2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
57 #include <trace/events/rcu.h>
61 /* Data structures. */
63 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 #define RCU_STATE_INITIALIZER(sname, cr) { \
66 .level = { &sname##_state.node[0] }, \
68 .fqs_state = RCU_GP_IDLE, \
71 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
72 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
73 .orphan_donetail = &sname##_state.orphan_donelist, \
74 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
75 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \
79 struct rcu_state rcu_sched_state =
80 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
81 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
83 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
84 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
86 static struct rcu_state *rcu_state;
87 LIST_HEAD(rcu_struct_flavors);
89 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
90 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
91 module_param(rcu_fanout_leaf, int, 0);
92 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
93 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
100 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
103 * The rcu_scheduler_active variable transitions from zero to one just
104 * before the first task is spawned. So when this variable is zero, RCU
105 * can assume that there is but one task, allowing RCU to (for example)
106 * optimized synchronize_sched() to a simple barrier(). When this variable
107 * is one, RCU must actually do all the hard work required to detect real
108 * grace periods. This variable is also used to suppress boot-time false
109 * positives from lockdep-RCU error checking.
111 int rcu_scheduler_active __read_mostly;
112 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
115 * The rcu_scheduler_fully_active variable transitions from zero to one
116 * during the early_initcall() processing, which is after the scheduler
117 * is capable of creating new tasks. So RCU processing (for example,
118 * creating tasks for RCU priority boosting) must be delayed until after
119 * rcu_scheduler_fully_active transitions from zero to one. We also
120 * currently delay invocation of any RCU callbacks until after this point.
122 * It might later prove better for people registering RCU callbacks during
123 * early boot to take responsibility for these callbacks, but one step at
126 static int rcu_scheduler_fully_active __read_mostly;
128 #ifdef CONFIG_RCU_BOOST
131 * Control variables for per-CPU and per-rcu_node kthreads. These
132 * handle all flavors of RCU.
134 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
135 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
136 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
137 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
138 DEFINE_PER_CPU(char, rcu_cpu_has_work);
140 #endif /* #ifdef CONFIG_RCU_BOOST */
142 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
143 static void invoke_rcu_core(void);
144 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
147 * Track the rcutorture test sequence number and the update version
148 * number within a given test. The rcutorture_testseq is incremented
149 * on every rcutorture module load and unload, so has an odd value
150 * when a test is running. The rcutorture_vernum is set to zero
151 * when rcutorture starts and is incremented on each rcutorture update.
152 * These variables enable correlating rcutorture output with the
153 * RCU tracing information.
155 unsigned long rcutorture_testseq;
156 unsigned long rcutorture_vernum;
159 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
160 * permit this function to be invoked without holding the root rcu_node
161 * structure's ->lock, but of course results can be subject to change.
163 static int rcu_gp_in_progress(struct rcu_state *rsp)
165 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
169 * Note a quiescent state. Because we do not need to know
170 * how many quiescent states passed, just if there was at least
171 * one since the start of the grace period, this just sets a flag.
172 * The caller must have disabled preemption.
174 void rcu_sched_qs(int cpu)
176 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
178 rdp->passed_quiesce_gpnum = rdp->gpnum;
180 if (rdp->passed_quiesce == 0)
181 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
182 rdp->passed_quiesce = 1;
185 void rcu_bh_qs(int cpu)
187 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
189 rdp->passed_quiesce_gpnum = rdp->gpnum;
191 if (rdp->passed_quiesce == 0)
192 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
193 rdp->passed_quiesce = 1;
197 * Note a context switch. This is a quiescent state for RCU-sched,
198 * and requires special handling for preemptible RCU.
199 * The caller must have disabled preemption.
201 void rcu_note_context_switch(int cpu)
203 trace_rcu_utilization("Start context switch");
205 rcu_preempt_note_context_switch(cpu);
206 trace_rcu_utilization("End context switch");
208 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
210 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
211 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
212 .dynticks = ATOMIC_INIT(1),
215 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
216 static int qhimark = 10000; /* If this many pending, ignore blimit. */
217 static int qlowmark = 100; /* Once only this many pending, use blimit. */
219 module_param(blimit, int, 0);
220 module_param(qhimark, int, 0);
221 module_param(qlowmark, int, 0);
223 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
224 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
226 module_param(rcu_cpu_stall_suppress, int, 0644);
227 module_param(rcu_cpu_stall_timeout, int, 0644);
229 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
230 static int rcu_pending(int cpu);
233 * Return the number of RCU-sched batches processed thus far for debug & stats.
235 long rcu_batches_completed_sched(void)
237 return rcu_sched_state.completed;
239 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
242 * Return the number of RCU BH batches processed thus far for debug & stats.
244 long rcu_batches_completed_bh(void)
246 return rcu_bh_state.completed;
248 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
251 * Force a quiescent state for RCU BH.
253 void rcu_bh_force_quiescent_state(void)
255 force_quiescent_state(&rcu_bh_state, 0);
257 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
260 * Record the number of times rcutorture tests have been initiated and
261 * terminated. This information allows the debugfs tracing stats to be
262 * correlated to the rcutorture messages, even when the rcutorture module
263 * is being repeatedly loaded and unloaded. In other words, we cannot
264 * store this state in rcutorture itself.
266 void rcutorture_record_test_transition(void)
268 rcutorture_testseq++;
269 rcutorture_vernum = 0;
271 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
274 * Record the number of writer passes through the current rcutorture test.
275 * This is also used to correlate debugfs tracing stats with the rcutorture
278 void rcutorture_record_progress(unsigned long vernum)
282 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
285 * Force a quiescent state for RCU-sched.
287 void rcu_sched_force_quiescent_state(void)
289 force_quiescent_state(&rcu_sched_state, 0);
291 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
294 * Does the CPU have callbacks ready to be invoked?
297 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
299 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
303 * Does the current CPU require a yet-as-unscheduled grace period?
306 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
308 return *rdp->nxttail[RCU_DONE_TAIL +
309 ACCESS_ONCE(rsp->completed) != rdp->completed] &&
310 !rcu_gp_in_progress(rsp);
314 * Return the root node of the specified rcu_state structure.
316 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
318 return &rsp->node[0];
322 * If the specified CPU is offline, tell the caller that it is in
323 * a quiescent state. Otherwise, whack it with a reschedule IPI.
324 * Grace periods can end up waiting on an offline CPU when that
325 * CPU is in the process of coming online -- it will be added to the
326 * rcu_node bitmasks before it actually makes it online. The same thing
327 * can happen while a CPU is in the process of coming online. Because this
328 * race is quite rare, we check for it after detecting that the grace
329 * period has been delayed rather than checking each and every CPU
330 * each and every time we start a new grace period.
332 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
335 * If the CPU is offline for more than a jiffy, it is in a quiescent
336 * state. We can trust its state not to change because interrupts
337 * are disabled. The reason for the jiffy's worth of slack is to
338 * handle CPUs initializing on the way up and finding their way
339 * to the idle loop on the way down.
341 if (cpu_is_offline(rdp->cpu) &&
342 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
343 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
351 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
353 * If the new value of the ->dynticks_nesting counter now is zero,
354 * we really have entered idle, and must do the appropriate accounting.
355 * The caller must have disabled interrupts.
357 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
359 trace_rcu_dyntick("Start", oldval, 0);
360 if (!is_idle_task(current)) {
361 struct task_struct *idle = idle_task(smp_processor_id());
363 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
364 ftrace_dump(DUMP_ORIG);
365 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
366 current->pid, current->comm,
367 idle->pid, idle->comm); /* must be idle task! */
369 rcu_prepare_for_idle(smp_processor_id());
370 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
371 smp_mb__before_atomic_inc(); /* See above. */
372 atomic_inc(&rdtp->dynticks);
373 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
374 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
377 * The idle task is not permitted to enter the idle loop while
378 * in an RCU read-side critical section.
380 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
381 "Illegal idle entry in RCU read-side critical section.");
382 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
383 "Illegal idle entry in RCU-bh read-side critical section.");
384 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
385 "Illegal idle entry in RCU-sched read-side critical section.");
389 * rcu_idle_enter - inform RCU that current CPU is entering idle
391 * Enter idle mode, in other words, -leave- the mode in which RCU
392 * read-side critical sections can occur. (Though RCU read-side
393 * critical sections can occur in irq handlers in idle, a possibility
394 * handled by irq_enter() and irq_exit().)
396 * We crowbar the ->dynticks_nesting field to zero to allow for
397 * the possibility of usermode upcalls having messed up our count
398 * of interrupt nesting level during the prior busy period.
400 void rcu_idle_enter(void)
404 struct rcu_dynticks *rdtp;
406 local_irq_save(flags);
407 rdtp = &__get_cpu_var(rcu_dynticks);
408 oldval = rdtp->dynticks_nesting;
409 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
410 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
411 rdtp->dynticks_nesting = 0;
413 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
414 rcu_idle_enter_common(rdtp, oldval);
415 local_irq_restore(flags);
417 EXPORT_SYMBOL_GPL(rcu_idle_enter);
420 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
422 * Exit from an interrupt handler, which might possibly result in entering
423 * idle mode, in other words, leaving the mode in which read-side critical
424 * sections can occur.
426 * This code assumes that the idle loop never does anything that might
427 * result in unbalanced calls to irq_enter() and irq_exit(). If your
428 * architecture violates this assumption, RCU will give you what you
429 * deserve, good and hard. But very infrequently and irreproducibly.
431 * Use things like work queues to work around this limitation.
433 * You have been warned.
435 void rcu_irq_exit(void)
439 struct rcu_dynticks *rdtp;
441 local_irq_save(flags);
442 rdtp = &__get_cpu_var(rcu_dynticks);
443 oldval = rdtp->dynticks_nesting;
444 rdtp->dynticks_nesting--;
445 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
446 if (rdtp->dynticks_nesting)
447 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
449 rcu_idle_enter_common(rdtp, oldval);
450 local_irq_restore(flags);
454 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
456 * If the new value of the ->dynticks_nesting counter was previously zero,
457 * we really have exited idle, and must do the appropriate accounting.
458 * The caller must have disabled interrupts.
460 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
462 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
463 atomic_inc(&rdtp->dynticks);
464 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
465 smp_mb__after_atomic_inc(); /* See above. */
466 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
467 rcu_cleanup_after_idle(smp_processor_id());
468 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
469 if (!is_idle_task(current)) {
470 struct task_struct *idle = idle_task(smp_processor_id());
472 trace_rcu_dyntick("Error on exit: not idle task",
473 oldval, rdtp->dynticks_nesting);
474 ftrace_dump(DUMP_ORIG);
475 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
476 current->pid, current->comm,
477 idle->pid, idle->comm); /* must be idle task! */
482 * rcu_idle_exit - inform RCU that current CPU is leaving idle
484 * Exit idle mode, in other words, -enter- the mode in which RCU
485 * read-side critical sections can occur.
487 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
488 * allow for the possibility of usermode upcalls messing up our count
489 * of interrupt nesting level during the busy period that is just
492 void rcu_idle_exit(void)
495 struct rcu_dynticks *rdtp;
498 local_irq_save(flags);
499 rdtp = &__get_cpu_var(rcu_dynticks);
500 oldval = rdtp->dynticks_nesting;
501 WARN_ON_ONCE(oldval < 0);
502 if (oldval & DYNTICK_TASK_NEST_MASK)
503 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
505 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
506 rcu_idle_exit_common(rdtp, oldval);
507 local_irq_restore(flags);
509 EXPORT_SYMBOL_GPL(rcu_idle_exit);
512 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
514 * Enter an interrupt handler, which might possibly result in exiting
515 * idle mode, in other words, entering the mode in which read-side critical
516 * sections can occur.
518 * Note that the Linux kernel is fully capable of entering an interrupt
519 * handler that it never exits, for example when doing upcalls to
520 * user mode! This code assumes that the idle loop never does upcalls to
521 * user mode. If your architecture does do upcalls from the idle loop (or
522 * does anything else that results in unbalanced calls to the irq_enter()
523 * and irq_exit() functions), RCU will give you what you deserve, good
524 * and hard. But very infrequently and irreproducibly.
526 * Use things like work queues to work around this limitation.
528 * You have been warned.
530 void rcu_irq_enter(void)
533 struct rcu_dynticks *rdtp;
536 local_irq_save(flags);
537 rdtp = &__get_cpu_var(rcu_dynticks);
538 oldval = rdtp->dynticks_nesting;
539 rdtp->dynticks_nesting++;
540 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
542 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
544 rcu_idle_exit_common(rdtp, oldval);
545 local_irq_restore(flags);
549 * rcu_nmi_enter - inform RCU of entry to NMI context
551 * If the CPU was idle with dynamic ticks active, and there is no
552 * irq handler running, this updates rdtp->dynticks_nmi to let the
553 * RCU grace-period handling know that the CPU is active.
555 void rcu_nmi_enter(void)
557 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
559 if (rdtp->dynticks_nmi_nesting == 0 &&
560 (atomic_read(&rdtp->dynticks) & 0x1))
562 rdtp->dynticks_nmi_nesting++;
563 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
564 atomic_inc(&rdtp->dynticks);
565 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
566 smp_mb__after_atomic_inc(); /* See above. */
567 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
571 * rcu_nmi_exit - inform RCU of exit from NMI context
573 * If the CPU was idle with dynamic ticks active, and there is no
574 * irq handler running, this updates rdtp->dynticks_nmi to let the
575 * RCU grace-period handling know that the CPU is no longer active.
577 void rcu_nmi_exit(void)
579 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
581 if (rdtp->dynticks_nmi_nesting == 0 ||
582 --rdtp->dynticks_nmi_nesting != 0)
584 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
585 smp_mb__before_atomic_inc(); /* See above. */
586 atomic_inc(&rdtp->dynticks);
587 smp_mb__after_atomic_inc(); /* Force delay to next write. */
588 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
592 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
594 * If the current CPU is in its idle loop and is neither in an interrupt
595 * or NMI handler, return true.
597 int rcu_is_cpu_idle(void)
602 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
606 EXPORT_SYMBOL(rcu_is_cpu_idle);
608 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
611 * Is the current CPU online? Disable preemption to avoid false positives
612 * that could otherwise happen due to the current CPU number being sampled,
613 * this task being preempted, its old CPU being taken offline, resuming
614 * on some other CPU, then determining that its old CPU is now offline.
615 * It is OK to use RCU on an offline processor during initial boot, hence
616 * the check for rcu_scheduler_fully_active. Note also that it is OK
617 * for a CPU coming online to use RCU for one jiffy prior to marking itself
618 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
619 * offline to continue to use RCU for one jiffy after marking itself
620 * offline in the cpu_online_mask. This leniency is necessary given the
621 * non-atomic nature of the online and offline processing, for example,
622 * the fact that a CPU enters the scheduler after completing the CPU_DYING
625 * This is also why RCU internally marks CPUs online during the
626 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
628 * Disable checking if in an NMI handler because we cannot safely report
629 * errors from NMI handlers anyway.
631 bool rcu_lockdep_current_cpu_online(void)
633 struct rcu_data *rdp;
634 struct rcu_node *rnp;
640 rdp = &__get_cpu_var(rcu_sched_data);
642 ret = (rdp->grpmask & rnp->qsmaskinit) ||
643 !rcu_scheduler_fully_active;
647 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
649 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
652 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
654 * If the current CPU is idle or running at a first-level (not nested)
655 * interrupt from idle, return true. The caller must have at least
656 * disabled preemption.
658 int rcu_is_cpu_rrupt_from_idle(void)
660 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
664 * Snapshot the specified CPU's dynticks counter so that we can later
665 * credit them with an implicit quiescent state. Return 1 if this CPU
666 * is in dynticks idle mode, which is an extended quiescent state.
668 static int dyntick_save_progress_counter(struct rcu_data *rdp)
670 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
671 return (rdp->dynticks_snap & 0x1) == 0;
675 * Return true if the specified CPU has passed through a quiescent
676 * state by virtue of being in or having passed through an dynticks
677 * idle state since the last call to dyntick_save_progress_counter()
680 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
685 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
686 snap = (unsigned int)rdp->dynticks_snap;
689 * If the CPU passed through or entered a dynticks idle phase with
690 * no active irq/NMI handlers, then we can safely pretend that the CPU
691 * already acknowledged the request to pass through a quiescent
692 * state. Either way, that CPU cannot possibly be in an RCU
693 * read-side critical section that started before the beginning
694 * of the current RCU grace period.
696 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
697 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
702 /* Go check for the CPU being offline. */
703 return rcu_implicit_offline_qs(rdp);
706 static int jiffies_till_stall_check(void)
708 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
711 * Limit check must be consistent with the Kconfig limits
712 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
714 if (till_stall_check < 3) {
715 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
716 till_stall_check = 3;
717 } else if (till_stall_check > 300) {
718 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
719 till_stall_check = 300;
721 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
724 static void record_gp_stall_check_time(struct rcu_state *rsp)
726 rsp->gp_start = jiffies;
727 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
730 static void print_other_cpu_stall(struct rcu_state *rsp)
736 struct rcu_node *rnp = rcu_get_root(rsp);
738 /* Only let one CPU complain about others per time interval. */
740 raw_spin_lock_irqsave(&rnp->lock, flags);
741 delta = jiffies - rsp->jiffies_stall;
742 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
743 raw_spin_unlock_irqrestore(&rnp->lock, flags);
746 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
747 raw_spin_unlock_irqrestore(&rnp->lock, flags);
750 * OK, time to rat on our buddy...
751 * See Documentation/RCU/stallwarn.txt for info on how to debug
752 * RCU CPU stall warnings.
754 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
756 print_cpu_stall_info_begin();
757 rcu_for_each_leaf_node(rsp, rnp) {
758 raw_spin_lock_irqsave(&rnp->lock, flags);
759 ndetected += rcu_print_task_stall(rnp);
760 raw_spin_unlock_irqrestore(&rnp->lock, flags);
761 if (rnp->qsmask == 0)
763 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
764 if (rnp->qsmask & (1UL << cpu)) {
765 print_cpu_stall_info(rsp, rnp->grplo + cpu);
771 * Now rat on any tasks that got kicked up to the root rcu_node
772 * due to CPU offlining.
774 rnp = rcu_get_root(rsp);
775 raw_spin_lock_irqsave(&rnp->lock, flags);
776 ndetected += rcu_print_task_stall(rnp);
777 raw_spin_unlock_irqrestore(&rnp->lock, flags);
779 print_cpu_stall_info_end();
780 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
781 smp_processor_id(), (long)(jiffies - rsp->gp_start));
783 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
784 else if (!trigger_all_cpu_backtrace())
787 /* If so configured, complain about tasks blocking the grace period. */
789 rcu_print_detail_task_stall(rsp);
791 force_quiescent_state(rsp, 0); /* Kick them all. */
794 static void print_cpu_stall(struct rcu_state *rsp)
797 struct rcu_node *rnp = rcu_get_root(rsp);
800 * OK, time to rat on ourselves...
801 * See Documentation/RCU/stallwarn.txt for info on how to debug
802 * RCU CPU stall warnings.
804 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
805 print_cpu_stall_info_begin();
806 print_cpu_stall_info(rsp, smp_processor_id());
807 print_cpu_stall_info_end();
808 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
809 if (!trigger_all_cpu_backtrace())
812 raw_spin_lock_irqsave(&rnp->lock, flags);
813 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
814 rsp->jiffies_stall = jiffies +
815 3 * jiffies_till_stall_check() + 3;
816 raw_spin_unlock_irqrestore(&rnp->lock, flags);
818 set_need_resched(); /* kick ourselves to get things going. */
821 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
825 struct rcu_node *rnp;
827 if (rcu_cpu_stall_suppress)
829 j = ACCESS_ONCE(jiffies);
830 js = ACCESS_ONCE(rsp->jiffies_stall);
832 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
834 /* We haven't checked in, so go dump stack. */
835 print_cpu_stall(rsp);
837 } else if (rcu_gp_in_progress(rsp) &&
838 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
840 /* They had a few time units to dump stack, so complain. */
841 print_other_cpu_stall(rsp);
845 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
847 rcu_cpu_stall_suppress = 1;
852 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
854 * Set the stall-warning timeout way off into the future, thus preventing
855 * any RCU CPU stall-warning messages from appearing in the current set of
858 * The caller must disable hard irqs.
860 void rcu_cpu_stall_reset(void)
862 struct rcu_state *rsp;
864 for_each_rcu_flavor(rsp)
865 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
868 static struct notifier_block rcu_panic_block = {
869 .notifier_call = rcu_panic,
872 static void __init check_cpu_stall_init(void)
874 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
878 * Update CPU-local rcu_data state to record the newly noticed grace period.
879 * This is used both when we started the grace period and when we notice
880 * that someone else started the grace period. The caller must hold the
881 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
882 * and must have irqs disabled.
884 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
886 if (rdp->gpnum != rnp->gpnum) {
888 * If the current grace period is waiting for this CPU,
889 * set up to detect a quiescent state, otherwise don't
890 * go looking for one.
892 rdp->gpnum = rnp->gpnum;
893 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
894 if (rnp->qsmask & rdp->grpmask) {
896 rdp->passed_quiesce = 0;
900 zero_cpu_stall_ticks(rdp);
904 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
907 struct rcu_node *rnp;
909 local_irq_save(flags);
911 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
912 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
913 local_irq_restore(flags);
916 __note_new_gpnum(rsp, rnp, rdp);
917 raw_spin_unlock_irqrestore(&rnp->lock, flags);
921 * Did someone else start a new RCU grace period start since we last
922 * checked? Update local state appropriately if so. Must be called
923 * on the CPU corresponding to rdp.
926 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
931 local_irq_save(flags);
932 if (rdp->gpnum != rsp->gpnum) {
933 note_new_gpnum(rsp, rdp);
936 local_irq_restore(flags);
941 * Initialize the specified rcu_data structure's callback list to empty.
943 static void init_callback_list(struct rcu_data *rdp)
948 for (i = 0; i < RCU_NEXT_SIZE; i++)
949 rdp->nxttail[i] = &rdp->nxtlist;
953 * Advance this CPU's callbacks, but only if the current grace period
954 * has ended. This may be called only from the CPU to whom the rdp
955 * belongs. In addition, the corresponding leaf rcu_node structure's
956 * ->lock must be held by the caller, with irqs disabled.
959 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
961 /* Did another grace period end? */
962 if (rdp->completed != rnp->completed) {
964 /* Advance callbacks. No harm if list empty. */
965 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
966 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
967 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
969 /* Remember that we saw this grace-period completion. */
970 rdp->completed = rnp->completed;
971 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
974 * If we were in an extended quiescent state, we may have
975 * missed some grace periods that others CPUs handled on
976 * our behalf. Catch up with this state to avoid noting
977 * spurious new grace periods. If another grace period
978 * has started, then rnp->gpnum will have advanced, so
979 * we will detect this later on.
981 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
982 rdp->gpnum = rdp->completed;
985 * If RCU does not need a quiescent state from this CPU,
986 * then make sure that this CPU doesn't go looking for one.
988 if ((rnp->qsmask & rdp->grpmask) == 0)
994 * Advance this CPU's callbacks, but only if the current grace period
995 * has ended. This may be called only from the CPU to whom the rdp
999 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1001 unsigned long flags;
1002 struct rcu_node *rnp;
1004 local_irq_save(flags);
1006 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1007 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1008 local_irq_restore(flags);
1011 __rcu_process_gp_end(rsp, rnp, rdp);
1012 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1016 * Do per-CPU grace-period initialization for running CPU. The caller
1017 * must hold the lock of the leaf rcu_node structure corresponding to
1021 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1023 /* Prior grace period ended, so advance callbacks for current CPU. */
1024 __rcu_process_gp_end(rsp, rnp, rdp);
1026 /* Set state so that this CPU will detect the next quiescent state. */
1027 __note_new_gpnum(rsp, rnp, rdp);
1031 * Body of kthread that handles grace periods.
1033 static int __noreturn rcu_gp_kthread(void *arg)
1035 unsigned long gp_duration;
1036 struct rcu_data *rdp;
1037 struct rcu_node *rnp;
1038 struct rcu_state *rsp = arg;
1042 /* Handle grace-period start. */
1043 rnp = rcu_get_root(rsp);
1045 wait_event_interruptible(rsp->gp_wq, rsp->gp_flags);
1048 flush_signals(current);
1050 raw_spin_lock_irq(&rnp->lock);
1052 rdp = this_cpu_ptr(rsp->rda);
1054 if (rcu_gp_in_progress(rsp)) {
1056 * A grace period is already in progress, so
1057 * don't start another one.
1059 raw_spin_unlock_irq(&rnp->lock);
1064 if (rsp->fqs_active) {
1066 * We need a grace period, but force_quiescent_state()
1067 * is running. Tell it to start one on our behalf.
1069 rsp->fqs_need_gp = 1;
1070 raw_spin_unlock_irq(&rnp->lock);
1075 /* Advance to a new grace period and initialize state. */
1077 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1078 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1079 rsp->fqs_state = RCU_GP_INIT; /* Stop force_quiescent_state. */
1080 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1081 record_gp_stall_check_time(rsp);
1082 raw_spin_unlock_irq(&rnp->lock);
1084 /* Exclude any concurrent CPU-hotplug operations. */
1088 * Set the quiescent-state-needed bits in all the rcu_node
1089 * structures for all currently online CPUs in breadth-first
1090 * order, starting from the root rcu_node structure.
1091 * This operation relies on the layout of the hierarchy
1092 * within the rsp->node[] array. Note that other CPUs will
1093 * access only the leaves of the hierarchy, which still
1094 * indicate that no grace period is in progress, at least
1095 * until the corresponding leaf node has been initialized.
1096 * In addition, we have excluded CPU-hotplug operations.
1098 rcu_for_each_node_breadth_first(rsp, rnp) {
1099 raw_spin_lock_irq(&rnp->lock);
1100 rcu_preempt_check_blocked_tasks(rnp);
1101 rnp->qsmask = rnp->qsmaskinit;
1102 rnp->gpnum = rsp->gpnum;
1103 rnp->completed = rsp->completed;
1104 if (rnp == rdp->mynode)
1105 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1106 rcu_preempt_boost_start_gp(rnp);
1107 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1108 rnp->level, rnp->grplo,
1109 rnp->grphi, rnp->qsmask);
1110 raw_spin_unlock_irq(&rnp->lock);
1114 rnp = rcu_get_root(rsp);
1115 raw_spin_lock_irq(&rnp->lock);
1116 /* force_quiescent_state() now OK. */
1117 rsp->fqs_state = RCU_SIGNAL_INIT;
1118 raw_spin_unlock_irq(&rnp->lock);
1121 /* Handle grace-period end. */
1122 rnp = rcu_get_root(rsp);
1124 wait_event_interruptible(rsp->gp_wq,
1125 !ACCESS_ONCE(rnp->qsmask) &&
1126 !rcu_preempt_blocked_readers_cgp(rnp));
1127 if (!ACCESS_ONCE(rnp->qsmask) &&
1128 !rcu_preempt_blocked_readers_cgp(rnp))
1130 flush_signals(current);
1133 raw_spin_lock_irq(&rnp->lock);
1134 gp_duration = jiffies - rsp->gp_start;
1135 if (gp_duration > rsp->gp_max)
1136 rsp->gp_max = gp_duration;
1139 * We know the grace period is complete, but to everyone else
1140 * it appears to still be ongoing. But it is also the case
1141 * that to everyone else it looks like there is nothing that
1142 * they can do to advance the grace period. It is therefore
1143 * safe for us to drop the lock in order to mark the grace
1144 * period as completed in all of the rcu_node structures.
1146 * But if this CPU needs another grace period, it will take
1147 * care of this while initializing the next grace period.
1148 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1149 * because the callbacks have not yet been advanced: Those
1150 * callbacks are waiting on the grace period that just now
1153 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1154 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1157 * Propagate new ->completed value to rcu_node
1158 * structures so that other CPUs don't have to
1159 * wait until the start of the next grace period
1160 * to process their callbacks.
1162 rcu_for_each_node_breadth_first(rsp, rnp) {
1163 /* irqs already disabled. */
1164 raw_spin_lock(&rnp->lock);
1165 rnp->completed = rsp->gpnum;
1166 /* irqs remain disabled. */
1167 raw_spin_unlock(&rnp->lock);
1169 rnp = rcu_get_root(rsp);
1170 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1173 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1174 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1175 rsp->fqs_state = RCU_GP_IDLE;
1176 if (cpu_needs_another_gp(rsp, rdp))
1178 raw_spin_unlock_irq(&rnp->lock);
1183 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1184 * in preparation for detecting the next grace period. The caller must hold
1185 * the root node's ->lock, which is released before return. Hard irqs must
1188 * Note that it is legal for a dying CPU (which is marked as offline) to
1189 * invoke this function. This can happen when the dying CPU reports its
1193 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1194 __releases(rcu_get_root(rsp)->lock)
1196 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1197 struct rcu_node *rnp = rcu_get_root(rsp);
1199 if (!rsp->gp_kthread ||
1200 !cpu_needs_another_gp(rsp, rdp)) {
1202 * Either we have not yet spawned the grace-period
1203 * task or this CPU does not need another grace period.
1204 * Either way, don't start a new grace period.
1206 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1211 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1212 wake_up(&rsp->gp_wq);
1216 * Report a full set of quiescent states to the specified rcu_state
1217 * data structure. This involves cleaning up after the prior grace
1218 * period and letting rcu_start_gp() start up the next grace period
1219 * if one is needed. Note that the caller must hold rnp->lock, as
1220 * required by rcu_start_gp(), which will release it.
1222 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1223 __releases(rcu_get_root(rsp)->lock)
1225 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1226 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1227 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1231 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1232 * Allows quiescent states for a group of CPUs to be reported at one go
1233 * to the specified rcu_node structure, though all the CPUs in the group
1234 * must be represented by the same rcu_node structure (which need not be
1235 * a leaf rcu_node structure, though it often will be). That structure's
1236 * lock must be held upon entry, and it is released before return.
1239 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1240 struct rcu_node *rnp, unsigned long flags)
1241 __releases(rnp->lock)
1243 struct rcu_node *rnp_c;
1245 /* Walk up the rcu_node hierarchy. */
1247 if (!(rnp->qsmask & mask)) {
1249 /* Our bit has already been cleared, so done. */
1250 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1253 rnp->qsmask &= ~mask;
1254 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1255 mask, rnp->qsmask, rnp->level,
1256 rnp->grplo, rnp->grphi,
1258 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1260 /* Other bits still set at this level, so done. */
1261 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1264 mask = rnp->grpmask;
1265 if (rnp->parent == NULL) {
1267 /* No more levels. Exit loop holding root lock. */
1271 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1274 raw_spin_lock_irqsave(&rnp->lock, flags);
1275 WARN_ON_ONCE(rnp_c->qsmask);
1279 * Get here if we are the last CPU to pass through a quiescent
1280 * state for this grace period. Invoke rcu_report_qs_rsp()
1281 * to clean up and start the next grace period if one is needed.
1283 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1287 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1288 * structure. This must be either called from the specified CPU, or
1289 * called when the specified CPU is known to be offline (and when it is
1290 * also known that no other CPU is concurrently trying to help the offline
1291 * CPU). The lastcomp argument is used to make sure we are still in the
1292 * grace period of interest. We don't want to end the current grace period
1293 * based on quiescent states detected in an earlier grace period!
1296 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1298 unsigned long flags;
1300 struct rcu_node *rnp;
1303 raw_spin_lock_irqsave(&rnp->lock, flags);
1304 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1307 * The grace period in which this quiescent state was
1308 * recorded has ended, so don't report it upwards.
1309 * We will instead need a new quiescent state that lies
1310 * within the current grace period.
1312 rdp->passed_quiesce = 0; /* need qs for new gp. */
1313 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1316 mask = rdp->grpmask;
1317 if ((rnp->qsmask & mask) == 0) {
1318 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1320 rdp->qs_pending = 0;
1323 * This GP can't end until cpu checks in, so all of our
1324 * callbacks can be processed during the next GP.
1326 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1328 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1333 * Check to see if there is a new grace period of which this CPU
1334 * is not yet aware, and if so, set up local rcu_data state for it.
1335 * Otherwise, see if this CPU has just passed through its first
1336 * quiescent state for this grace period, and record that fact if so.
1339 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1341 /* If there is now a new grace period, record and return. */
1342 if (check_for_new_grace_period(rsp, rdp))
1346 * Does this CPU still need to do its part for current grace period?
1347 * If no, return and let the other CPUs do their part as well.
1349 if (!rdp->qs_pending)
1353 * Was there a quiescent state since the beginning of the grace
1354 * period? If no, then exit and wait for the next call.
1356 if (!rdp->passed_quiesce)
1360 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1363 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1366 #ifdef CONFIG_HOTPLUG_CPU
1369 * Send the specified CPU's RCU callbacks to the orphanage. The
1370 * specified CPU must be offline, and the caller must hold the
1374 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1375 struct rcu_node *rnp, struct rcu_data *rdp)
1378 * Orphan the callbacks. First adjust the counts. This is safe
1379 * because ->onofflock excludes _rcu_barrier()'s adoption of
1380 * the callbacks, thus no memory barrier is required.
1382 if (rdp->nxtlist != NULL) {
1383 rsp->qlen_lazy += rdp->qlen_lazy;
1384 rsp->qlen += rdp->qlen;
1385 rdp->n_cbs_orphaned += rdp->qlen;
1387 ACCESS_ONCE(rdp->qlen) = 0;
1391 * Next, move those callbacks still needing a grace period to
1392 * the orphanage, where some other CPU will pick them up.
1393 * Some of the callbacks might have gone partway through a grace
1394 * period, but that is too bad. They get to start over because we
1395 * cannot assume that grace periods are synchronized across CPUs.
1396 * We don't bother updating the ->nxttail[] array yet, instead
1397 * we just reset the whole thing later on.
1399 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1400 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1401 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1402 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1406 * Then move the ready-to-invoke callbacks to the orphanage,
1407 * where some other CPU will pick them up. These will not be
1408 * required to pass though another grace period: They are done.
1410 if (rdp->nxtlist != NULL) {
1411 *rsp->orphan_donetail = rdp->nxtlist;
1412 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1415 /* Finally, initialize the rcu_data structure's list to empty. */
1416 init_callback_list(rdp);
1420 * Adopt the RCU callbacks from the specified rcu_state structure's
1421 * orphanage. The caller must hold the ->onofflock.
1423 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1426 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1429 * If there is an rcu_barrier() operation in progress, then
1430 * only the task doing that operation is permitted to adopt
1431 * callbacks. To do otherwise breaks rcu_barrier() and friends
1432 * by causing them to fail to wait for the callbacks in the
1435 if (rsp->rcu_barrier_in_progress &&
1436 rsp->rcu_barrier_in_progress != current)
1439 /* Do the accounting first. */
1440 rdp->qlen_lazy += rsp->qlen_lazy;
1441 rdp->qlen += rsp->qlen;
1442 rdp->n_cbs_adopted += rsp->qlen;
1443 if (rsp->qlen_lazy != rsp->qlen)
1444 rcu_idle_count_callbacks_posted();
1449 * We do not need a memory barrier here because the only way we
1450 * can get here if there is an rcu_barrier() in flight is if
1451 * we are the task doing the rcu_barrier().
1454 /* First adopt the ready-to-invoke callbacks. */
1455 if (rsp->orphan_donelist != NULL) {
1456 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1457 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1458 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1459 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1460 rdp->nxttail[i] = rsp->orphan_donetail;
1461 rsp->orphan_donelist = NULL;
1462 rsp->orphan_donetail = &rsp->orphan_donelist;
1465 /* And then adopt the callbacks that still need a grace period. */
1466 if (rsp->orphan_nxtlist != NULL) {
1467 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1468 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1469 rsp->orphan_nxtlist = NULL;
1470 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1475 * Trace the fact that this CPU is going offline.
1477 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1479 RCU_TRACE(unsigned long mask);
1480 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1481 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1483 RCU_TRACE(mask = rdp->grpmask);
1484 trace_rcu_grace_period(rsp->name,
1485 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1490 * The CPU has been completely removed, and some other CPU is reporting
1491 * this fact from process context. Do the remainder of the cleanup,
1492 * including orphaning the outgoing CPU's RCU callbacks, and also
1493 * adopting them, if there is no _rcu_barrier() instance running.
1494 * There can only be one CPU hotplug operation at a time, so no other
1495 * CPU can be attempting to update rcu_cpu_kthread_task.
1497 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1499 unsigned long flags;
1501 int need_report = 0;
1502 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1503 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1505 /* Adjust any no-longer-needed kthreads. */
1506 rcu_stop_cpu_kthread(cpu);
1507 rcu_node_kthread_setaffinity(rnp, -1);
1509 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1511 /* Exclude any attempts to start a new grace period. */
1512 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1514 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1515 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1516 rcu_adopt_orphan_cbs(rsp);
1518 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1519 mask = rdp->grpmask; /* rnp->grplo is constant. */
1521 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1522 rnp->qsmaskinit &= ~mask;
1523 if (rnp->qsmaskinit != 0) {
1524 if (rnp != rdp->mynode)
1525 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1528 if (rnp == rdp->mynode)
1529 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1531 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1532 mask = rnp->grpmask;
1534 } while (rnp != NULL);
1537 * We still hold the leaf rcu_node structure lock here, and
1538 * irqs are still disabled. The reason for this subterfuge is
1539 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1540 * held leads to deadlock.
1542 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1544 if (need_report & RCU_OFL_TASKS_NORM_GP)
1545 rcu_report_unblock_qs_rnp(rnp, flags);
1547 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1548 if (need_report & RCU_OFL_TASKS_EXP_GP)
1549 rcu_report_exp_rnp(rsp, rnp, true);
1550 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1551 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1552 cpu, rdp->qlen, rdp->nxtlist);
1555 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1557 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1561 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1565 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1569 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1572 * Invoke any RCU callbacks that have made it to the end of their grace
1573 * period. Thottle as specified by rdp->blimit.
1575 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1577 unsigned long flags;
1578 struct rcu_head *next, *list, **tail;
1579 int bl, count, count_lazy, i;
1581 /* If no callbacks are ready, just return.*/
1582 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1583 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1584 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1585 need_resched(), is_idle_task(current),
1586 rcu_is_callbacks_kthread());
1591 * Extract the list of ready callbacks, disabling to prevent
1592 * races with call_rcu() from interrupt handlers.
1594 local_irq_save(flags);
1595 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1597 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1598 list = rdp->nxtlist;
1599 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1600 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1601 tail = rdp->nxttail[RCU_DONE_TAIL];
1602 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1603 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1604 rdp->nxttail[i] = &rdp->nxtlist;
1605 local_irq_restore(flags);
1607 /* Invoke callbacks. */
1608 count = count_lazy = 0;
1612 debug_rcu_head_unqueue(list);
1613 if (__rcu_reclaim(rsp->name, list))
1616 /* Stop only if limit reached and CPU has something to do. */
1617 if (++count >= bl &&
1619 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1623 local_irq_save(flags);
1624 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1625 is_idle_task(current),
1626 rcu_is_callbacks_kthread());
1628 /* Update count, and requeue any remaining callbacks. */
1630 *tail = rdp->nxtlist;
1631 rdp->nxtlist = list;
1632 for (i = 0; i < RCU_NEXT_SIZE; i++)
1633 if (&rdp->nxtlist == rdp->nxttail[i])
1634 rdp->nxttail[i] = tail;
1638 smp_mb(); /* List handling before counting for rcu_barrier(). */
1639 rdp->qlen_lazy -= count_lazy;
1640 ACCESS_ONCE(rdp->qlen) -= count;
1641 rdp->n_cbs_invoked += count;
1643 /* Reinstate batch limit if we have worked down the excess. */
1644 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1645 rdp->blimit = blimit;
1647 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1648 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1649 rdp->qlen_last_fqs_check = 0;
1650 rdp->n_force_qs_snap = rsp->n_force_qs;
1651 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1652 rdp->qlen_last_fqs_check = rdp->qlen;
1653 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1655 local_irq_restore(flags);
1657 /* Re-invoke RCU core processing if there are callbacks remaining. */
1658 if (cpu_has_callbacks_ready_to_invoke(rdp))
1663 * Check to see if this CPU is in a non-context-switch quiescent state
1664 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1665 * Also schedule RCU core processing.
1667 * This function must be called from hardirq context. It is normally
1668 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1669 * false, there is no point in invoking rcu_check_callbacks().
1671 void rcu_check_callbacks(int cpu, int user)
1673 trace_rcu_utilization("Start scheduler-tick");
1674 increment_cpu_stall_ticks();
1675 if (user || rcu_is_cpu_rrupt_from_idle()) {
1678 * Get here if this CPU took its interrupt from user
1679 * mode or from the idle loop, and if this is not a
1680 * nested interrupt. In this case, the CPU is in
1681 * a quiescent state, so note it.
1683 * No memory barrier is required here because both
1684 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1685 * variables that other CPUs neither access nor modify,
1686 * at least not while the corresponding CPU is online.
1692 } else if (!in_softirq()) {
1695 * Get here if this CPU did not take its interrupt from
1696 * softirq, in other words, if it is not interrupting
1697 * a rcu_bh read-side critical section. This is an _bh
1698 * critical section, so note it.
1703 rcu_preempt_check_callbacks(cpu);
1704 if (rcu_pending(cpu))
1706 trace_rcu_utilization("End scheduler-tick");
1710 * Scan the leaf rcu_node structures, processing dyntick state for any that
1711 * have not yet encountered a quiescent state, using the function specified.
1712 * Also initiate boosting for any threads blocked on the root rcu_node.
1714 * The caller must have suppressed start of new grace periods.
1716 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1720 unsigned long flags;
1722 struct rcu_node *rnp;
1724 rcu_for_each_leaf_node(rsp, rnp) {
1726 raw_spin_lock_irqsave(&rnp->lock, flags);
1727 if (!rcu_gp_in_progress(rsp)) {
1728 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1731 if (rnp->qsmask == 0) {
1732 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1737 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1738 if ((rnp->qsmask & bit) != 0 &&
1739 f(per_cpu_ptr(rsp->rda, cpu)))
1744 /* rcu_report_qs_rnp() releases rnp->lock. */
1745 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1748 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1750 rnp = rcu_get_root(rsp);
1751 if (rnp->qsmask == 0) {
1752 raw_spin_lock_irqsave(&rnp->lock, flags);
1753 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1758 * Force quiescent states on reluctant CPUs, and also detect which
1759 * CPUs are in dyntick-idle mode.
1761 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1763 unsigned long flags;
1764 struct rcu_node *rnp = rcu_get_root(rsp);
1766 trace_rcu_utilization("Start fqs");
1767 if (!rcu_gp_in_progress(rsp)) {
1768 trace_rcu_utilization("End fqs");
1769 return; /* No grace period in progress, nothing to force. */
1771 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1772 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1773 trace_rcu_utilization("End fqs");
1774 return; /* Someone else is already on the job. */
1776 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1777 goto unlock_fqs_ret; /* no emergency and done recently. */
1779 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1780 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1781 if(!rcu_gp_in_progress(rsp)) {
1782 rsp->n_force_qs_ngp++;
1783 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1784 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1786 rsp->fqs_active = 1;
1787 switch (rsp->fqs_state) {
1791 break; /* grace period idle or initializing, ignore. */
1793 case RCU_SAVE_DYNTICK:
1795 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1797 /* Record dyntick-idle state. */
1798 force_qs_rnp(rsp, dyntick_save_progress_counter);
1799 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1800 if (rcu_gp_in_progress(rsp))
1801 rsp->fqs_state = RCU_FORCE_QS;
1806 /* Check dyntick-idle state, send IPI to laggarts. */
1807 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1808 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1810 /* Leave state in case more forcing is required. */
1812 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1815 rsp->fqs_active = 0;
1816 if (rsp->fqs_need_gp) {
1817 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1818 rsp->fqs_need_gp = 0;
1819 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1820 trace_rcu_utilization("End fqs");
1823 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1825 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1826 trace_rcu_utilization("End fqs");
1830 * This does the RCU core processing work for the specified rcu_state
1831 * and rcu_data structures. This may be called only from the CPU to
1832 * whom the rdp belongs.
1835 __rcu_process_callbacks(struct rcu_state *rsp)
1837 unsigned long flags;
1838 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1840 WARN_ON_ONCE(rdp->beenonline == 0);
1843 * If an RCU GP has gone long enough, go check for dyntick
1844 * idle CPUs and, if needed, send resched IPIs.
1846 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1847 force_quiescent_state(rsp, 1);
1850 * Advance callbacks in response to end of earlier grace
1851 * period that some other CPU ended.
1853 rcu_process_gp_end(rsp, rdp);
1855 /* Update RCU state based on any recent quiescent states. */
1856 rcu_check_quiescent_state(rsp, rdp);
1858 /* Does this CPU require a not-yet-started grace period? */
1859 if (cpu_needs_another_gp(rsp, rdp)) {
1860 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1861 rcu_start_gp(rsp, flags); /* releases above lock */
1864 /* If there are callbacks ready, invoke them. */
1865 if (cpu_has_callbacks_ready_to_invoke(rdp))
1866 invoke_rcu_callbacks(rsp, rdp);
1870 * Do RCU core processing for the current CPU.
1872 static void rcu_process_callbacks(struct softirq_action *unused)
1874 struct rcu_state *rsp;
1876 trace_rcu_utilization("Start RCU core");
1877 for_each_rcu_flavor(rsp)
1878 __rcu_process_callbacks(rsp);
1879 trace_rcu_utilization("End RCU core");
1883 * Schedule RCU callback invocation. If the specified type of RCU
1884 * does not support RCU priority boosting, just do a direct call,
1885 * otherwise wake up the per-CPU kernel kthread. Note that because we
1886 * are running on the current CPU with interrupts disabled, the
1887 * rcu_cpu_kthread_task cannot disappear out from under us.
1889 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1891 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1893 if (likely(!rsp->boost)) {
1894 rcu_do_batch(rsp, rdp);
1897 invoke_rcu_callbacks_kthread();
1900 static void invoke_rcu_core(void)
1902 raise_softirq(RCU_SOFTIRQ);
1906 * Handle any core-RCU processing required by a call_rcu() invocation.
1908 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1909 struct rcu_head *head, unsigned long flags)
1912 * If called from an extended quiescent state, invoke the RCU
1913 * core in order to force a re-evaluation of RCU's idleness.
1915 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1918 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1919 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1923 * Force the grace period if too many callbacks or too long waiting.
1924 * Enforce hysteresis, and don't invoke force_quiescent_state()
1925 * if some other CPU has recently done so. Also, don't bother
1926 * invoking force_quiescent_state() if the newly enqueued callback
1927 * is the only one waiting for a grace period to complete.
1929 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1931 /* Are we ignoring a completed grace period? */
1932 rcu_process_gp_end(rsp, rdp);
1933 check_for_new_grace_period(rsp, rdp);
1935 /* Start a new grace period if one not already started. */
1936 if (!rcu_gp_in_progress(rsp)) {
1937 unsigned long nestflag;
1938 struct rcu_node *rnp_root = rcu_get_root(rsp);
1940 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1941 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1943 /* Give the grace period a kick. */
1944 rdp->blimit = LONG_MAX;
1945 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1946 *rdp->nxttail[RCU_DONE_TAIL] != head)
1947 force_quiescent_state(rsp, 0);
1948 rdp->n_force_qs_snap = rsp->n_force_qs;
1949 rdp->qlen_last_fqs_check = rdp->qlen;
1951 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1952 force_quiescent_state(rsp, 1);
1956 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1957 struct rcu_state *rsp, bool lazy)
1959 unsigned long flags;
1960 struct rcu_data *rdp;
1962 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1963 debug_rcu_head_queue(head);
1967 smp_mb(); /* Ensure RCU update seen before callback registry. */
1970 * Opportunistically note grace-period endings and beginnings.
1971 * Note that we might see a beginning right after we see an
1972 * end, but never vice versa, since this CPU has to pass through
1973 * a quiescent state betweentimes.
1975 local_irq_save(flags);
1976 rdp = this_cpu_ptr(rsp->rda);
1978 /* Add the callback to our list. */
1979 ACCESS_ONCE(rdp->qlen)++;
1983 rcu_idle_count_callbacks_posted();
1984 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1985 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1986 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1988 if (__is_kfree_rcu_offset((unsigned long)func))
1989 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1990 rdp->qlen_lazy, rdp->qlen);
1992 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1994 /* Go handle any RCU core processing required. */
1995 __call_rcu_core(rsp, rdp, head, flags);
1996 local_irq_restore(flags);
2000 * Queue an RCU-sched callback for invocation after a grace period.
2002 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2004 __call_rcu(head, func, &rcu_sched_state, 0);
2006 EXPORT_SYMBOL_GPL(call_rcu_sched);
2009 * Queue an RCU callback for invocation after a quicker grace period.
2011 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2013 __call_rcu(head, func, &rcu_bh_state, 0);
2015 EXPORT_SYMBOL_GPL(call_rcu_bh);
2018 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2019 * any blocking grace-period wait automatically implies a grace period
2020 * if there is only one CPU online at any point time during execution
2021 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2022 * occasionally incorrectly indicate that there are multiple CPUs online
2023 * when there was in fact only one the whole time, as this just adds
2024 * some overhead: RCU still operates correctly.
2026 static inline int rcu_blocking_is_gp(void)
2030 might_sleep(); /* Check for RCU read-side critical section. */
2032 ret = num_online_cpus() <= 1;
2038 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2040 * Control will return to the caller some time after a full rcu-sched
2041 * grace period has elapsed, in other words after all currently executing
2042 * rcu-sched read-side critical sections have completed. These read-side
2043 * critical sections are delimited by rcu_read_lock_sched() and
2044 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2045 * local_irq_disable(), and so on may be used in place of
2046 * rcu_read_lock_sched().
2048 * This means that all preempt_disable code sequences, including NMI and
2049 * hardware-interrupt handlers, in progress on entry will have completed
2050 * before this primitive returns. However, this does not guarantee that
2051 * softirq handlers will have completed, since in some kernels, these
2052 * handlers can run in process context, and can block.
2054 * This primitive provides the guarantees made by the (now removed)
2055 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2056 * guarantees that rcu_read_lock() sections will have completed.
2057 * In "classic RCU", these two guarantees happen to be one and
2058 * the same, but can differ in realtime RCU implementations.
2060 void synchronize_sched(void)
2062 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2063 !lock_is_held(&rcu_lock_map) &&
2064 !lock_is_held(&rcu_sched_lock_map),
2065 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2066 if (rcu_blocking_is_gp())
2068 wait_rcu_gp(call_rcu_sched);
2070 EXPORT_SYMBOL_GPL(synchronize_sched);
2073 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2075 * Control will return to the caller some time after a full rcu_bh grace
2076 * period has elapsed, in other words after all currently executing rcu_bh
2077 * read-side critical sections have completed. RCU read-side critical
2078 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2079 * and may be nested.
2081 void synchronize_rcu_bh(void)
2083 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2084 !lock_is_held(&rcu_lock_map) &&
2085 !lock_is_held(&rcu_sched_lock_map),
2086 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2087 if (rcu_blocking_is_gp())
2089 wait_rcu_gp(call_rcu_bh);
2091 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2093 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2094 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2096 static int synchronize_sched_expedited_cpu_stop(void *data)
2099 * There must be a full memory barrier on each affected CPU
2100 * between the time that try_stop_cpus() is called and the
2101 * time that it returns.
2103 * In the current initial implementation of cpu_stop, the
2104 * above condition is already met when the control reaches
2105 * this point and the following smp_mb() is not strictly
2106 * necessary. Do smp_mb() anyway for documentation and
2107 * robustness against future implementation changes.
2109 smp_mb(); /* See above comment block. */
2114 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2116 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2117 * approach to force the grace period to end quickly. This consumes
2118 * significant time on all CPUs and is unfriendly to real-time workloads,
2119 * so is thus not recommended for any sort of common-case code. In fact,
2120 * if you are using synchronize_sched_expedited() in a loop, please
2121 * restructure your code to batch your updates, and then use a single
2122 * synchronize_sched() instead.
2124 * Note that it is illegal to call this function while holding any lock
2125 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2126 * to call this function from a CPU-hotplug notifier. Failing to observe
2127 * these restriction will result in deadlock.
2129 * This implementation can be thought of as an application of ticket
2130 * locking to RCU, with sync_sched_expedited_started and
2131 * sync_sched_expedited_done taking on the roles of the halves
2132 * of the ticket-lock word. Each task atomically increments
2133 * sync_sched_expedited_started upon entry, snapshotting the old value,
2134 * then attempts to stop all the CPUs. If this succeeds, then each
2135 * CPU will have executed a context switch, resulting in an RCU-sched
2136 * grace period. We are then done, so we use atomic_cmpxchg() to
2137 * update sync_sched_expedited_done to match our snapshot -- but
2138 * only if someone else has not already advanced past our snapshot.
2140 * On the other hand, if try_stop_cpus() fails, we check the value
2141 * of sync_sched_expedited_done. If it has advanced past our
2142 * initial snapshot, then someone else must have forced a grace period
2143 * some time after we took our snapshot. In this case, our work is
2144 * done for us, and we can simply return. Otherwise, we try again,
2145 * but keep our initial snapshot for purposes of checking for someone
2146 * doing our work for us.
2148 * If we fail too many times in a row, we fall back to synchronize_sched().
2150 void synchronize_sched_expedited(void)
2152 int firstsnap, s, snap, trycount = 0;
2154 /* Note that atomic_inc_return() implies full memory barrier. */
2155 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2157 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2160 * Each pass through the following loop attempts to force a
2161 * context switch on each CPU.
2163 while (try_stop_cpus(cpu_online_mask,
2164 synchronize_sched_expedited_cpu_stop,
2168 /* No joy, try again later. Or just synchronize_sched(). */
2169 if (trycount++ < 10) {
2170 udelay(trycount * num_online_cpus());
2172 synchronize_sched();
2176 /* Check to see if someone else did our work for us. */
2177 s = atomic_read(&sync_sched_expedited_done);
2178 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2179 smp_mb(); /* ensure test happens before caller kfree */
2184 * Refetching sync_sched_expedited_started allows later
2185 * callers to piggyback on our grace period. We subtract
2186 * 1 to get the same token that the last incrementer got.
2187 * We retry after they started, so our grace period works
2188 * for them, and they started after our first try, so their
2189 * grace period works for us.
2192 snap = atomic_read(&sync_sched_expedited_started);
2193 smp_mb(); /* ensure read is before try_stop_cpus(). */
2197 * Everyone up to our most recent fetch is covered by our grace
2198 * period. Update the counter, but only if our work is still
2199 * relevant -- which it won't be if someone who started later
2200 * than we did beat us to the punch.
2203 s = atomic_read(&sync_sched_expedited_done);
2204 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2205 smp_mb(); /* ensure test happens before caller kfree */
2208 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2212 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2215 * Check to see if there is any immediate RCU-related work to be done
2216 * by the current CPU, for the specified type of RCU, returning 1 if so.
2217 * The checks are in order of increasing expense: checks that can be
2218 * carried out against CPU-local state are performed first. However,
2219 * we must check for CPU stalls first, else we might not get a chance.
2221 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2223 struct rcu_node *rnp = rdp->mynode;
2225 rdp->n_rcu_pending++;
2227 /* Check for CPU stalls, if enabled. */
2228 check_cpu_stall(rsp, rdp);
2230 /* Is the RCU core waiting for a quiescent state from this CPU? */
2231 if (rcu_scheduler_fully_active &&
2232 rdp->qs_pending && !rdp->passed_quiesce) {
2235 * If force_quiescent_state() coming soon and this CPU
2236 * needs a quiescent state, and this is either RCU-sched
2237 * or RCU-bh, force a local reschedule.
2239 rdp->n_rp_qs_pending++;
2240 if (!rdp->preemptible &&
2241 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2244 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2245 rdp->n_rp_report_qs++;
2249 /* Does this CPU have callbacks ready to invoke? */
2250 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2251 rdp->n_rp_cb_ready++;
2255 /* Has RCU gone idle with this CPU needing another grace period? */
2256 if (cpu_needs_another_gp(rsp, rdp)) {
2257 rdp->n_rp_cpu_needs_gp++;
2261 /* Has another RCU grace period completed? */
2262 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2263 rdp->n_rp_gp_completed++;
2267 /* Has a new RCU grace period started? */
2268 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2269 rdp->n_rp_gp_started++;
2273 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2274 if (rcu_gp_in_progress(rsp) &&
2275 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2276 rdp->n_rp_need_fqs++;
2281 rdp->n_rp_need_nothing++;
2286 * Check to see if there is any immediate RCU-related work to be done
2287 * by the current CPU, returning 1 if so. This function is part of the
2288 * RCU implementation; it is -not- an exported member of the RCU API.
2290 static int rcu_pending(int cpu)
2292 struct rcu_state *rsp;
2294 for_each_rcu_flavor(rsp)
2295 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2301 * Check to see if any future RCU-related work will need to be done
2302 * by the current CPU, even if none need be done immediately, returning
2305 static int rcu_cpu_has_callbacks(int cpu)
2307 struct rcu_state *rsp;
2309 /* RCU callbacks either ready or pending? */
2310 for_each_rcu_flavor(rsp)
2311 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2317 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2318 * the compiler is expected to optimize this away.
2320 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2321 int cpu, unsigned long done)
2323 trace_rcu_barrier(rsp->name, s, cpu,
2324 atomic_read(&rsp->barrier_cpu_count), done);
2328 * RCU callback function for _rcu_barrier(). If we are last, wake
2329 * up the task executing _rcu_barrier().
2331 static void rcu_barrier_callback(struct rcu_head *rhp)
2333 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2334 struct rcu_state *rsp = rdp->rsp;
2336 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2337 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2338 complete(&rsp->barrier_completion);
2340 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2345 * Called with preemption disabled, and from cross-cpu IRQ context.
2347 static void rcu_barrier_func(void *type)
2349 struct rcu_state *rsp = type;
2350 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2352 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2353 atomic_inc(&rsp->barrier_cpu_count);
2354 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2358 * Orchestrate the specified type of RCU barrier, waiting for all
2359 * RCU callbacks of the specified type to complete.
2361 static void _rcu_barrier(struct rcu_state *rsp)
2364 unsigned long flags;
2365 struct rcu_data *rdp;
2367 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2368 unsigned long snap_done;
2370 init_rcu_head_on_stack(&rd.barrier_head);
2371 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2373 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2374 mutex_lock(&rsp->barrier_mutex);
2377 * Ensure that all prior references, including to ->n_barrier_done,
2378 * are ordered before the _rcu_barrier() machinery.
2380 smp_mb(); /* See above block comment. */
2383 * Recheck ->n_barrier_done to see if others did our work for us.
2384 * This means checking ->n_barrier_done for an even-to-odd-to-even
2385 * transition. The "if" expression below therefore rounds the old
2386 * value up to the next even number and adds two before comparing.
2388 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2389 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2390 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2391 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2392 smp_mb(); /* caller's subsequent code after above check. */
2393 mutex_unlock(&rsp->barrier_mutex);
2398 * Increment ->n_barrier_done to avoid duplicate work. Use
2399 * ACCESS_ONCE() to prevent the compiler from speculating
2400 * the increment to precede the early-exit check.
2402 ACCESS_ONCE(rsp->n_barrier_done)++;
2403 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2404 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2405 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2408 * Initialize the count to one rather than to zero in order to
2409 * avoid a too-soon return to zero in case of a short grace period
2410 * (or preemption of this task). Also flag this task as doing
2411 * an rcu_barrier(). This will prevent anyone else from adopting
2412 * orphaned callbacks, which could cause otherwise failure if a
2413 * CPU went offline and quickly came back online. To see this,
2414 * consider the following sequence of events:
2416 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2417 * 2. CPU 1 goes offline, orphaning its callbacks.
2418 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2419 * 4. CPU 1 comes back online.
2420 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2421 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2422 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2424 init_completion(&rsp->barrier_completion);
2425 atomic_set(&rsp->barrier_cpu_count, 1);
2426 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2427 rsp->rcu_barrier_in_progress = current;
2428 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2431 * Force every CPU with callbacks to register a new callback
2432 * that will tell us when all the preceding callbacks have
2433 * been invoked. If an offline CPU has callbacks, wait for
2434 * it to either come back online or to finish orphaning those
2437 for_each_possible_cpu(cpu) {
2439 rdp = per_cpu_ptr(rsp->rda, cpu);
2440 if (cpu_is_offline(cpu)) {
2441 _rcu_barrier_trace(rsp, "Offline", cpu,
2442 rsp->n_barrier_done);
2444 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2445 schedule_timeout_interruptible(1);
2446 } else if (ACCESS_ONCE(rdp->qlen)) {
2447 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2448 rsp->n_barrier_done);
2449 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2452 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2453 rsp->n_barrier_done);
2459 * Now that all online CPUs have rcu_barrier_callback() callbacks
2460 * posted, we can adopt all of the orphaned callbacks and place
2461 * an rcu_barrier_callback() callback after them. When that is done,
2462 * we are guaranteed to have an rcu_barrier_callback() callback
2463 * following every callback that could possibly have been
2464 * registered before _rcu_barrier() was called.
2466 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2467 rcu_adopt_orphan_cbs(rsp);
2468 rsp->rcu_barrier_in_progress = NULL;
2469 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2470 atomic_inc(&rsp->barrier_cpu_count);
2471 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2473 rsp->call(&rd.barrier_head, rcu_barrier_callback);
2476 * Now that we have an rcu_barrier_callback() callback on each
2477 * CPU, and thus each counted, remove the initial count.
2479 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2480 complete(&rsp->barrier_completion);
2482 /* Increment ->n_barrier_done to prevent duplicate work. */
2483 smp_mb(); /* Keep increment after above mechanism. */
2484 ACCESS_ONCE(rsp->n_barrier_done)++;
2485 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2486 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2487 smp_mb(); /* Keep increment before caller's subsequent code. */
2489 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2490 wait_for_completion(&rsp->barrier_completion);
2492 /* Other rcu_barrier() invocations can now safely proceed. */
2493 mutex_unlock(&rsp->barrier_mutex);
2495 destroy_rcu_head_on_stack(&rd.barrier_head);
2499 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2501 void rcu_barrier_bh(void)
2503 _rcu_barrier(&rcu_bh_state);
2505 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2508 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2510 void rcu_barrier_sched(void)
2512 _rcu_barrier(&rcu_sched_state);
2514 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2517 * Do boot-time initialization of a CPU's per-CPU RCU data.
2520 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2522 unsigned long flags;
2523 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2524 struct rcu_node *rnp = rcu_get_root(rsp);
2526 /* Set up local state, ensuring consistent view of global state. */
2527 raw_spin_lock_irqsave(&rnp->lock, flags);
2528 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2529 init_callback_list(rdp);
2531 ACCESS_ONCE(rdp->qlen) = 0;
2532 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2533 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2534 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2537 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2541 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2542 * offline event can be happening at a given time. Note also that we
2543 * can accept some slop in the rsp->completed access due to the fact
2544 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2546 static void __cpuinit
2547 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2549 unsigned long flags;
2551 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2552 struct rcu_node *rnp = rcu_get_root(rsp);
2554 /* Set up local state, ensuring consistent view of global state. */
2555 raw_spin_lock_irqsave(&rnp->lock, flags);
2556 rdp->beenonline = 1; /* We have now been online. */
2557 rdp->preemptible = preemptible;
2558 rdp->qlen_last_fqs_check = 0;
2559 rdp->n_force_qs_snap = rsp->n_force_qs;
2560 rdp->blimit = blimit;
2561 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2562 atomic_set(&rdp->dynticks->dynticks,
2563 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2564 rcu_prepare_for_idle_init(cpu);
2565 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2568 * A new grace period might start here. If so, we won't be part
2569 * of it, but that is OK, as we are currently in a quiescent state.
2572 /* Exclude any attempts to start a new GP on large systems. */
2573 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2575 /* Add CPU to rcu_node bitmasks. */
2577 mask = rdp->grpmask;
2579 /* Exclude any attempts to start a new GP on small systems. */
2580 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2581 rnp->qsmaskinit |= mask;
2582 mask = rnp->grpmask;
2583 if (rnp == rdp->mynode) {
2585 * If there is a grace period in progress, we will
2586 * set up to wait for it next time we run the
2589 rdp->gpnum = rnp->completed;
2590 rdp->completed = rnp->completed;
2591 rdp->passed_quiesce = 0;
2592 rdp->qs_pending = 0;
2593 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2594 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2596 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2598 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2600 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2603 static void __cpuinit rcu_prepare_cpu(int cpu)
2605 struct rcu_state *rsp;
2607 for_each_rcu_flavor(rsp)
2608 rcu_init_percpu_data(cpu, rsp,
2609 strcmp(rsp->name, "rcu_preempt") == 0);
2613 * Handle CPU online/offline notification events.
2615 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2616 unsigned long action, void *hcpu)
2618 long cpu = (long)hcpu;
2619 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2620 struct rcu_node *rnp = rdp->mynode;
2621 struct rcu_state *rsp;
2623 trace_rcu_utilization("Start CPU hotplug");
2625 case CPU_UP_PREPARE:
2626 case CPU_UP_PREPARE_FROZEN:
2627 rcu_prepare_cpu(cpu);
2628 rcu_prepare_kthreads(cpu);
2631 case CPU_DOWN_FAILED:
2632 rcu_node_kthread_setaffinity(rnp, -1);
2633 rcu_cpu_kthread_setrt(cpu, 1);
2635 case CPU_DOWN_PREPARE:
2636 rcu_node_kthread_setaffinity(rnp, cpu);
2637 rcu_cpu_kthread_setrt(cpu, 0);
2640 case CPU_DYING_FROZEN:
2642 * The whole machine is "stopped" except this CPU, so we can
2643 * touch any data without introducing corruption. We send the
2644 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2646 for_each_rcu_flavor(rsp)
2647 rcu_cleanup_dying_cpu(rsp);
2648 rcu_cleanup_after_idle(cpu);
2651 case CPU_DEAD_FROZEN:
2652 case CPU_UP_CANCELED:
2653 case CPU_UP_CANCELED_FROZEN:
2654 for_each_rcu_flavor(rsp)
2655 rcu_cleanup_dead_cpu(cpu, rsp);
2660 trace_rcu_utilization("End CPU hotplug");
2665 * Spawn the kthread that handles this RCU flavor's grace periods.
2667 static int __init rcu_spawn_gp_kthread(void)
2669 unsigned long flags;
2670 struct rcu_node *rnp;
2671 struct rcu_state *rsp;
2672 struct task_struct *t;
2674 for_each_rcu_flavor(rsp) {
2675 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2677 rnp = rcu_get_root(rsp);
2678 raw_spin_lock_irqsave(&rnp->lock, flags);
2679 rsp->gp_kthread = t;
2680 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2684 early_initcall(rcu_spawn_gp_kthread);
2687 * This function is invoked towards the end of the scheduler's initialization
2688 * process. Before this is called, the idle task might contain
2689 * RCU read-side critical sections (during which time, this idle
2690 * task is booting the system). After this function is called, the
2691 * idle tasks are prohibited from containing RCU read-side critical
2692 * sections. This function also enables RCU lockdep checking.
2694 void rcu_scheduler_starting(void)
2696 WARN_ON(num_online_cpus() != 1);
2697 WARN_ON(nr_context_switches() > 0);
2698 rcu_scheduler_active = 1;
2702 * Compute the per-level fanout, either using the exact fanout specified
2703 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2705 #ifdef CONFIG_RCU_FANOUT_EXACT
2706 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2710 for (i = rcu_num_lvls - 1; i > 0; i--)
2711 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2712 rsp->levelspread[0] = rcu_fanout_leaf;
2714 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2715 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2722 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2723 ccur = rsp->levelcnt[i];
2724 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2728 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2731 * Helper function for rcu_init() that initializes one rcu_state structure.
2733 static void __init rcu_init_one(struct rcu_state *rsp,
2734 struct rcu_data __percpu *rda)
2736 static char *buf[] = { "rcu_node_level_0",
2739 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2743 struct rcu_node *rnp;
2745 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2747 /* Initialize the level-tracking arrays. */
2749 for (i = 0; i < rcu_num_lvls; i++)
2750 rsp->levelcnt[i] = num_rcu_lvl[i];
2751 for (i = 1; i < rcu_num_lvls; i++)
2752 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2753 rcu_init_levelspread(rsp);
2755 /* Initialize the elements themselves, starting from the leaves. */
2757 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2758 cpustride *= rsp->levelspread[i];
2759 rnp = rsp->level[i];
2760 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2761 raw_spin_lock_init(&rnp->lock);
2762 lockdep_set_class_and_name(&rnp->lock,
2763 &rcu_node_class[i], buf[i]);
2766 rnp->qsmaskinit = 0;
2767 rnp->grplo = j * cpustride;
2768 rnp->grphi = (j + 1) * cpustride - 1;
2769 if (rnp->grphi >= NR_CPUS)
2770 rnp->grphi = NR_CPUS - 1;
2776 rnp->grpnum = j % rsp->levelspread[i - 1];
2777 rnp->grpmask = 1UL << rnp->grpnum;
2778 rnp->parent = rsp->level[i - 1] +
2779 j / rsp->levelspread[i - 1];
2782 INIT_LIST_HEAD(&rnp->blkd_tasks);
2787 init_waitqueue_head(&rsp->gp_wq);
2788 rnp = rsp->level[rcu_num_lvls - 1];
2789 for_each_possible_cpu(i) {
2790 while (i > rnp->grphi)
2792 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2793 rcu_boot_init_percpu_data(i, rsp);
2795 list_add(&rsp->flavors, &rcu_struct_flavors);
2799 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2800 * replace the definitions in rcutree.h because those are needed to size
2801 * the ->node array in the rcu_state structure.
2803 static void __init rcu_init_geometry(void)
2808 int rcu_capacity[MAX_RCU_LVLS + 1];
2810 /* If the compile-time values are accurate, just leave. */
2811 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2815 * Compute number of nodes that can be handled an rcu_node tree
2816 * with the given number of levels. Setting rcu_capacity[0] makes
2817 * some of the arithmetic easier.
2819 rcu_capacity[0] = 1;
2820 rcu_capacity[1] = rcu_fanout_leaf;
2821 for (i = 2; i <= MAX_RCU_LVLS; i++)
2822 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2825 * The boot-time rcu_fanout_leaf parameter is only permitted
2826 * to increase the leaf-level fanout, not decrease it. Of course,
2827 * the leaf-level fanout cannot exceed the number of bits in
2828 * the rcu_node masks. Finally, the tree must be able to accommodate
2829 * the configured number of CPUs. Complain and fall back to the
2830 * compile-time values if these limits are exceeded.
2832 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2833 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2834 n > rcu_capacity[MAX_RCU_LVLS]) {
2839 /* Calculate the number of rcu_nodes at each level of the tree. */
2840 for (i = 1; i <= MAX_RCU_LVLS; i++)
2841 if (n <= rcu_capacity[i]) {
2842 for (j = 0; j <= i; j++)
2844 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2846 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2851 /* Calculate the total number of rcu_node structures. */
2853 for (i = 0; i <= MAX_RCU_LVLS; i++)
2854 rcu_num_nodes += num_rcu_lvl[i];
2858 void __init rcu_init(void)
2862 rcu_bootup_announce();
2863 rcu_init_geometry();
2864 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2865 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2866 __rcu_init_preempt();
2867 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2870 * We don't need protection against CPU-hotplug here because
2871 * this is called early in boot, before either interrupts
2872 * or the scheduler are operational.
2874 cpu_notifier(rcu_cpu_notify, 0);
2875 for_each_online_cpu(cpu)
2876 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2877 check_cpu_stall_init();
2880 #include "rcutree_plugin.h"