2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
58 #include <trace/events/rcu.h>
62 /* Data structures. */
64 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
65 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
67 #define RCU_STATE_INITIALIZER(sname, cr) { \
68 .level = { &sname##_state.node[0] }, \
70 .fqs_state = RCU_GP_IDLE, \
73 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
74 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
75 .orphan_donetail = &sname##_state.orphan_donelist, \
76 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
80 struct rcu_state rcu_sched_state =
81 RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
82 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
84 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
85 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
87 static struct rcu_state *rcu_state;
88 LIST_HEAD(rcu_struct_flavors);
90 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
91 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
92 module_param(rcu_fanout_leaf, int, 0444);
93 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
94 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
101 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
104 * The rcu_scheduler_active variable transitions from zero to one just
105 * before the first task is spawned. So when this variable is zero, RCU
106 * can assume that there is but one task, allowing RCU to (for example)
107 * optimized synchronize_sched() to a simple barrier(). When this variable
108 * is one, RCU must actually do all the hard work required to detect real
109 * grace periods. This variable is also used to suppress boot-time false
110 * positives from lockdep-RCU error checking.
112 int rcu_scheduler_active __read_mostly;
113 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
116 * The rcu_scheduler_fully_active variable transitions from zero to one
117 * during the early_initcall() processing, which is after the scheduler
118 * is capable of creating new tasks. So RCU processing (for example,
119 * creating tasks for RCU priority boosting) must be delayed until after
120 * rcu_scheduler_fully_active transitions from zero to one. We also
121 * currently delay invocation of any RCU callbacks until after this point.
123 * It might later prove better for people registering RCU callbacks during
124 * early boot to take responsibility for these callbacks, but one step at
127 static int rcu_scheduler_fully_active __read_mostly;
129 #ifdef CONFIG_RCU_BOOST
132 * Control variables for per-CPU and per-rcu_node kthreads. These
133 * handle all flavors of RCU.
135 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
136 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
137 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
138 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
139 DEFINE_PER_CPU(char, rcu_cpu_has_work);
141 #endif /* #ifdef CONFIG_RCU_BOOST */
143 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
144 static void invoke_rcu_core(void);
145 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
148 * Track the rcutorture test sequence number and the update version
149 * number within a given test. The rcutorture_testseq is incremented
150 * on every rcutorture module load and unload, so has an odd value
151 * when a test is running. The rcutorture_vernum is set to zero
152 * when rcutorture starts and is incremented on each rcutorture update.
153 * These variables enable correlating rcutorture output with the
154 * RCU tracing information.
156 unsigned long rcutorture_testseq;
157 unsigned long rcutorture_vernum;
160 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
161 * permit this function to be invoked without holding the root rcu_node
162 * structure's ->lock, but of course results can be subject to change.
164 static int rcu_gp_in_progress(struct rcu_state *rsp)
166 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
170 * Note a quiescent state. Because we do not need to know
171 * how many quiescent states passed, just if there was at least
172 * one since the start of the grace period, this just sets a flag.
173 * The caller must have disabled preemption.
175 void rcu_sched_qs(int cpu)
177 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
179 rdp->passed_quiesce_gpnum = rdp->gpnum;
181 if (rdp->passed_quiesce == 0)
182 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
183 rdp->passed_quiesce = 1;
186 void rcu_bh_qs(int cpu)
188 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
190 rdp->passed_quiesce_gpnum = rdp->gpnum;
192 if (rdp->passed_quiesce == 0)
193 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
194 rdp->passed_quiesce = 1;
198 * Note a context switch. This is a quiescent state for RCU-sched,
199 * and requires special handling for preemptible RCU.
200 * The caller must have disabled preemption.
202 void rcu_note_context_switch(int cpu)
204 trace_rcu_utilization("Start context switch");
206 rcu_preempt_note_context_switch(cpu);
207 trace_rcu_utilization("End context switch");
209 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
211 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
212 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
213 .dynticks = ATOMIC_INIT(1),
216 static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
217 static int qhimark = 10000; /* If this many pending, ignore blimit. */
218 static int qlowmark = 100; /* Once only this many pending, use blimit. */
220 module_param(blimit, int, 0444);
221 module_param(qhimark, int, 0444);
222 module_param(qlowmark, int, 0444);
224 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
225 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
227 module_param(rcu_cpu_stall_suppress, int, 0644);
228 module_param(rcu_cpu_stall_timeout, int, 0644);
230 static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
231 static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
233 module_param(jiffies_till_first_fqs, ulong, 0644);
234 module_param(jiffies_till_next_fqs, ulong, 0644);
236 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
237 static void force_quiescent_state(struct rcu_state *rsp);
238 static int rcu_pending(int cpu);
241 * Return the number of RCU-sched batches processed thus far for debug & stats.
243 long rcu_batches_completed_sched(void)
245 return rcu_sched_state.completed;
247 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
250 * Return the number of RCU BH batches processed thus far for debug & stats.
252 long rcu_batches_completed_bh(void)
254 return rcu_bh_state.completed;
256 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
259 * Force a quiescent state for RCU BH.
261 void rcu_bh_force_quiescent_state(void)
263 force_quiescent_state(&rcu_bh_state);
265 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
268 * Record the number of times rcutorture tests have been initiated and
269 * terminated. This information allows the debugfs tracing stats to be
270 * correlated to the rcutorture messages, even when the rcutorture module
271 * is being repeatedly loaded and unloaded. In other words, we cannot
272 * store this state in rcutorture itself.
274 void rcutorture_record_test_transition(void)
276 rcutorture_testseq++;
277 rcutorture_vernum = 0;
279 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
282 * Record the number of writer passes through the current rcutorture test.
283 * This is also used to correlate debugfs tracing stats with the rcutorture
286 void rcutorture_record_progress(unsigned long vernum)
290 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
293 * Force a quiescent state for RCU-sched.
295 void rcu_sched_force_quiescent_state(void)
297 force_quiescent_state(&rcu_sched_state);
299 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
302 * Does the CPU have callbacks ready to be invoked?
305 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
307 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
311 * Does the current CPU require a yet-as-unscheduled grace period?
314 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
316 return *rdp->nxttail[RCU_DONE_TAIL +
317 ACCESS_ONCE(rsp->completed) != rdp->completed] &&
318 !rcu_gp_in_progress(rsp);
322 * Return the root node of the specified rcu_state structure.
324 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
326 return &rsp->node[0];
330 * If the specified CPU is offline, tell the caller that it is in
331 * a quiescent state. Otherwise, whack it with a reschedule IPI.
332 * Grace periods can end up waiting on an offline CPU when that
333 * CPU is in the process of coming online -- it will be added to the
334 * rcu_node bitmasks before it actually makes it online. The same thing
335 * can happen while a CPU is in the process of coming online. Because this
336 * race is quite rare, we check for it after detecting that the grace
337 * period has been delayed rather than checking each and every CPU
338 * each and every time we start a new grace period.
340 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
343 * If the CPU is offline for more than a jiffy, it is in a quiescent
344 * state. We can trust its state not to change because interrupts
345 * are disabled. The reason for the jiffy's worth of slack is to
346 * handle CPUs initializing on the way up and finding their way
347 * to the idle loop on the way down.
349 if (cpu_is_offline(rdp->cpu) &&
350 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
351 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
359 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
361 * If the new value of the ->dynticks_nesting counter now is zero,
362 * we really have entered idle, and must do the appropriate accounting.
363 * The caller must have disabled interrupts.
365 static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
367 trace_rcu_dyntick("Start", oldval, 0);
368 if (!is_idle_task(current)) {
369 struct task_struct *idle = idle_task(smp_processor_id());
371 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
372 ftrace_dump(DUMP_ORIG);
373 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
374 current->pid, current->comm,
375 idle->pid, idle->comm); /* must be idle task! */
377 rcu_prepare_for_idle(smp_processor_id());
378 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
379 smp_mb__before_atomic_inc(); /* See above. */
380 atomic_inc(&rdtp->dynticks);
381 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
382 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
385 * The idle task is not permitted to enter the idle loop while
386 * in an RCU read-side critical section.
388 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
389 "Illegal idle entry in RCU read-side critical section.");
390 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
391 "Illegal idle entry in RCU-bh read-side critical section.");
392 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
393 "Illegal idle entry in RCU-sched read-side critical section.");
397 * rcu_idle_enter - inform RCU that current CPU is entering idle
399 * Enter idle mode, in other words, -leave- the mode in which RCU
400 * read-side critical sections can occur. (Though RCU read-side
401 * critical sections can occur in irq handlers in idle, a possibility
402 * handled by irq_enter() and irq_exit().)
404 * We crowbar the ->dynticks_nesting field to zero to allow for
405 * the possibility of usermode upcalls having messed up our count
406 * of interrupt nesting level during the prior busy period.
408 void rcu_idle_enter(void)
412 struct rcu_dynticks *rdtp;
414 local_irq_save(flags);
415 rdtp = &__get_cpu_var(rcu_dynticks);
416 oldval = rdtp->dynticks_nesting;
417 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
418 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
419 rdtp->dynticks_nesting = 0;
421 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
422 rcu_idle_enter_common(rdtp, oldval);
423 local_irq_restore(flags);
425 EXPORT_SYMBOL_GPL(rcu_idle_enter);
428 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
430 * Exit from an interrupt handler, which might possibly result in entering
431 * idle mode, in other words, leaving the mode in which read-side critical
432 * sections can occur.
434 * This code assumes that the idle loop never does anything that might
435 * result in unbalanced calls to irq_enter() and irq_exit(). If your
436 * architecture violates this assumption, RCU will give you what you
437 * deserve, good and hard. But very infrequently and irreproducibly.
439 * Use things like work queues to work around this limitation.
441 * You have been warned.
443 void rcu_irq_exit(void)
447 struct rcu_dynticks *rdtp;
449 local_irq_save(flags);
450 rdtp = &__get_cpu_var(rcu_dynticks);
451 oldval = rdtp->dynticks_nesting;
452 rdtp->dynticks_nesting--;
453 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
454 if (rdtp->dynticks_nesting)
455 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
457 rcu_idle_enter_common(rdtp, oldval);
458 local_irq_restore(flags);
462 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
464 * If the new value of the ->dynticks_nesting counter was previously zero,
465 * we really have exited idle, and must do the appropriate accounting.
466 * The caller must have disabled interrupts.
468 static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
470 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
471 atomic_inc(&rdtp->dynticks);
472 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
473 smp_mb__after_atomic_inc(); /* See above. */
474 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
475 rcu_cleanup_after_idle(smp_processor_id());
476 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
477 if (!is_idle_task(current)) {
478 struct task_struct *idle = idle_task(smp_processor_id());
480 trace_rcu_dyntick("Error on exit: not idle task",
481 oldval, rdtp->dynticks_nesting);
482 ftrace_dump(DUMP_ORIG);
483 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
484 current->pid, current->comm,
485 idle->pid, idle->comm); /* must be idle task! */
490 * rcu_idle_exit - inform RCU that current CPU is leaving idle
492 * Exit idle mode, in other words, -enter- the mode in which RCU
493 * read-side critical sections can occur.
495 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
496 * allow for the possibility of usermode upcalls messing up our count
497 * of interrupt nesting level during the busy period that is just
500 void rcu_idle_exit(void)
503 struct rcu_dynticks *rdtp;
506 local_irq_save(flags);
507 rdtp = &__get_cpu_var(rcu_dynticks);
508 oldval = rdtp->dynticks_nesting;
509 WARN_ON_ONCE(oldval < 0);
510 if (oldval & DYNTICK_TASK_NEST_MASK)
511 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
513 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
514 rcu_idle_exit_common(rdtp, oldval);
515 local_irq_restore(flags);
517 EXPORT_SYMBOL_GPL(rcu_idle_exit);
520 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
522 * Enter an interrupt handler, which might possibly result in exiting
523 * idle mode, in other words, entering the mode in which read-side critical
524 * sections can occur.
526 * Note that the Linux kernel is fully capable of entering an interrupt
527 * handler that it never exits, for example when doing upcalls to
528 * user mode! This code assumes that the idle loop never does upcalls to
529 * user mode. If your architecture does do upcalls from the idle loop (or
530 * does anything else that results in unbalanced calls to the irq_enter()
531 * and irq_exit() functions), RCU will give you what you deserve, good
532 * and hard. But very infrequently and irreproducibly.
534 * Use things like work queues to work around this limitation.
536 * You have been warned.
538 void rcu_irq_enter(void)
541 struct rcu_dynticks *rdtp;
544 local_irq_save(flags);
545 rdtp = &__get_cpu_var(rcu_dynticks);
546 oldval = rdtp->dynticks_nesting;
547 rdtp->dynticks_nesting++;
548 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
550 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
552 rcu_idle_exit_common(rdtp, oldval);
553 local_irq_restore(flags);
557 * rcu_nmi_enter - inform RCU of entry to NMI context
559 * If the CPU was idle with dynamic ticks active, and there is no
560 * irq handler running, this updates rdtp->dynticks_nmi to let the
561 * RCU grace-period handling know that the CPU is active.
563 void rcu_nmi_enter(void)
565 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
567 if (rdtp->dynticks_nmi_nesting == 0 &&
568 (atomic_read(&rdtp->dynticks) & 0x1))
570 rdtp->dynticks_nmi_nesting++;
571 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
572 atomic_inc(&rdtp->dynticks);
573 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
574 smp_mb__after_atomic_inc(); /* See above. */
575 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
579 * rcu_nmi_exit - inform RCU of exit from NMI context
581 * If the CPU was idle with dynamic ticks active, and there is no
582 * irq handler running, this updates rdtp->dynticks_nmi to let the
583 * RCU grace-period handling know that the CPU is no longer active.
585 void rcu_nmi_exit(void)
587 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
589 if (rdtp->dynticks_nmi_nesting == 0 ||
590 --rdtp->dynticks_nmi_nesting != 0)
592 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
593 smp_mb__before_atomic_inc(); /* See above. */
594 atomic_inc(&rdtp->dynticks);
595 smp_mb__after_atomic_inc(); /* Force delay to next write. */
596 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
600 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
602 * If the current CPU is in its idle loop and is neither in an interrupt
603 * or NMI handler, return true.
605 int rcu_is_cpu_idle(void)
610 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
614 EXPORT_SYMBOL(rcu_is_cpu_idle);
616 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
619 * Is the current CPU online? Disable preemption to avoid false positives
620 * that could otherwise happen due to the current CPU number being sampled,
621 * this task being preempted, its old CPU being taken offline, resuming
622 * on some other CPU, then determining that its old CPU is now offline.
623 * It is OK to use RCU on an offline processor during initial boot, hence
624 * the check for rcu_scheduler_fully_active. Note also that it is OK
625 * for a CPU coming online to use RCU for one jiffy prior to marking itself
626 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
627 * offline to continue to use RCU for one jiffy after marking itself
628 * offline in the cpu_online_mask. This leniency is necessary given the
629 * non-atomic nature of the online and offline processing, for example,
630 * the fact that a CPU enters the scheduler after completing the CPU_DYING
633 * This is also why RCU internally marks CPUs online during the
634 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
636 * Disable checking if in an NMI handler because we cannot safely report
637 * errors from NMI handlers anyway.
639 bool rcu_lockdep_current_cpu_online(void)
641 struct rcu_data *rdp;
642 struct rcu_node *rnp;
648 rdp = &__get_cpu_var(rcu_sched_data);
650 ret = (rdp->grpmask & rnp->qsmaskinit) ||
651 !rcu_scheduler_fully_active;
655 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
657 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
660 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
662 * If the current CPU is idle or running at a first-level (not nested)
663 * interrupt from idle, return true. The caller must have at least
664 * disabled preemption.
666 int rcu_is_cpu_rrupt_from_idle(void)
668 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
672 * Snapshot the specified CPU's dynticks counter so that we can later
673 * credit them with an implicit quiescent state. Return 1 if this CPU
674 * is in dynticks idle mode, which is an extended quiescent state.
676 static int dyntick_save_progress_counter(struct rcu_data *rdp)
678 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
679 return (rdp->dynticks_snap & 0x1) == 0;
683 * Return true if the specified CPU has passed through a quiescent
684 * state by virtue of being in or having passed through an dynticks
685 * idle state since the last call to dyntick_save_progress_counter()
688 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
693 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
694 snap = (unsigned int)rdp->dynticks_snap;
697 * If the CPU passed through or entered a dynticks idle phase with
698 * no active irq/NMI handlers, then we can safely pretend that the CPU
699 * already acknowledged the request to pass through a quiescent
700 * state. Either way, that CPU cannot possibly be in an RCU
701 * read-side critical section that started before the beginning
702 * of the current RCU grace period.
704 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
705 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
710 /* Go check for the CPU being offline. */
711 return rcu_implicit_offline_qs(rdp);
714 static int jiffies_till_stall_check(void)
716 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
719 * Limit check must be consistent with the Kconfig limits
720 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
722 if (till_stall_check < 3) {
723 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
724 till_stall_check = 3;
725 } else if (till_stall_check > 300) {
726 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
727 till_stall_check = 300;
729 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
732 static void record_gp_stall_check_time(struct rcu_state *rsp)
734 rsp->gp_start = jiffies;
735 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
738 static void print_other_cpu_stall(struct rcu_state *rsp)
744 struct rcu_node *rnp = rcu_get_root(rsp);
746 /* Only let one CPU complain about others per time interval. */
748 raw_spin_lock_irqsave(&rnp->lock, flags);
749 delta = jiffies - rsp->jiffies_stall;
750 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
751 raw_spin_unlock_irqrestore(&rnp->lock, flags);
754 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
755 raw_spin_unlock_irqrestore(&rnp->lock, flags);
758 * OK, time to rat on our buddy...
759 * See Documentation/RCU/stallwarn.txt for info on how to debug
760 * RCU CPU stall warnings.
762 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
764 print_cpu_stall_info_begin();
765 rcu_for_each_leaf_node(rsp, rnp) {
766 raw_spin_lock_irqsave(&rnp->lock, flags);
767 ndetected += rcu_print_task_stall(rnp);
768 raw_spin_unlock_irqrestore(&rnp->lock, flags);
769 if (rnp->qsmask == 0)
771 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
772 if (rnp->qsmask & (1UL << cpu)) {
773 print_cpu_stall_info(rsp, rnp->grplo + cpu);
779 * Now rat on any tasks that got kicked up to the root rcu_node
780 * due to CPU offlining.
782 rnp = rcu_get_root(rsp);
783 raw_spin_lock_irqsave(&rnp->lock, flags);
784 ndetected += rcu_print_task_stall(rnp);
785 raw_spin_unlock_irqrestore(&rnp->lock, flags);
787 print_cpu_stall_info_end();
788 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
789 smp_processor_id(), (long)(jiffies - rsp->gp_start));
791 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
792 else if (!trigger_all_cpu_backtrace())
795 /* Complain about tasks blocking the grace period. */
797 rcu_print_detail_task_stall(rsp);
799 force_quiescent_state(rsp); /* Kick them all. */
802 static void print_cpu_stall(struct rcu_state *rsp)
805 struct rcu_node *rnp = rcu_get_root(rsp);
808 * OK, time to rat on ourselves...
809 * See Documentation/RCU/stallwarn.txt for info on how to debug
810 * RCU CPU stall warnings.
812 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
813 print_cpu_stall_info_begin();
814 print_cpu_stall_info(rsp, smp_processor_id());
815 print_cpu_stall_info_end();
816 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
817 if (!trigger_all_cpu_backtrace())
820 raw_spin_lock_irqsave(&rnp->lock, flags);
821 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
822 rsp->jiffies_stall = jiffies +
823 3 * jiffies_till_stall_check() + 3;
824 raw_spin_unlock_irqrestore(&rnp->lock, flags);
826 set_need_resched(); /* kick ourselves to get things going. */
829 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
833 struct rcu_node *rnp;
835 if (rcu_cpu_stall_suppress)
837 j = ACCESS_ONCE(jiffies);
838 js = ACCESS_ONCE(rsp->jiffies_stall);
840 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
842 /* We haven't checked in, so go dump stack. */
843 print_cpu_stall(rsp);
845 } else if (rcu_gp_in_progress(rsp) &&
846 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
848 /* They had a few time units to dump stack, so complain. */
849 print_other_cpu_stall(rsp);
853 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
855 rcu_cpu_stall_suppress = 1;
860 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
862 * Set the stall-warning timeout way off into the future, thus preventing
863 * any RCU CPU stall-warning messages from appearing in the current set of
866 * The caller must disable hard irqs.
868 void rcu_cpu_stall_reset(void)
870 struct rcu_state *rsp;
872 for_each_rcu_flavor(rsp)
873 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
876 static struct notifier_block rcu_panic_block = {
877 .notifier_call = rcu_panic,
880 static void __init check_cpu_stall_init(void)
882 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
886 * Update CPU-local rcu_data state to record the newly noticed grace period.
887 * This is used both when we started the grace period and when we notice
888 * that someone else started the grace period. The caller must hold the
889 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
890 * and must have irqs disabled.
892 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
894 if (rdp->gpnum != rnp->gpnum) {
896 * If the current grace period is waiting for this CPU,
897 * set up to detect a quiescent state, otherwise don't
898 * go looking for one.
900 rdp->gpnum = rnp->gpnum;
901 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
902 if (rnp->qsmask & rdp->grpmask) {
904 rdp->passed_quiesce = 0;
908 zero_cpu_stall_ticks(rdp);
912 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
915 struct rcu_node *rnp;
917 local_irq_save(flags);
919 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
920 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
921 local_irq_restore(flags);
924 __note_new_gpnum(rsp, rnp, rdp);
925 raw_spin_unlock_irqrestore(&rnp->lock, flags);
929 * Did someone else start a new RCU grace period start since we last
930 * checked? Update local state appropriately if so. Must be called
931 * on the CPU corresponding to rdp.
934 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
939 local_irq_save(flags);
940 if (rdp->gpnum != rsp->gpnum) {
941 note_new_gpnum(rsp, rdp);
944 local_irq_restore(flags);
949 * Initialize the specified rcu_data structure's callback list to empty.
951 static void init_callback_list(struct rcu_data *rdp)
956 for (i = 0; i < RCU_NEXT_SIZE; i++)
957 rdp->nxttail[i] = &rdp->nxtlist;
961 * Advance this CPU's callbacks, but only if the current grace period
962 * has ended. This may be called only from the CPU to whom the rdp
963 * belongs. In addition, the corresponding leaf rcu_node structure's
964 * ->lock must be held by the caller, with irqs disabled.
967 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
969 /* Did another grace period end? */
970 if (rdp->completed != rnp->completed) {
972 /* Advance callbacks. No harm if list empty. */
973 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
974 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
975 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
977 /* Remember that we saw this grace-period completion. */
978 rdp->completed = rnp->completed;
979 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
982 * If we were in an extended quiescent state, we may have
983 * missed some grace periods that others CPUs handled on
984 * our behalf. Catch up with this state to avoid noting
985 * spurious new grace periods. If another grace period
986 * has started, then rnp->gpnum will have advanced, so
987 * we will detect this later on.
989 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
990 rdp->gpnum = rdp->completed;
993 * If RCU does not need a quiescent state from this CPU,
994 * then make sure that this CPU doesn't go looking for one.
996 if ((rnp->qsmask & rdp->grpmask) == 0)
1002 * Advance this CPU's callbacks, but only if the current grace period
1003 * has ended. This may be called only from the CPU to whom the rdp
1007 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1009 unsigned long flags;
1010 struct rcu_node *rnp;
1012 local_irq_save(flags);
1014 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1015 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1016 local_irq_restore(flags);
1019 __rcu_process_gp_end(rsp, rnp, rdp);
1020 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1024 * Do per-CPU grace-period initialization for running CPU. The caller
1025 * must hold the lock of the leaf rcu_node structure corresponding to
1029 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1031 /* Prior grace period ended, so advance callbacks for current CPU. */
1032 __rcu_process_gp_end(rsp, rnp, rdp);
1034 /* Set state so that this CPU will detect the next quiescent state. */
1035 __note_new_gpnum(rsp, rnp, rdp);
1039 * Initialize a new grace period.
1041 static int rcu_gp_init(struct rcu_state *rsp)
1043 struct rcu_data *rdp;
1044 struct rcu_node *rnp = rcu_get_root(rsp);
1046 raw_spin_lock_irq(&rnp->lock);
1047 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1049 if (rcu_gp_in_progress(rsp)) {
1050 /* Grace period already in progress, don't start another. */
1051 raw_spin_unlock_irq(&rnp->lock);
1055 /* Advance to a new grace period and initialize state. */
1057 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1058 record_gp_stall_check_time(rsp);
1059 raw_spin_unlock_irq(&rnp->lock);
1061 /* Exclude any concurrent CPU-hotplug operations. */
1065 * Set the quiescent-state-needed bits in all the rcu_node
1066 * structures for all currently online CPUs in breadth-first order,
1067 * starting from the root rcu_node structure, relying on the layout
1068 * of the tree within the rsp->node[] array. Note that other CPUs
1069 * will access only the leaves of the hierarchy, thus seeing that no
1070 * grace period is in progress, at least until the corresponding
1071 * leaf node has been initialized. In addition, we have excluded
1072 * CPU-hotplug operations.
1074 * The grace period cannot complete until the initialization
1075 * process finishes, because this kthread handles both.
1077 rcu_for_each_node_breadth_first(rsp, rnp) {
1078 raw_spin_lock_irq(&rnp->lock);
1079 rdp = this_cpu_ptr(rsp->rda);
1080 rcu_preempt_check_blocked_tasks(rnp);
1081 rnp->qsmask = rnp->qsmaskinit;
1082 rnp->gpnum = rsp->gpnum;
1083 rnp->completed = rsp->completed;
1084 if (rnp == rdp->mynode)
1085 rcu_start_gp_per_cpu(rsp, rnp, rdp);
1086 rcu_preempt_boost_start_gp(rnp);
1087 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1088 rnp->level, rnp->grplo,
1089 rnp->grphi, rnp->qsmask);
1090 raw_spin_unlock_irq(&rnp->lock);
1091 #ifdef CONFIG_PROVE_RCU_DELAY
1092 if ((random32() % (rcu_num_nodes * 8)) == 0)
1093 schedule_timeout_uninterruptible(2);
1094 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1103 * Do one round of quiescent-state forcing.
1105 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1107 int fqs_state = fqs_state_in;
1108 struct rcu_node *rnp = rcu_get_root(rsp);
1111 if (fqs_state == RCU_SAVE_DYNTICK) {
1112 /* Collect dyntick-idle snapshots. */
1113 force_qs_rnp(rsp, dyntick_save_progress_counter);
1114 fqs_state = RCU_FORCE_QS;
1116 /* Handle dyntick-idle and offline CPUs. */
1117 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1119 /* Clear flag to prevent immediate re-entry. */
1120 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1121 raw_spin_lock_irq(&rnp->lock);
1122 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1123 raw_spin_unlock_irq(&rnp->lock);
1129 * Clean up after the old grace period.
1131 static void rcu_gp_cleanup(struct rcu_state *rsp)
1133 unsigned long gp_duration;
1134 struct rcu_data *rdp;
1135 struct rcu_node *rnp = rcu_get_root(rsp);
1137 raw_spin_lock_irq(&rnp->lock);
1138 gp_duration = jiffies - rsp->gp_start;
1139 if (gp_duration > rsp->gp_max)
1140 rsp->gp_max = gp_duration;
1143 * We know the grace period is complete, but to everyone else
1144 * it appears to still be ongoing. But it is also the case
1145 * that to everyone else it looks like there is nothing that
1146 * they can do to advance the grace period. It is therefore
1147 * safe for us to drop the lock in order to mark the grace
1148 * period as completed in all of the rcu_node structures.
1150 raw_spin_unlock_irq(&rnp->lock);
1153 * Propagate new ->completed value to rcu_node structures so
1154 * that other CPUs don't have to wait until the start of the next
1155 * grace period to process their callbacks. This also avoids
1156 * some nasty RCU grace-period initialization races by forcing
1157 * the end of the current grace period to be completely recorded in
1158 * all of the rcu_node structures before the beginning of the next
1159 * grace period is recorded in any of the rcu_node structures.
1161 rcu_for_each_node_breadth_first(rsp, rnp) {
1162 raw_spin_lock_irq(&rnp->lock);
1163 rnp->completed = rsp->gpnum;
1164 raw_spin_unlock_irq(&rnp->lock);
1167 rnp = rcu_get_root(rsp);
1168 raw_spin_lock_irq(&rnp->lock);
1170 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1171 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1172 rsp->fqs_state = RCU_GP_IDLE;
1173 rdp = this_cpu_ptr(rsp->rda);
1174 if (cpu_needs_another_gp(rsp, rdp))
1176 raw_spin_unlock_irq(&rnp->lock);
1180 * Body of kthread that handles grace periods.
1182 static int __noreturn rcu_gp_kthread(void *arg)
1187 struct rcu_state *rsp = arg;
1188 struct rcu_node *rnp = rcu_get_root(rsp);
1192 /* Handle grace-period start. */
1194 wait_event_interruptible(rsp->gp_wq,
1197 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1201 flush_signals(current);
1204 /* Handle quiescent-state forcing. */
1205 fqs_state = RCU_SAVE_DYNTICK;
1206 j = jiffies_till_first_fqs;
1209 jiffies_till_first_fqs = HZ;
1212 rsp->jiffies_force_qs = jiffies + j;
1213 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1214 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1215 (!ACCESS_ONCE(rnp->qsmask) &&
1216 !rcu_preempt_blocked_readers_cgp(rnp)),
1218 /* If grace period done, leave loop. */
1219 if (!ACCESS_ONCE(rnp->qsmask) &&
1220 !rcu_preempt_blocked_readers_cgp(rnp))
1222 /* If time for quiescent-state forcing, do it. */
1223 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1224 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1227 /* Deal with stray signal. */
1229 flush_signals(current);
1231 j = jiffies_till_next_fqs;
1234 jiffies_till_next_fqs = HZ;
1237 jiffies_till_next_fqs = 1;
1241 /* Handle grace-period end. */
1242 rcu_gp_cleanup(rsp);
1247 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1248 * in preparation for detecting the next grace period. The caller must hold
1249 * the root node's ->lock, which is released before return. Hard irqs must
1252 * Note that it is legal for a dying CPU (which is marked as offline) to
1253 * invoke this function. This can happen when the dying CPU reports its
1257 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1258 __releases(rcu_get_root(rsp)->lock)
1260 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1261 struct rcu_node *rnp = rcu_get_root(rsp);
1263 if (!rsp->gp_kthread ||
1264 !cpu_needs_another_gp(rsp, rdp)) {
1266 * Either we have not yet spawned the grace-period
1267 * task or this CPU does not need another grace period.
1268 * Either way, don't start a new grace period.
1270 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1274 rsp->gp_flags = RCU_GP_FLAG_INIT;
1275 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1276 wake_up(&rsp->gp_wq);
1280 * Report a full set of quiescent states to the specified rcu_state
1281 * data structure. This involves cleaning up after the prior grace
1282 * period and letting rcu_start_gp() start up the next grace period
1283 * if one is needed. Note that the caller must hold rnp->lock, as
1284 * required by rcu_start_gp(), which will release it.
1286 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1287 __releases(rcu_get_root(rsp)->lock)
1289 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1290 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1291 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1295 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1296 * Allows quiescent states for a group of CPUs to be reported at one go
1297 * to the specified rcu_node structure, though all the CPUs in the group
1298 * must be represented by the same rcu_node structure (which need not be
1299 * a leaf rcu_node structure, though it often will be). That structure's
1300 * lock must be held upon entry, and it is released before return.
1303 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1304 struct rcu_node *rnp, unsigned long flags)
1305 __releases(rnp->lock)
1307 struct rcu_node *rnp_c;
1309 /* Walk up the rcu_node hierarchy. */
1311 if (!(rnp->qsmask & mask)) {
1313 /* Our bit has already been cleared, so done. */
1314 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1317 rnp->qsmask &= ~mask;
1318 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1319 mask, rnp->qsmask, rnp->level,
1320 rnp->grplo, rnp->grphi,
1322 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1324 /* Other bits still set at this level, so done. */
1325 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1328 mask = rnp->grpmask;
1329 if (rnp->parent == NULL) {
1331 /* No more levels. Exit loop holding root lock. */
1335 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1338 raw_spin_lock_irqsave(&rnp->lock, flags);
1339 WARN_ON_ONCE(rnp_c->qsmask);
1343 * Get here if we are the last CPU to pass through a quiescent
1344 * state for this grace period. Invoke rcu_report_qs_rsp()
1345 * to clean up and start the next grace period if one is needed.
1347 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1351 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1352 * structure. This must be either called from the specified CPU, or
1353 * called when the specified CPU is known to be offline (and when it is
1354 * also known that no other CPU is concurrently trying to help the offline
1355 * CPU). The lastcomp argument is used to make sure we are still in the
1356 * grace period of interest. We don't want to end the current grace period
1357 * based on quiescent states detected in an earlier grace period!
1360 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1362 unsigned long flags;
1364 struct rcu_node *rnp;
1367 raw_spin_lock_irqsave(&rnp->lock, flags);
1368 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1371 * The grace period in which this quiescent state was
1372 * recorded has ended, so don't report it upwards.
1373 * We will instead need a new quiescent state that lies
1374 * within the current grace period.
1376 rdp->passed_quiesce = 0; /* need qs for new gp. */
1377 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1380 mask = rdp->grpmask;
1381 if ((rnp->qsmask & mask) == 0) {
1382 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1384 rdp->qs_pending = 0;
1387 * This GP can't end until cpu checks in, so all of our
1388 * callbacks can be processed during the next GP.
1390 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1392 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1397 * Check to see if there is a new grace period of which this CPU
1398 * is not yet aware, and if so, set up local rcu_data state for it.
1399 * Otherwise, see if this CPU has just passed through its first
1400 * quiescent state for this grace period, and record that fact if so.
1403 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1405 /* If there is now a new grace period, record and return. */
1406 if (check_for_new_grace_period(rsp, rdp))
1410 * Does this CPU still need to do its part for current grace period?
1411 * If no, return and let the other CPUs do their part as well.
1413 if (!rdp->qs_pending)
1417 * Was there a quiescent state since the beginning of the grace
1418 * period? If no, then exit and wait for the next call.
1420 if (!rdp->passed_quiesce)
1424 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1427 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1430 #ifdef CONFIG_HOTPLUG_CPU
1433 * Send the specified CPU's RCU callbacks to the orphanage. The
1434 * specified CPU must be offline, and the caller must hold the
1438 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1439 struct rcu_node *rnp, struct rcu_data *rdp)
1442 * Orphan the callbacks. First adjust the counts. This is safe
1443 * because ->onofflock excludes _rcu_barrier()'s adoption of
1444 * the callbacks, thus no memory barrier is required.
1446 if (rdp->nxtlist != NULL) {
1447 rsp->qlen_lazy += rdp->qlen_lazy;
1448 rsp->qlen += rdp->qlen;
1449 rdp->n_cbs_orphaned += rdp->qlen;
1451 ACCESS_ONCE(rdp->qlen) = 0;
1455 * Next, move those callbacks still needing a grace period to
1456 * the orphanage, where some other CPU will pick them up.
1457 * Some of the callbacks might have gone partway through a grace
1458 * period, but that is too bad. They get to start over because we
1459 * cannot assume that grace periods are synchronized across CPUs.
1460 * We don't bother updating the ->nxttail[] array yet, instead
1461 * we just reset the whole thing later on.
1463 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1464 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1465 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1466 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1470 * Then move the ready-to-invoke callbacks to the orphanage,
1471 * where some other CPU will pick them up. These will not be
1472 * required to pass though another grace period: They are done.
1474 if (rdp->nxtlist != NULL) {
1475 *rsp->orphan_donetail = rdp->nxtlist;
1476 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1479 /* Finally, initialize the rcu_data structure's list to empty. */
1480 init_callback_list(rdp);
1484 * Adopt the RCU callbacks from the specified rcu_state structure's
1485 * orphanage. The caller must hold the ->onofflock.
1487 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1490 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1493 * If there is an rcu_barrier() operation in progress, then
1494 * only the task doing that operation is permitted to adopt
1495 * callbacks. To do otherwise breaks rcu_barrier() and friends
1496 * by causing them to fail to wait for the callbacks in the
1499 if (rsp->rcu_barrier_in_progress &&
1500 rsp->rcu_barrier_in_progress != current)
1503 /* Do the accounting first. */
1504 rdp->qlen_lazy += rsp->qlen_lazy;
1505 rdp->qlen += rsp->qlen;
1506 rdp->n_cbs_adopted += rsp->qlen;
1507 if (rsp->qlen_lazy != rsp->qlen)
1508 rcu_idle_count_callbacks_posted();
1513 * We do not need a memory barrier here because the only way we
1514 * can get here if there is an rcu_barrier() in flight is if
1515 * we are the task doing the rcu_barrier().
1518 /* First adopt the ready-to-invoke callbacks. */
1519 if (rsp->orphan_donelist != NULL) {
1520 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1521 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1522 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1523 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1524 rdp->nxttail[i] = rsp->orphan_donetail;
1525 rsp->orphan_donelist = NULL;
1526 rsp->orphan_donetail = &rsp->orphan_donelist;
1529 /* And then adopt the callbacks that still need a grace period. */
1530 if (rsp->orphan_nxtlist != NULL) {
1531 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1532 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1533 rsp->orphan_nxtlist = NULL;
1534 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1539 * Trace the fact that this CPU is going offline.
1541 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1543 RCU_TRACE(unsigned long mask);
1544 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1545 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1547 RCU_TRACE(mask = rdp->grpmask);
1548 trace_rcu_grace_period(rsp->name,
1549 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1554 * The CPU has been completely removed, and some other CPU is reporting
1555 * this fact from process context. Do the remainder of the cleanup,
1556 * including orphaning the outgoing CPU's RCU callbacks, and also
1557 * adopting them, if there is no _rcu_barrier() instance running.
1558 * There can only be one CPU hotplug operation at a time, so no other
1559 * CPU can be attempting to update rcu_cpu_kthread_task.
1561 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1563 unsigned long flags;
1565 int need_report = 0;
1566 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1567 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1569 /* Adjust any no-longer-needed kthreads. */
1570 rcu_stop_cpu_kthread(cpu);
1571 rcu_node_kthread_setaffinity(rnp, -1);
1573 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1575 /* Exclude any attempts to start a new grace period. */
1576 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1578 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1579 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1580 rcu_adopt_orphan_cbs(rsp);
1582 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1583 mask = rdp->grpmask; /* rnp->grplo is constant. */
1585 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1586 rnp->qsmaskinit &= ~mask;
1587 if (rnp->qsmaskinit != 0) {
1588 if (rnp != rdp->mynode)
1589 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1592 if (rnp == rdp->mynode)
1593 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1595 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1596 mask = rnp->grpmask;
1598 } while (rnp != NULL);
1601 * We still hold the leaf rcu_node structure lock here, and
1602 * irqs are still disabled. The reason for this subterfuge is
1603 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1604 * held leads to deadlock.
1606 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1608 if (need_report & RCU_OFL_TASKS_NORM_GP)
1609 rcu_report_unblock_qs_rnp(rnp, flags);
1611 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1612 if (need_report & RCU_OFL_TASKS_EXP_GP)
1613 rcu_report_exp_rnp(rsp, rnp, true);
1614 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1615 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1616 cpu, rdp->qlen, rdp->nxtlist);
1619 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1621 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1625 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1629 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1633 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1636 * Invoke any RCU callbacks that have made it to the end of their grace
1637 * period. Thottle as specified by rdp->blimit.
1639 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1641 unsigned long flags;
1642 struct rcu_head *next, *list, **tail;
1643 int bl, count, count_lazy, i;
1645 /* If no callbacks are ready, just return.*/
1646 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1647 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1648 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1649 need_resched(), is_idle_task(current),
1650 rcu_is_callbacks_kthread());
1655 * Extract the list of ready callbacks, disabling to prevent
1656 * races with call_rcu() from interrupt handlers.
1658 local_irq_save(flags);
1659 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1661 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1662 list = rdp->nxtlist;
1663 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1664 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1665 tail = rdp->nxttail[RCU_DONE_TAIL];
1666 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1667 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1668 rdp->nxttail[i] = &rdp->nxtlist;
1669 local_irq_restore(flags);
1671 /* Invoke callbacks. */
1672 count = count_lazy = 0;
1676 debug_rcu_head_unqueue(list);
1677 if (__rcu_reclaim(rsp->name, list))
1680 /* Stop only if limit reached and CPU has something to do. */
1681 if (++count >= bl &&
1683 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1687 local_irq_save(flags);
1688 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1689 is_idle_task(current),
1690 rcu_is_callbacks_kthread());
1692 /* Update count, and requeue any remaining callbacks. */
1694 *tail = rdp->nxtlist;
1695 rdp->nxtlist = list;
1696 for (i = 0; i < RCU_NEXT_SIZE; i++)
1697 if (&rdp->nxtlist == rdp->nxttail[i])
1698 rdp->nxttail[i] = tail;
1702 smp_mb(); /* List handling before counting for rcu_barrier(). */
1703 rdp->qlen_lazy -= count_lazy;
1704 ACCESS_ONCE(rdp->qlen) -= count;
1705 rdp->n_cbs_invoked += count;
1707 /* Reinstate batch limit if we have worked down the excess. */
1708 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1709 rdp->blimit = blimit;
1711 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1712 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1713 rdp->qlen_last_fqs_check = 0;
1714 rdp->n_force_qs_snap = rsp->n_force_qs;
1715 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1716 rdp->qlen_last_fqs_check = rdp->qlen;
1717 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1719 local_irq_restore(flags);
1721 /* Re-invoke RCU core processing if there are callbacks remaining. */
1722 if (cpu_has_callbacks_ready_to_invoke(rdp))
1727 * Check to see if this CPU is in a non-context-switch quiescent state
1728 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1729 * Also schedule RCU core processing.
1731 * This function must be called from hardirq context. It is normally
1732 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1733 * false, there is no point in invoking rcu_check_callbacks().
1735 void rcu_check_callbacks(int cpu, int user)
1737 trace_rcu_utilization("Start scheduler-tick");
1738 increment_cpu_stall_ticks();
1739 if (user || rcu_is_cpu_rrupt_from_idle()) {
1742 * Get here if this CPU took its interrupt from user
1743 * mode or from the idle loop, and if this is not a
1744 * nested interrupt. In this case, the CPU is in
1745 * a quiescent state, so note it.
1747 * No memory barrier is required here because both
1748 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1749 * variables that other CPUs neither access nor modify,
1750 * at least not while the corresponding CPU is online.
1756 } else if (!in_softirq()) {
1759 * Get here if this CPU did not take its interrupt from
1760 * softirq, in other words, if it is not interrupting
1761 * a rcu_bh read-side critical section. This is an _bh
1762 * critical section, so note it.
1767 rcu_preempt_check_callbacks(cpu);
1768 if (rcu_pending(cpu))
1770 trace_rcu_utilization("End scheduler-tick");
1774 * Scan the leaf rcu_node structures, processing dyntick state for any that
1775 * have not yet encountered a quiescent state, using the function specified.
1776 * Also initiate boosting for any threads blocked on the root rcu_node.
1778 * The caller must have suppressed start of new grace periods.
1780 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1784 unsigned long flags;
1786 struct rcu_node *rnp;
1788 rcu_for_each_leaf_node(rsp, rnp) {
1791 raw_spin_lock_irqsave(&rnp->lock, flags);
1792 if (!rcu_gp_in_progress(rsp)) {
1793 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1796 if (rnp->qsmask == 0) {
1797 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1802 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1803 if ((rnp->qsmask & bit) != 0 &&
1804 f(per_cpu_ptr(rsp->rda, cpu)))
1809 /* rcu_report_qs_rnp() releases rnp->lock. */
1810 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1813 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1815 rnp = rcu_get_root(rsp);
1816 if (rnp->qsmask == 0) {
1817 raw_spin_lock_irqsave(&rnp->lock, flags);
1818 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1823 * Force quiescent states on reluctant CPUs, and also detect which
1824 * CPUs are in dyntick-idle mode.
1826 static void force_quiescent_state(struct rcu_state *rsp)
1828 unsigned long flags;
1830 struct rcu_node *rnp;
1831 struct rcu_node *rnp_old = NULL;
1833 /* Funnel through hierarchy to reduce memory contention. */
1834 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1835 for (; rnp != NULL; rnp = rnp->parent) {
1836 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1837 !raw_spin_trylock(&rnp->fqslock);
1838 if (rnp_old != NULL)
1839 raw_spin_unlock(&rnp_old->fqslock);
1841 rsp->n_force_qs_lh++;
1846 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1848 /* Reached the root of the rcu_node tree, acquire lock. */
1849 raw_spin_lock_irqsave(&rnp_old->lock, flags);
1850 raw_spin_unlock(&rnp_old->fqslock);
1851 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1852 rsp->n_force_qs_lh++;
1853 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1854 return; /* Someone beat us to it. */
1856 rsp->gp_flags |= RCU_GP_FLAG_FQS;
1857 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1858 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1862 * This does the RCU core processing work for the specified rcu_state
1863 * and rcu_data structures. This may be called only from the CPU to
1864 * whom the rdp belongs.
1867 __rcu_process_callbacks(struct rcu_state *rsp)
1869 unsigned long flags;
1870 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1872 WARN_ON_ONCE(rdp->beenonline == 0);
1875 * Advance callbacks in response to end of earlier grace
1876 * period that some other CPU ended.
1878 rcu_process_gp_end(rsp, rdp);
1880 /* Update RCU state based on any recent quiescent states. */
1881 rcu_check_quiescent_state(rsp, rdp);
1883 /* Does this CPU require a not-yet-started grace period? */
1884 if (cpu_needs_another_gp(rsp, rdp)) {
1885 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1886 rcu_start_gp(rsp, flags); /* releases above lock */
1889 /* If there are callbacks ready, invoke them. */
1890 if (cpu_has_callbacks_ready_to_invoke(rdp))
1891 invoke_rcu_callbacks(rsp, rdp);
1895 * Do RCU core processing for the current CPU.
1897 static void rcu_process_callbacks(struct softirq_action *unused)
1899 struct rcu_state *rsp;
1901 if (cpu_is_offline(smp_processor_id()))
1903 trace_rcu_utilization("Start RCU core");
1904 for_each_rcu_flavor(rsp)
1905 __rcu_process_callbacks(rsp);
1906 trace_rcu_utilization("End RCU core");
1910 * Schedule RCU callback invocation. If the specified type of RCU
1911 * does not support RCU priority boosting, just do a direct call,
1912 * otherwise wake up the per-CPU kernel kthread. Note that because we
1913 * are running on the current CPU with interrupts disabled, the
1914 * rcu_cpu_kthread_task cannot disappear out from under us.
1916 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1918 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1920 if (likely(!rsp->boost)) {
1921 rcu_do_batch(rsp, rdp);
1924 invoke_rcu_callbacks_kthread();
1927 static void invoke_rcu_core(void)
1929 raise_softirq(RCU_SOFTIRQ);
1933 * Handle any core-RCU processing required by a call_rcu() invocation.
1935 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1936 struct rcu_head *head, unsigned long flags)
1939 * If called from an extended quiescent state, invoke the RCU
1940 * core in order to force a re-evaluation of RCU's idleness.
1942 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1945 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1946 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1950 * Force the grace period if too many callbacks or too long waiting.
1951 * Enforce hysteresis, and don't invoke force_quiescent_state()
1952 * if some other CPU has recently done so. Also, don't bother
1953 * invoking force_quiescent_state() if the newly enqueued callback
1954 * is the only one waiting for a grace period to complete.
1956 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1958 /* Are we ignoring a completed grace period? */
1959 rcu_process_gp_end(rsp, rdp);
1960 check_for_new_grace_period(rsp, rdp);
1962 /* Start a new grace period if one not already started. */
1963 if (!rcu_gp_in_progress(rsp)) {
1964 unsigned long nestflag;
1965 struct rcu_node *rnp_root = rcu_get_root(rsp);
1967 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1968 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1970 /* Give the grace period a kick. */
1971 rdp->blimit = LONG_MAX;
1972 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1973 *rdp->nxttail[RCU_DONE_TAIL] != head)
1974 force_quiescent_state(rsp);
1975 rdp->n_force_qs_snap = rsp->n_force_qs;
1976 rdp->qlen_last_fqs_check = rdp->qlen;
1982 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1983 struct rcu_state *rsp, bool lazy)
1985 unsigned long flags;
1986 struct rcu_data *rdp;
1988 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1989 debug_rcu_head_queue(head);
1993 smp_mb(); /* Ensure RCU update seen before callback registry. */
1996 * Opportunistically note grace-period endings and beginnings.
1997 * Note that we might see a beginning right after we see an
1998 * end, but never vice versa, since this CPU has to pass through
1999 * a quiescent state betweentimes.
2001 local_irq_save(flags);
2002 rdp = this_cpu_ptr(rsp->rda);
2004 /* Add the callback to our list. */
2005 ACCESS_ONCE(rdp->qlen)++;
2009 rcu_idle_count_callbacks_posted();
2010 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2011 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2012 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2014 if (__is_kfree_rcu_offset((unsigned long)func))
2015 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2016 rdp->qlen_lazy, rdp->qlen);
2018 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2020 /* Go handle any RCU core processing required. */
2021 __call_rcu_core(rsp, rdp, head, flags);
2022 local_irq_restore(flags);
2026 * Queue an RCU-sched callback for invocation after a grace period.
2028 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2030 __call_rcu(head, func, &rcu_sched_state, 0);
2032 EXPORT_SYMBOL_GPL(call_rcu_sched);
2035 * Queue an RCU callback for invocation after a quicker grace period.
2037 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2039 __call_rcu(head, func, &rcu_bh_state, 0);
2041 EXPORT_SYMBOL_GPL(call_rcu_bh);
2044 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2045 * any blocking grace-period wait automatically implies a grace period
2046 * if there is only one CPU online at any point time during execution
2047 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2048 * occasionally incorrectly indicate that there are multiple CPUs online
2049 * when there was in fact only one the whole time, as this just adds
2050 * some overhead: RCU still operates correctly.
2052 static inline int rcu_blocking_is_gp(void)
2056 might_sleep(); /* Check for RCU read-side critical section. */
2058 ret = num_online_cpus() <= 1;
2064 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2066 * Control will return to the caller some time after a full rcu-sched
2067 * grace period has elapsed, in other words after all currently executing
2068 * rcu-sched read-side critical sections have completed. These read-side
2069 * critical sections are delimited by rcu_read_lock_sched() and
2070 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2071 * local_irq_disable(), and so on may be used in place of
2072 * rcu_read_lock_sched().
2074 * This means that all preempt_disable code sequences, including NMI and
2075 * hardware-interrupt handlers, in progress on entry will have completed
2076 * before this primitive returns. However, this does not guarantee that
2077 * softirq handlers will have completed, since in some kernels, these
2078 * handlers can run in process context, and can block.
2080 * This primitive provides the guarantees made by the (now removed)
2081 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2082 * guarantees that rcu_read_lock() sections will have completed.
2083 * In "classic RCU", these two guarantees happen to be one and
2084 * the same, but can differ in realtime RCU implementations.
2086 void synchronize_sched(void)
2088 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2089 !lock_is_held(&rcu_lock_map) &&
2090 !lock_is_held(&rcu_sched_lock_map),
2091 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2092 if (rcu_blocking_is_gp())
2094 wait_rcu_gp(call_rcu_sched);
2096 EXPORT_SYMBOL_GPL(synchronize_sched);
2099 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2101 * Control will return to the caller some time after a full rcu_bh grace
2102 * period has elapsed, in other words after all currently executing rcu_bh
2103 * read-side critical sections have completed. RCU read-side critical
2104 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2105 * and may be nested.
2107 void synchronize_rcu_bh(void)
2109 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2110 !lock_is_held(&rcu_lock_map) &&
2111 !lock_is_held(&rcu_sched_lock_map),
2112 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2113 if (rcu_blocking_is_gp())
2115 wait_rcu_gp(call_rcu_bh);
2117 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2119 static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2120 static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2122 static int synchronize_sched_expedited_cpu_stop(void *data)
2125 * There must be a full memory barrier on each affected CPU
2126 * between the time that try_stop_cpus() is called and the
2127 * time that it returns.
2129 * In the current initial implementation of cpu_stop, the
2130 * above condition is already met when the control reaches
2131 * this point and the following smp_mb() is not strictly
2132 * necessary. Do smp_mb() anyway for documentation and
2133 * robustness against future implementation changes.
2135 smp_mb(); /* See above comment block. */
2140 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2142 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2143 * approach to force the grace period to end quickly. This consumes
2144 * significant time on all CPUs and is unfriendly to real-time workloads,
2145 * so is thus not recommended for any sort of common-case code. In fact,
2146 * if you are using synchronize_sched_expedited() in a loop, please
2147 * restructure your code to batch your updates, and then use a single
2148 * synchronize_sched() instead.
2150 * Note that it is illegal to call this function while holding any lock
2151 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2152 * to call this function from a CPU-hotplug notifier. Failing to observe
2153 * these restriction will result in deadlock.
2155 * This implementation can be thought of as an application of ticket
2156 * locking to RCU, with sync_sched_expedited_started and
2157 * sync_sched_expedited_done taking on the roles of the halves
2158 * of the ticket-lock word. Each task atomically increments
2159 * sync_sched_expedited_started upon entry, snapshotting the old value,
2160 * then attempts to stop all the CPUs. If this succeeds, then each
2161 * CPU will have executed a context switch, resulting in an RCU-sched
2162 * grace period. We are then done, so we use atomic_cmpxchg() to
2163 * update sync_sched_expedited_done to match our snapshot -- but
2164 * only if someone else has not already advanced past our snapshot.
2166 * On the other hand, if try_stop_cpus() fails, we check the value
2167 * of sync_sched_expedited_done. If it has advanced past our
2168 * initial snapshot, then someone else must have forced a grace period
2169 * some time after we took our snapshot. In this case, our work is
2170 * done for us, and we can simply return. Otherwise, we try again,
2171 * but keep our initial snapshot for purposes of checking for someone
2172 * doing our work for us.
2174 * If we fail too many times in a row, we fall back to synchronize_sched().
2176 void synchronize_sched_expedited(void)
2178 int firstsnap, s, snap, trycount = 0;
2180 /* Note that atomic_inc_return() implies full memory barrier. */
2181 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2183 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2186 * Each pass through the following loop attempts to force a
2187 * context switch on each CPU.
2189 while (try_stop_cpus(cpu_online_mask,
2190 synchronize_sched_expedited_cpu_stop,
2194 /* No joy, try again later. Or just synchronize_sched(). */
2195 if (trycount++ < 10) {
2196 udelay(trycount * num_online_cpus());
2198 synchronize_sched();
2202 /* Check to see if someone else did our work for us. */
2203 s = atomic_read(&sync_sched_expedited_done);
2204 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2205 smp_mb(); /* ensure test happens before caller kfree */
2210 * Refetching sync_sched_expedited_started allows later
2211 * callers to piggyback on our grace period. We subtract
2212 * 1 to get the same token that the last incrementer got.
2213 * We retry after they started, so our grace period works
2214 * for them, and they started after our first try, so their
2215 * grace period works for us.
2218 snap = atomic_read(&sync_sched_expedited_started);
2219 smp_mb(); /* ensure read is before try_stop_cpus(). */
2223 * Everyone up to our most recent fetch is covered by our grace
2224 * period. Update the counter, but only if our work is still
2225 * relevant -- which it won't be if someone who started later
2226 * than we did beat us to the punch.
2229 s = atomic_read(&sync_sched_expedited_done);
2230 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2231 smp_mb(); /* ensure test happens before caller kfree */
2234 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2238 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2241 * Check to see if there is any immediate RCU-related work to be done
2242 * by the current CPU, for the specified type of RCU, returning 1 if so.
2243 * The checks are in order of increasing expense: checks that can be
2244 * carried out against CPU-local state are performed first. However,
2245 * we must check for CPU stalls first, else we might not get a chance.
2247 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2249 struct rcu_node *rnp = rdp->mynode;
2251 rdp->n_rcu_pending++;
2253 /* Check for CPU stalls, if enabled. */
2254 check_cpu_stall(rsp, rdp);
2256 /* Is the RCU core waiting for a quiescent state from this CPU? */
2257 if (rcu_scheduler_fully_active &&
2258 rdp->qs_pending && !rdp->passed_quiesce) {
2259 rdp->n_rp_qs_pending++;
2260 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2261 rdp->n_rp_report_qs++;
2265 /* Does this CPU have callbacks ready to invoke? */
2266 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2267 rdp->n_rp_cb_ready++;
2271 /* Has RCU gone idle with this CPU needing another grace period? */
2272 if (cpu_needs_another_gp(rsp, rdp)) {
2273 rdp->n_rp_cpu_needs_gp++;
2277 /* Has another RCU grace period completed? */
2278 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2279 rdp->n_rp_gp_completed++;
2283 /* Has a new RCU grace period started? */
2284 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2285 rdp->n_rp_gp_started++;
2290 rdp->n_rp_need_nothing++;
2295 * Check to see if there is any immediate RCU-related work to be done
2296 * by the current CPU, returning 1 if so. This function is part of the
2297 * RCU implementation; it is -not- an exported member of the RCU API.
2299 static int rcu_pending(int cpu)
2301 struct rcu_state *rsp;
2303 for_each_rcu_flavor(rsp)
2304 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2310 * Check to see if any future RCU-related work will need to be done
2311 * by the current CPU, even if none need be done immediately, returning
2314 static int rcu_cpu_has_callbacks(int cpu)
2316 struct rcu_state *rsp;
2318 /* RCU callbacks either ready or pending? */
2319 for_each_rcu_flavor(rsp)
2320 if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2326 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2327 * the compiler is expected to optimize this away.
2329 static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2330 int cpu, unsigned long done)
2332 trace_rcu_barrier(rsp->name, s, cpu,
2333 atomic_read(&rsp->barrier_cpu_count), done);
2337 * RCU callback function for _rcu_barrier(). If we are last, wake
2338 * up the task executing _rcu_barrier().
2340 static void rcu_barrier_callback(struct rcu_head *rhp)
2342 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2343 struct rcu_state *rsp = rdp->rsp;
2345 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2346 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2347 complete(&rsp->barrier_completion);
2349 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2354 * Called with preemption disabled, and from cross-cpu IRQ context.
2356 static void rcu_barrier_func(void *type)
2358 struct rcu_state *rsp = type;
2359 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2361 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2362 atomic_inc(&rsp->barrier_cpu_count);
2363 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2367 * Orchestrate the specified type of RCU barrier, waiting for all
2368 * RCU callbacks of the specified type to complete.
2370 static void _rcu_barrier(struct rcu_state *rsp)
2373 unsigned long flags;
2374 struct rcu_data *rdp;
2376 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2377 unsigned long snap_done;
2379 init_rcu_head_on_stack(&rd.barrier_head);
2380 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2382 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2383 mutex_lock(&rsp->barrier_mutex);
2386 * Ensure that all prior references, including to ->n_barrier_done,
2387 * are ordered before the _rcu_barrier() machinery.
2389 smp_mb(); /* See above block comment. */
2392 * Recheck ->n_barrier_done to see if others did our work for us.
2393 * This means checking ->n_barrier_done for an even-to-odd-to-even
2394 * transition. The "if" expression below therefore rounds the old
2395 * value up to the next even number and adds two before comparing.
2397 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2398 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2399 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2400 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2401 smp_mb(); /* caller's subsequent code after above check. */
2402 mutex_unlock(&rsp->barrier_mutex);
2407 * Increment ->n_barrier_done to avoid duplicate work. Use
2408 * ACCESS_ONCE() to prevent the compiler from speculating
2409 * the increment to precede the early-exit check.
2411 ACCESS_ONCE(rsp->n_barrier_done)++;
2412 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2413 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2414 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2417 * Initialize the count to one rather than to zero in order to
2418 * avoid a too-soon return to zero in case of a short grace period
2419 * (or preemption of this task). Also flag this task as doing
2420 * an rcu_barrier(). This will prevent anyone else from adopting
2421 * orphaned callbacks, which could cause otherwise failure if a
2422 * CPU went offline and quickly came back online. To see this,
2423 * consider the following sequence of events:
2425 * 1. We cause CPU 0 to post an rcu_barrier_callback() callback.
2426 * 2. CPU 1 goes offline, orphaning its callbacks.
2427 * 3. CPU 0 adopts CPU 1's orphaned callbacks.
2428 * 4. CPU 1 comes back online.
2429 * 5. We cause CPU 1 to post an rcu_barrier_callback() callback.
2430 * 6. Both rcu_barrier_callback() callbacks are invoked, awakening
2431 * us -- but before CPU 1's orphaned callbacks are invoked!!!
2433 init_completion(&rsp->barrier_completion);
2434 atomic_set(&rsp->barrier_cpu_count, 1);
2435 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2436 rsp->rcu_barrier_in_progress = current;
2437 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2440 * Force every CPU with callbacks to register a new callback
2441 * that will tell us when all the preceding callbacks have
2442 * been invoked. If an offline CPU has callbacks, wait for
2443 * it to either come back online or to finish orphaning those
2446 for_each_possible_cpu(cpu) {
2448 rdp = per_cpu_ptr(rsp->rda, cpu);
2449 if (cpu_is_offline(cpu)) {
2450 _rcu_barrier_trace(rsp, "Offline", cpu,
2451 rsp->n_barrier_done);
2453 while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2454 schedule_timeout_interruptible(1);
2455 } else if (ACCESS_ONCE(rdp->qlen)) {
2456 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2457 rsp->n_barrier_done);
2458 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2461 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2462 rsp->n_barrier_done);
2468 * Now that all online CPUs have rcu_barrier_callback() callbacks
2469 * posted, we can adopt all of the orphaned callbacks and place
2470 * an rcu_barrier_callback() callback after them. When that is done,
2471 * we are guaranteed to have an rcu_barrier_callback() callback
2472 * following every callback that could possibly have been
2473 * registered before _rcu_barrier() was called.
2475 raw_spin_lock_irqsave(&rsp->onofflock, flags);
2476 rcu_adopt_orphan_cbs(rsp);
2477 rsp->rcu_barrier_in_progress = NULL;
2478 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2479 atomic_inc(&rsp->barrier_cpu_count);
2480 smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2482 rsp->call(&rd.barrier_head, rcu_barrier_callback);
2485 * Now that we have an rcu_barrier_callback() callback on each
2486 * CPU, and thus each counted, remove the initial count.
2488 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2489 complete(&rsp->barrier_completion);
2491 /* Increment ->n_barrier_done to prevent duplicate work. */
2492 smp_mb(); /* Keep increment after above mechanism. */
2493 ACCESS_ONCE(rsp->n_barrier_done)++;
2494 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2495 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2496 smp_mb(); /* Keep increment before caller's subsequent code. */
2498 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2499 wait_for_completion(&rsp->barrier_completion);
2501 /* Other rcu_barrier() invocations can now safely proceed. */
2502 mutex_unlock(&rsp->barrier_mutex);
2504 destroy_rcu_head_on_stack(&rd.barrier_head);
2508 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2510 void rcu_barrier_bh(void)
2512 _rcu_barrier(&rcu_bh_state);
2514 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2517 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2519 void rcu_barrier_sched(void)
2521 _rcu_barrier(&rcu_sched_state);
2523 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2526 * Do boot-time initialization of a CPU's per-CPU RCU data.
2529 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2531 unsigned long flags;
2532 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2533 struct rcu_node *rnp = rcu_get_root(rsp);
2535 /* Set up local state, ensuring consistent view of global state. */
2536 raw_spin_lock_irqsave(&rnp->lock, flags);
2537 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2538 init_callback_list(rdp);
2540 ACCESS_ONCE(rdp->qlen) = 0;
2541 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2542 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2543 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2546 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2550 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2551 * offline event can be happening at a given time. Note also that we
2552 * can accept some slop in the rsp->completed access due to the fact
2553 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2555 static void __cpuinit
2556 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2558 unsigned long flags;
2560 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2561 struct rcu_node *rnp = rcu_get_root(rsp);
2563 /* Set up local state, ensuring consistent view of global state. */
2564 raw_spin_lock_irqsave(&rnp->lock, flags);
2565 rdp->beenonline = 1; /* We have now been online. */
2566 rdp->preemptible = preemptible;
2567 rdp->qlen_last_fqs_check = 0;
2568 rdp->n_force_qs_snap = rsp->n_force_qs;
2569 rdp->blimit = blimit;
2570 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2571 atomic_set(&rdp->dynticks->dynticks,
2572 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2573 rcu_prepare_for_idle_init(cpu);
2574 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2577 * A new grace period might start here. If so, we won't be part
2578 * of it, but that is OK, as we are currently in a quiescent state.
2581 /* Exclude any attempts to start a new GP on large systems. */
2582 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2584 /* Add CPU to rcu_node bitmasks. */
2586 mask = rdp->grpmask;
2588 /* Exclude any attempts to start a new GP on small systems. */
2589 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2590 rnp->qsmaskinit |= mask;
2591 mask = rnp->grpmask;
2592 if (rnp == rdp->mynode) {
2594 * If there is a grace period in progress, we will
2595 * set up to wait for it next time we run the
2598 rdp->gpnum = rnp->completed;
2599 rdp->completed = rnp->completed;
2600 rdp->passed_quiesce = 0;
2601 rdp->qs_pending = 0;
2602 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2603 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2605 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2607 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2609 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2612 static void __cpuinit rcu_prepare_cpu(int cpu)
2614 struct rcu_state *rsp;
2616 for_each_rcu_flavor(rsp)
2617 rcu_init_percpu_data(cpu, rsp,
2618 strcmp(rsp->name, "rcu_preempt") == 0);
2622 * Handle CPU online/offline notification events.
2624 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2625 unsigned long action, void *hcpu)
2627 long cpu = (long)hcpu;
2628 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2629 struct rcu_node *rnp = rdp->mynode;
2630 struct rcu_state *rsp;
2632 trace_rcu_utilization("Start CPU hotplug");
2634 case CPU_UP_PREPARE:
2635 case CPU_UP_PREPARE_FROZEN:
2636 rcu_prepare_cpu(cpu);
2637 rcu_prepare_kthreads(cpu);
2640 case CPU_DOWN_FAILED:
2641 rcu_node_kthread_setaffinity(rnp, -1);
2642 rcu_cpu_kthread_setrt(cpu, 1);
2644 case CPU_DOWN_PREPARE:
2645 rcu_node_kthread_setaffinity(rnp, cpu);
2646 rcu_cpu_kthread_setrt(cpu, 0);
2649 case CPU_DYING_FROZEN:
2651 * The whole machine is "stopped" except this CPU, so we can
2652 * touch any data without introducing corruption. We send the
2653 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2655 for_each_rcu_flavor(rsp)
2656 rcu_cleanup_dying_cpu(rsp);
2657 rcu_cleanup_after_idle(cpu);
2660 case CPU_DEAD_FROZEN:
2661 case CPU_UP_CANCELED:
2662 case CPU_UP_CANCELED_FROZEN:
2663 for_each_rcu_flavor(rsp)
2664 rcu_cleanup_dead_cpu(cpu, rsp);
2669 trace_rcu_utilization("End CPU hotplug");
2674 * Spawn the kthread that handles this RCU flavor's grace periods.
2676 static int __init rcu_spawn_gp_kthread(void)
2678 unsigned long flags;
2679 struct rcu_node *rnp;
2680 struct rcu_state *rsp;
2681 struct task_struct *t;
2683 for_each_rcu_flavor(rsp) {
2684 t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2686 rnp = rcu_get_root(rsp);
2687 raw_spin_lock_irqsave(&rnp->lock, flags);
2688 rsp->gp_kthread = t;
2689 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2693 early_initcall(rcu_spawn_gp_kthread);
2696 * This function is invoked towards the end of the scheduler's initialization
2697 * process. Before this is called, the idle task might contain
2698 * RCU read-side critical sections (during which time, this idle
2699 * task is booting the system). After this function is called, the
2700 * idle tasks are prohibited from containing RCU read-side critical
2701 * sections. This function also enables RCU lockdep checking.
2703 void rcu_scheduler_starting(void)
2705 WARN_ON(num_online_cpus() != 1);
2706 WARN_ON(nr_context_switches() > 0);
2707 rcu_scheduler_active = 1;
2711 * Compute the per-level fanout, either using the exact fanout specified
2712 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2714 #ifdef CONFIG_RCU_FANOUT_EXACT
2715 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2719 for (i = rcu_num_lvls - 1; i > 0; i--)
2720 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2721 rsp->levelspread[0] = rcu_fanout_leaf;
2723 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2724 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2731 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2732 ccur = rsp->levelcnt[i];
2733 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2737 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2740 * Helper function for rcu_init() that initializes one rcu_state structure.
2742 static void __init rcu_init_one(struct rcu_state *rsp,
2743 struct rcu_data __percpu *rda)
2745 static char *buf[] = { "rcu_node_0",
2748 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
2749 static char *fqs[] = { "rcu_node_fqs_0",
2752 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
2756 struct rcu_node *rnp;
2758 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2760 /* Initialize the level-tracking arrays. */
2762 for (i = 0; i < rcu_num_lvls; i++)
2763 rsp->levelcnt[i] = num_rcu_lvl[i];
2764 for (i = 1; i < rcu_num_lvls; i++)
2765 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2766 rcu_init_levelspread(rsp);
2768 /* Initialize the elements themselves, starting from the leaves. */
2770 for (i = rcu_num_lvls - 1; i >= 0; i--) {
2771 cpustride *= rsp->levelspread[i];
2772 rnp = rsp->level[i];
2773 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2774 raw_spin_lock_init(&rnp->lock);
2775 lockdep_set_class_and_name(&rnp->lock,
2776 &rcu_node_class[i], buf[i]);
2777 raw_spin_lock_init(&rnp->fqslock);
2778 lockdep_set_class_and_name(&rnp->fqslock,
2779 &rcu_fqs_class[i], fqs[i]);
2782 rnp->qsmaskinit = 0;
2783 rnp->grplo = j * cpustride;
2784 rnp->grphi = (j + 1) * cpustride - 1;
2785 if (rnp->grphi >= NR_CPUS)
2786 rnp->grphi = NR_CPUS - 1;
2792 rnp->grpnum = j % rsp->levelspread[i - 1];
2793 rnp->grpmask = 1UL << rnp->grpnum;
2794 rnp->parent = rsp->level[i - 1] +
2795 j / rsp->levelspread[i - 1];
2798 INIT_LIST_HEAD(&rnp->blkd_tasks);
2803 init_waitqueue_head(&rsp->gp_wq);
2804 rnp = rsp->level[rcu_num_lvls - 1];
2805 for_each_possible_cpu(i) {
2806 while (i > rnp->grphi)
2808 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2809 rcu_boot_init_percpu_data(i, rsp);
2811 list_add(&rsp->flavors, &rcu_struct_flavors);
2815 * Compute the rcu_node tree geometry from kernel parameters. This cannot
2816 * replace the definitions in rcutree.h because those are needed to size
2817 * the ->node array in the rcu_state structure.
2819 static void __init rcu_init_geometry(void)
2824 int rcu_capacity[MAX_RCU_LVLS + 1];
2826 /* If the compile-time values are accurate, just leave. */
2827 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2831 * Compute number of nodes that can be handled an rcu_node tree
2832 * with the given number of levels. Setting rcu_capacity[0] makes
2833 * some of the arithmetic easier.
2835 rcu_capacity[0] = 1;
2836 rcu_capacity[1] = rcu_fanout_leaf;
2837 for (i = 2; i <= MAX_RCU_LVLS; i++)
2838 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2841 * The boot-time rcu_fanout_leaf parameter is only permitted
2842 * to increase the leaf-level fanout, not decrease it. Of course,
2843 * the leaf-level fanout cannot exceed the number of bits in
2844 * the rcu_node masks. Finally, the tree must be able to accommodate
2845 * the configured number of CPUs. Complain and fall back to the
2846 * compile-time values if these limits are exceeded.
2848 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2849 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2850 n > rcu_capacity[MAX_RCU_LVLS]) {
2855 /* Calculate the number of rcu_nodes at each level of the tree. */
2856 for (i = 1; i <= MAX_RCU_LVLS; i++)
2857 if (n <= rcu_capacity[i]) {
2858 for (j = 0; j <= i; j++)
2860 DIV_ROUND_UP(n, rcu_capacity[i - j]);
2862 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2867 /* Calculate the total number of rcu_node structures. */
2869 for (i = 0; i <= MAX_RCU_LVLS; i++)
2870 rcu_num_nodes += num_rcu_lvl[i];
2874 void __init rcu_init(void)
2878 rcu_bootup_announce();
2879 rcu_init_geometry();
2880 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2881 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2882 __rcu_init_preempt();
2883 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2886 * We don't need protection against CPU-hotplug here because
2887 * this is called early in boot, before either interrupts
2888 * or the scheduler are operational.
2890 cpu_notifier(rcu_cpu_notify, 0);
2891 for_each_online_cpu(cpu)
2892 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2893 check_cpu_stall_init();
2896 #include "rcutree_plugin.h"