be part
2265 * of it, but that is OK, as we are currently in a quiescent state.
2268 /* Exclude any attempts to start a new GP on large systems. */
2269 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2271 /* Add CPU to rcu_node bitmasks. */
2273 mask = rdp->grpmask;
2275 /* Exclude any attempts to start a new GP on small systems. */
2276 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2277 rnp->qsmaskinit |= mask;
2278 mask = rnp->grpmask;
2279 if (rnp == rdp->mynode) {
2281 * If there is a grace period in progress, we will
2282 * set up to wait for it next time we run the
2285 rdp->gpnum = rnp->completed;
2286 rdp->completed = rnp->completed;
2287 rdp->passed_quiesce = 0;
2288 rdp->qs_pending = 0;
2289 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2290 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2292 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2294 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2296 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2299 static void __cpuinit rcu_prepare_cpu(int cpu)
2301 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2302 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2303 rcu_preempt_init_percpu_data(cpu);
2307 * Handle CPU online/offline notification events.
2309 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2310 unsigned long action, void *hcpu)
2312 long cpu = (long)hcpu;
2313 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2314 struct rcu_node *rnp = rdp->mynode;
2316 trace_rcu_utilization("Start CPU hotplug");
2318 case CPU_UP_PREPARE:
2319 case CPU_UP_PREPARE_FROZEN:
2320 rcu_prepare_cpu(cpu);
2321 rcu_prepare_kthreads(cpu);
2324 case CPU_DOWN_FAILED:
2325 rcu_node_kthread_setaffinity(rnp, -1);
2326 rcu_cpu_kthread_setrt(cpu, 1);
2328 case CPU_DOWN_PREPARE:
2329 rcu_node_kthread_setaffinity(rnp, cpu);
2330 rcu_cpu_kthread_setrt(cpu, 0);
2333 case CPU_DYING_FROZEN:
2335 * The whole machine is "stopped" except this CPU, so we can
2336 * touch any data without introducing corruption. We send the
2337 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2339 rcu_cleanup_dying_cpu(&rcu_bh_state);
2340 rcu_cleanup_dying_cpu(&rcu_sched_state);
2341 rcu_preempt_cleanup_dying_cpu();
2342 rcu_cleanup_after_idle(cpu);
2345 case CPU_DEAD_FROZEN:
2346 case CPU_UP_CANCELED:
2347 case CPU_UP_CANCELED_FROZEN:
2348 rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2349 rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2350 rcu_preempt_cleanup_dead_cpu(cpu);
2355 trace_rcu_utilization("End CPU hotplug");
2360 * This function is invoked towards the end of the scheduler's initialization
2361 * process. Before this is called, the idle task might contain
2362 * RCU read-side critical sections (during which time, this idle
2363 * task is booting the system). After this function is called, the
2364 * idle tasks are prohibited from containing RCU read-side critical
2365 * sections. This function also enables RCU lockdep checking.
2367 void rcu_scheduler_starting(void)
2369 WARN_ON(num_online_cpus() != 1);
2370 WARN_ON(nr_context_switches() > 0);
2371 rcu_scheduler_active = 1;
2375 * Compute the per-level fanout, either using the exact fanout specified
2376 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2378 #ifdef CONFIG_RCU_FANOUT_EXACT
2379 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2383 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2384 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2385 rsp->levelspread[0] = RCU_FANOUT_LEAF;
2387 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2388 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2395 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2396 ccur = rsp->levelcnt[i];
2397 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2401 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2404 * Helper function for rcu_init() that initializes one rcu_state structure.
2406 static void __init rcu_init_one(struct rcu_state *rsp,
2407 struct rcu_data __percpu *rda)
2409 static char *buf[] = { "rcu_node_level_0",
2412 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2416 struct rcu_node *rnp;
2418 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2420 /* Initialize the level-tracking arrays. */
2422 for (i = 1; i < NUM_RCU_LVLS; i++)
2423 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2424 rcu_init_levelspread(rsp);
2426 /* Initialize the elements themselves, starting from the leaves. */
2428 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2429 cpustride *= rsp->levelspread[i];
2430 rnp = rsp->level[i];
2431 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2432 raw_spin_lock_init(&rnp->lock);
2433 lockdep_set_class_and_name(&rnp->lock,
2434 &rcu_node_class[i], buf[i]);
2437 rnp->qsmaskinit = 0;
2438 rnp->grplo = j * cpustride;
2439 rnp->grphi = (j + 1) * cpustride - 1;
2440 if (rnp->grphi >= NR_CPUS)
2441 rnp->grphi = NR_CPUS - 1;
2447 rnp->grpnum = j % rsp->levelspread[i - 1];
2448 rnp->grpmask = 1UL << rnp->grpnum;
2449 rnp->parent = rsp->level[i - 1] +
2450 j / rsp->levelspread[i - 1];
2453 INIT_LIST_HEAD(&rnp->blkd_tasks);
2458 rnp = rsp->level[NUM_RCU_LVLS - 1];
2459 for_each_possible_cpu(i) {
2460 while (i > rnp->grphi)
2462 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2463 rcu_boot_init_percpu_data(i, rsp);
2467 void __init rcu_init(void)
2471 rcu_bootup_announce();
2472 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2473 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2474 __rcu_init_preempt();
2475 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2478 * We don't need protection against CPU-hotplug here because
2479 * this is called early in boot, before either interrupts
2480 * or the scheduler are operational.
2482 cpu_notifier(rcu_cpu_notify, 0);
2483 for_each_online_cpu(cpu)
2484 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2485 check_cpu_stall_init();
2488 #include "rcutree_plugin.h"