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] && !rcu_gp_in_progress(rsp);
312 * Return the root node of the specified rcu_state structure.
314 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
316 return &rsp->node[0];
320 * If the specified CPU is offline, tell the caller that it is in
321 * a quiescent state. Otherwise, whack it with a reschedule IPI.
322 * Grace periods can end up waiting on an offline CPU when that
323 * CPU is in the process of coming online -- it will be added to the
324 * rcu_node bitmasks before it actually makes it online. The same thing
325 * can happen while a CPU is in the process of coming online. Because this
326 * race is quite rare, we check for it after detecting that the grace
327 * period has been delayed rather than checking each and every CPU
328 * each and every time we start a new grace period.
330 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
333 * If the CPU is offline for more than a jiffy, it is in a quiescent
334 * state. We can trust its state not to change because interrupts
335 * are disabled. The reason for the jiffy's worth of slack is to
336 * handle CPUs initializing on the way up and finding their way
337 * to the idle loop on the way down.
339 if (cpu_is_offline(rdp->cpu) &&
340 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
341 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
349 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
351 * If the new value of the ->dynticks_nesting counter now is zero,
352 * we really have entered idle, and must do the appropriate accounting.
353 * The caller must have disabled interrupts.
355 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
357 trace_rcu_dyntick("Start", oldval, 0);
358 if (!is_idle_task(current)) {
359 struct task_struct *idle = idle_task(smp_processor_id());
361 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
362 ftrace_dump(DUMP_ALL);
363 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
364 current->pid, current->comm,
365 idle->pid, idle->comm); /* must be idle task! */
367 rcu_prepare_for_idle(smp_processor_id());
368 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
369 smp_mb__before_atomic_inc(); /* See above. */
370 atomic_inc(&rdtp->dynticks);
371 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
372 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
375 * The idle task is not permitted to enter the idle loop while
376 * in an RCU read-side critical section.
378 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
379 "Illegal idle entry in RCU read-side critical section.");
380 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
381 "Illegal idle entry in RCU-bh read-side critical section.");
382 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
383 "Illegal idle entry in RCU-sched read-side critical section.");
387 * rcu_idle_enter - inform RCU that current CPU is entering idle
389 * Enter idle mode, in other words, -leave- the mode in which RCU
390 * read-side critical sections can occur. (Though RCU read-side
391 * critical sections can occur in irq handlers in idle, a possibility
392 * handled by irq_enter() and irq_exit().)
394 * We crowbar the ->dynticks_nesting field to zero to allow for
395 * the possibility of usermode upcalls having messed up our count
396 * of interrupt nesting level during the prior busy period.
398 void rcu_idle_enter(void)
402 struct rcu_dynticks *rdtp;
404 local_irq_save(flags);
405 rdtp = &__get_cpu_var(rcu_dynticks);
406 oldval = rdtp->dynticks_nesting;
407 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
408 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
409 rdtp->dynticks_nesting = 0;
411 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
412 rcu_idle_enter_common(rdtp, oldval);
413 local_irq_restore(flags);
415 EXPORT_SYMBOL_GPL(rcu_idle_enter);
418 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
420 * Exit from an interrupt handler, which might possibly result in entering
421 * idle mode, in other words, leaving the mode in which read-side critical
422 * sections can occur.
424 * This code assumes that the idle loop never does anything that might
425 * result in unbalanced calls to irq_enter() and irq_exit(). If your
426 * architecture violates this assumption, RCU will give you what you
427 * deserve, good and hard. But very infrequently and irreproducibly.
429 * Use things like work queues to work around this limitation.
431 * You have been warned.
433 void rcu_irq_exit(void)
437 struct rcu_dynticks *rdtp;
439 local_irq_save(flags);
440 rdtp = &__get_cpu_var(rcu_dynticks);
441 oldval = rdtp->dynticks_nesting;
442 rdtp->dynticks_nesting--;
443 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
444 if (rdtp->dynticks_nesting)
445 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
447 rcu_idle_enter_common(rdtp, oldval);
448 local_irq_restore(flags);
452 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
454 * If the new value of the ->dynticks_nesting counter was previously zero,
455 * we really have exited idle, and must do the appropriate accounting.
456 * The caller must have disabled interrupts.
458 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
460 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
461 atomic_inc(&rdtp->dynticks);
462 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
463 smp_mb__after_atomic_inc(); /* See above. */
464 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
465 rcu_cleanup_after_idle(smp_processor_id());
466 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
467 if (!is_idle_task(current)) {
468 struct task_struct *idle = idle_task(smp_processor_id());
470 trace_rcu_dyntick("Error on exit: not idle task",
471 oldval, rdtp->dynticks_nesting);
472 ftrace_dump(DUMP_ALL);
473 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
474 current->pid, current->comm,
475 idle->pid, idle->comm); /* must be idle task! */
480 * rcu_idle_exit - inform RCU that current CPU is leaving idle
482 * Exit idle mode, in other words, -enter- the mode in which RCU
483 * read-side critical sections can occur.
485 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
486 * allow for the possibility of usermode upcalls messing up our count
487 * of interrupt nesting level during the busy period that is just
490 void rcu_idle_exit(void)
493 struct rcu_dynticks *rdtp;
496 local_irq_save(flags);
497 rdtp = &__get_cpu_var(rcu_dynticks);
498 oldval = rdtp->dynticks_nesting;
499 WARN_ON_ONCE(oldval < 0);
500 if (oldval & DYNTICK_TASK_NEST_MASK)
501 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
503 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
504 rcu_idle_exit_common(rdtp, oldval);
505 local_irq_restore(flags);
507 EXPORT_SYMBOL_GPL(rcu_idle_exit);
510 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
512 * Enter an interrupt handler, which might possibly result in exiting
513 * idle mode, in other words, entering the mode in which read-side critical
514 * sections can occur.
516 * Note that the Linux kernel is fully capable of entering an interrupt
517 * handler that it never exits, for example when doing upcalls to
518 * user mode! This code assumes that the idle loop never does upcalls to
519 * user mode. If your architecture does do upcalls from the idle loop (or
520 * does anything else that results in unbalanced calls to the irq_enter()
521 * and irq_exit() functions), RCU will give you what you deserve, good
522 * and hard. But very infrequently and irreproducibly.
524 * Use things like work queues to work around this limitation.
526 * You have been warned.
528 void rcu_irq_enter(void)
531 struct rcu_dynticks *rdtp;
534 local_irq_save(flags);
535 rdtp = &__get_cpu_var(rcu_dynticks);
536 oldval = rdtp->dynticks_nesting;
537 rdtp->dynticks_nesting++;
538 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
540 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
542 rcu_idle_exit_common(rdtp, oldval);
543 local_irq_restore(flags);
547 * rcu_nmi_enter - inform RCU of entry to NMI context
549 * If the CPU was idle with dynamic ticks active, and there is no
550 * irq handler running, this updates rdtp->dynticks_nmi to let the
551 * RCU grace-period handling know that the CPU is active.
553 void rcu_nmi_enter(void)
555 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
557 if (rdtp->dynticks_nmi_nesting == 0 &&
558 (atomic_read(&rdtp->dynticks) & 0x1))
560 rdtp->dynticks_nmi_nesting++;
561 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
562 atomic_inc(&rdtp->dynticks);
563 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
564 smp_mb__after_atomic_inc(); /* See above. */
565 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
569 * rcu_nmi_exit - inform RCU of exit from NMI context
571 * If the CPU was idle with dynamic ticks active, and there is no
572 * irq handler running, this updates rdtp->dynticks_nmi to let the
573 * RCU grace-period handling know that the CPU is no longer active.
575 void rcu_nmi_exit(void)
577 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
579 if (rdtp->dynticks_nmi_nesting == 0 ||
580 --rdtp->dynticks_nmi_nesting != 0)
582 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
583 smp_mb__before_atomic_inc(); /* See above. */
584 atomic_inc(&rdtp->dynticks);
585 smp_mb__after_atomic_inc(); /* Force delay to next write. */
586 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
589 #ifdef CONFIG_PROVE_RCU
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 #ifdef 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 /* #ifdef CONFIG_HOTPLUG_CPU */
651 #endif /* #ifdef CONFIG_PROVE_RCU */
654 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
656 * If the current CPU is idle or running at a first-level (not nested)
657 * interrupt from idle, return true. The caller must have at least
658 * disabled preemption.
660 int rcu_is_cpu_rrupt_from_idle(void)
662 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
666 * Snapshot the specified CPU's dynticks counter so that we can later
667 * credit them with an implicit quiescent state. Return 1 if this CPU
668 * is in dynticks idle mode, which is an extended quiescent state.
670 static int dyntick_save_progress_counter(struct rcu_data *rdp)
672 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
673 return (rdp->dynticks_snap & 0x1) == 0;
677 * Return true if the specified CPU has passed through a quiescent
678 * state by virtue of being in or having passed through an dynticks
679 * idle state since the last call to dyntick_save_progress_counter()
682 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
687 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
688 snap = (unsigned int)rdp->dynticks_snap;
691 * If the CPU passed through or entered a dynticks idle phase with
692 * no active irq/NMI handlers, then we can safely pretend that the CPU
693 * already acknowledged the request to pass through a quiescent
694 * state. Either way, that CPU cannot possibly be in an RCU
695 * read-side critical section that started before the beginning
696 * of the current RCU grace period.
698 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
699 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
704 /* Go check for the CPU being offline. */
705 return rcu_implicit_offline_qs(rdp);
708 static int jiffies_till_stall_check(void)
710 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
713 * Limit check must be consistent with the Kconfig limits
714 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
716 if (till_stall_check < 3) {
717 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
718 till_stall_check = 3;
719 } else if (till_stall_check > 300) {
720 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
721 till_stall_check = 300;
723 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
726 static void record_gp_stall_check_time(struct rcu_state *rsp)
728 rsp->gp_start = jiffies;
729 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
732 static void print_other_cpu_stall(struct rcu_state *rsp)
738 struct rcu_node *rnp = rcu_get_root(rsp);
740 /* Only let one CPU complain about others per time interval. */
742 raw_spin_lock_irqsave(&rnp->lock, flags);
743 delta = jiffies - rsp->jiffies_stall;
744 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
745 raw_spin_unlock_irqrestore(&rnp->lock, flags);
748 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
749 raw_spin_unlock_irqrestore(&rnp->lock, flags);
752 * OK, time to rat on our buddy...
753 * See Documentation/RCU/stallwarn.txt for info on how to debug
754 * RCU CPU stall warnings.
756 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
758 print_cpu_stall_info_begin();
759 rcu_for_each_leaf_node(rsp, rnp) {
760 raw_spin_lock_irqsave(&rnp->lock, flags);
761 ndetected += rcu_print_task_stall(rnp);
762 raw_spin_unlock_irqrestore(&rnp->lock, flags);
763 if (rnp->qsmask == 0)
765 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
766 if (rnp->qsmask & (1UL << cpu)) {
767 print_cpu_stall_info(rsp, rnp->grplo + cpu);
773 * Now rat on any tasks that got kicked up to the root rcu_node
774 * due to CPU offlining.
776 rnp = rcu_get_root(rsp);
777 raw_spin_lock_irqsave(&rnp->lock, flags);
778 ndetected = rcu_print_task_stall(rnp);
779 raw_spin_unlock_irqrestore(&rnp->lock, flags);
781 print_cpu_stall_info_end();
782 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
783 smp_processor_id(), (long)(jiffies - rsp->gp_start));
785 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
786 else if (!trigger_all_cpu_backtrace())
789 /* If so configured, complain about tasks blocking the grace period. */
791 rcu_print_detail_task_stall(rsp);
793 force_quiescent_state(rsp, 0); /* Kick them all. */
796 static void print_cpu_stall(struct rcu_state *rsp)
799 struct rcu_node *rnp = rcu_get_root(rsp);
802 * OK, time to rat on ourselves...
803 * See Documentation/RCU/stallwarn.txt for info on how to debug
804 * RCU CPU stall warnings.
806 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
807 print_cpu_stall_info_begin();
808 print_cpu_stall_info(rsp, smp_processor_id());
809 print_cpu_stall_info_end();
810 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
811 if (!trigger_all_cpu_backtrace())
814 raw_spin_lock_irqsave(&rnp->lock, flags);
815 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
816 rsp->jiffies_stall = jiffies +
817 3 * jiffies_till_stall_check() + 3;
818 raw_spin_unlock_irqrestore(&rnp->lock, flags);
820 set_need_resched(); /* kick ourselves to get things going. */
823 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
827 struct rcu_node *rnp;
829 if (rcu_cpu_stall_suppress)
831 j = ACCESS_ONCE(jiffies);
832 js = ACCESS_ONCE(rsp->jiffies_stall);
834 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
836 /* We haven't checked in, so go dump stack. */
837 print_cpu_stall(rsp);
839 } else if (rcu_gp_in_progress(rsp) &&
840 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
842 /* They had a few time units to dump stack, so complain. */
843 print_other_cpu_stall(rsp);
847 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
849 rcu_cpu_stall_suppress = 1;
854 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
856 * Set the stall-warning timeout way off into the future, thus preventing
857 * any RCU CPU stall-warning messages from appearing in the current set of
860 * The caller must disable hard irqs.
862 void rcu_cpu_stall_reset(void)
864 struct rcu_state *rsp;
866 for_each_rcu_flavor(rsp)
867 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
870 static struct notifier_block rcu_panic_block = {
871 .notifier_call = rcu_panic,
874 static void __init check_cpu_stall_init(void)
876 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
880 * Update CPU-local rcu_data state to record the newly noticed grace period.
881 * This is used both when we started the grace period and when we notice
882 * that someone else started the grace period. The caller must hold the
883 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
884 * and must have irqs disabled.
886 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
888 if (rdp->gpnum != rnp->gpnum) {
890 * If the current grace period is waiting for this CPU,
891 * set up to detect a quiescent state, otherwise don't
892 * go looking for one.
894 rdp->gpnum = rnp->gpnum;
895 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
896 if (rnp->qsmask & rdp->grpmask) {
898 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 * Advance this CPU's callbacks, but only if the current grace period
943 * has ended. This may be called only from the CPU to whom the rdp
944 * belongs. In addition, the corresponding leaf rcu_node structure's
945 * ->lock must be held by the caller, with irqs disabled.
948 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
950 /* Did another grace period end? */
951 if (rdp->completed != rnp->completed) {
953 /* Advance callbacks. No harm if list empty. */
954 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
955 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
956 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
958 /* Remember that we saw this grace-period completion. */
959 rdp->completed = rnp->completed;
960 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
963 * If we were in an extended quiescent state, we may have
964 * missed some grace periods that others CPUs handled on
965 * our behalf. Catch up with this state to avoid noting
966 * spurious new grace periods. If another grace period
967 * has started, then rnp->gpnum will have advanced, so
968 * we will detect this later on.
970 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
971 rdp->gpnum = rdp->completed;
974 * If RCU does not need a quiescent state from this CPU,
975 * then make sure that this CPU doesn't go looking for one.
977 if ((rnp->qsmask & rdp->grpmask) == 0)
983 * Advance this CPU's callbacks, but only if the current grace period
984 * has ended. This may be called only from the CPU to whom the rdp
988 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
991 struct rcu_node *rnp;
993 local_irq_save(flags);
995 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
996 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
997 local_irq_restore(flags);
1000 __rcu_process_gp_end(rsp, rnp, rdp);
1001 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1005 * Do per-CPU grace-period initialization for running CPU. The caller
1006 * must hold the lock of the leaf rcu_node structure corresponding to
1010 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1012 /* Prior grace period ended, so advance callbacks for current CPU. */
1013 __rcu_process_gp_end(rsp, rnp, rdp);
1016 * Because this CPU just now started the new grace period, we know
1017 * that all of its callbacks will be covered by this upcoming grace
1018 * period, even the ones that were registered arbitrarily recently.
1019 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1021 * Other CPUs cannot be sure exactly when the grace period started.
1022 * Therefore, their recently registered callbacks must pass through
1023 * an additional RCU_NEXT_READY stage, so that they will be handled
1024 * by the next RCU grace period.
1026 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1027 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1029 /* Set state so that this CPU will detect the next quiescent state. */
1030 __note_new_gpnum(rsp, rnp, rdp);
1034 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1035 * in preparation for detecting the next grace period. The caller must hold
1036 * the root node's ->lock, which is released before return. Hard irqs must
1039 * Note that it is legal for a dying CPU (which is marked as offline) to
1040 * invoke this function. This can happen when the dying CPU reports its
1044 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1045 __releases(rcu_get_root(rsp)->lock)
1047 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1048 struct rcu_node *rnp = rcu_get_root(rsp);
1050 if (!rcu_scheduler_fully_active ||
1051 !cpu_needs_another_gp(rsp, rdp)) {
1053 * Either the scheduler hasn't yet spawned the first
1054 * non-idle task or this CPU does not need another
1055 * grace period. Either way, don't start a new grace
1058 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1062 if (rsp->fqs_active) {
1064 * This CPU needs a grace period, but force_quiescent_state()
1065 * is running. Tell it to start one on this CPU's behalf.
1067 rsp->fqs_need_gp = 1;
1068 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1072 /* Advance to a new grace period and initialize state. */
1074 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1075 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1076 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1077 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1078 record_gp_stall_check_time(rsp);
1079 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
1081 /* Exclude any concurrent CPU-hotplug operations. */
1082 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1085 * Set the quiescent-state-needed bits in all the rcu_node
1086 * structures for all currently online CPUs in breadth-first
1087 * order, starting from the root rcu_node structure. This
1088 * operation relies on the layout of the hierarchy within the
1089 * rsp->node[] array. Note that other CPUs will access only
1090 * the leaves of the hierarchy, which still indicate that no
1091 * grace period is in progress, at least until the corresponding
1092 * leaf node has been initialized. In addition, we have excluded
1093 * CPU-hotplug operations.
1095 * Note that the grace period cannot complete until we finish
1096 * the initialization process, as there will be at least one
1097 * qsmask bit set in the root node until that time, namely the
1098 * one corresponding to this CPU, due to the fact that we have
1101 rcu_for_each_node_breadth_first(rsp, rnp) {
1102 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1103 rcu_preempt_check_blocked_tasks(rnp);
1104 rnp->qsmask = rnp->qsmaskinit;
1105 rnp->gpnum = rsp->gpnum;
1106 rnp->completed = rsp->completed;
1107 if (rnp == rdp->mynode)
1108 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1109 rcu_preempt_boost_start_gp(rnp);
1110 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1111 rnp->level, rnp->grplo,
1112 rnp->grphi, rnp->qsmask);
1113 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1116 rnp = rcu_get_root(rsp);
1117 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1118 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1119 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1120 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1124 * Report a full set of quiescent states to the specified rcu_state
1125 * data structure. This involves cleaning up after the prior grace
1126 * period and letting rcu_start_gp() start up the next grace period
1127 * if one is needed. Note that the caller must hold rnp->lock, as
1128 * required by rcu_start_gp(), which will release it.
1130 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1131 __releases(rcu_get_root(rsp)->lock)
1133 unsigned long gp_duration;
1134 struct rcu_node *rnp = rcu_get_root(rsp);
1135 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1137 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1140 * Ensure that all grace-period and pre-grace-period activity
1141 * is seen before the assignment to rsp->completed.
1143 smp_mb(); /* See above block comment. */
1144 gp_duration = jiffies - rsp->gp_start;
1145 if (gp_duration > rsp->gp_max)
1146 rsp->gp_max = gp_duration;
1149 * We know the grace period is complete, but to everyone else
1150 * it appears to still be ongoing. But it is also the case
1151 * that to everyone else it looks like there is nothing that
1152 * they can do to advance the grace period. It is therefore
1153 * safe for us to drop the lock in order to mark the grace
1154 * period as completed in all of the rcu_node structures.
1156 * But if this CPU needs another grace period, it will take
1157 * care of this while initializing the next grace period.
1158 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1159 * because the callbacks have not yet been advanced: Those
1160 * callbacks are waiting on the grace period that just now
1163 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1164 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1167 * Propagate new ->completed value to rcu_node structures
1168 * so that other CPUs don't have to wait until the start
1169 * of the next grace period to process their callbacks.
1171 rcu_for_each_node_breadth_first(rsp, rnp) {
1172 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1173 rnp->completed = rsp->gpnum;
1174 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1176 rnp = rcu_get_root(rsp);
1177 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1180 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
1181 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1182 rsp->fqs_state = RCU_GP_IDLE;
1183 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1187 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1188 * Allows quiescent states for a group of CPUs to be reported at one go
1189 * to the specified rcu_node structure, though all the CPUs in the group
1190 * must be represented by the same rcu_node structure (which need not be
1191 * a leaf rcu_node structure, though it often will be). That structure's
1192 * lock must be held upon entry, and it is released before return.
1195 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1196 struct rcu_node *rnp, unsigned long flags)
1197 __releases(rnp->lock)
1199 struct rcu_node *rnp_c;
1201 /* Walk up the rcu_node hierarchy. */
1203 if (!(rnp->qsmask & mask)) {
1205 /* Our bit has already been cleared, so done. */
1206 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1209 rnp->qsmask &= ~mask;
1210 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1211 mask, rnp->qsmask, rnp->level,
1212 rnp->grplo, rnp->grphi,
1214 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1216 /* Other bits still set at this level, so done. */
1217 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1220 mask = rnp->grpmask;
1221 if (rnp->parent == NULL) {
1223 /* No more levels. Exit loop holding root lock. */
1227 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1230 raw_spin_lock_irqsave(&rnp->lock, flags);
1231 WARN_ON_ONCE(rnp_c->qsmask);
1235 * Get here if we are the last CPU to pass through a quiescent
1236 * state for this grace period. Invoke rcu_report_qs_rsp()
1237 * to clean up and start the next grace period if one is needed.
1239 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1243 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1244 * structure. This must be either called from the specified CPU, or
1245 * called when the specified CPU is known to be offline (and when it is
1246 * also known that no other CPU is concurrently trying to help the offline
1247 * CPU). The lastcomp argument is used to make sure we are still in the
1248 * grace period of interest. We don't want to end the current grace period
1249 * based on quiescent states detected in an earlier grace period!
1252 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1254 unsigned long flags;
1256 struct rcu_node *rnp;
1259 raw_spin_lock_irqsave(&rnp->lock, flags);
1260 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1263 * The grace period in which this quiescent state was
1264 * recorded has ended, so don't report it upwards.
1265 * We will instead need a new quiescent state that lies
1266 * within the current grace period.
1268 rdp->passed_quiesce = 0; /* need qs for new gp. */
1269 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1272 mask = rdp->grpmask;
1273 if ((rnp->qsmask & mask) == 0) {
1274 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1276 rdp->qs_pending = 0;
1279 * This GP can't end until cpu checks in, so all of our
1280 * callbacks can be processed during the next GP.
1282 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1284 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1289 * Check to see if there is a new grace period of which this CPU
1290 * is not yet aware, and if so, set up local rcu_data state for it.
1291 * Otherwise, see if this CPU has just passed through its first
1292 * quiescent state for this grace period, and record that fact if so.
1295 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1297 /* If there is now a new grace period, record and return. */
1298 if (check_for_new_grace_period(rsp, rdp))
1302 * Does this CPU still need to do its part for current grace period?
1303 * If no, return and let the other CPUs do their part as well.
1305 if (!rdp->qs_pending)
1309 * Was there a quiescent state since the beginning of the grace
1310 * period? If no, then exit and wait for the next call.
1312 if (!rdp->passed_quiesce)
1316 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1319 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1322 #ifdef CONFIG_HOTPLUG_CPU
1325 * Send the specified CPU's RCU callbacks to the orphanage. The
1326 * specified CPU must be offline, and the caller must hold the
1330 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1331 struct rcu_node *rnp, struct rcu_data *rdp)
1336 * Orphan the callbacks. First adjust the counts. This is safe
1337 * because ->onofflock excludes _rcu_barrier()'s adoption of
1338 * the callbacks, thus no memory barrier is required.
1340 if (rdp->nxtlist != NULL) {
1341 rsp->qlen_lazy += rdp->qlen_lazy;
1342 rsp->qlen += rdp->qlen;
1343 rdp->n_cbs_orphaned += rdp->qlen;
1349 * Next, move those callbacks still needing a grace period to
1350 * the orphanage, where some other CPU will pick them up.
1351 * Some of the callbacks might have gone partway through a grace
1352 * period, but that is too bad. They get to start over because we
1353 * cannot assume that grace periods are synchronized across CPUs.
1354 * We don't bother updating the ->nxttail[] array yet, instead
1355 * we just reset the whole thing later on.
1357 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1358 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1359 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1360 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1364 * Then move the ready-to-invoke callbacks to the orphanage,
1365 * where some other CPU will pick them up. These will not be
1366 * required to pass though another grace period: They are done.
1368 if (rdp->nxtlist != NULL) {
1369 *rsp->orphan_donetail = rdp->nxtlist;
1370 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1373 /* Finally, initialize the rcu_data structure's list to empty. */
1374 rdp->nxtlist = NULL;
1375 for (i = 0; i < RCU_NEXT_SIZE; i++)
1376 rdp->nxttail[i] = &rdp->nxtlist;
1380 * Adopt the RCU callbacks from the specified rcu_state structure's
1381 * orphanage. The caller must hold the ->onofflock.
1383 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1386 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1389 * If there is an rcu_barrier() operation in progress, then
1390 * only the task doing that operation is permitted to adopt
1391 * callbacks. To do otherwise breaks rcu_barrier() and friends
1392 * by causing them to fail to wait for the callbacks in the
1395 if (rsp->rcu_barrier_in_progress &&
1396 rsp->rcu_barrier_in_progress != current)
1399 /* Do the accounting first. */
1400 rdp->qlen_lazy += rsp->qlen_lazy;
1401 rdp->qlen += rsp->qlen;
1402 rdp->n_cbs_adopted += rsp->qlen;
1403 if (rsp->qlen_lazy != rsp->qlen)
1404 rcu_idle_count_callbacks_posted();
1409 * We do not need a memory barrier here because the only way we
1410 * can get here if there is an rcu_barrier() in flight is if
1411 * we are the task doing the rcu_barrier().
1414 /* First adopt the ready-to-invoke callbacks. */
1415 if (rsp->orphan_donelist != NULL) {
1416 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1417 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1418 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1419 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1420 rdp->nxttail[i] = rsp->orphan_donetail;
1421 rsp->orphan_donelist = NULL;
1422 rsp->orphan_donetail = &rsp->orphan_donelist;
1425 /* And then adopt the callbacks that still need a grace period. */
1426 if (rsp->orphan_nxtlist != NULL) {
1427 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1428 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1429 rsp->orphan_nxtlist = NULL;
1430 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1435 * Trace the fact that this CPU is going offline.
1437 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1439 RCU_TRACE(unsigned long mask);
1440 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1441 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1443 RCU_TRACE(mask = rdp->grpmask);
1444 trace_rcu_grace_period(rsp->name,
1445 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1450 * The CPU has been completely removed, and some other CPU is reporting
1451 * this fact from process context. Do the remainder of the cleanup,
1452 * including orphaning the outgoing CPU's RCU callbacks, and also
1453 * adopting them, if there is no _rcu_barrier() instance running.
1454 * There can only be one CPU hotplug operation at a time, so no other
1455 * CPU can be attempting to update rcu_cpu_kthread_task.
1457 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1459 unsigned long flags;
1461 int need_report = 0;
1462 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1463 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1465 /* Adjust any no-longer-needed kthreads. */
1466 rcu_stop_cpu_kthread(cpu);
1467 rcu_node_kthread_setaffinity(rnp, -1);
1469 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1471 /* Exclude any attempts to start a new grace period. */
1472 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1474 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1475 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1476 rcu_adopt_orphan_cbs(rsp);
1478 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1479 mask = rdp->grpmask; /* rnp->grplo is constant. */
1481 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1482 rnp->qsmaskinit &= ~mask;
1483 if (rnp->qsmaskinit != 0) {
1484 if (rnp != rdp->mynode)
1485 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1488 if (rnp == rdp->mynode)
1489 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1491 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1492 mask = rnp->grpmask;
1494 } while (rnp != NULL);
1497 * We still hold the leaf rcu_node structure lock here, and
1498 * irqs are still disabled. The reason for this subterfuge is
1499 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1500 * held leads to deadlock.
1502 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1504 if (need_report & RCU_OFL_TASKS_NORM_GP)
1505 rcu_report_unblock_qs_rnp(rnp, flags);
1507 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1508 if (need_report & RCU_OFL_TASKS_EXP_GP)
1509 rcu_report_exp_rnp(rsp, rnp, true);
1512 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1514 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1518 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1522 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1526 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1529 * Invoke any RCU callbacks that have made it to the end of their grace
1530 * period. Thottle as specified by rdp->blimit.
1532 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1534 unsigned long flags;
1535 struct rcu_head *next, *list, **tail;
1536 int bl, count, count_lazy, i;
1538 /* If no callbacks are ready, just return.*/
1539 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1540 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1541 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1542 need_resched(), is_idle_task(current),
1543 rcu_is_callbacks_kthread());
1548 * Extract the list of ready callbacks, disabling to prevent
1549 * races with call_rcu() from interrupt handlers.
1551 local_irq_save(flags);
1552 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1554 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1555 list = rdp->nxtlist;
1556 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1557 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1558 tail = rdp->nxttail[RCU_DONE_TAIL];
1559 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1560 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1561 rdp->nxttail[i] = &rdp->nxtlist;
1562 local_irq_restore(flags);
1564 /* Invoke callbacks. */
1565 count = count_lazy = 0;
1569 debug_rcu_head_unqueue(list);
1570 if (__rcu_reclaim(rsp->name, list))
1573 /* Stop only if limit reached and CPU has something to do. */
1574 if (++count >= bl &&
1576 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1580 local_irq_save(flags);
1581 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1582 is_idle_task(current),
1583 rcu_is_callbacks_kthread());
1585 /* Update count, and requeue any remaining callbacks. */
1587 *tail = rdp->nxtlist;
1588 rdp->nxtlist = list;
1589 for (i = 0; i < RCU_NEXT_SIZE; i++)
1590 if (&rdp->nxtlist == rdp->nxttail[i])
1591 rdp->nxttail[i] = tail;
1595 smp_mb(); /* List handling before counting for rcu_barrier(). */
1596 rdp->qlen_lazy -= count_lazy;
1598 rdp->n_cbs_invoked += count;
1600 /* Reinstate batch limit if we have worked down the excess. */
1601 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1602 rdp->blimit = blimit;
1604 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1605 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1606 rdp->qlen_last_fqs_check = 0;
1607 rdp->n_force_qs_snap = rsp->n_force_qs;
1608 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1609 rdp->qlen_last_fqs_check = rdp->qlen;
1611 local_irq_restore(flags);
1613 /* Re-invoke RCU core processing if there are callbacks remaining. */
1614 if (cpu_has_callbacks_ready_to_invoke(rdp))
1619 * Check to see if this CPU is in a non-context-switch quiescent state
1620 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1621 * Also schedule RCU core processing.
1623 * This function must be called from hardirq context. It is normally
1624 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1625 * false, there is no point in invoking rcu_check_callbacks().
1627 void rcu_check_callbacks(int cpu, int user)
1629 trace_rcu_utilization("Start scheduler-tick");
1630 increment_cpu_stall_ticks();
1631 if (user || rcu_is_cpu_rrupt_from_idle()) {
1634 * Get here if this CPU took its interrupt from user
1635 * mode or from the idle loop, and if this is not a
1636 * nested interrupt. In this case, the CPU is in
1637 * a quiescent state, so note it.
1639 * No memory barrier is required here because both
1640 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1641 * variables that other CPUs neither access nor modify,
1642 * at least not while the corresponding CPU is online.
1648 } else if (!in_softirq()) {
1651 * Get here if this CPU did not take its interrupt from
1652 * softirq, in other words, if it is not interrupting
1653 * a rcu_bh read-side critical section. This is an _bh
1654 * critical section, so note it.
1659 rcu_preempt_check_callbacks(cpu);
1660 if (rcu_pending(cpu))
1662 trace_rcu_utilization("End scheduler-tick");
1666 * Scan the leaf rcu_node structures, processing dyntick state for any that
1667 * have not yet encountered a quiescent state, using the function specified.
1668 * Also initiate boosting for any threads blocked on the root rcu_node.
1670 * The caller must have suppressed start of new grace periods.
1672 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1676 unsigned long flags;
1678 struct rcu_node *rnp;
1680 rcu_for_each_leaf_node(rsp, rnp) {
1682 raw_spin_lock_irqsave(&rnp->lock, flags);
1683 if (!rcu_gp_in_progress(rsp)) {
1684 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1687 if (rnp->qsmask == 0) {
1688 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1693 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1694 if ((rnp->qsmask & bit) != 0 &&
1695 f(per_cpu_ptr(rsp->rda, cpu)))
1700 /* rcu_report_qs_rnp() releases rnp->lock. */
1701 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1704 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1706 rnp = rcu_get_root(rsp);
1707 if (rnp->qsmask == 0) {
1708 raw_spin_lock_irqsave(&rnp->lock, flags);
1709 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1714 * Force quiescent states on reluctant CPUs, and also detect which
1715 * CPUs are in dyntick-idle mode.
1717 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1719 unsigned long flags;
1720 struct rcu_node *rnp = rcu_get_root(rsp);
1722 trace_rcu_utilization("Start fqs");
1723 if (!rcu_gp_in_progress(rsp)) {
1724 trace_rcu_utilization("End fqs");
1725 return; /* No grace period in progress, nothing to force. */
1727 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1728 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1729 trace_rcu_utilization("End fqs");
1730 return; /* Someone else is already on the job. */
1732 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1733 goto unlock_fqs_ret; /* no emergency and done recently. */
1735 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1736 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1737 if(!rcu_gp_in_progress(rsp)) {
1738 rsp->n_force_qs_ngp++;
1739 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1740 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1742 rsp->fqs_active = 1;
1743 switch (rsp->fqs_state) {
1747 break; /* grace period idle or initializing, ignore. */
1749 case RCU_SAVE_DYNTICK:
1751 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1753 /* Record dyntick-idle state. */
1754 force_qs_rnp(rsp, dyntick_save_progress_counter);
1755 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1756 if (rcu_gp_in_progress(rsp))
1757 rsp->fqs_state = RCU_FORCE_QS;
1762 /* Check dyntick-idle state, send IPI to laggarts. */
1763 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1764 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1766 /* Leave state in case more forcing is required. */
1768 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1771 rsp->fqs_active = 0;
1772 if (rsp->fqs_need_gp) {
1773 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1774 rsp->fqs_need_gp = 0;
1775 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1776 trace_rcu_utilization("End fqs");
1779 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1781 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1782 trace_rcu_utilization("End fqs");
1786 * This does the RCU core processing work for the specified rcu_state
1787 * and rcu_data structures. This may be called only from the CPU to
1788 * whom the rdp belongs.
1791 __rcu_process_callbacks(struct rcu_state *rsp)
1793 unsigned long flags;
1794 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1796 WARN_ON_ONCE(rdp->beenonline == 0);
1799 * If an RCU GP has gone long enough, go check for dyntick
1800 * idle CPUs and, if needed, send resched IPIs.
1802 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1803 force_quiescent_state(rsp, 1);
1806 * Advance callbacks in response to end of earlier grace
1807 * period that some other CPU ended.
1809 rcu_process_gp_end(rsp, rdp);
1811 /* Update RCU state based on any recent quiescent states. */
1812 rcu_check_quiescent_state(rsp, rdp);
1814 /* Does this CPU require a not-yet-started grace period? */
1815 if (cpu_needs_another_gp(rsp, rdp)) {
1816 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1817 rcu_start_gp(rsp, flags); /* releases above lock */
1820 /* If there are callbacks ready, invoke them. */
1821 if (cpu_has_callbacks_ready_to_invoke(rdp))
1822 invoke_rcu_callbacks(rsp, rdp);
1826 * Do RCU core processing for the current CPU.
1828 static void rcu_process_callbacks(struct softirq_action *unused)
1830 struct rcu_state *rsp;
1832 trace_rcu_utilization("Start RCU core");
1833 for_each_rcu_flavor(rsp)
1834 __rcu_process_callbacks(rsp);
1835 trace_rcu_utilization("End RCU core");
1839 * Schedule RCU callback invocation. If the specified type of RCU
1840 * does not support RCU priority boosting, just do a direct call,
1841 * otherwise wake up the per-CPU kernel kthread. Note that because we
1842 * are running on the current CPU with interrupts disabled, the
1843 * rcu_cpu_kthread_task cannot disappear out from under us.
1845 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1847 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1849 if (likely(!rsp->boost)) {
1850 rcu_do_batch(rsp, rdp);
1853 invoke_rcu_callbacks_kthread();
1856 static void invoke_rcu_core(void)
1858 raise_softirq(RCU_SOFTIRQ);
1862 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1863 struct rcu_state *rsp, bool lazy)
1865 unsigned long flags;
1866 struct rcu_data *rdp;
1868 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1869 debug_rcu_head_queue(head);
1873 smp_mb(); /* Ensure RCU update seen before callback registry. */
1876 * Opportunistically note grace-period endings and beginnings.
1877 * Note that we might see a beginning right after we see an
1878 * end, but never vice versa, since this CPU has to pass through
1879 * a quiescent state betweentimes.
1881 local_irq_save(flags);
1882 rdp = this_cpu_ptr(rsp->rda);
1884 /* Add the callback to our list. */
1889 rcu_idle_count_callbacks_posted();
1890 smp_mb(); /* Count before adding callback for rcu_barrier(). */
1891 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1892 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1894 if (__is_kfree_rcu_offset((unsigned long)func))
1895 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1896 rdp->qlen_lazy, rdp->qlen);
1898 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1900 /* If interrupts were disabled, don't dive into RCU core. */
1901 if (irqs_disabled_flags(flags)) {
1902 local_irq_restore(flags);
1907 * Force the grace period if too many callbacks or too long waiting.
1908 * Enforce hysteresis, and don't invoke force_quiescent_state()
1909 * if some other CPU has recently done so. Also, don't bother
1910 * invoking force_quiescent_state() if the newly enqueued callback
1911 * is the only one waiting for a grace period to complete.
1913 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1915 /* Are we ignoring a completed grace period? */
1916 rcu_process_gp_end(rsp, rdp);
1917 check_for_new_grace_period(rsp, rdp);
1919 /* Start a new grace period if one not already started. */
1920 if (!rcu_gp_in_progress(rsp)) {
1921 unsigned long nestflag;
1922 struct rcu_node *rnp_root = rcu_get_root(rsp);
1924 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1925 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1927 /* Give the grace period a kick. */
1928 rdp->blimit = LONG_MAX;
1929 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1930 *rdp->nxttail[RCU_DONE_TAIL] != head)
1931 force_quiescent_state(rsp, 0);
1932 rdp->n_force_qs_snap = rsp->n_force_qs;
1933 rdp->qlen_last_fqs_check = rdp->qlen;
1935 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1936 force_quiescent_state(rsp, 1);
1937 local_irq_restore(flags);
1941 * Queue an RCU-sched callback for invocation after a grace period.
1943 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1945 __call_rcu(head, func, &rcu_sched_state, 0);
1947 EXPORT_SYMBOL_GPL(call_rcu_sched);
1950 * Queue an RCU callback for invocation after a quicker grace period.
1952 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1954 __call_rcu(head, func, &rcu_bh_state, 0);
1956 EXPORT_SYMBOL_GPL(call_rcu_bh);
1959 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1960 * any blocking grace-period wait automatically implies a grace period
1961 * if there is only one CPU online at any point time during execution
1962 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
1963 * occasionally incorrectly indicate that there are multiple CPUs online
1964 * when there was in fact only one the whole time, as this just adds
1965 * some overhead: RCU still operates correctly.
1967 * Of course, sampling num_online_cpus() with preemption enabled can
1968 * give erroneous results if there are concurrent CPU-hotplug operations.
1969 * For example, given a demonic sequence of preemptions in num_online_cpus()
1970 * and CPU-hotplug operations, there could be two or more CPUs online at
1971 * all times, but num_online_cpus() might well return one (or even zero).
1973 * However, all such demonic sequences require at least one CPU-offline
1974 * operation. Furthermore, rcu_blocking_is_gp() giving the wrong answer
1975 * is only a problem if there is an RCU read-side critical section executing
1976 * throughout. But RCU-sched and RCU-bh read-side critical sections
1977 * disable either preemption or bh, which prevents a CPU from going offline.
1978 * Therefore, the only way that rcu_blocking_is_gp() can incorrectly return
1979 * that there is only one CPU when in fact there was more than one throughout
1980 * is when there were no RCU readers in the system. If there are no
1981 * RCU readers, the grace period by definition can be of zero length,
1982 * regardless of the number of online CPUs.
1984 static inline int rcu_blocking_is_gp(void)
1986 might_sleep(); /* Check for RCU read-side critical section. */
1987 return num_online_cpus() <= 1;
1991 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1993 * Control will return to the caller some time after a full rcu-sched
1994 * grace period has elapsed, in other words after all currently executing
1995 * rcu-sched read-side critical sections have completed. These read-side
1996 * critical sections are delimited by rcu_read_lock_sched() and
1997 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1998 * local_irq_disable(), and so on may be used in place of
1999 * rcu_read_lock_sched().
2001 * This means that all preempt_disable code sequences, including NMI and
2002 * hardware-interrupt handlers, in progress on entry will have completed
2003 * before this primitive returns. However, this does not guarantee that
2004 * softirq handlers will have completed, since in some kernels, these
2005 * handlers can run in process context, and can block.
2007 * This primitive provides the guarantees made by the (now removed)
2008 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2009 * guarantees that rcu_read_lock() sections will have completed.
2010 * In "classic RCU", these two guarantees happen to be one and
2011 * the same, but can differ in realtime RCU implementations.
2013 void synchronize_sched(void)
2015 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2016 !lock_is_held(&rcu_lock_map) &&
2017 !lock_is_held(&rcu_sched_lock_map),
2018 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2019 if (rcu_blocking_is_gp())
2021 wait_rcu_gp(call_rcu_sched);
2023 EXPORT_SYMBOL_GPL(synchronize_sched);
2026 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2028 * Control will return to the caller some time after a full rcu_bh grace
2029 * period has elapsed, in other words after all currently executing rcu_bh
2030 * read-side critical sections have completed. RCU read-side critical
2031 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2032 * and may be nested.
2034 void synchronize_rcu_bh(void)
2036 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2037 !lock_is_held(&rcu_lock_map) &&
2038 !lock_is_held(&rcu_sched_lock_map),
2039 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2040 if (rcu_blocking_is_gp())
2042 wait_rcu_gp(call_rcu_bh);
2044 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2046 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2047 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2049 static int synchronize_sched_expedited_cpu_stop(void *data)
2052 * There must be a full memory barrier on each affected CPU
2053 * between the time that try_stop_cpus() is called and the
2054 * time that it returns.
2056 * In the current initial implementation of cpu_stop, the
2057 * above condition is already met when the control reaches
2058 * this point and the following smp_mb() is not strictly
2059 * necessary. Do smp_mb() anyway for documentation and
2060 * robustness against future implementation changes.
2062 smp_mb(); /* See above comment block. */
2067 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2069 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2070 * approach to force the grace period to end quickly. This consumes
2071 * significant time on all CPUs and is unfriendly to real-time workloads,
2072 * so is thus not recommended for any sort of common-case code. In fact,
2073 * if you are using synchronize_sched_expedited() in a loop, please
2074 * restructure your code to batch your updates, and then use a single
2075 * synchronize_sched() instead.
2077 * Note that it is illegal to call this function while holding any lock
2078 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2079 * to call this function from a CPU-hotplug notifier. Failing to observe
2080 * these restriction will result in deadlock.
2082 * This implementation can be thought of as an application of ticket
2083 * locking to RCU, with sync_sched_expedited_started and
2084 * sync_sched_expedited_done taking on the roles of the halves
2085 * of the ticket-lock word. Each task atomically increments
2086 * sync_sched_expedited_started upon entry, snapshotting the old value,
2087 * then attempts to stop all the CPUs. If this succeeds, then each
2088 * CPU will have executed a context switch, resulting in an RCU-sched
2089 * grace period. We are then done, so we use atomic_cmpxchg() to
2090 * update sync_sched_expedited_done to match our snapshot -- but
2091 * only if someone else has not already advanced past our snapshot.
2093 * On the other hand, if try_stop_cpus() fails, we check the value
2094 * of sync_sched_expedited_done. If it has advanced past our
2095 * initial snapshot, then someone else must have forced a grace period
2096 * some time after we took our snapshot. In this case, our work is
2097 * done for us, and we can simply return. Otherwise, we try again,
2098 * but keep our initial snapshot for purposes of checking for someone
2099 * doing our work for us.
2101 * If we fail too many times in a row, we fall back to synchronize_sched().
2103 void synchronize_sched_expedited(void)
2105 int firstsnap, s, snap, trycount = 0;
2107 /* Note that atomic_inc_return() implies full memory barrier. */
2108 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2110 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2113 * Each pass through the following loop attempts to force a
2114 * context switch on each CPU.
2116 while (try_stop_cpus(cpu_online_mask,
2117 synchronize_sched_expedited_cpu_stop,
2121 /* No joy, try again later. Or just synchronize_sched(). */
2122 if (trycount++ < 10)
2123 udelay(trycount * num_online_cpus());
2125 synchronize_sched();
2129 /* Check to see if someone else did our work for us. */
2130 s = atomic_read(&sync_sched_expedited_done);
2131 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2132 smp_mb(); /* ensure test happens before caller kfree */
2137 * Refetching sync_sched_expedited_started allows later
2138 * callers to piggyback on our grace period. We subtract
2139 * 1 to get the same token that the last incrementer got.
2140 * We retry after they started, so our grace period works
2141 * for them, and they started after our first try, so their
2142 * grace period works for us.
2145 snap = atomic_read(&sync_sched_expedited_started);
2146 smp_mb(); /* ensure read is before try_stop_cpus(). */
2150 * Everyone up to our most recent fetch is covered by our grace
2151 * period. Update the counter, but only if our work is still
2152 * relevant -- which it won't be if someone who started later
2153 * than we did beat us to the punch.
2156 s = atomic_read(&sync_sched_expedited_done);
2157 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2158 smp_mb(); /* ensure test happens before caller kfree */
2161 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2165 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2168 * Check to see if there is any immediate RCU-related work to be done
2169 * by the current CPU, for the specified type of RCU, returning 1 if so.
2170 * The checks are in order of increasing expense: checks that can be
2171 * carried out against CPU-local state are performed first. However,
2172 * we must check for CPU stalls first, else we might not get a chance.
2174 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2176 struct rcu_node *rnp = rdp->mynode;
2178 rdp->n_rcu_pending++;
2180 /* Check for CPU stalls, if enabled. */
2181 check_cpu_stall(rsp, rdp);
2183 /* Is the RCU core waiting for a quiescent state from this CPU? */
2184 if (rcu_scheduler_fully_active &&
2185 rdp->qs_pending && !rdp->passed_quiesce) {
2188 * If force_quiescent_state() coming soon and this CPU
2189 * needs a quiescent state, and this is either RCU-sched
2190 * or RCU-bh, force a local reschedule.
2192 rdp->n_rp_qs_pending++;
2193 if (!rdp->preemptible &&
2194 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2197 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2198 rdp->n_rp_report_qs++;
2202 /* Does this CPU have callbacks ready to invoke? */
2203 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2204 rdp->n_rp_cb_ready++;
2208 /* Has RCU gone idle with this CPU needing another grace period? */
2209 if (cpu_needs_another_gp(rsp, rdp)) {
2210 rdp->n_rp_cpu_needs_gp++;
2214 /* Has another RCU grace period completed? */
2215 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2216 rdp->n_rp_gp_completed++;
2220 /* Has a new RCU grace period started? */
2221 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2222 rdp->n_rp_gp_started++;
2226 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2227 if (rcu_gp_in_progress(rsp) &&
2228 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2229 rdp->n_rp_need_fqs++;
2234 rdp->n_rp_need_nothing++;
2239 * Check to see if there is any immediate RCU-related work to be done
2240 * by the current CPU, returning 1 if so. This function is part of the
2241 * RCU implementation; it is -not- an exported member of the RCU API.
2243 static int rcu_pending(int cpu)
2245 struct rcu_state *rsp;
2247 for_each_rcu_flavor(rsp)
2248 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2254 * Check to see if any future RCU-related work will need to be done
2255 * by the current CPU, even if none need be done immediately, returning
2258 static int rcu_cpu_has_callbacks(int cpu)
2260 struct rcu_state *rsp;
2262 /* RCU callbacks either ready or pending? */
2263 for_each_rcu_flavor(rsp)
2264 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2270 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2271 * the compiler is expected to optimize this away.
2273 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2274 int cpu, unsigned long done)
2276 trace_rcu_barrier(rsp->name, s, cpu,
2277 atomic_read(&rsp->barrier_cpu_count), done);
2281 * RCU callback function for _rcu_barrier(). If we are last, wake
2282 * up the task executing _rcu_barrier().
2284 static void rcu_barrier_callback(struct rcu_head *rhp)
2286 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2287 struct rcu_state *rsp = rdp->rsp;
2289 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2290 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2291 complete(&rsp->barrier_completion);
2293 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2298 * Called with preemption disabled, and from cross-cpu IRQ context.
2300 static void rcu_barrier_func(void *type)
2302 struct rcu_state *rsp = type;
2303 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2305 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2306 atomic_inc(&rsp->barrier_cpu_count);
2307 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2311 * Orchestrate the specified type of RCU barrier, waiting for all
2312 * RCU callbacks of the specified type to complete.
2314 static void _rcu_barrier(struct rcu_state *rsp)
2317 unsigned long flags;
2318 struct rcu_data *rdp;
2320 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2321 unsigned long snap_done;
2323 init_rcu_head_on_stack(&rd.barrier_head);
2324 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2326 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2327 mutex_lock(&rsp->barrier_mutex);
2330 * Ensure that all prior references, including to ->n_barrier_done,
2331 * are ordered before the _rcu_barrier() machinery.
2333 smp_mb(); /* See above block comment. */
2336 * Recheck ->n_barrier_done to see if others did our work for us.
2337 * This means checking ->n_barrier_done for an even-to-odd-to-even
2338 * transition. The "if" expression below therefore rounds the old
2339 * value up to the next even number and adds two before comparing.
2341 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2342 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2343 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2344 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2345 smp_mb(); /* caller's subsequent code after above check. */
2346 mutex_unlock(&rsp->barrier_mutex);
2351 * Increment ->n_barrier_done to avoid duplicate work. Use
2352 * ACCESS_ONCE() to prevent the compiler from speculating
2353 * the increment to precede the early-exit check.
2355 ACCESS_ONCE(rsp->n_barrier_done)++;
2356 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2357 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2358 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2361 * Initialize the count to one rather than to zero in order to
2362 * avoid a too-soon return to zero in case of a short grace period
2363 * (or preemption of this task). Also flag this task as doing
2364 * an rcu_barrier(). This will prevent anyone else from adopting
2365 * orphaned callbacks, which could cause otherwise failure if a
2366 * CPU went offline and quickly came back online. To see this,
2367 * consider the following sequence of events:
2369 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2370 * 2. CPU 1 goes offline, orphaning its callbacks.
2371 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2372 * 4. CPU 1 comes back online.
2373 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2374 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2375 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2377 init_completion(&rsp->barrier_completion);
2378 atomic_set(&rsp->barrier_cpu_count, 1);
2379 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2380 rsp->rcu_barrier_in_progress = current;
2381 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2384 * Force every CPU with callbacks to register a new callback
2385 * that will tell us when all the preceding callbacks have
2386 * been invoked. If an offline CPU has callbacks, wait for
2387 * it to either come back online or to finish orphaning those
2390 for_each_possible_cpu(cpu) {
2392 rdp = per_cpu_ptr(rsp->rda, cpu);
2393 if (cpu_is_offline(cpu)) {
2394 _rcu_barrier_trace(rsp, "Offline", cpu,
2395 rsp->n_barrier_done);
2397 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2398 schedule_timeout_interruptible(1);
2399 } else if (ACCESS_ONCE(rdp->qlen)) {
2400 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2401 rsp->n_barrier_done);
2402 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2405 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2406 rsp->n_barrier_done);
2412 * Now that all online CPUs have rcu_barrier_callback() callbacks
2413 * posted, we can adopt all of the orphaned callbacks and place
2414 * an rcu_barrier_callback() callback after them. When that is done,
2415 * we are guaranteed to have an rcu_barrier_callback() callback
2416 * following every callback that could possibly have been
2417 * registered before _rcu_barrier() was called.
2419 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2420 rcu_adopt_orphan_cbs(rsp);
2421 rsp->rcu_barrier_in_progress = NULL;
2422 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2423 atomic_inc(&rsp->barrier_cpu_count);
2424 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2426 rsp->call(&rd.barrier_head, rcu_barrier_callback);
2429 * Now that we have an rcu_barrier_callback() callback on each
2430 * CPU, and thus each counted, remove the initial count.
2432 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2433 complete(&rsp->barrier_completion);
2435 /* Increment ->n_barrier_done to prevent duplicate work. */
2436 smp_mb(); /* Keep increment after above mechanism. */
2437 ACCESS_ONCE(rsp->n_barrier_done)++;
2438 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2439 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2440 smp_mb(); /* Keep increment before caller's subsequent code. */
2442 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2443 wait_for_completion(&rsp->barrier_completion);
2445 /* Other rcu_barrier() invocations can now safely proceed. */
2446 mutex_unlock(&rsp->barrier_mutex);
2448 destroy_rcu_head_on_stack(&rd.barrier_head);
2452 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2454 void rcu_barrier_bh(void)
2456 _rcu_barrier(&rcu_bh_state);
2458 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2461 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2463 void rcu_barrier_sched(void)
2465 _rcu_barrier(&rcu_sched_state);
2467 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2470 * Do boot-time initialization of a CPU's per-CPU RCU data.
2473 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2475 unsigned long flags;
2477 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2478 struct rcu_node *rnp = rcu_get_root(rsp);
2480 /* Set up local state, ensuring consistent view of global state. */
2481 raw_spin_lock_irqsave(&rnp->lock, flags);
2482 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2483 rdp->nxtlist = NULL;
2484 for (i = 0; i < RCU_NEXT_SIZE; i++)
2485 rdp->nxttail[i] = &rdp->nxtlist;
2488 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2489 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2490 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2493 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2497 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2498 * offline event can be happening at a given time. Note also that we
2499 * can accept some slop in the rsp->completed access due to the fact
2500 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2502 static void __cpuinit
2503 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2505 unsigned long flags;
2507 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2508 struct rcu_node *rnp = rcu_get_root(rsp);
2510 /* Set up local state, ensuring consistent view of global state. */
2511 raw_spin_lock_irqsave(&rnp->lock, flags);
2512 rdp->beenonline = 1; /* We have now been online. */
2513 rdp->preemptible = preemptible;
2514 rdp->qlen_last_fqs_check = 0;
2515 rdp->n_force_qs_snap = rsp->n_force_qs;
2516 rdp->blimit = blimit;
2517 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2518 atomic_set(&rdp->dynticks->dynticks,
2519 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2520 rcu_prepare_for_idle_init(cpu);
2521 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2524 * A new grace period might start here. If so, we won't be part
2525 * of it, but that is OK, as we are currently in a quiescent state.
2528 /* Exclude any attempts to start a new GP on large systems. */
2529 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2531 /* Add CPU to rcu_node bitmasks. */
2533 mask = rdp->grpmask;
2535 /* Exclude any attempts to start a new GP on small systems. */
2536 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2537 rnp->qsmaskinit |= mask;
2538 mask = rnp->grpmask;
2539 if (rnp == rdp->mynode) {
2541 * If there is a grace period in progress, we will
2542 * set up to wait for it next time we run the
2545 rdp->gpnum = rnp->completed;
2546 rdp->completed = rnp->completed;
2547 rdp->passed_quiesce = 0;
2548 rdp->qs_pending = 0;
2549 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2550 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2552 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2554 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2556 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2559 static void __cpuinit rcu_prepare_cpu(int cpu)
2561 struct rcu_state *rsp;
2563 for_each_rcu_flavor(rsp)
2564 rcu_init_percpu_data(cpu, rsp,
2565 strcmp(rsp->name, "rcu_preempt") == 0);
2569 * Handle CPU online/offline notification events.
2571 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2572 unsigned long action, void *hcpu)
2574 long cpu = (long)hcpu;
2575 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2576 struct rcu_node *rnp = rdp->mynode;
2577 struct rcu_state *rsp;
2579 trace_rcu_utilization("Start CPU hotplug");
2581 case CPU_UP_PREPARE:
2582 case CPU_UP_PREPARE_FROZEN:
2583 rcu_prepare_cpu(cpu);
2584 rcu_prepare_kthreads(cpu);
2587 case CPU_DOWN_FAILED:
2588 rcu_node_kthread_setaffinity(rnp, -1);
2589 rcu_cpu_kthread_setrt(cpu, 1);
2591 case CPU_DOWN_PREPARE:
2592 rcu_node_kthread_setaffinity(rnp, cpu);
2593 rcu_cpu_kthread_setrt(cpu, 0);
2596 case CPU_DYING_FROZEN:
2598 * The whole machine is "stopped" except this CPU, so we can
2599 * touch any data without introducing corruption. We send the
2600 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2602 for_each_rcu_flavor(rsp)
2603 rcu_cleanup_dying_cpu(rsp);
2604 rcu_cleanup_after_idle(cpu);
2607 case CPU_DEAD_FROZEN:
2608 case CPU_UP_CANCELED:
2609 case CPU_UP_CANCELED_FROZEN:
2610 for_each_rcu_flavor(rsp)
2611 rcu_cleanup_dead_cpu(cpu, rsp);
2616 trace_rcu_utilization("End CPU hotplug");
2621 * This function is invoked towards the end of the scheduler's initialization
2622 * process. Before this is called, the idle task might contain
2623 * RCU read-side critical sections (during which time, this idle
2624 * task is booting the system). After this function is called, the
2625 * idle tasks are prohibited from containing RCU read-side critical
2626 * sections. This function also enables RCU lockdep checking.
2628 void rcu_scheduler_starting(void)
2630 WARN_ON(num_online_cpus() != 1);
2631 WARN_ON(nr_context_switches() > 0);
2632 rcu_scheduler_active = 1;
2636 * Compute the per-level fanout, either using the exact fanout specified
2637 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2639 #ifdef CONFIG_RCU_FANOUT_EXACT
2640 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2644 for (i = rcu_num_lvls - 1; i > 0; i--)
2645 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2646 rsp->levelspread[0] = rcu_fanout_leaf;
2648 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2649 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2656 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2657 ccur = rsp->levelcnt[i];
2658 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2662 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2665 * Helper function for rcu_init() that initializes one rcu_state structure.
2667 static void __init rcu_init_one(struct rcu_state *rsp,
2668 struct rcu_data __percpu *rda)
2670 static char *buf[] = { "rcu_node_level_0",
2673 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2677 struct rcu_node *rnp;
2679 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2681 /* Initialize the level-tracking arrays. */
2683 for (i = 0; i < rcu_num_lvls; i++)
2684 rsp->levelcnt[i] = num_rcu_lvl[i];
2685 for (i = 1; i < rcu_num_lvls; i++)
2686 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2687 rcu_init_levelspread(rsp);
2689 /* Initialize the elements themselves, starting from the leaves. */
2691 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2692 cpustride *= rsp->levelspread[i];
2693 rnp = rsp->level[i];
2694 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2695 raw_spin_lock_init(&rnp->lock);
2696 lockdep_set_class_and_name(&rnp->lock,
2697 &rcu_node_class[i], buf[i]);
2700 rnp->qsmaskinit = 0;
2701 rnp->grplo = j * cpustride;
2702 rnp->grphi = (j + 1) * cpustride - 1;
2703 if (rnp->grphi >= NR_CPUS)
2704 rnp->grphi = NR_CPUS - 1;
2710 rnp->grpnum = j % rsp->levelspread[i - 1];
2711 rnp->grpmask = 1UL << rnp->grpnum;
2712 rnp->parent = rsp->level[i - 1] +
2713 j / rsp->levelspread[i - 1];
2716 INIT_LIST_HEAD(&rnp->blkd_tasks);
2721 rnp = rsp->level[rcu_num_lvls - 1];
2722 for_each_possible_cpu(i) {
2723 while (i > rnp->grphi)
2725 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2726 rcu_boot_init_percpu_data(i, rsp);
2728 list_add(&rsp->flavors, &rcu_struct_flavors);
2732 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2733 * replace the definitions in rcutree.h because those are needed to size
2734 * the ->node array in the rcu_state structure.
2736 static void __init rcu_init_geometry(void)
2741 int rcu_capacity[MAX_RCU_LVLS + 1];
2743 /* If the compile-time values are accurate, just leave. */
2744 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2748 * Compute number of nodes that can be handled an rcu_node tree
2749 * with the given number of levels. Setting rcu_capacity[0] makes
2750 * some of the arithmetic easier.
2752 rcu_capacity[0] = 1;
2753 rcu_capacity[1] = rcu_fanout_leaf;
2754 for (i = 2; i <= MAX_RCU_LVLS; i++)
2755 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2758 * The boot-time rcu_fanout_leaf parameter is only permitted
2759 * to increase the leaf-level fanout, not decrease it. Of course,
2760 * the leaf-level fanout cannot exceed the number of bits in
2761 * the rcu_node masks. Finally, the tree must be able to accommodate
2762 * the configured number of CPUs. Complain and fall back to the
2763 * compile-time values if these limits are exceeded.
2765 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2766 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2767 n > rcu_capacity[MAX_RCU_LVLS]) {
2772 /* Calculate the number of rcu_nodes at each level of the tree. */
2773 for (i = 1; i <= MAX_RCU_LVLS; i++)
2774 if (n <= rcu_capacity[i]) {
2775 for (j = 0; j <= i; j++)
2777 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2779 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2784 /* Calculate the total number of rcu_node structures. */
2786 for (i = 0; i <= MAX_RCU_LVLS; i++)
2787 rcu_num_nodes += num_rcu_lvl[i];
2791 void __init rcu_init(void)
2795 rcu_bootup_announce();
2796 rcu_init_geometry();
2797 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2798 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2799 __rcu_init_preempt();
2800 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2803 * We don't need protection against CPU-hotplug here because
2804 * this is called early in boot, before either interrupts
2805 * or the scheduler are operational.
2807 cpu_notifier(rcu_cpu_notify, 0);
2808 for_each_online_cpu(cpu)
2809 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2810 check_cpu_stall_init();
2813 #include "rcutree_plugin.h"