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];
64 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
66 #define RCU_STATE_INITIALIZER(sname, cr) { \
67 .level = { &sname##_state.node[0] }, \
69 .fqs_state = RCU_GP_IDLE, \
72 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
73 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
74 .orphan_donetail = &sname##_state.orphan_donelist, \
75 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
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_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
230 static void force_quiescent_state(struct rcu_state *rsp);
231 static int rcu_pending(int cpu);
234 * Return the number of RCU-sched batches processed thus far for debug & stats.
236 long rcu_batches_completed_sched(void)
238 return rcu_sched_state.completed;
240 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
243 * Return the number of RCU BH batches processed thus far for debug & stats.
245 long rcu_batches_completed_bh(void)
247 return rcu_bh_state.completed;
249 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
252 * Force a quiescent state for RCU BH.
254 void rcu_bh_force_quiescent_state(void)
256 force_quiescent_state(&rcu_bh_state);
258 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
261 * Record the number of times rcutorture tests have been initiated and
262 * terminated. This information allows the debugfs tracing stats to be
263 * correlated to the rcutorture messages, even when the rcutorture module
264 * is being repeatedly loaded and unloaded. In other words, we cannot
265 * store this state in rcutorture itself.
267 void rcutorture_record_test_transition(void)
269 rcutorture_testseq++;
270 rcutorture_vernum = 0;
272 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
275 * Record the number of writer passes through the current rcutorture test.
276 * This is also used to correlate debugfs tracing stats with the rcutorture
279 void rcutorture_record_progress(unsigned long vernum)
283 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
286 * Force a quiescent state for RCU-sched.
288 void rcu_sched_force_quiescent_state(void)
290 force_quiescent_state(&rcu_sched_state);
292 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
295 * Does the CPU have callbacks ready to be invoked?
298 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
300 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
304 * Does the current CPU require a yet-as-unscheduled grace period?
307 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
309 return *rdp->nxttail[RCU_DONE_TAIL +
310 ACCESS_ONCE(rsp->completed) != rdp->completed] &&
311 !rcu_gp_in_progress(rsp);
315 * Return the root node of the specified rcu_state structure.
317 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
319 return &rsp->node[0];
323 * If the specified CPU is offline, tell the caller that it is in
324 * a quiescent state. Otherwise, whack it with a reschedule IPI.
325 * Grace periods can end up waiting on an offline CPU when that
326 * CPU is in the process of coming online -- it will be added to the
327 * rcu_node bitmasks before it actually makes it online. The same thing
328 * can happen while a CPU is in the process of coming online. Because this
329 * race is quite rare, we check for it after detecting that the grace
330 * period has been delayed rather than checking each and every CPU
331 * each and every time we start a new grace period.
333 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
336 * If the CPU is offline for more than a jiffy, it is in a quiescent
337 * state. We can trust its state not to change because interrupts
338 * are disabled. The reason for the jiffy's worth of slack is to
339 * handle CPUs initializing on the way up and finding their way
340 * to the idle loop on the way down.
342 if (cpu_is_offline(rdp->cpu) &&
343 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
344 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
352 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
354 * If the new value of the ->dynticks_nesting counter now is zero,
355 * we really have entered idle, and must do the appropriate accounting.
356 * The caller must have disabled interrupts.
358 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
360 trace_rcu_dyntick("Start", oldval, 0);
361 if (!is_idle_task(current)) {
362 struct task_struct *idle = idle_task(smp_processor_id());
364 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
365 ftrace_dump(DUMP_ORIG);
366 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
367 current->pid, current->comm,
368 idle->pid, idle->comm); /* must be idle task! */
370 rcu_prepare_for_idle(smp_processor_id());
371 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
372 smp_mb__before_atomic_inc(); /* See above. */
373 atomic_inc(&rdtp->dynticks);
374 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
375 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
378 * The idle task is not permitted to enter the idle loop while
379 * in an RCU read-side critical section.
381 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
382 "Illegal idle entry in RCU read-side critical section.");
383 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
384 "Illegal idle entry in RCU-bh read-side critical section.");
385 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
386 "Illegal idle entry in RCU-sched read-side critical section.");
390 * rcu_idle_enter - inform RCU that current CPU is entering idle
392 * Enter idle mode, in other words, -leave- the mode in which RCU
393 * read-side critical sections can occur. (Though RCU read-side
394 * critical sections can occur in irq handlers in idle, a possibility
395 * handled by irq_enter() and irq_exit().)
397 * We crowbar the ->dynticks_nesting field to zero to allow for
398 * the possibility of usermode upcalls having messed up our count
399 * of interrupt nesting level during the prior busy period.
401 void rcu_idle_enter(void)
405 struct rcu_dynticks *rdtp;
407 local_irq_save(flags);
408 rdtp = &__get_cpu_var(rcu_dynticks);
409 oldval = rdtp->dynticks_nesting;
410 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
411 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
412 rdtp->dynticks_nesting = 0;
414 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
415 rcu_idle_enter_common(rdtp, oldval);
416 local_irq_restore(flags);
418 EXPORT_SYMBOL_GPL(rcu_idle_enter);
421 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
423 * Exit from an interrupt handler, which might possibly result in entering
424 * idle mode, in other words, leaving the mode in which read-side critical
425 * sections can occur.
427 * This code assumes that the idle loop never does anything that might
428 * result in unbalanced calls to irq_enter() and irq_exit(). If your
429 * architecture violates this assumption, RCU will give you what you
430 * deserve, good and hard. But very infrequently and irreproducibly.
432 * Use things like work queues to work around this limitation.
434 * You have been warned.
436 void rcu_irq_exit(void)
440 struct rcu_dynticks *rdtp;
442 local_irq_save(flags);
443 rdtp = &__get_cpu_var(rcu_dynticks);
444 oldval = rdtp->dynticks_nesting;
445 rdtp->dynticks_nesting--;
446 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
447 if (rdtp->dynticks_nesting)
448 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
450 rcu_idle_enter_common(rdtp, oldval);
451 local_irq_restore(flags);
455 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
457 * If the new value of the ->dynticks_nesting counter was previously zero,
458 * we really have exited idle, and must do the appropriate accounting.
459 * The caller must have disabled interrupts.
461 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
463 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
464 atomic_inc(&rdtp->dynticks);
465 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
466 smp_mb__after_atomic_inc(); /* See above. */
467 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
468 rcu_cleanup_after_idle(smp_processor_id());
469 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
470 if (!is_idle_task(current)) {
471 struct task_struct *idle = idle_task(smp_processor_id());
473 trace_rcu_dyntick("Error on exit: not idle task",
474 oldval, rdtp->dynticks_nesting);
475 ftrace_dump(DUMP_ORIG);
476 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
477 current->pid, current->comm,
478 idle->pid, idle->comm); /* must be idle task! */
483 * rcu_idle_exit - inform RCU that current CPU is leaving idle
485 * Exit idle mode, in other words, -enter- the mode in which RCU
486 * read-side critical sections can occur.
488 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
489 * allow for the possibility of usermode upcalls messing up our count
490 * of interrupt nesting level during the busy period that is just
493 void rcu_idle_exit(void)
496 struct rcu_dynticks *rdtp;
499 local_irq_save(flags);
500 rdtp = &__get_cpu_var(rcu_dynticks);
501 oldval = rdtp->dynticks_nesting;
502 WARN_ON_ONCE(oldval < 0);
503 if (oldval & DYNTICK_TASK_NEST_MASK)
504 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
506 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
507 rcu_idle_exit_common(rdtp, oldval);
508 local_irq_restore(flags);
510 EXPORT_SYMBOL_GPL(rcu_idle_exit);
513 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
515 * Enter an interrupt handler, which might possibly result in exiting
516 * idle mode, in other words, entering the mode in which read-side critical
517 * sections can occur.
519 * Note that the Linux kernel is fully capable of entering an interrupt
520 * handler that it never exits, for example when doing upcalls to
521 * user mode! This code assumes that the idle loop never does upcalls to
522 * user mode. If your architecture does do upcalls from the idle loop (or
523 * does anything else that results in unbalanced calls to the irq_enter()
524 * and irq_exit() functions), RCU will give you what you deserve, good
525 * and hard. But very infrequently and irreproducibly.
527 * Use things like work queues to work around this limitation.
529 * You have been warned.
531 void rcu_irq_enter(void)
534 struct rcu_dynticks *rdtp;
537 local_irq_save(flags);
538 rdtp = &__get_cpu_var(rcu_dynticks);
539 oldval = rdtp->dynticks_nesting;
540 rdtp->dynticks_nesting++;
541 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
543 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
545 rcu_idle_exit_common(rdtp, oldval);
546 local_irq_restore(flags);
550 * rcu_nmi_enter - inform RCU of entry to NMI context
552 * If the CPU was idle with dynamic ticks active, and there is no
553 * irq handler running, this updates rdtp->dynticks_nmi to let the
554 * RCU grace-period handling know that the CPU is active.
556 void rcu_nmi_enter(void)
558 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
560 if (rdtp->dynticks_nmi_nesting == 0 &&
561 (atomic_read(&rdtp->dynticks) & 0x1))
563 rdtp->dynticks_nmi_nesting++;
564 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
565 atomic_inc(&rdtp->dynticks);
566 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
567 smp_mb__after_atomic_inc(); /* See above. */
568 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
572 * rcu_nmi_exit - inform RCU of exit from NMI context
574 * If the CPU was idle with dynamic ticks active, and there is no
575 * irq handler running, this updates rdtp->dynticks_nmi to let the
576 * RCU grace-period handling know that the CPU is no longer active.
578 void rcu_nmi_exit(void)
580 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
582 if (rdtp->dynticks_nmi_nesting == 0 ||
583 --rdtp->dynticks_nmi_nesting != 0)
585 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
586 smp_mb__before_atomic_inc(); /* See above. */
587 atomic_inc(&rdtp->dynticks);
588 smp_mb__after_atomic_inc(); /* Force delay to next write. */
589 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
593 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
595 * If the current CPU is in its idle loop and is neither in an interrupt
596 * or NMI handler, return true.
598 int rcu_is_cpu_idle(void)
603 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
607 EXPORT_SYMBOL(rcu_is_cpu_idle);
609 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
612 * Is the current CPU online? Disable preemption to avoid false positives
613 * that could otherwise happen due to the current CPU number being sampled,
614 * this task being preempted, its old CPU being taken offline, resuming
615 * on some other CPU, then determining that its old CPU is now offline.
616 * It is OK to use RCU on an offline processor during initial boot, hence
617 * the check for rcu_scheduler_fully_active. Note also that it is OK
618 * for a CPU coming online to use RCU for one jiffy prior to marking itself
619 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
620 * offline to continue to use RCU for one jiffy after marking itself
621 * offline in the cpu_online_mask. This leniency is necessary given the
622 * non-atomic nature of the online and offline processing, for example,
623 * the fact that a CPU enters the scheduler after completing the CPU_DYING
626 * This is also why RCU internally marks CPUs online during the
627 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
629 * Disable checking if in an NMI handler because we cannot safely report
630 * errors from NMI handlers anyway.
632 bool rcu_lockdep_current_cpu_online(void)
634 struct rcu_data *rdp;
635 struct rcu_node *rnp;
641 rdp = &__get_cpu_var(rcu_sched_data);
643 ret = (rdp->grpmask & rnp->qsmaskinit) ||
644 !rcu_scheduler_fully_active;
648 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
650 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
653 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
655 * If the current CPU is idle or running at a first-level (not nested)
656 * interrupt from idle, return true. The caller must have at least
657 * disabled preemption.
659 int rcu_is_cpu_rrupt_from_idle(void)
661 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
665 * Snapshot the specified CPU's dynticks counter so that we can later
666 * credit them with an implicit quiescent state. Return 1 if this CPU
667 * is in dynticks idle mode, which is an extended quiescent state.
669 static int dyntick_save_progress_counter(struct rcu_data *rdp)
671 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
672 return (rdp->dynticks_snap & 0x1) == 0;
676 * Return true if the specified CPU has passed through a quiescent
677 * state by virtue of being in or having passed through an dynticks
678 * idle state since the last call to dyntick_save_progress_counter()
681 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
686 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
687 snap = (unsigned int)rdp->dynticks_snap;
690 * If the CPU passed through or entered a dynticks idle phase with
691 * no active irq/NMI handlers, then we can safely pretend that the CPU
692 * already acknowledged the request to pass through a quiescent
693 * state. Either way, that CPU cannot possibly be in an RCU
694 * read-side critical section that started before the beginning
695 * of the current RCU grace period.
697 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
698 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
703 /* Go check for the CPU being offline. */
704 return rcu_implicit_offline_qs(rdp);
707 static int jiffies_till_stall_check(void)
709 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
712 * Limit check must be consistent with the Kconfig limits
713 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
715 if (till_stall_check < 3) {
716 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
717 till_stall_check = 3;
718 } else if (till_stall_check > 300) {
719 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
720 till_stall_check = 300;
722 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
725 static void record_gp_stall_check_time(struct rcu_state *rsp)
727 rsp->gp_start = jiffies;
728 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
731 static void print_other_cpu_stall(struct rcu_state *rsp)
737 struct rcu_node *rnp = rcu_get_root(rsp);
739 /* Only let one CPU complain about others per time interval. */
741 raw_spin_lock_irqsave(&rnp->lock, flags);
742 delta = jiffies - rsp->jiffies_stall;
743 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
744 raw_spin_unlock_irqrestore(&rnp->lock, flags);
747 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
748 raw_spin_unlock_irqrestore(&rnp->lock, flags);
751 * OK, time to rat on our buddy...
752 * See Documentation/RCU/stallwarn.txt for info on how to debug
753 * RCU CPU stall warnings.
755 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
757 print_cpu_stall_info_begin();
758 rcu_for_each_leaf_node(rsp, rnp) {
759 raw_spin_lock_irqsave(&rnp->lock, flags);
760 ndetected += rcu_print_task_stall(rnp);
761 raw_spin_unlock_irqrestore(&rnp->lock, flags);
762 if (rnp->qsmask == 0)
764 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
765 if (rnp->qsmask & (1UL << cpu)) {
766 print_cpu_stall_info(rsp, rnp->grplo + cpu);
772 * Now rat on any tasks that got kicked up to the root rcu_node
773 * due to CPU offlining.
775 rnp = rcu_get_root(rsp);
776 raw_spin_lock_irqsave(&rnp->lock, flags);
777 ndetected += rcu_print_task_stall(rnp);
778 raw_spin_unlock_irqrestore(&rnp->lock, flags);
780 print_cpu_stall_info_end();
781 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
782 smp_processor_id(), (long)(jiffies - rsp->gp_start));
784 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
785 else if (!trigger_all_cpu_backtrace())
788 /* Complain about tasks blocking the grace period. */
790 rcu_print_detail_task_stall(rsp);
792 force_quiescent_state(rsp); /* Kick them all. */
795 static void print_cpu_stall(struct rcu_state *rsp)
798 struct rcu_node *rnp = rcu_get_root(rsp);
801 * OK, time to rat on ourselves...
802 * See Documentation/RCU/stallwarn.txt for info on how to debug
803 * RCU CPU stall warnings.
805 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
806 print_cpu_stall_info_begin();
807 print_cpu_stall_info(rsp, smp_processor_id());
808 print_cpu_stall_info_end();
809 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
810 if (!trigger_all_cpu_backtrace())
813 raw_spin_lock_irqsave(&rnp->lock, flags);
814 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
815 rsp->jiffies_stall = jiffies +
816 3 * jiffies_till_stall_check() + 3;
817 raw_spin_unlock_irqrestore(&rnp->lock, flags);
819 set_need_resched(); /* kick ourselves to get things going. */
822 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
826 struct rcu_node *rnp;
828 if (rcu_cpu_stall_suppress)
830 j = ACCESS_ONCE(jiffies);
831 js = ACCESS_ONCE(rsp->jiffies_stall);
833 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
835 /* We haven't checked in, so go dump stack. */
836 print_cpu_stall(rsp);
838 } else if (rcu_gp_in_progress(rsp) &&
839 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
841 /* They had a few time units to dump stack, so complain. */
842 print_other_cpu_stall(rsp);
846 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
848 rcu_cpu_stall_suppress = 1;
853 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
855 * Set the stall-warning timeout way off into the future, thus preventing
856 * any RCU CPU stall-warning messages from appearing in the current set of
859 * The caller must disable hard irqs.
861 void rcu_cpu_stall_reset(void)
863 struct rcu_state *rsp;
865 for_each_rcu_flavor(rsp)
866 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
869 static struct notifier_block rcu_panic_block = {
870 .notifier_call = rcu_panic,
873 static void __init check_cpu_stall_init(void)
875 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
879 * Update CPU-local rcu_data state to record the newly noticed grace period.
880 * This is used both when we started the grace period and when we notice
881 * that someone else started the grace period. The caller must hold the
882 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
883 * and must have irqs disabled.
885 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
887 if (rdp->gpnum != rnp->gpnum) {
889 * If the current grace period is waiting for this CPU,
890 * set up to detect a quiescent state, otherwise don't
891 * go looking for one.
893 rdp->gpnum = rnp->gpnum;
894 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
895 if (rnp->qsmask & rdp->grpmask) {
897 rdp->passed_quiesce = 0;
901 zero_cpu_stall_ticks(rdp);
905 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
908 struct rcu_node *rnp;
910 local_irq_save(flags);
912 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
913 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
914 local_irq_restore(flags);
917 __note_new_gpnum(rsp, rnp, rdp);
918 raw_spin_unlock_irqrestore(&rnp->lock, flags);
922 * Did someone else start a new RCU grace period start since we last
923 * checked? Update local state appropriately if so. Must be called
924 * on the CPU corresponding to rdp.
927 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
932 local_irq_save(flags);
933 if (rdp->gpnum != rsp->gpnum) {
934 note_new_gpnum(rsp, rdp);
937 local_irq_restore(flags);
942 * Initialize the specified rcu_data structure's callback list to empty.
944 static void init_callback_list(struct rcu_data *rdp)
949 for (i = 0; i < RCU_NEXT_SIZE; i++)
950 rdp->nxttail[i] = &rdp->nxtlist;
954 * Advance this CPU's callbacks, but only if the current grace period
955 * has ended. This may be called only from the CPU to whom the rdp
956 * belongs. In addition, the corresponding leaf rcu_node structure's
957 * ->lock must be held by the caller, with irqs disabled.
960 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
962 /* Did another grace period end? */
963 if (rdp->completed != rnp->completed) {
965 /* Advance callbacks. No harm if list empty. */
966 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
967 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
968 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
970 /* Remember that we saw this grace-period completion. */
971 rdp->completed = rnp->completed;
972 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
975 * If we were in an extended quiescent state, we may have
976 * missed some grace periods that others CPUs handled on
977 * our behalf. Catch up with this state to avoid noting
978 * spurious new grace periods. If another grace period
979 * has started, then rnp->gpnum will have advanced, so
980 * we will detect this later on.
982 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
983 rdp->gpnum = rdp->completed;
986 * If RCU does not need a quiescent state from this CPU,
987 * then make sure that this CPU doesn't go looking for one.
989 if ((rnp->qsmask & rdp->grpmask) == 0)
995 * Advance this CPU's callbacks, but only if the current grace period
996 * has ended. This may be called only from the CPU to whom the rdp
1000 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1002 unsigned long flags;
1003 struct rcu_node *rnp;
1005 local_irq_save(flags);
1007 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1008 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1009 local_irq_restore(flags);
1012 __rcu_process_gp_end(rsp, rnp, rdp);
1013 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1017 * Do per-CPU grace-period initialization for running CPU. The caller
1018 * must hold the lock of the leaf rcu_node structure corresponding to
1022 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1024 /* Prior grace period ended, so advance callbacks for current CPU. */
1025 __rcu_process_gp_end(rsp, rnp, rdp);
1027 /* Set state so that this CPU will detect the next quiescent state. */
1028 __note_new_gpnum(rsp, rnp, rdp);
1032 * Initialize a new grace period.
1034 static int rcu_gp_init(struct rcu_state *rsp)
1036 struct rcu_data *rdp;
1037 struct rcu_node *rnp = rcu_get_root(rsp);
1039 raw_spin_lock_irq(&rnp->lock);
1040 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1042 if (rcu_gp_in_progress(rsp)) {
1043 /* Grace period already in progress, don't start another. */
1044 raw_spin_unlock_irq(&rnp->lock);
1048 /* Advance to a new grace period and initialize state. */
1050 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1051 record_gp_stall_check_time(rsp);
1052 raw_spin_unlock_irq(&rnp->lock);
1054 /* Exclude any concurrent CPU-hotplug operations. */
1058 * Set the quiescent-state-needed bits in all the rcu_node
1059 * structures for all currently online CPUs in breadth-first order,
1060 * starting from the root rcu_node structure, relying on the layout
1061 * of the tree within the rsp->node[] array. Note that other CPUs
1062 * will access only the leaves of the hierarchy, thus seeing that no
1063 * grace period is in progress, at least until the corresponding
1064 * leaf node has been initialized. In addition, we have excluded
1065 * CPU-hotplug operations.
1067 * The grace period cannot complete until the initialization
1068 * process finishes, because this kthread handles both.
1070 rcu_for_each_node_breadth_first(rsp, rnp) {
1071 raw_spin_lock_irq(&rnp->lock);
1072 rdp = this_cpu_ptr(rsp->rda);
1073 rcu_preempt_check_blocked_tasks(rnp);
1074 rnp->qsmask = rnp->qsmaskinit;
1075 rnp->gpnum = rsp->gpnum;
1076 rnp->completed = rsp->completed;
1077 if (rnp == rdp->mynode)
1078 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1079 rcu_preempt_boost_start_gp(rnp);
1080 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1081 rnp->level, rnp->grplo,
1082 rnp->grphi, rnp->qsmask);
1083 raw_spin_unlock_irq(&rnp->lock);
1092 * Do one round of quiescent-state forcing.
1094 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1096 int fqs_state = fqs_state_in;
1097 struct rcu_node *rnp = rcu_get_root(rsp);
1100 if (fqs_state == RCU_SAVE_DYNTICK) {
1101 /* Collect dyntick-idle snapshots. */
1102 force_qs_rnp(rsp, dyntick_save_progress_counter);
1103 fqs_state = RCU_FORCE_QS;
1105 /* Handle dyntick-idle and offline CPUs. */
1106 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1108 /* Clear flag to prevent immediate re-entry. */
1109 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1110 raw_spin_lock_irq(&rnp->lock);
1111 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1112 raw_spin_unlock_irq(&rnp->lock);
1118 * Clean up after the old grace period.
1120 static void rcu_gp_cleanup(struct rcu_state *rsp)
1122 unsigned long gp_duration;
1123 struct rcu_data *rdp;
1124 struct rcu_node *rnp = rcu_get_root(rsp);
1126 raw_spin_lock_irq(&rnp->lock);
1127 gp_duration = jiffies - rsp->gp_start;
1128 if (gp_duration > rsp->gp_max)
1129 rsp->gp_max = gp_duration;
1132 * We know the grace period is complete, but to everyone else
1133 * it appears to still be ongoing. But it is also the case
1134 * that to everyone else it looks like there is nothing that
1135 * they can do to advance the grace period. It is therefore
1136 * safe for us to drop the lock in order to mark the grace
1137 * period as completed in all of the rcu_node structures.
1139 * But if this CPU needs another grace period, it will take
1140 * care of this while initializing the next grace period.
1141 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1142 * because the callbacks have not yet been advanced: Those
1143 * callbacks are waiting on the grace period that just now
1146 rdp = this_cpu_ptr(rsp->rda);
1147 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1148 raw_spin_unlock_irq(&rnp->lock);
1151 * Propagate new ->completed value to rcu_node
1152 * structures so that other CPUs don't have to
1153 * wait until the start of the next grace period
1154 * to process their callbacks.
1156 rcu_for_each_node_breadth_first(rsp, rnp) {
1157 raw_spin_lock_irq(&rnp->lock);
1158 rnp->completed = rsp->gpnum;
1159 raw_spin_unlock_irq(&rnp->lock);
1162 rnp = rcu_get_root(rsp);
1163 raw_spin_lock_irq(&rnp->lock);
1166 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1167 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1168 rsp->fqs_state = RCU_GP_IDLE;
1169 if (cpu_needs_another_gp(rsp, rdp))
1171 raw_spin_unlock_irq(&rnp->lock);
1175 * Body of kthread that handles grace periods.
1177 static int __noreturn rcu_gp_kthread(void *arg)
1181 struct rcu_state *rsp = arg;
1182 struct rcu_node *rnp = rcu_get_root(rsp);
1186 /* Handle grace-period start. */
1188 wait_event_interruptible(rsp->gp_wq,
1191 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1195 flush_signals(current);
1198 /* Handle quiescent-state forcing. */
1199 fqs_state = RCU_SAVE_DYNTICK;
1201 rsp->jiffies_force_qs = jiffies +
1202 RCU_JIFFIES_TILL_FORCE_QS;
1203 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1204 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1205 (!ACCESS_ONCE(rnp->qsmask) &&
1206 !rcu_preempt_blocked_readers_cgp(rnp)),
1207 RCU_JIFFIES_TILL_FORCE_QS);
1208 /* If grace period done, leave loop. */
1209 if (!ACCESS_ONCE(rnp->qsmask) &&
1210 !rcu_preempt_blocked_readers_cgp(rnp))
1212 /* If time for quiescent-state forcing, do it. */
1213 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1214 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1217 /* Deal with stray signal. */
1219 flush_signals(current);
1223 /* Handle grace-period end. */
1224 rcu_gp_cleanup(rsp);
1229 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1230 * in preparation for detecting the next grace period. The caller must hold
1231 * the root node's ->lock, which is released before return. Hard irqs must
1234 * Note that it is legal for a dying CPU (which is marked as offline) to
1235 * invoke this function. This can happen when the dying CPU reports its
1239 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1240 __releases(rcu_get_root(rsp)->lock)
1242 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1243 struct rcu_node *rnp = rcu_get_root(rsp);
1245 if (!rsp->gp_kthread ||
1246 !cpu_needs_another_gp(rsp, rdp)) {
1248 * Either we have not yet spawned the grace-period
1249 * task or this CPU does not need another grace period.
1250 * Either way, don't start a new grace period.
1252 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1256 rsp->gp_flags = RCU_GP_FLAG_INIT;
1257 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1258 wake_up(&rsp->gp_wq);
1262 * Report a full set of quiescent states to the specified rcu_state
1263 * data structure. This involves cleaning up after the prior grace
1264 * period and letting rcu_start_gp() start up the next grace period
1265 * if one is needed. Note that the caller must hold rnp->lock, as
1266 * required by rcu_start_gp(), which will release it.
1268 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1269 __releases(rcu_get_root(rsp)->lock)
1271 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1272 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1273 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1277 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1278 * Allows quiescent states for a group of CPUs to be reported at one go
1279 * to the specified rcu_node structure, though all the CPUs in the group
1280 * must be represented by the same rcu_node structure (which need not be
1281 * a leaf rcu_node structure, though it often will be). That structure's
1282 * lock must be held upon entry, and it is released before return.
1285 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1286 struct rcu_node *rnp, unsigned long flags)
1287 __releases(rnp->lock)
1289 struct rcu_node *rnp_c;
1291 /* Walk up the rcu_node hierarchy. */
1293 if (!(rnp->qsmask & mask)) {
1295 /* Our bit has already been cleared, so done. */
1296 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1299 rnp->qsmask &= ~mask;
1300 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1301 mask, rnp->qsmask, rnp->level,
1302 rnp->grplo, rnp->grphi,
1304 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1306 /* Other bits still set at this level, so done. */
1307 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1310 mask = rnp->grpmask;
1311 if (rnp->parent == NULL) {
1313 /* No more levels. Exit loop holding root lock. */
1317 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1320 raw_spin_lock_irqsave(&rnp->lock, flags);
1321 WARN_ON_ONCE(rnp_c->qsmask);
1325 * Get here if we are the last CPU to pass through a quiescent
1326 * state for this grace period. Invoke rcu_report_qs_rsp()
1327 * to clean up and start the next grace period if one is needed.
1329 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1333 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1334 * structure. This must be either called from the specified CPU, or
1335 * called when the specified CPU is known to be offline (and when it is
1336 * also known that no other CPU is concurrently trying to help the offline
1337 * CPU). The lastcomp argument is used to make sure we are still in the
1338 * grace period of interest. We don't want to end the current grace period
1339 * based on quiescent states detected in an earlier grace period!
1342 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1344 unsigned long flags;
1346 struct rcu_node *rnp;
1349 raw_spin_lock_irqsave(&rnp->lock, flags);
1350 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1353 * The grace period in which this quiescent state was
1354 * recorded has ended, so don't report it upwards.
1355 * We will instead need a new quiescent state that lies
1356 * within the current grace period.
1358 rdp->passed_quiesce = 0; /* need qs for new gp. */
1359 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1362 mask = rdp->grpmask;
1363 if ((rnp->qsmask & mask) == 0) {
1364 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1366 rdp->qs_pending = 0;
1369 * This GP can't end until cpu checks in, so all of our
1370 * callbacks can be processed during the next GP.
1372 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1374 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1379 * Check to see if there is a new grace period of which this CPU
1380 * is not yet aware, and if so, set up local rcu_data state for it.
1381 * Otherwise, see if this CPU has just passed through its first
1382 * quiescent state for this grace period, and record that fact if so.
1385 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1387 /* If there is now a new grace period, record and return. */
1388 if (check_for_new_grace_period(rsp, rdp))
1392 * Does this CPU still need to do its part for current grace period?
1393 * If no, return and let the other CPUs do their part as well.
1395 if (!rdp->qs_pending)
1399 * Was there a quiescent state since the beginning of the grace
1400 * period? If no, then exit and wait for the next call.
1402 if (!rdp->passed_quiesce)
1406 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1409 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1412 #ifdef CONFIG_HOTPLUG_CPU
1415 * Send the specified CPU's RCU callbacks to the orphanage. The
1416 * specified CPU must be offline, and the caller must hold the
1420 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1421 struct rcu_node *rnp, struct rcu_data *rdp)
1424 * Orphan the callbacks. First adjust the counts. This is safe
1425 * because ->onofflock excludes _rcu_barrier()'s adoption of
1426 * the callbacks, thus no memory barrier is required.
1428 if (rdp->nxtlist != NULL) {
1429 rsp->qlen_lazy += rdp->qlen_lazy;
1430 rsp->qlen += rdp->qlen;
1431 rdp->n_cbs_orphaned += rdp->qlen;
1433 ACCESS_ONCE(rdp->qlen) = 0;
1437 * Next, move those callbacks still needing a grace period to
1438 * the orphanage, where some other CPU will pick them up.
1439 * Some of the callbacks might have gone partway through a grace
1440 * period, but that is too bad. They get to start over because we
1441 * cannot assume that grace periods are synchronized across CPUs.
1442 * We don't bother updating the ->nxttail[] array yet, instead
1443 * we just reset the whole thing later on.
1445 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1446 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1447 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1448 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1452 * Then move the ready-to-invoke callbacks to the orphanage,
1453 * where some other CPU will pick them up. These will not be
1454 * required to pass though another grace period: They are done.
1456 if (rdp->nxtlist != NULL) {
1457 *rsp->orphan_donetail = rdp->nxtlist;
1458 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1461 /* Finally, initialize the rcu_data structure's list to empty. */
1462 init_callback_list(rdp);
1466 * Adopt the RCU callbacks from the specified rcu_state structure's
1467 * orphanage. The caller must hold the ->onofflock.
1469 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1472 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1475 * If there is an rcu_barrier() operation in progress, then
1476 * only the task doing that operation is permitted to adopt
1477 * callbacks. To do otherwise breaks rcu_barrier() and friends
1478 * by causing them to fail to wait for the callbacks in the
1481 if (rsp->rcu_barrier_in_progress &&
1482 rsp->rcu_barrier_in_progress != current)
1485 /* Do the accounting first. */
1486 rdp->qlen_lazy += rsp->qlen_lazy;
1487 rdp->qlen += rsp->qlen;
1488 rdp->n_cbs_adopted += rsp->qlen;
1489 if (rsp->qlen_lazy != rsp->qlen)
1490 rcu_idle_count_callbacks_posted();
1495 * We do not need a memory barrier here because the only way we
1496 * can get here if there is an rcu_barrier() in flight is if
1497 * we are the task doing the rcu_barrier().
1500 /* First adopt the ready-to-invoke callbacks. */
1501 if (rsp->orphan_donelist != NULL) {
1502 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1503 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1504 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1505 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1506 rdp->nxttail[i] = rsp->orphan_donetail;
1507 rsp->orphan_donelist = NULL;
1508 rsp->orphan_donetail = &rsp->orphan_donelist;
1511 /* And then adopt the callbacks that still need a grace period. */
1512 if (rsp->orphan_nxtlist != NULL) {
1513 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1514 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1515 rsp->orphan_nxtlist = NULL;
1516 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1521 * Trace the fact that this CPU is going offline.
1523 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1525 RCU_TRACE(unsigned long mask);
1526 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1527 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1529 RCU_TRACE(mask = rdp->grpmask);
1530 trace_rcu_grace_period(rsp->name,
1531 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1536 * The CPU has been completely removed, and some other CPU is reporting
1537 * this fact from process context. Do the remainder of the cleanup,
1538 * including orphaning the outgoing CPU's RCU callbacks, and also
1539 * adopting them, if there is no _rcu_barrier() instance running.
1540 * There can only be one CPU hotplug operation at a time, so no other
1541 * CPU can be attempting to update rcu_cpu_kthread_task.
1543 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1545 unsigned long flags;
1547 int need_report = 0;
1548 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1549 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1551 /* Adjust any no-longer-needed kthreads. */
1552 rcu_stop_cpu_kthread(cpu);
1553 rcu_node_kthread_setaffinity(rnp, -1);
1555 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1557 /* Exclude any attempts to start a new grace period. */
1558 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1560 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1561 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1562 rcu_adopt_orphan_cbs(rsp);
1564 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1565 mask = rdp->grpmask; /* rnp->grplo is constant. */
1567 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1568 rnp->qsmaskinit &= ~mask;
1569 if (rnp->qsmaskinit != 0) {
1570 if (rnp != rdp->mynode)
1571 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1574 if (rnp == rdp->mynode)
1575 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1577 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1578 mask = rnp->grpmask;
1580 } while (rnp != NULL);
1583 * We still hold the leaf rcu_node structure lock here, and
1584 * irqs are still disabled. The reason for this subterfuge is
1585 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1586 * held leads to deadlock.
1588 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1590 if (need_report & RCU_OFL_TASKS_NORM_GP)
1591 rcu_report_unblock_qs_rnp(rnp, flags);
1593 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1594 if (need_report & RCU_OFL_TASKS_EXP_GP)
1595 rcu_report_exp_rnp(rsp, rnp, true);
1596 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1597 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1598 cpu, rdp->qlen, rdp->nxtlist);
1601 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1603 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1607 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1611 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1615 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1618 * Invoke any RCU callbacks that have made it to the end of their grace
1619 * period. Thottle as specified by rdp->blimit.
1621 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1623 unsigned long flags;
1624 struct rcu_head *next, *list, **tail;
1625 int bl, count, count_lazy, i;
1627 /* If no callbacks are ready, just return.*/
1628 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1629 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1630 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1631 need_resched(), is_idle_task(current),
1632 rcu_is_callbacks_kthread());
1637 * Extract the list of ready callbacks, disabling to prevent
1638 * races with call_rcu() from interrupt handlers.
1640 local_irq_save(flags);
1641 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1643 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1644 list = rdp->nxtlist;
1645 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1646 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1647 tail = rdp->nxttail[RCU_DONE_TAIL];
1648 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1649 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1650 rdp->nxttail[i] = &rdp->nxtlist;
1651 local_irq_restore(flags);
1653 /* Invoke callbacks. */
1654 count = count_lazy = 0;
1658 debug_rcu_head_unqueue(list);
1659 if (__rcu_reclaim(rsp->name, list))
1662 /* Stop only if limit reached and CPU has something to do. */
1663 if (++count >= bl &&
1665 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1669 local_irq_save(flags);
1670 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1671 is_idle_task(current),
1672 rcu_is_callbacks_kthread());
1674 /* Update count, and requeue any remaining callbacks. */
1676 *tail = rdp->nxtlist;
1677 rdp->nxtlist = list;
1678 for (i = 0; i < RCU_NEXT_SIZE; i++)
1679 if (&rdp->nxtlist == rdp->nxttail[i])
1680 rdp->nxttail[i] = tail;
1684 smp_mb(); /* List handling before counting for rcu_barrier(). */
1685 rdp->qlen_lazy -= count_lazy;
1686 ACCESS_ONCE(rdp->qlen) -= count;
1687 rdp->n_cbs_invoked += count;
1689 /* Reinstate batch limit if we have worked down the excess. */
1690 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1691 rdp->blimit = blimit;
1693 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1694 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1695 rdp->qlen_last_fqs_check = 0;
1696 rdp->n_force_qs_snap = rsp->n_force_qs;
1697 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1698 rdp->qlen_last_fqs_check = rdp->qlen;
1699 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1701 local_irq_restore(flags);
1703 /* Re-invoke RCU core processing if there are callbacks remaining. */
1704 if (cpu_has_callbacks_ready_to_invoke(rdp))
1709 * Check to see if this CPU is in a non-context-switch quiescent state
1710 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1711 * Also schedule RCU core processing.
1713 * This function must be called from hardirq context. It is normally
1714 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1715 * false, there is no point in invoking rcu_check_callbacks().
1717 void rcu_check_callbacks(int cpu, int user)
1719 trace_rcu_utilization("Start scheduler-tick");
1720 increment_cpu_stall_ticks();
1721 if (user || rcu_is_cpu_rrupt_from_idle()) {
1724 * Get here if this CPU took its interrupt from user
1725 * mode or from the idle loop, and if this is not a
1726 * nested interrupt. In this case, the CPU is in
1727 * a quiescent state, so note it.
1729 * No memory barrier is required here because both
1730 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1731 * variables that other CPUs neither access nor modify,
1732 * at least not while the corresponding CPU is online.
1738 } else if (!in_softirq()) {
1741 * Get here if this CPU did not take its interrupt from
1742 * softirq, in other words, if it is not interrupting
1743 * a rcu_bh read-side critical section. This is an _bh
1744 * critical section, so note it.
1749 rcu_preempt_check_callbacks(cpu);
1750 if (rcu_pending(cpu))
1752 trace_rcu_utilization("End scheduler-tick");
1756 * Scan the leaf rcu_node structures, processing dyntick state for any that
1757 * have not yet encountered a quiescent state, using the function specified.
1758 * Also initiate boosting for any threads blocked on the root rcu_node.
1760 * The caller must have suppressed start of new grace periods.
1762 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1766 unsigned long flags;
1768 struct rcu_node *rnp;
1770 rcu_for_each_leaf_node(rsp, rnp) {
1773 raw_spin_lock_irqsave(&rnp->lock, flags);
1774 if (!rcu_gp_in_progress(rsp)) {
1775 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1778 if (rnp->qsmask == 0) {
1779 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1784 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1785 if ((rnp->qsmask & bit) != 0 &&
1786 f(per_cpu_ptr(rsp->rda, cpu)))
1791 /* rcu_report_qs_rnp() releases rnp->lock. */
1792 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1795 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1797 rnp = rcu_get_root(rsp);
1798 if (rnp->qsmask == 0) {
1799 raw_spin_lock_irqsave(&rnp->lock, flags);
1800 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1805 * Force quiescent states on reluctant CPUs, and also detect which
1806 * CPUs are in dyntick-idle mode.
1808 static void force_quiescent_state(struct rcu_state *rsp)
1810 unsigned long flags;
1812 struct rcu_node *rnp;
1813 struct rcu_node *rnp_old = NULL;
1815 /* Funnel through hierarchy to reduce memory contention. */
1816 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1817 for (; rnp != NULL; rnp = rnp->parent) {
1818 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1819 !raw_spin_trylock(&rnp->fqslock);
1820 if (rnp_old != NULL)
1821 raw_spin_unlock(&rnp_old->fqslock);
1823 rsp->n_force_qs_lh++;
1828 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1830 /* Reached the root of the rcu_node tree, acquire lock. */
1831 raw_spin_lock_irqsave(&rnp_old->lock, flags);
1832 raw_spin_unlock(&rnp_old->fqslock);
1833 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1834 rsp->n_force_qs_lh++;
1835 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1836 return; /* Someone beat us to it. */
1838 rsp->gp_flags |= RCU_GP_FLAG_FQS;
1839 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1840 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1844 * This does the RCU core processing work for the specified rcu_state
1845 * and rcu_data structures. This may be called only from the CPU to
1846 * whom the rdp belongs.
1849 __rcu_process_callbacks(struct rcu_state *rsp)
1851 unsigned long flags;
1852 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1854 WARN_ON_ONCE(rdp->beenonline == 0);
1857 * Advance callbacks in response to end of earlier grace
1858 * period that some other CPU ended.
1860 rcu_process_gp_end(rsp, rdp);
1862 /* Update RCU state based on any recent quiescent states. */
1863 rcu_check_quiescent_state(rsp, rdp);
1865 /* Does this CPU require a not-yet-started grace period? */
1866 if (cpu_needs_another_gp(rsp, rdp)) {
1867 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1868 rcu_start_gp(rsp, flags); /* releases above lock */
1871 /* If there are callbacks ready, invoke them. */
1872 if (cpu_has_callbacks_ready_to_invoke(rdp))
1873 invoke_rcu_callbacks(rsp, rdp);
1877 * Do RCU core processing for the current CPU.
1879 static void rcu_process_callbacks(struct softirq_action *unused)
1881 struct rcu_state *rsp;
1883 if (cpu_is_offline(smp_processor_id()))
1885 trace_rcu_utilization("Start RCU core");
1886 for_each_rcu_flavor(rsp)
1887 __rcu_process_callbacks(rsp);
1888 trace_rcu_utilization("End RCU core");
1892 * Schedule RCU callback invocation. If the specified type of RCU
1893 * does not support RCU priority boosting, just do a direct call,
1894 * otherwise wake up the per-CPU kernel kthread. Note that because we
1895 * are running on the current CPU with interrupts disabled, the
1896 * rcu_cpu_kthread_task cannot disappear out from under us.
1898 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1900 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1902 if (likely(!rsp->boost)) {
1903 rcu_do_batch(rsp, rdp);
1906 invoke_rcu_callbacks_kthread();
1909 static void invoke_rcu_core(void)
1911 raise_softirq(RCU_SOFTIRQ);
1915 * Handle any core-RCU processing required by a call_rcu() invocation.
1917 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1918 struct rcu_head *head, unsigned long flags)
1921 * If called from an extended quiescent state, invoke the RCU
1922 * core in order to force a re-evaluation of RCU's idleness.
1924 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1927 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1928 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1932 * Force the grace period if too many callbacks or too long waiting.
1933 * Enforce hysteresis, and don't invoke force_quiescent_state()
1934 * if some other CPU has recently done so. Also, don't bother
1935 * invoking force_quiescent_state() if the newly enqueued callback
1936 * is the only one waiting for a grace period to complete.
1938 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1940 /* Are we ignoring a completed grace period? */
1941 rcu_process_gp_end(rsp, rdp);
1942 check_for_new_grace_period(rsp, rdp);
1944 /* Start a new grace period if one not already started. */
1945 if (!rcu_gp_in_progress(rsp)) {
1946 unsigned long nestflag;
1947 struct rcu_node *rnp_root = rcu_get_root(rsp);
1949 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1950 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1952 /* Give the grace period a kick. */
1953 rdp->blimit = LONG_MAX;
1954 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1955 *rdp->nxttail[RCU_DONE_TAIL] != head)
1956 force_quiescent_state(rsp);
1957 rdp->n_force_qs_snap = rsp->n_force_qs;
1958 rdp->qlen_last_fqs_check = rdp->qlen;
1964 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1965 struct rcu_state *rsp, bool lazy)
1967 unsigned long flags;
1968 struct rcu_data *rdp;
1970 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1971 debug_rcu_head_queue(head);
1975 smp_mb(); /* Ensure RCU update seen before callback registry. */
1978 * Opportunistically note grace-period endings and beginnings.
1979 * Note that we might see a beginning right after we see an
1980 * end, but never vice versa, since this CPU has to pass through
1981 * a quiescent state betweentimes.
1983 local_irq_save(flags);
1984 rdp = this_cpu_ptr(rsp->rda);
1986 /* Add the callback to our list. */
1987 ACCESS_ONCE(rdp->qlen)++;
1991 rcu_idle_count_callbacks_posted();
1992 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1993 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1994 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1996 if (__is_kfree_rcu_offset((unsigned long)func))
1997 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1998 rdp->qlen_lazy, rdp->qlen);
2000 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2002 /* Go handle any RCU core processing required. */
2003 __call_rcu_core(rsp, rdp, head, flags);
2004 local_irq_restore(flags);
2008 * Queue an RCU-sched callback for invocation after a grace period.
2010 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2012 __call_rcu(head, func, &rcu_sched_state, 0);
2014 EXPORT_SYMBOL_GPL(call_rcu_sched);
2017 * Queue an RCU callback for invocation after a quicker grace period.
2019 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2021 __call_rcu(head, func, &rcu_bh_state, 0);
2023 EXPORT_SYMBOL_GPL(call_rcu_bh);
2026 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2027 * any blocking grace-period wait automatically implies a grace period
2028 * if there is only one CPU online at any point time during execution
2029 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2030 * occasionally incorrectly indicate that there are multiple CPUs online
2031 * when there was in fact only one the whole time, as this just adds
2032 * some overhead: RCU still operates correctly.
2034 static inline int rcu_blocking_is_gp(void)
2038 might_sleep(); /* Check for RCU read-side critical section. */
2040 ret = num_online_cpus() <= 1;
2046 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2048 * Control will return to the caller some time after a full rcu-sched
2049 * grace period has elapsed, in other words after all currently executing
2050 * rcu-sched read-side critical sections have completed. These read-side
2051 * critical sections are delimited by rcu_read_lock_sched() and
2052 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2053 * local_irq_disable(), and so on may be used in place of
2054 * rcu_read_lock_sched().
2056 * This means that all preempt_disable code sequences, including NMI and
2057 * hardware-interrupt handlers, in progress on entry will have completed
2058 * before this primitive returns. However, this does not guarantee that
2059 * softirq handlers will have completed, since in some kernels, these
2060 * handlers can run in process context, and can block.
2062 * This primitive provides the guarantees made by the (now removed)
2063 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2064 * guarantees that rcu_read_lock() sections will have completed.
2065 * In "classic RCU", these two guarantees happen to be one and
2066 * the same, but can differ in realtime RCU implementations.
2068 void synchronize_sched(void)
2070 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2071 !lock_is_held(&rcu_lock_map) &&
2072 !lock_is_held(&rcu_sched_lock_map),
2073 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2074 if (rcu_blocking_is_gp())
2076 wait_rcu_gp(call_rcu_sched);
2078 EXPORT_SYMBOL_GPL(synchronize_sched);
2081 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2083 * Control will return to the caller some time after a full rcu_bh grace
2084 * period has elapsed, in other words after all currently executing rcu_bh
2085 * read-side critical sections have completed. RCU read-side critical
2086 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2087 * and may be nested.
2089 void synchronize_rcu_bh(void)
2091 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2092 !lock_is_held(&rcu_lock_map) &&
2093 !lock_is_held(&rcu_sched_lock_map),
2094 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2095 if (rcu_blocking_is_gp())
2097 wait_rcu_gp(call_rcu_bh);
2099 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2101 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2102 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2104 static int synchronize_sched_expedited_cpu_stop(void *data)
2107 * There must be a full memory barrier on each affected CPU
2108 * between the time that try_stop_cpus() is called and the
2109 * time that it returns.
2111 * In the current initial implementation of cpu_stop, the
2112 * above condition is already met when the control reaches
2113 * this point and the following smp_mb() is not strictly
2114 * necessary. Do smp_mb() anyway for documentation and
2115 * robustness against future implementation changes.
2117 smp_mb(); /* See above comment block. */
2122 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2124 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2125 * approach to force the grace period to end quickly. This consumes
2126 * significant time on all CPUs and is unfriendly to real-time workloads,
2127 * so is thus not recommended for any sort of common-case code. In fact,
2128 * if you are using synchronize_sched_expedited() in a loop, please
2129 * restructure your code to batch your updates, and then use a single
2130 * synchronize_sched() instead.
2132 * Note that it is illegal to call this function while holding any lock
2133 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2134 * to call this function from a CPU-hotplug notifier. Failing to observe
2135 * these restriction will result in deadlock.
2137 * This implementation can be thought of as an application of ticket
2138 * locking to RCU, with sync_sched_expedited_started and
2139 * sync_sched_expedited_done taking on the roles of the halves
2140 * of the ticket-lock word. Each task atomically increments
2141 * sync_sched_expedited_started upon entry, snapshotting the old value,
2142 * then attempts to stop all the CPUs. If this succeeds, then each
2143 * CPU will have executed a context switch, resulting in an RCU-sched
2144 * grace period. We are then done, so we use atomic_cmpxchg() to
2145 * update sync_sched_expedited_done to match our snapshot -- but
2146 * only if someone else has not already advanced past our snapshot.
2148 * On the other hand, if try_stop_cpus() fails, we check the value
2149 * of sync_sched_expedited_done. If it has advanced past our
2150 * initial snapshot, then someone else must have forced a grace period
2151 * some time after we took our snapshot. In this case, our work is
2152 * done for us, and we can simply return. Otherwise, we try again,
2153 * but keep our initial snapshot for purposes of checking for someone
2154 * doing our work for us.
2156 * If we fail too many times in a row, we fall back to synchronize_sched().
2158 void synchronize_sched_expedited(void)
2160 int firstsnap, s, snap, trycount = 0;
2162 /* Note that atomic_inc_return() implies full memory barrier. */
2163 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2165 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2168 * Each pass through the following loop attempts to force a
2169 * context switch on each CPU.
2171 while (try_stop_cpus(cpu_online_mask,
2172 synchronize_sched_expedited_cpu_stop,
2176 /* No joy, try again later. Or just synchronize_sched(). */
2177 if (trycount++ < 10) {
2178 udelay(trycount * num_online_cpus());
2180 synchronize_sched();
2184 /* Check to see if someone else did our work for us. */
2185 s = atomic_read(&sync_sched_expedited_done);
2186 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2187 smp_mb(); /* ensure test happens before caller kfree */
2192 * Refetching sync_sched_expedited_started allows later
2193 * callers to piggyback on our grace period. We subtract
2194 * 1 to get the same token that the last incrementer got.
2195 * We retry after they started, so our grace period works
2196 * for them, and they started after our first try, so their
2197 * grace period works for us.
2200 snap = atomic_read(&sync_sched_expedited_started);
2201 smp_mb(); /* ensure read is before try_stop_cpus(). */
2205 * Everyone up to our most recent fetch is covered by our grace
2206 * period. Update the counter, but only if our work is still
2207 * relevant -- which it won't be if someone who started later
2208 * than we did beat us to the punch.
2211 s = atomic_read(&sync_sched_expedited_done);
2212 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2213 smp_mb(); /* ensure test happens before caller kfree */
2216 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2220 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2223 * Check to see if there is any immediate RCU-related work to be done
2224 * by the current CPU, for the specified type of RCU, returning 1 if so.
2225 * The checks are in order of increasing expense: checks that can be
2226 * carried out against CPU-local state are performed first. However,
2227 * we must check for CPU stalls first, else we might not get a chance.
2229 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2231 struct rcu_node *rnp = rdp->mynode;
2233 rdp->n_rcu_pending++;
2235 /* Check for CPU stalls, if enabled. */
2236 check_cpu_stall(rsp, rdp);
2238 /* Is the RCU core waiting for a quiescent state from this CPU? */
2239 if (rcu_scheduler_fully_active &&
2240 rdp->qs_pending && !rdp->passed_quiesce) {
2241 rdp->n_rp_qs_pending++;
2242 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2243 rdp->n_rp_report_qs++;
2247 /* Does this CPU have callbacks ready to invoke? */
2248 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2249 rdp->n_rp_cb_ready++;
2253 /* Has RCU gone idle with this CPU needing another grace period? */
2254 if (cpu_needs_another_gp(rsp, rdp)) {
2255 rdp->n_rp_cpu_needs_gp++;
2259 /* Has another RCU grace period completed? */
2260 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2261 rdp->n_rp_gp_completed++;
2265 /* Has a new RCU grace period started? */
2266 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2267 rdp->n_rp_gp_started++;
2272 rdp->n_rp_need_nothing++;
2277 * Check to see if there is any immediate RCU-related work to be done
2278 * by the current CPU, returning 1 if so. This function is part of the
2279 * RCU implementation; it is -not- an exported member of the RCU API.
2281 static int rcu_pending(int cpu)
2283 struct rcu_state *rsp;
2285 for_each_rcu_flavor(rsp)
2286 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2292 * Check to see if any future RCU-related work will need to be done
2293 * by the current CPU, even if none need be done immediately, returning
2296 static int rcu_cpu_has_callbacks(int cpu)
2298 struct rcu_state *rsp;
2300 /* RCU callbacks either ready or pending? */
2301 for_each_rcu_flavor(rsp)
2302 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2308 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2309 * the compiler is expected to optimize this away.
2311 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2312 int cpu, unsigned long done)
2314 trace_rcu_barrier(rsp->name, s, cpu,
2315 atomic_read(&rsp->barrier_cpu_count), done);
2319 * RCU callback function for _rcu_barrier(). If we are last, wake
2320 * up the task executing _rcu_barrier().
2322 static void rcu_barrier_callback(struct rcu_head *rhp)
2324 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2325 struct rcu_state *rsp = rdp->rsp;
2327 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2328 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2329 complete(&rsp->barrier_completion);
2331 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2336 * Called with preemption disabled, and from cross-cpu IRQ context.
2338 static void rcu_barrier_func(void *type)
2340 struct rcu_state *rsp = type;
2341 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2343 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2344 atomic_inc(&rsp->barrier_cpu_count);
2345 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2349 * Orchestrate the specified type of RCU barrier, waiting for all
2350 * RCU callbacks of the specified type to complete.
2352 static void _rcu_barrier(struct rcu_state *rsp)
2355 unsigned long flags;
2356 struct rcu_data *rdp;
2358 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2359 unsigned long snap_done;
2361 init_rcu_head_on_stack(&rd.barrier_head);
2362 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2364 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2365 mutex_lock(&rsp->barrier_mutex);
2368 * Ensure that all prior references, including to ->n_barrier_done,
2369 * are ordered before the _rcu_barrier() machinery.
2371 smp_mb(); /* See above block comment. */
2374 * Recheck ->n_barrier_done to see if others did our work for us.
2375 * This means checking ->n_barrier_done for an even-to-odd-to-even
2376 * transition. The "if" expression below therefore rounds the old
2377 * value up to the next even number and adds two before comparing.
2379 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2380 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2381 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2382 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2383 smp_mb(); /* caller's subsequent code after above check. */
2384 mutex_unlock(&rsp->barrier_mutex);
2389 * Increment ->n_barrier_done to avoid duplicate work. Use
2390 * ACCESS_ONCE() to prevent the compiler from speculating
2391 * the increment to precede the early-exit check.
2393 ACCESS_ONCE(rsp->n_barrier_done)++;
2394 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2395 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2396 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2399 * Initialize the count to one rather than to zero in order to
2400 * avoid a too-soon return to zero in case of a short grace period
2401 * (or preemption of this task). Also flag this task as doing
2402 * an rcu_barrier(). This will prevent anyone else from adopting
2403 * orphaned callbacks, which could cause otherwise failure if a
2404 * CPU went offline and quickly came back online. To see this,
2405 * consider the following sequence of events:
2407 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2408 * 2. CPU 1 goes offline, orphaning its callbacks.
2409 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2410 * 4. CPU 1 comes back online.
2411 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2412 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2413 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2415 init_completion(&rsp->barrier_completion);
2416 atomic_set(&rsp->barrier_cpu_count, 1);
2417 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2418 rsp->rcu_barrier_in_progress = current;
2419 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2422 * Force every CPU with callbacks to register a new callback
2423 * that will tell us when all the preceding callbacks have
2424 * been invoked. If an offline CPU has callbacks, wait for
2425 * it to either come back online or to finish orphaning those
2428 for_each_possible_cpu(cpu) {
2430 rdp = per_cpu_ptr(rsp->rda, cpu);
2431 if (cpu_is_offline(cpu)) {
2432 _rcu_barrier_trace(rsp, "Offline", cpu,
2433 rsp->n_barrier_done);
2435 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2436 schedule_timeout_interruptible(1);
2437 } else if (ACCESS_ONCE(rdp->qlen)) {
2438 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2439 rsp->n_barrier_done);
2440 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2443 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2444 rsp->n_barrier_done);
2450 * Now that all online CPUs have rcu_barrier_callback() callbacks
2451 * posted, we can adopt all of the orphaned callbacks and place
2452 * an rcu_barrier_callback() callback after them. When that is done,
2453 * we are guaranteed to have an rcu_barrier_callback() callback
2454 * following every callback that could possibly have been
2455 * registered before _rcu_barrier() was called.
2457 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2458 rcu_adopt_orphan_cbs(rsp);
2459 rsp->rcu_barrier_in_progress = NULL;
2460 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2461 atomic_inc(&rsp->barrier_cpu_count);
2462 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2464 rsp->call(&rd.barrier_head, rcu_barrier_callback);
2467 * Now that we have an rcu_barrier_callback() callback on each
2468 * CPU, and thus each counted, remove the initial count.
2470 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2471 complete(&rsp->barrier_completion);
2473 /* Increment ->n_barrier_done to prevent duplicate work. */
2474 smp_mb(); /* Keep increment after above mechanism. */
2475 ACCESS_ONCE(rsp->n_barrier_done)++;
2476 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2477 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2478 smp_mb(); /* Keep increment before caller's subsequent code. */
2480 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2481 wait_for_completion(&rsp->barrier_completion);
2483 /* Other rcu_barrier() invocations can now safely proceed. */
2484 mutex_unlock(&rsp->barrier_mutex);
2486 destroy_rcu_head_on_stack(&rd.barrier_head);
2490 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2492 void rcu_barrier_bh(void)
2494 _rcu_barrier(&rcu_bh_state);
2496 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2499 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2501 void rcu_barrier_sched(void)
2503 _rcu_barrier(&rcu_sched_state);
2505 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2508 * Do boot-time initialization of a CPU's per-CPU RCU data.
2511 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2513 unsigned long flags;
2514 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2515 struct rcu_node *rnp = rcu_get_root(rsp);
2517 /* Set up local state, ensuring consistent view of global state. */
2518 raw_spin_lock_irqsave(&rnp->lock, flags);
2519 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2520 init_callback_list(rdp);
2522 ACCESS_ONCE(rdp->qlen) = 0;
2523 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2524 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2525 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2528 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2532 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2533 * offline event can be happening at a given time. Note also that we
2534 * can accept some slop in the rsp->completed access due to the fact
2535 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2537 static void __cpuinit
2538 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2540 unsigned long flags;
2542 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2543 struct rcu_node *rnp = rcu_get_root(rsp);
2545 /* Set up local state, ensuring consistent view of global state. */
2546 raw_spin_lock_irqsave(&rnp->lock, flags);
2547 rdp->beenonline = 1; /* We have now been online. */
2548 rdp->preemptible = preemptible;
2549 rdp->qlen_last_fqs_check = 0;
2550 rdp->n_force_qs_snap = rsp->n_force_qs;
2551 rdp->blimit = blimit;
2552 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2553 atomic_set(&rdp->dynticks->dynticks,
2554 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2555 rcu_prepare_for_idle_init(cpu);
2556 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2559 * A new grace period might start here. If so, we won't be part
2560 * of it, but that is OK, as we are currently in a quiescent state.
2563 /* Exclude any attempts to start a new GP on large systems. */
2564 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2566 /* Add CPU to rcu_node bitmasks. */
2568 mask = rdp->grpmask;
2570 /* Exclude any attempts to start a new GP on small systems. */
2571 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2572 rnp->qsmaskinit |= mask;
2573 mask = rnp->grpmask;
2574 if (rnp == rdp->mynode) {
2576 * If there is a grace period in progress, we will
2577 * set up to wait for it next time we run the
2580 rdp->gpnum = rnp->completed;
2581 rdp->completed = rnp->completed;
2582 rdp->passed_quiesce = 0;
2583 rdp->qs_pending = 0;
2584 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2585 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2587 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2589 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2591 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2594 static void __cpuinit rcu_prepare_cpu(int cpu)
2596 struct rcu_state *rsp;
2598 for_each_rcu_flavor(rsp)
2599 rcu_init_percpu_data(cpu, rsp,
2600 strcmp(rsp->name, "rcu_preempt") == 0);
2604 * Handle CPU online/offline notification events.
2606 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2607 unsigned long action, void *hcpu)
2609 long cpu = (long)hcpu;
2610 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2611 struct rcu_node *rnp = rdp->mynode;
2612 struct rcu_state *rsp;
2614 trace_rcu_utilization("Start CPU hotplug");
2616 case CPU_UP_PREPARE:
2617 case CPU_UP_PREPARE_FROZEN:
2618 rcu_prepare_cpu(cpu);
2619 rcu_prepare_kthreads(cpu);
2622 case CPU_DOWN_FAILED:
2623 rcu_node_kthread_setaffinity(rnp, -1);
2624 rcu_cpu_kthread_setrt(cpu, 1);
2626 case CPU_DOWN_PREPARE:
2627 rcu_node_kthread_setaffinity(rnp, cpu);
2628 rcu_cpu_kthread_setrt(cpu, 0);
2631 case CPU_DYING_FROZEN:
2633 * The whole machine is "stopped" except this CPU, so we can
2634 * touch any data without introducing corruption. We send the
2635 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2637 for_each_rcu_flavor(rsp)
2638 rcu_cleanup_dying_cpu(rsp);
2639 rcu_cleanup_after_idle(cpu);
2642 case CPU_DEAD_FROZEN:
2643 case CPU_UP_CANCELED:
2644 case CPU_UP_CANCELED_FROZEN:
2645 for_each_rcu_flavor(rsp)
2646 rcu_cleanup_dead_cpu(cpu, rsp);
2651 trace_rcu_utilization("End CPU hotplug");
2656 * Spawn the kthread that handles this RCU flavor's grace periods.
2658 static int __init rcu_spawn_gp_kthread(void)
2660 unsigned long flags;
2661 struct rcu_node *rnp;
2662 struct rcu_state *rsp;
2663 struct task_struct *t;
2665 for_each_rcu_flavor(rsp) {
2666 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2668 rnp = rcu_get_root(rsp);
2669 raw_spin_lock_irqsave(&rnp->lock, flags);
2670 rsp->gp_kthread = t;
2671 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2675 early_initcall(rcu_spawn_gp_kthread);
2678 * This function is invoked towards the end of the scheduler's initialization
2679 * process. Before this is called, the idle task might contain
2680 * RCU read-side critical sections (during which time, this idle
2681 * task is booting the system). After this function is called, the
2682 * idle tasks are prohibited from containing RCU read-side critical
2683 * sections. This function also enables RCU lockdep checking.
2685 void rcu_scheduler_starting(void)
2687 WARN_ON(num_online_cpus() != 1);
2688 WARN_ON(nr_context_switches() > 0);
2689 rcu_scheduler_active = 1;
2693 * Compute the per-level fanout, either using the exact fanout specified
2694 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2696 #ifdef CONFIG_RCU_FANOUT_EXACT
2697 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2701 for (i = rcu_num_lvls - 1; i > 0; i--)
2702 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2703 rsp->levelspread[0] = rcu_fanout_leaf;
2705 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2706 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2713 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2714 ccur = rsp->levelcnt[i];
2715 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2719 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2722 * Helper function for rcu_init() that initializes one rcu_state structure.
2724 static void __init rcu_init_one(struct rcu_state *rsp,
2725 struct rcu_data __percpu *rda)
2727 static char *buf[] = { "rcu_node_0",
2730 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2731 static char *fqs[] = { "rcu_node_fqs_0",
2734 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2738 struct rcu_node *rnp;
2740 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2742 /* Initialize the level-tracking arrays. */
2744 for (i = 0; i < rcu_num_lvls; i++)
2745 rsp->levelcnt[i] = num_rcu_lvl[i];
2746 for (i = 1; i < rcu_num_lvls; i++)
2747 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2748 rcu_init_levelspread(rsp);
2750 /* Initialize the elements themselves, starting from the leaves. */
2752 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2753 cpustride *= rsp->levelspread[i];
2754 rnp = rsp->level[i];
2755 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2756 raw_spin_lock_init(&rnp->lock);
2757 lockdep_set_class_and_name(&rnp->lock,
2758 &rcu_node_class[i], buf[i]);
2759 raw_spin_lock_init(&rnp->fqslock);
2760 lockdep_set_class_and_name(&rnp->fqslock,
2761 &rcu_fqs_class[i], fqs[i]);
2764 rnp->qsmaskinit = 0;
2765 rnp->grplo = j * cpustride;
2766 rnp->grphi = (j + 1) * cpustride - 1;
2767 if (rnp->grphi >= NR_CPUS)
2768 rnp->grphi = NR_CPUS - 1;
2774 rnp->grpnum = j % rsp->levelspread[i - 1];
2775 rnp->grpmask = 1UL << rnp->grpnum;
2776 rnp->parent = rsp->level[i - 1] +
2777 j / rsp->levelspread[i - 1];
2780 INIT_LIST_HEAD(&rnp->blkd_tasks);
2785 init_waitqueue_head(&rsp->gp_wq);
2786 rnp = rsp->level[rcu_num_lvls - 1];
2787 for_each_possible_cpu(i) {
2788 while (i > rnp->grphi)
2790 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2791 rcu_boot_init_percpu_data(i, rsp);
2793 list_add(&rsp->flavors, &rcu_struct_flavors);
2797 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2798 * replace the definitions in rcutree.h because those are needed to size
2799 * the ->node array in the rcu_state structure.
2801 static void __init rcu_init_geometry(void)
2806 int rcu_capacity[MAX_RCU_LVLS + 1];
2808 /* If the compile-time values are accurate, just leave. */
2809 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2813 * Compute number of nodes that can be handled an rcu_node tree
2814 * with the given number of levels. Setting rcu_capacity[0] makes
2815 * some of the arithmetic easier.
2817 rcu_capacity[0] = 1;
2818 rcu_capacity[1] = rcu_fanout_leaf;
2819 for (i = 2; i <= MAX_RCU_LVLS; i++)
2820 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2823 * The boot-time rcu_fanout_leaf parameter is only permitted
2824 * to increase the leaf-level fanout, not decrease it. Of course,
2825 * the leaf-level fanout cannot exceed the number of bits in
2826 * the rcu_node masks. Finally, the tree must be able to accommodate
2827 * the configured number of CPUs. Complain and fall back to the
2828 * compile-time values if these limits are exceeded.
2830 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2831 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2832 n > rcu_capacity[MAX_RCU_LVLS]) {
2837 /* Calculate the number of rcu_nodes at each level of the tree. */
2838 for (i = 1; i <= MAX_RCU_LVLS; i++)
2839 if (n <= rcu_capacity[i]) {
2840 for (j = 0; j <= i; j++)
2842 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2844 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2849 /* Calculate the total number of rcu_node structures. */
2851 for (i = 0; i <= MAX_RCU_LVLS; i++)
2852 rcu_num_nodes += num_rcu_lvl[i];
2856 void __init rcu_init(void)
2860 rcu_bootup_announce();
2861 rcu_init_geometry();
2862 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2863 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2864 __rcu_init_preempt();
2865 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2868 * We don't need protection against CPU-hotplug here because
2869 * this is called early in boot, before either interrupts
2870 * or the scheduler are operational.
2872 cpu_notifier(rcu_cpu_notify, 0);
2873 for_each_online_cpu(cpu)
2874 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2875 check_cpu_stall_init();
2878 #include "rcutree_plugin.h"