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, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
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/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
82 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83 static char sname##_varname[] = #sname; \
84 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85 struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
99 DEFINE_PER_CPU(struct rcu_data, sname##_data)
101 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
102 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
104 static struct rcu_state *rcu_state;
105 LIST_HEAD(rcu_struct_flavors);
107 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
109 module_param(rcu_fanout_leaf, int, 0444);
110 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
111 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
118 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
129 int rcu_scheduler_active __read_mostly;
130 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
144 static int rcu_scheduler_fully_active __read_mostly;
146 #ifdef CONFIG_RCU_BOOST
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
152 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
153 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
155 DEFINE_PER_CPU(char, rcu_cpu_has_work);
157 #endif /* #ifdef CONFIG_RCU_BOOST */
159 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
172 unsigned long rcutorture_testseq;
173 unsigned long rcutorture_vernum;
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
180 static int rcu_gp_in_progress(struct rcu_state *rsp)
182 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
191 void rcu_sched_qs(int cpu)
193 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
195 if (rdp->passed_quiesce == 0)
196 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
197 rdp->passed_quiesce = 1;
200 void rcu_bh_qs(int cpu)
202 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
204 if (rdp->passed_quiesce == 0)
205 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
206 rdp->passed_quiesce = 1;
210 * Note a context switch. This is a quiescent state for RCU-sched,
211 * and requires special handling for preemptible RCU.
212 * The caller must have disabled preemption.
214 void rcu_note_context_switch(int cpu)
216 trace_rcu_utilization(TPS("Start context switch"));
218 rcu_preempt_note_context_switch(cpu);
219 trace_rcu_utilization(TPS("End context switch"));
221 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
223 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
224 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
225 .dynticks = ATOMIC_INIT(1),
226 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
227 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
228 .dynticks_idle = ATOMIC_INIT(1),
229 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
232 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
233 static long qhimark = 10000; /* If this many pending, ignore blimit. */
234 static long qlowmark = 100; /* Once only this many pending, use blimit. */
236 module_param(blimit, long, 0444);
237 module_param(qhimark, long, 0444);
238 module_param(qlowmark, long, 0444);
240 static ulong jiffies_till_first_fqs = ULONG_MAX;
241 static ulong jiffies_till_next_fqs = ULONG_MAX;
243 module_param(jiffies_till_first_fqs, ulong, 0644);
244 module_param(jiffies_till_next_fqs, ulong, 0644);
246 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
247 struct rcu_data *rdp);
248 static void force_qs_rnp(struct rcu_state *rsp,
249 int (*f)(struct rcu_data *rsp, bool *isidle,
250 unsigned long *maxj),
251 bool *isidle, unsigned long *maxj);
252 static void force_quiescent_state(struct rcu_state *rsp);
253 static int rcu_pending(int cpu);
256 * Return the number of RCU-sched batches processed thus far for debug & stats.
258 long rcu_batches_completed_sched(void)
260 return rcu_sched_state.completed;
262 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
265 * Return the number of RCU BH batches processed thus far for debug & stats.
267 long rcu_batches_completed_bh(void)
269 return rcu_bh_state.completed;
271 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
274 * Force a quiescent state for RCU BH.
276 void rcu_bh_force_quiescent_state(void)
278 force_quiescent_state(&rcu_bh_state);
280 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
283 * Record the number of times rcutorture tests have been initiated and
284 * terminated. This information allows the debugfs tracing stats to be
285 * correlated to the rcutorture messages, even when the rcutorture module
286 * is being repeatedly loaded and unloaded. In other words, we cannot
287 * store this state in rcutorture itself.
289 void rcutorture_record_test_transition(void)
291 rcutorture_testseq++;
292 rcutorture_vernum = 0;
294 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
297 * Record the number of writer passes through the current rcutorture test.
298 * This is also used to correlate debugfs tracing stats with the rcutorture
301 void rcutorture_record_progress(unsigned long vernum)
305 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
308 * Force a quiescent state for RCU-sched.
310 void rcu_sched_force_quiescent_state(void)
312 force_quiescent_state(&rcu_sched_state);
314 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
317 * Does the CPU have callbacks ready to be invoked?
320 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
322 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
323 rdp->nxttail[RCU_DONE_TAIL] != NULL;
327 * Return the root node of the specified rcu_state structure.
329 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
331 return &rsp->node[0];
335 * Is there any need for future grace periods?
336 * Interrupts must be disabled. If the caller does not hold the root
337 * rnp_node structure's ->lock, the results are advisory only.
339 static int rcu_future_needs_gp(struct rcu_state *rsp)
341 struct rcu_node *rnp = rcu_get_root(rsp);
342 int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1;
343 int *fp = &rnp->need_future_gp[idx];
345 return ACCESS_ONCE(*fp);
349 * Does the current CPU require a not-yet-started grace period?
350 * The caller must have disabled interrupts to prevent races with
351 * normal callback registry.
354 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
358 if (rcu_gp_in_progress(rsp))
359 return 0; /* No, a grace period is already in progress. */
360 if (rcu_future_needs_gp(rsp))
361 return 1; /* Yes, a no-CBs CPU needs one. */
362 if (!rdp->nxttail[RCU_NEXT_TAIL])
363 return 0; /* No, this is a no-CBs (or offline) CPU. */
364 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
365 return 1; /* Yes, this CPU has newly registered callbacks. */
366 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
367 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
368 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
369 rdp->nxtcompleted[i]))
370 return 1; /* Yes, CBs for future grace period. */
371 return 0; /* No grace period needed. */
375 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
377 * If the new value of the ->dynticks_nesting counter now is zero,
378 * we really have entered idle, and must do the appropriate accounting.
379 * The caller must have disabled interrupts.
381 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
384 struct rcu_state *rsp;
385 struct rcu_data *rdp;
387 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
388 if (!user && !is_idle_task(current)) {
389 struct task_struct *idle __maybe_unused =
390 idle_task(smp_processor_id());
392 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
393 ftrace_dump(DUMP_ORIG);
394 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
395 current->pid, current->comm,
396 idle->pid, idle->comm); /* must be idle task! */
398 for_each_rcu_flavor(rsp) {
399 rdp = this_cpu_ptr(rsp->rda);
400 do_nocb_deferred_wakeup(rdp);
402 rcu_prepare_for_idle(smp_processor_id());
403 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
404 smp_mb__before_atomic_inc(); /* See above. */
405 atomic_inc(&rdtp->dynticks);
406 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
407 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
410 * It is illegal to enter an extended quiescent state while
411 * in an RCU read-side critical section.
413 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
414 "Illegal idle entry in RCU read-side critical section.");
415 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
416 "Illegal idle entry in RCU-bh read-side critical section.");
417 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
418 "Illegal idle entry in RCU-sched read-side critical section.");
422 * Enter an RCU extended quiescent state, which can be either the
423 * idle loop or adaptive-tickless usermode execution.
425 static void rcu_eqs_enter(bool user)
428 struct rcu_dynticks *rdtp;
430 rdtp = this_cpu_ptr(&rcu_dynticks);
431 oldval = rdtp->dynticks_nesting;
432 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
433 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
434 rdtp->dynticks_nesting = 0;
435 rcu_eqs_enter_common(rdtp, oldval, user);
437 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
442 * rcu_idle_enter - inform RCU that current CPU is entering idle
444 * Enter idle mode, in other words, -leave- the mode in which RCU
445 * read-side critical sections can occur. (Though RCU read-side
446 * critical sections can occur in irq handlers in idle, a possibility
447 * handled by irq_enter() and irq_exit().)
449 * We crowbar the ->dynticks_nesting field to zero to allow for
450 * the possibility of usermode upcalls having messed up our count
451 * of interrupt nesting level during the prior busy period.
453 void rcu_idle_enter(void)
457 local_irq_save(flags);
458 rcu_eqs_enter(false);
459 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
460 local_irq_restore(flags);
462 EXPORT_SYMBOL_GPL(rcu_idle_enter);
464 #ifdef CONFIG_RCU_USER_QS
466 * rcu_user_enter - inform RCU that we are resuming userspace.
468 * Enter RCU idle mode right before resuming userspace. No use of RCU
469 * is permitted between this call and rcu_user_exit(). This way the
470 * CPU doesn't need to maintain the tick for RCU maintenance purposes
471 * when the CPU runs in userspace.
473 void rcu_user_enter(void)
477 #endif /* CONFIG_RCU_USER_QS */
480 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
482 * Exit from an interrupt handler, which might possibly result in entering
483 * idle mode, in other words, leaving the mode in which read-side critical
484 * sections can occur.
486 * This code assumes that the idle loop never does anything that might
487 * result in unbalanced calls to irq_enter() and irq_exit(). If your
488 * architecture violates this assumption, RCU will give you what you
489 * deserve, good and hard. But very infrequently and irreproducibly.
491 * Use things like work queues to work around this limitation.
493 * You have been warned.
495 void rcu_irq_exit(void)
499 struct rcu_dynticks *rdtp;
501 local_irq_save(flags);
502 rdtp = this_cpu_ptr(&rcu_dynticks);
503 oldval = rdtp->dynticks_nesting;
504 rdtp->dynticks_nesting--;
505 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
506 if (rdtp->dynticks_nesting)
507 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
509 rcu_eqs_enter_common(rdtp, oldval, true);
510 rcu_sysidle_enter(rdtp, 1);
511 local_irq_restore(flags);
515 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
517 * If the new value of the ->dynticks_nesting counter was previously zero,
518 * we really have exited idle, and must do the appropriate accounting.
519 * The caller must have disabled interrupts.
521 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
524 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
525 atomic_inc(&rdtp->dynticks);
526 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
527 smp_mb__after_atomic_inc(); /* See above. */
528 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
529 rcu_cleanup_after_idle(smp_processor_id());
530 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
531 if (!user && !is_idle_task(current)) {
532 struct task_struct *idle __maybe_unused =
533 idle_task(smp_processor_id());
535 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
536 oldval, rdtp->dynticks_nesting);
537 ftrace_dump(DUMP_ORIG);
538 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
539 current->pid, current->comm,
540 idle->pid, idle->comm); /* must be idle task! */
545 * Exit an RCU extended quiescent state, which can be either the
546 * idle loop or adaptive-tickless usermode execution.
548 static void rcu_eqs_exit(bool user)
550 struct rcu_dynticks *rdtp;
553 rdtp = this_cpu_ptr(&rcu_dynticks);
554 oldval = rdtp->dynticks_nesting;
555 WARN_ON_ONCE(oldval < 0);
556 if (oldval & DYNTICK_TASK_NEST_MASK) {
557 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
559 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
560 rcu_eqs_exit_common(rdtp, oldval, user);
565 * rcu_idle_exit - inform RCU that current CPU is leaving idle
567 * Exit idle mode, in other words, -enter- the mode in which RCU
568 * read-side critical sections can occur.
570 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
571 * allow for the possibility of usermode upcalls messing up our count
572 * of interrupt nesting level during the busy period that is just
575 void rcu_idle_exit(void)
579 local_irq_save(flags);
581 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
582 local_irq_restore(flags);
584 EXPORT_SYMBOL_GPL(rcu_idle_exit);
586 #ifdef CONFIG_RCU_USER_QS
588 * rcu_user_exit - inform RCU that we are exiting userspace.
590 * Exit RCU idle mode while entering the kernel because it can
591 * run a RCU read side critical section anytime.
593 void rcu_user_exit(void)
597 #endif /* CONFIG_RCU_USER_QS */
600 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
602 * Enter an interrupt handler, which might possibly result in exiting
603 * idle mode, in other words, entering the mode in which read-side critical
604 * sections can occur.
606 * Note that the Linux kernel is fully capable of entering an interrupt
607 * handler that it never exits, for example when doing upcalls to
608 * user mode! This code assumes that the idle loop never does upcalls to
609 * user mode. If your architecture does do upcalls from the idle loop (or
610 * does anything else that results in unbalanced calls to the irq_enter()
611 * and irq_exit() functions), RCU will give you what you deserve, good
612 * and hard. But very infrequently and irreproducibly.
614 * Use things like work queues to work around this limitation.
616 * You have been warned.
618 void rcu_irq_enter(void)
621 struct rcu_dynticks *rdtp;
624 local_irq_save(flags);
625 rdtp = this_cpu_ptr(&rcu_dynticks);
626 oldval = rdtp->dynticks_nesting;
627 rdtp->dynticks_nesting++;
628 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
630 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
632 rcu_eqs_exit_common(rdtp, oldval, true);
633 rcu_sysidle_exit(rdtp, 1);
634 local_irq_restore(flags);
638 * rcu_nmi_enter - inform RCU of entry to NMI context
640 * If the CPU was idle with dynamic ticks active, and there is no
641 * irq handler running, this updates rdtp->dynticks_nmi to let the
642 * RCU grace-period handling know that the CPU is active.
644 void rcu_nmi_enter(void)
646 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
648 if (rdtp->dynticks_nmi_nesting == 0 &&
649 (atomic_read(&rdtp->dynticks) & 0x1))
651 rdtp->dynticks_nmi_nesting++;
652 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
653 atomic_inc(&rdtp->dynticks);
654 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
655 smp_mb__after_atomic_inc(); /* See above. */
656 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
660 * rcu_nmi_exit - inform RCU of exit from NMI context
662 * If the CPU was idle with dynamic ticks active, and there is no
663 * irq handler running, this updates rdtp->dynticks_nmi to let the
664 * RCU grace-period handling know that the CPU is no longer active.
666 void rcu_nmi_exit(void)
668 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
670 if (rdtp->dynticks_nmi_nesting == 0 ||
671 --rdtp->dynticks_nmi_nesting != 0)
673 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
674 smp_mb__before_atomic_inc(); /* See above. */
675 atomic_inc(&rdtp->dynticks);
676 smp_mb__after_atomic_inc(); /* Force delay to next write. */
677 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
681 * __rcu_is_watching - are RCU read-side critical sections safe?
683 * Return true if RCU is watching the running CPU, which means that
684 * this CPU can safely enter RCU read-side critical sections. Unlike
685 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
686 * least disabled preemption.
688 bool notrace __rcu_is_watching(void)
690 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
694 * rcu_is_watching - see if RCU thinks that the current CPU is idle
696 * If the current CPU is in its idle loop and is neither in an interrupt
697 * or NMI handler, return true.
699 bool notrace rcu_is_watching(void)
704 ret = __rcu_is_watching();
708 EXPORT_SYMBOL_GPL(rcu_is_watching);
710 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
713 * Is the current CPU online? Disable preemption to avoid false positives
714 * that could otherwise happen due to the current CPU number being sampled,
715 * this task being preempted, its old CPU being taken offline, resuming
716 * on some other CPU, then determining that its old CPU is now offline.
717 * It is OK to use RCU on an offline processor during initial boot, hence
718 * the check for rcu_scheduler_fully_active. Note also that it is OK
719 * for a CPU coming online to use RCU for one jiffy prior to marking itself
720 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
721 * offline to continue to use RCU for one jiffy after marking itself
722 * offline in the cpu_online_mask. This leniency is necessary given the
723 * non-atomic nature of the online and offline processing, for example,
724 * the fact that a CPU enters the scheduler after completing the CPU_DYING
727 * This is also why RCU internally marks CPUs online during the
728 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
730 * Disable checking if in an NMI handler because we cannot safely report
731 * errors from NMI handlers anyway.
733 bool rcu_lockdep_current_cpu_online(void)
735 struct rcu_data *rdp;
736 struct rcu_node *rnp;
742 rdp = this_cpu_ptr(&rcu_sched_data);
744 ret = (rdp->grpmask & rnp->qsmaskinit) ||
745 !rcu_scheduler_fully_active;
749 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
751 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
754 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
756 * If the current CPU is idle or running at a first-level (not nested)
757 * interrupt from idle, return true. The caller must have at least
758 * disabled preemption.
760 static int rcu_is_cpu_rrupt_from_idle(void)
762 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
766 * Snapshot the specified CPU's dynticks counter so that we can later
767 * credit them with an implicit quiescent state. Return 1 if this CPU
768 * is in dynticks idle mode, which is an extended quiescent state.
770 static int dyntick_save_progress_counter(struct rcu_data *rdp,
771 bool *isidle, unsigned long *maxj)
773 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
774 rcu_sysidle_check_cpu(rdp, isidle, maxj);
775 return (rdp->dynticks_snap & 0x1) == 0;
779 * This function really isn't for public consumption, but RCU is special in
780 * that context switches can allow the state machine to make progress.
782 extern void resched_cpu(int cpu);
785 * Return true if the specified CPU has passed through a quiescent
786 * state by virtue of being in or having passed through an dynticks
787 * idle state since the last call to dyntick_save_progress_counter()
788 * for this same CPU, or by virtue of having been offline.
790 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
791 bool *isidle, unsigned long *maxj)
796 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
797 snap = (unsigned int)rdp->dynticks_snap;
800 * If the CPU passed through or entered a dynticks idle phase with
801 * no active irq/NMI handlers, then we can safely pretend that the CPU
802 * already acknowledged the request to pass through a quiescent
803 * state. Either way, that CPU cannot possibly be in an RCU
804 * read-side critical section that started before the beginning
805 * of the current RCU grace period.
807 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
808 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
814 * Check for the CPU being offline, but only if the grace period
815 * is old enough. We don't need to worry about the CPU changing
816 * state: If we see it offline even once, it has been through a
819 * The reason for insisting that the grace period be at least
820 * one jiffy old is that CPUs that are not quite online and that
821 * have just gone offline can still execute RCU read-side critical
824 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
825 return 0; /* Grace period is not old enough. */
827 if (cpu_is_offline(rdp->cpu)) {
828 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
834 * There is a possibility that a CPU in adaptive-ticks state
835 * might run in the kernel with the scheduling-clock tick disabled
836 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
837 * force the CPU to restart the scheduling-clock tick in this
838 * CPU is in this state.
840 rcu_kick_nohz_cpu(rdp->cpu);
843 * Alternatively, the CPU might be running in the kernel
844 * for an extended period of time without a quiescent state.
845 * Attempt to force the CPU through the scheduler to gain the
846 * needed quiescent state, but only if the grace period has gone
847 * on for an uncommonly long time. If there are many stuck CPUs,
848 * we will beat on the first one until it gets unstuck, then move
849 * to the next. Only do this for the primary flavor of RCU.
851 if (rdp->rsp == rcu_state &&
852 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
853 rdp->rsp->jiffies_resched += 5;
854 resched_cpu(rdp->cpu);
860 static void record_gp_stall_check_time(struct rcu_state *rsp)
862 unsigned long j = jiffies;
866 smp_wmb(); /* Record start time before stall time. */
867 j1 = rcu_jiffies_till_stall_check();
868 rsp->jiffies_stall = j + j1;
869 rsp->jiffies_resched = j + j1 / 2;
873 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
874 * for architectures that do not implement trigger_all_cpu_backtrace().
875 * The NMI-triggered stack traces are more accurate because they are
876 * printed by the target CPU.
878 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
882 struct rcu_node *rnp;
884 rcu_for_each_leaf_node(rsp, rnp) {
885 raw_spin_lock_irqsave(&rnp->lock, flags);
886 if (rnp->qsmask != 0) {
887 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
888 if (rnp->qsmask & (1UL << cpu))
889 dump_cpu_task(rnp->grplo + cpu);
891 raw_spin_unlock_irqrestore(&rnp->lock, flags);
895 static void print_other_cpu_stall(struct rcu_state *rsp)
901 struct rcu_node *rnp = rcu_get_root(rsp);
904 /* Only let one CPU complain about others per time interval. */
906 raw_spin_lock_irqsave(&rnp->lock, flags);
907 delta = jiffies - rsp->jiffies_stall;
908 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
909 raw_spin_unlock_irqrestore(&rnp->lock, flags);
912 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
913 raw_spin_unlock_irqrestore(&rnp->lock, flags);
916 * OK, time to rat on our buddy...
917 * See Documentation/RCU/stallwarn.txt for info on how to debug
918 * RCU CPU stall warnings.
920 pr_err("INFO: %s detected stalls on CPUs/tasks:",
922 print_cpu_stall_info_begin();
923 rcu_for_each_leaf_node(rsp, rnp) {
924 raw_spin_lock_irqsave(&rnp->lock, flags);
925 ndetected += rcu_print_task_stall(rnp);
926 if (rnp->qsmask != 0) {
927 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
928 if (rnp->qsmask & (1UL << cpu)) {
929 print_cpu_stall_info(rsp,
934 raw_spin_unlock_irqrestore(&rnp->lock, flags);
938 * Now rat on any tasks that got kicked up to the root rcu_node
939 * due to CPU offlining.
941 rnp = rcu_get_root(rsp);
942 raw_spin_lock_irqsave(&rnp->lock, flags);
943 ndetected += rcu_print_task_stall(rnp);
944 raw_spin_unlock_irqrestore(&rnp->lock, flags);
946 print_cpu_stall_info_end();
947 for_each_possible_cpu(cpu)
948 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
949 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
950 smp_processor_id(), (long)(jiffies - rsp->gp_start),
951 (long)rsp->gpnum, (long)rsp->completed, totqlen);
953 pr_err("INFO: Stall ended before state dump start\n");
954 else if (!trigger_all_cpu_backtrace())
955 rcu_dump_cpu_stacks(rsp);
957 /* Complain about tasks blocking the grace period. */
959 rcu_print_detail_task_stall(rsp);
961 force_quiescent_state(rsp); /* Kick them all. */
965 * This function really isn't for public consumption, but RCU is special in
966 * that context switches can allow the state machine to make progress.
968 extern void resched_cpu(int cpu);
970 static void print_cpu_stall(struct rcu_state *rsp)
974 struct rcu_node *rnp = rcu_get_root(rsp);
978 * OK, time to rat on ourselves...
979 * See Documentation/RCU/stallwarn.txt for info on how to debug
980 * RCU CPU stall warnings.
982 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
983 print_cpu_stall_info_begin();
984 print_cpu_stall_info(rsp, smp_processor_id());
985 print_cpu_stall_info_end();
986 for_each_possible_cpu(cpu)
987 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
988 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
989 jiffies - rsp->gp_start,
990 (long)rsp->gpnum, (long)rsp->completed, totqlen);
991 if (!trigger_all_cpu_backtrace())
994 raw_spin_lock_irqsave(&rnp->lock, flags);
995 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
996 rsp->jiffies_stall = jiffies +
997 3 * rcu_jiffies_till_stall_check() + 3;
998 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1001 * Attempt to revive the RCU machinery by forcing a context switch.
1003 * A context switch would normally allow the RCU state machine to make
1004 * progress and it could be we're stuck in kernel space without context
1005 * switches for an entirely unreasonable amount of time.
1007 resched_cpu(smp_processor_id());
1010 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1012 unsigned long completed;
1013 unsigned long gpnum;
1017 struct rcu_node *rnp;
1019 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1024 * Lots of memory barriers to reject false positives.
1026 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1027 * then rsp->gp_start, and finally rsp->completed. These values
1028 * are updated in the opposite order with memory barriers (or
1029 * equivalent) during grace-period initialization and cleanup.
1030 * Now, a false positive can occur if we get an new value of
1031 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1032 * the memory barriers, the only way that this can happen is if one
1033 * grace period ends and another starts between these two fetches.
1034 * Detect this by comparing rsp->completed with the previous fetch
1037 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1038 * and rsp->gp_start suffice to forestall false positives.
1040 gpnum = ACCESS_ONCE(rsp->gpnum);
1041 smp_rmb(); /* Pick up ->gpnum first... */
1042 js = ACCESS_ONCE(rsp->jiffies_stall);
1043 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1044 gps = ACCESS_ONCE(rsp->gp_start);
1045 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1046 completed = ACCESS_ONCE(rsp->completed);
1047 if (ULONG_CMP_GE(completed, gpnum) ||
1048 ULONG_CMP_LT(j, js) ||
1049 ULONG_CMP_GE(gps, js))
1050 return; /* No stall or GP completed since entering function. */
1052 if (rcu_gp_in_progress(rsp) &&
1053 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1055 /* We haven't checked in, so go dump stack. */
1056 print_cpu_stall(rsp);
1058 } else if (rcu_gp_in_progress(rsp) &&
1059 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1061 /* They had a few time units to dump stack, so complain. */
1062 print_other_cpu_stall(rsp);
1067 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1069 * Set the stall-warning timeout way off into the future, thus preventing
1070 * any RCU CPU stall-warning messages from appearing in the current set of
1071 * RCU grace periods.
1073 * The caller must disable hard irqs.
1075 void rcu_cpu_stall_reset(void)
1077 struct rcu_state *rsp;
1079 for_each_rcu_flavor(rsp)
1080 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1084 * Initialize the specified rcu_data structure's callback list to empty.
1086 static void init_callback_list(struct rcu_data *rdp)
1090 if (init_nocb_callback_list(rdp))
1092 rdp->nxtlist = NULL;
1093 for (i = 0; i < RCU_NEXT_SIZE; i++)
1094 rdp->nxttail[i] = &rdp->nxtlist;
1098 * Determine the value that ->completed will have at the end of the
1099 * next subsequent grace period. This is used to tag callbacks so that
1100 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1101 * been dyntick-idle for an extended period with callbacks under the
1102 * influence of RCU_FAST_NO_HZ.
1104 * The caller must hold rnp->lock with interrupts disabled.
1106 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1107 struct rcu_node *rnp)
1110 * If RCU is idle, we just wait for the next grace period.
1111 * But we can only be sure that RCU is idle if we are looking
1112 * at the root rcu_node structure -- otherwise, a new grace
1113 * period might have started, but just not yet gotten around
1114 * to initializing the current non-root rcu_node structure.
1116 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1117 return rnp->completed + 1;
1120 * Otherwise, wait for a possible partial grace period and
1121 * then the subsequent full grace period.
1123 return rnp->completed + 2;
1127 * Trace-event helper function for rcu_start_future_gp() and
1128 * rcu_nocb_wait_gp().
1130 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1131 unsigned long c, const char *s)
1133 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1134 rnp->completed, c, rnp->level,
1135 rnp->grplo, rnp->grphi, s);
1139 * Start some future grace period, as needed to handle newly arrived
1140 * callbacks. The required future grace periods are recorded in each
1141 * rcu_node structure's ->need_future_gp field.
1143 * The caller must hold the specified rcu_node structure's ->lock.
1145 static unsigned long __maybe_unused
1146 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1150 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1153 * Pick up grace-period number for new callbacks. If this
1154 * grace period is already marked as needed, return to the caller.
1156 c = rcu_cbs_completed(rdp->rsp, rnp);
1157 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1158 if (rnp->need_future_gp[c & 0x1]) {
1159 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1164 * If either this rcu_node structure or the root rcu_node structure
1165 * believe that a grace period is in progress, then we must wait
1166 * for the one following, which is in "c". Because our request
1167 * will be noticed at the end of the current grace period, we don't
1168 * need to explicitly start one.
1170 if (rnp->gpnum != rnp->completed ||
1171 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1172 rnp->need_future_gp[c & 0x1]++;
1173 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1178 * There might be no grace period in progress. If we don't already
1179 * hold it, acquire the root rcu_node structure's lock in order to
1180 * start one (if needed).
1182 if (rnp != rnp_root) {
1183 raw_spin_lock(&rnp_root->lock);
1184 smp_mb__after_unlock_lock();
1188 * Get a new grace-period number. If there really is no grace
1189 * period in progress, it will be smaller than the one we obtained
1190 * earlier. Adjust callbacks as needed. Note that even no-CBs
1191 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1193 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1194 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1195 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1196 rdp->nxtcompleted[i] = c;
1199 * If the needed for the required grace period is already
1200 * recorded, trace and leave.
1202 if (rnp_root->need_future_gp[c & 0x1]) {
1203 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1207 /* Record the need for the future grace period. */
1208 rnp_root->need_future_gp[c & 0x1]++;
1210 /* If a grace period is not already in progress, start one. */
1211 if (rnp_root->gpnum != rnp_root->completed) {
1212 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1214 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1215 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1218 if (rnp != rnp_root)
1219 raw_spin_unlock(&rnp_root->lock);
1224 * Clean up any old requests for the just-ended grace period. Also return
1225 * whether any additional grace periods have been requested. Also invoke
1226 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1227 * waiting for this grace period to complete.
1229 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1231 int c = rnp->completed;
1233 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1235 rcu_nocb_gp_cleanup(rsp, rnp);
1236 rnp->need_future_gp[c & 0x1] = 0;
1237 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1238 trace_rcu_future_gp(rnp, rdp, c,
1239 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1244 * If there is room, assign a ->completed number to any callbacks on
1245 * this CPU that have not already been assigned. Also accelerate any
1246 * callbacks that were previously assigned a ->completed number that has
1247 * since proven to be too conservative, which can happen if callbacks get
1248 * assigned a ->completed number while RCU is idle, but with reference to
1249 * a non-root rcu_node structure. This function is idempotent, so it does
1250 * not hurt to call it repeatedly.
1252 * The caller must hold rnp->lock with interrupts disabled.
1254 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1255 struct rcu_data *rdp)
1260 /* If the CPU has no callbacks, nothing to do. */
1261 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1265 * Starting from the sublist containing the callbacks most
1266 * recently assigned a ->completed number and working down, find the
1267 * first sublist that is not assignable to an upcoming grace period.
1268 * Such a sublist has something in it (first two tests) and has
1269 * a ->completed number assigned that will complete sooner than
1270 * the ->completed number for newly arrived callbacks (last test).
1272 * The key point is that any later sublist can be assigned the
1273 * same ->completed number as the newly arrived callbacks, which
1274 * means that the callbacks in any of these later sublist can be
1275 * grouped into a single sublist, whether or not they have already
1276 * been assigned a ->completed number.
1278 c = rcu_cbs_completed(rsp, rnp);
1279 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1280 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1281 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1285 * If there are no sublist for unassigned callbacks, leave.
1286 * At the same time, advance "i" one sublist, so that "i" will
1287 * index into the sublist where all the remaining callbacks should
1290 if (++i >= RCU_NEXT_TAIL)
1294 * Assign all subsequent callbacks' ->completed number to the next
1295 * full grace period and group them all in the sublist initially
1298 for (; i <= RCU_NEXT_TAIL; i++) {
1299 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1300 rdp->nxtcompleted[i] = c;
1302 /* Record any needed additional grace periods. */
1303 rcu_start_future_gp(rnp, rdp);
1305 /* Trace depending on how much we were able to accelerate. */
1306 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1307 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1309 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1313 * Move any callbacks whose grace period has completed to the
1314 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1315 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1316 * sublist. This function is idempotent, so it does not hurt to
1317 * invoke it repeatedly. As long as it is not invoked -too- often...
1319 * The caller must hold rnp->lock with interrupts disabled.
1321 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1322 struct rcu_data *rdp)
1326 /* If the CPU has no callbacks, nothing to do. */
1327 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1331 * Find all callbacks whose ->completed numbers indicate that they
1332 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1334 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1335 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1337 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1339 /* Clean up any sublist tail pointers that were misordered above. */
1340 for (j = RCU_WAIT_TAIL; j < i; j++)
1341 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1343 /* Copy down callbacks to fill in empty sublists. */
1344 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1345 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1347 rdp->nxttail[j] = rdp->nxttail[i];
1348 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1351 /* Classify any remaining callbacks. */
1352 rcu_accelerate_cbs(rsp, rnp, rdp);
1356 * Update CPU-local rcu_data state to record the beginnings and ends of
1357 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1358 * structure corresponding to the current CPU, and must have irqs disabled.
1360 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1362 /* Handle the ends of any preceding grace periods first. */
1363 if (rdp->completed == rnp->completed) {
1365 /* No grace period end, so just accelerate recent callbacks. */
1366 rcu_accelerate_cbs(rsp, rnp, rdp);
1370 /* Advance callbacks. */
1371 rcu_advance_cbs(rsp, rnp, rdp);
1373 /* Remember that we saw this grace-period completion. */
1374 rdp->completed = rnp->completed;
1375 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1378 if (rdp->gpnum != rnp->gpnum) {
1380 * If the current grace period is waiting for this CPU,
1381 * set up to detect a quiescent state, otherwise don't
1382 * go looking for one.
1384 rdp->gpnum = rnp->gpnum;
1385 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1386 rdp->passed_quiesce = 0;
1387 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1388 zero_cpu_stall_ticks(rdp);
1392 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1394 unsigned long flags;
1395 struct rcu_node *rnp;
1397 local_irq_save(flags);
1399 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1400 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1401 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1402 local_irq_restore(flags);
1405 smp_mb__after_unlock_lock();
1406 __note_gp_changes(rsp, rnp, rdp);
1407 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1411 * Initialize a new grace period. Return 0 if no grace period required.
1413 static int rcu_gp_init(struct rcu_state *rsp)
1415 struct rcu_data *rdp;
1416 struct rcu_node *rnp = rcu_get_root(rsp);
1418 rcu_bind_gp_kthread();
1419 raw_spin_lock_irq(&rnp->lock);
1420 smp_mb__after_unlock_lock();
1421 if (!ACCESS_ONCE(rsp->gp_flags)) {
1422 /* Spurious wakeup, tell caller to go back to sleep. */
1423 raw_spin_unlock_irq(&rnp->lock);
1426 ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */
1428 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1430 * Grace period already in progress, don't start another.
1431 * Not supposed to be able to happen.
1433 raw_spin_unlock_irq(&rnp->lock);
1437 /* Advance to a new grace period and initialize state. */
1438 record_gp_stall_check_time(rsp);
1439 /* Record GP times before starting GP, hence smp_store_release(). */
1440 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1441 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1442 raw_spin_unlock_irq(&rnp->lock);
1444 /* Exclude any concurrent CPU-hotplug operations. */
1445 mutex_lock(&rsp->onoff_mutex);
1446 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1449 * Set the quiescent-state-needed bits in all the rcu_node
1450 * structures for all currently online CPUs in breadth-first order,
1451 * starting from the root rcu_node structure, relying on the layout
1452 * of the tree within the rsp->node[] array. Note that other CPUs
1453 * will access only the leaves of the hierarchy, thus seeing that no
1454 * grace period is in progress, at least until the corresponding
1455 * leaf node has been initialized. In addition, we have excluded
1456 * CPU-hotplug operations.
1458 * The grace period cannot complete until the initialization
1459 * process finishes, because this kthread handles both.
1461 rcu_for_each_node_breadth_first(rsp, rnp) {
1462 raw_spin_lock_irq(&rnp->lock);
1463 smp_mb__after_unlock_lock();
1464 rdp = this_cpu_ptr(rsp->rda);
1465 rcu_preempt_check_blocked_tasks(rnp);
1466 rnp->qsmask = rnp->qsmaskinit;
1467 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1468 WARN_ON_ONCE(rnp->completed != rsp->completed);
1469 ACCESS_ONCE(rnp->completed) = rsp->completed;
1470 if (rnp == rdp->mynode)
1471 __note_gp_changes(rsp, rnp, rdp);
1472 rcu_preempt_boost_start_gp(rnp);
1473 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1474 rnp->level, rnp->grplo,
1475 rnp->grphi, rnp->qsmask);
1476 raw_spin_unlock_irq(&rnp->lock);
1477 #ifdef CONFIG_PROVE_RCU_DELAY
1478 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1479 system_state == SYSTEM_RUNNING)
1481 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1485 mutex_unlock(&rsp->onoff_mutex);
1490 * Do one round of quiescent-state forcing.
1492 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1494 int fqs_state = fqs_state_in;
1495 bool isidle = false;
1497 struct rcu_node *rnp = rcu_get_root(rsp);
1500 if (fqs_state == RCU_SAVE_DYNTICK) {
1501 /* Collect dyntick-idle snapshots. */
1502 if (is_sysidle_rcu_state(rsp)) {
1504 maxj = jiffies - ULONG_MAX / 4;
1506 force_qs_rnp(rsp, dyntick_save_progress_counter,
1508 rcu_sysidle_report_gp(rsp, isidle, maxj);
1509 fqs_state = RCU_FORCE_QS;
1511 /* Handle dyntick-idle and offline CPUs. */
1513 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1515 /* Clear flag to prevent immediate re-entry. */
1516 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1517 raw_spin_lock_irq(&rnp->lock);
1518 smp_mb__after_unlock_lock();
1519 ACCESS_ONCE(rsp->gp_flags) &= ~RCU_GP_FLAG_FQS;
1520 raw_spin_unlock_irq(&rnp->lock);
1526 * Clean up after the old grace period.
1528 static void rcu_gp_cleanup(struct rcu_state *rsp)
1530 unsigned long gp_duration;
1532 struct rcu_data *rdp;
1533 struct rcu_node *rnp = rcu_get_root(rsp);
1535 raw_spin_lock_irq(&rnp->lock);
1536 smp_mb__after_unlock_lock();
1537 gp_duration = jiffies - rsp->gp_start;
1538 if (gp_duration > rsp->gp_max)
1539 rsp->gp_max = gp_duration;
1542 * We know the grace period is complete, but to everyone else
1543 * it appears to still be ongoing. But it is also the case
1544 * that to everyone else it looks like there is nothing that
1545 * they can do to advance the grace period. It is therefore
1546 * safe for us to drop the lock in order to mark the grace
1547 * period as completed in all of the rcu_node structures.
1549 raw_spin_unlock_irq(&rnp->lock);
1552 * Propagate new ->completed value to rcu_node structures so
1553 * that other CPUs don't have to wait until the start of the next
1554 * grace period to process their callbacks. This also avoids
1555 * some nasty RCU grace-period initialization races by forcing
1556 * the end of the current grace period to be completely recorded in
1557 * all of the rcu_node structures before the beginning of the next
1558 * grace period is recorded in any of the rcu_node structures.
1560 rcu_for_each_node_breadth_first(rsp, rnp) {
1561 raw_spin_lock_irq(&rnp->lock);
1562 smp_mb__after_unlock_lock();
1563 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1564 rdp = this_cpu_ptr(rsp->rda);
1565 if (rnp == rdp->mynode)
1566 __note_gp_changes(rsp, rnp, rdp);
1567 /* smp_mb() provided by prior unlock-lock pair. */
1568 nocb += rcu_future_gp_cleanup(rsp, rnp);
1569 raw_spin_unlock_irq(&rnp->lock);
1572 rnp = rcu_get_root(rsp);
1573 raw_spin_lock_irq(&rnp->lock);
1574 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1575 rcu_nocb_gp_set(rnp, nocb);
1577 /* Declare grace period done. */
1578 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1579 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1580 rsp->fqs_state = RCU_GP_IDLE;
1581 rdp = this_cpu_ptr(rsp->rda);
1582 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1583 if (cpu_needs_another_gp(rsp, rdp)) {
1584 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1585 trace_rcu_grace_period(rsp->name,
1586 ACCESS_ONCE(rsp->gpnum),
1589 raw_spin_unlock_irq(&rnp->lock);
1593 * Body of kthread that handles grace periods.
1595 static int __noreturn rcu_gp_kthread(void *arg)
1601 struct rcu_state *rsp = arg;
1602 struct rcu_node *rnp = rcu_get_root(rsp);
1606 /* Handle grace-period start. */
1608 trace_rcu_grace_period(rsp->name,
1609 ACCESS_ONCE(rsp->gpnum),
1611 wait_event_interruptible(rsp->gp_wq,
1612 ACCESS_ONCE(rsp->gp_flags) &
1614 /* Locking provides needed memory barrier. */
1615 if (rcu_gp_init(rsp))
1618 flush_signals(current);
1619 trace_rcu_grace_period(rsp->name,
1620 ACCESS_ONCE(rsp->gpnum),
1624 /* Handle quiescent-state forcing. */
1625 fqs_state = RCU_SAVE_DYNTICK;
1626 j = jiffies_till_first_fqs;
1629 jiffies_till_first_fqs = HZ;
1634 rsp->jiffies_force_qs = jiffies + j;
1635 trace_rcu_grace_period(rsp->name,
1636 ACCESS_ONCE(rsp->gpnum),
1638 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1639 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1641 (!ACCESS_ONCE(rnp->qsmask) &&
1642 !rcu_preempt_blocked_readers_cgp(rnp)),
1644 /* Locking provides needed memory barriers. */
1645 /* If grace period done, leave loop. */
1646 if (!ACCESS_ONCE(rnp->qsmask) &&
1647 !rcu_preempt_blocked_readers_cgp(rnp))
1649 /* If time for quiescent-state forcing, do it. */
1650 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1651 (gf & RCU_GP_FLAG_FQS)) {
1652 trace_rcu_grace_period(rsp->name,
1653 ACCESS_ONCE(rsp->gpnum),
1655 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1656 trace_rcu_grace_period(rsp->name,
1657 ACCESS_ONCE(rsp->gpnum),
1661 /* Deal with stray signal. */
1663 flush_signals(current);
1664 trace_rcu_grace_period(rsp->name,
1665 ACCESS_ONCE(rsp->gpnum),
1668 j = jiffies_till_next_fqs;
1671 jiffies_till_next_fqs = HZ;
1674 jiffies_till_next_fqs = 1;
1678 /* Handle grace-period end. */
1679 rcu_gp_cleanup(rsp);
1683 static void rsp_wakeup(struct irq_work *work)
1685 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1687 /* Wake up rcu_gp_kthread() to start the grace period. */
1688 wake_up(&rsp->gp_wq);
1689 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1694 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1695 * in preparation for detecting the next grace period. The caller must hold
1696 * the root node's ->lock and hard irqs must be disabled.
1698 * Note that it is legal for a dying CPU (which is marked as offline) to
1699 * invoke this function. This can happen when the dying CPU reports its
1703 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1704 struct rcu_data *rdp)
1706 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1708 * Either we have not yet spawned the grace-period
1709 * task, this CPU does not need another grace period,
1710 * or a grace period is already in progress.
1711 * Either way, don't start a new grace period.
1715 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1716 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1720 * We can't do wakeups while holding the rnp->lock, as that
1721 * could cause possible deadlocks with the rq->lock. Defer
1722 * the wakeup to interrupt context. And don't bother waking
1723 * up the running kthread.
1725 if (current != rsp->gp_kthread) {
1726 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1728 irq_work_queue(&rsp->wakeup_work);
1733 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1734 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1735 * is invoked indirectly from rcu_advance_cbs(), which would result in
1736 * endless recursion -- or would do so if it wasn't for the self-deadlock
1737 * that is encountered beforehand.
1740 rcu_start_gp(struct rcu_state *rsp)
1742 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1743 struct rcu_node *rnp = rcu_get_root(rsp);
1746 * If there is no grace period in progress right now, any
1747 * callbacks we have up to this point will be satisfied by the
1748 * next grace period. Also, advancing the callbacks reduces the
1749 * probability of false positives from cpu_needs_another_gp()
1750 * resulting in pointless grace periods. So, advance callbacks
1751 * then start the grace period!
1753 rcu_advance_cbs(rsp, rnp, rdp);
1754 rcu_start_gp_advanced(rsp, rnp, rdp);
1758 * Report a full set of quiescent states to the specified rcu_state
1759 * data structure. This involves cleaning up after the prior grace
1760 * period and letting rcu_start_gp() start up the next grace period
1761 * if one is needed. Note that the caller must hold rnp->lock, which
1762 * is released before return.
1764 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1765 __releases(rcu_get_root(rsp)->lock)
1767 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1768 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1769 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1773 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1774 * Allows quiescent states for a group of CPUs to be reported at one go
1775 * to the specified rcu_node structure, though all the CPUs in the group
1776 * must be represented by the same rcu_node structure (which need not be
1777 * a leaf rcu_node structure, though it often will be). That structure's
1778 * lock must be held upon entry, and it is released before return.
1781 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1782 struct rcu_node *rnp, unsigned long flags)
1783 __releases(rnp->lock)
1785 struct rcu_node *rnp_c;
1787 /* Walk up the rcu_node hierarchy. */
1789 if (!(rnp->qsmask & mask)) {
1791 /* Our bit has already been cleared, so done. */
1792 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1795 rnp->qsmask &= ~mask;
1796 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1797 mask, rnp->qsmask, rnp->level,
1798 rnp->grplo, rnp->grphi,
1800 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1802 /* Other bits still set at this level, so done. */
1803 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1806 mask = rnp->grpmask;
1807 if (rnp->parent == NULL) {
1809 /* No more levels. Exit loop holding root lock. */
1813 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1816 raw_spin_lock_irqsave(&rnp->lock, flags);
1817 smp_mb__after_unlock_lock();
1818 WARN_ON_ONCE(rnp_c->qsmask);
1822 * Get here if we are the last CPU to pass through a quiescent
1823 * state for this grace period. Invoke rcu_report_qs_rsp()
1824 * to clean up and start the next grace period if one is needed.
1826 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1830 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1831 * structure. This must be either called from the specified CPU, or
1832 * called when the specified CPU is known to be offline (and when it is
1833 * also known that no other CPU is concurrently trying to help the offline
1834 * CPU). The lastcomp argument is used to make sure we are still in the
1835 * grace period of interest. We don't want to end the current grace period
1836 * based on quiescent states detected in an earlier grace period!
1839 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1841 unsigned long flags;
1843 struct rcu_node *rnp;
1846 raw_spin_lock_irqsave(&rnp->lock, flags);
1847 smp_mb__after_unlock_lock();
1848 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1849 rnp->completed == rnp->gpnum) {
1852 * The grace period in which this quiescent state was
1853 * recorded has ended, so don't report it upwards.
1854 * We will instead need a new quiescent state that lies
1855 * within the current grace period.
1857 rdp->passed_quiesce = 0; /* need qs for new gp. */
1858 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1861 mask = rdp->grpmask;
1862 if ((rnp->qsmask & mask) == 0) {
1863 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1865 rdp->qs_pending = 0;
1868 * This GP can't end until cpu checks in, so all of our
1869 * callbacks can be processed during the next GP.
1871 rcu_accelerate_cbs(rsp, rnp, rdp);
1873 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1878 * Check to see if there is a new grace period of which this CPU
1879 * is not yet aware, and if so, set up local rcu_data state for it.
1880 * Otherwise, see if this CPU has just passed through its first
1881 * quiescent state for this grace period, and record that fact if so.
1884 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1886 /* Check for grace-period ends and beginnings. */
1887 note_gp_changes(rsp, rdp);
1890 * Does this CPU still need to do its part for current grace period?
1891 * If no, return and let the other CPUs do their part as well.
1893 if (!rdp->qs_pending)
1897 * Was there a quiescent state since the beginning of the grace
1898 * period? If no, then exit and wait for the next call.
1900 if (!rdp->passed_quiesce)
1904 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1907 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1910 #ifdef CONFIG_HOTPLUG_CPU
1913 * Send the specified CPU's RCU callbacks to the orphanage. The
1914 * specified CPU must be offline, and the caller must hold the
1918 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1919 struct rcu_node *rnp, struct rcu_data *rdp)
1921 /* No-CBs CPUs do not have orphanable callbacks. */
1922 if (rcu_is_nocb_cpu(rdp->cpu))
1926 * Orphan the callbacks. First adjust the counts. This is safe
1927 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1928 * cannot be running now. Thus no memory barrier is required.
1930 if (rdp->nxtlist != NULL) {
1931 rsp->qlen_lazy += rdp->qlen_lazy;
1932 rsp->qlen += rdp->qlen;
1933 rdp->n_cbs_orphaned += rdp->qlen;
1935 ACCESS_ONCE(rdp->qlen) = 0;
1939 * Next, move those callbacks still needing a grace period to
1940 * the orphanage, where some other CPU will pick them up.
1941 * Some of the callbacks might have gone partway through a grace
1942 * period, but that is too bad. They get to start over because we
1943 * cannot assume that grace periods are synchronized across CPUs.
1944 * We don't bother updating the ->nxttail[] array yet, instead
1945 * we just reset the whole thing later on.
1947 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1948 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1949 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1950 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1954 * Then move the ready-to-invoke callbacks to the orphanage,
1955 * where some other CPU will pick them up. These will not be
1956 * required to pass though another grace period: They are done.
1958 if (rdp->nxtlist != NULL) {
1959 *rsp->orphan_donetail = rdp->nxtlist;
1960 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1963 /* Finally, initialize the rcu_data structure's list to empty. */
1964 init_callback_list(rdp);
1968 * Adopt the RCU callbacks from the specified rcu_state structure's
1969 * orphanage. The caller must hold the ->orphan_lock.
1971 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
1974 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1976 /* No-CBs CPUs are handled specially. */
1977 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
1980 /* Do the accounting first. */
1981 rdp->qlen_lazy += rsp->qlen_lazy;
1982 rdp->qlen += rsp->qlen;
1983 rdp->n_cbs_adopted += rsp->qlen;
1984 if (rsp->qlen_lazy != rsp->qlen)
1985 rcu_idle_count_callbacks_posted();
1990 * We do not need a memory barrier here because the only way we
1991 * can get here if there is an rcu_barrier() in flight is if
1992 * we are the task doing the rcu_barrier().
1995 /* First adopt the ready-to-invoke callbacks. */
1996 if (rsp->orphan_donelist != NULL) {
1997 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1998 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1999 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2000 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2001 rdp->nxttail[i] = rsp->orphan_donetail;
2002 rsp->orphan_donelist = NULL;
2003 rsp->orphan_donetail = &rsp->orphan_donelist;
2006 /* And then adopt the callbacks that still need a grace period. */
2007 if (rsp->orphan_nxtlist != NULL) {
2008 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2009 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2010 rsp->orphan_nxtlist = NULL;
2011 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2016 * Trace the fact that this CPU is going offline.
2018 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2020 RCU_TRACE(unsigned long mask);
2021 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2022 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2024 RCU_TRACE(mask = rdp->grpmask);
2025 trace_rcu_grace_period(rsp->name,
2026 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2031 * The CPU has been completely removed, and some other CPU is reporting
2032 * this fact from process context. Do the remainder of the cleanup,
2033 * including orphaning the outgoing CPU's RCU callbacks, and also
2034 * adopting them. There can only be one CPU hotplug operation at a time,
2035 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2037 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2039 unsigned long flags;
2041 int need_report = 0;
2042 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2043 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2045 /* Adjust any no-longer-needed kthreads. */
2046 rcu_boost_kthread_setaffinity(rnp, -1);
2048 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2050 /* Exclude any attempts to start a new grace period. */
2051 mutex_lock(&rsp->onoff_mutex);
2052 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2054 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2055 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2056 rcu_adopt_orphan_cbs(rsp, flags);
2058 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2059 mask = rdp->grpmask; /* rnp->grplo is constant. */
2061 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2062 smp_mb__after_unlock_lock();
2063 rnp->qsmaskinit &= ~mask;
2064 if (rnp->qsmaskinit != 0) {
2065 if (rnp != rdp->mynode)
2066 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2069 if (rnp == rdp->mynode)
2070 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2072 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2073 mask = rnp->grpmask;
2075 } while (rnp != NULL);
2078 * We still hold the leaf rcu_node structure lock here, and
2079 * irqs are still disabled. The reason for this subterfuge is
2080 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2081 * held leads to deadlock.
2083 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2085 if (need_report & RCU_OFL_TASKS_NORM_GP)
2086 rcu_report_unblock_qs_rnp(rnp, flags);
2088 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2089 if (need_report & RCU_OFL_TASKS_EXP_GP)
2090 rcu_report_exp_rnp(rsp, rnp, true);
2091 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2092 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2093 cpu, rdp->qlen, rdp->nxtlist);
2094 init_callback_list(rdp);
2095 /* Disallow further callbacks on this CPU. */
2096 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2097 mutex_unlock(&rsp->onoff_mutex);
2100 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2102 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2106 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2110 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2113 * Invoke any RCU callbacks that have made it to the end of their grace
2114 * period. Thottle as specified by rdp->blimit.
2116 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2118 unsigned long flags;
2119 struct rcu_head *next, *list, **tail;
2120 long bl, count, count_lazy;
2123 /* If no callbacks are ready, just return. */
2124 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2125 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2126 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2127 need_resched(), is_idle_task(current),
2128 rcu_is_callbacks_kthread());
2133 * Extract the list of ready callbacks, disabling to prevent
2134 * races with call_rcu() from interrupt handlers.
2136 local_irq_save(flags);
2137 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2139 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2140 list = rdp->nxtlist;
2141 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2142 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2143 tail = rdp->nxttail[RCU_DONE_TAIL];
2144 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2145 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2146 rdp->nxttail[i] = &rdp->nxtlist;
2147 local_irq_restore(flags);
2149 /* Invoke callbacks. */
2150 count = count_lazy = 0;
2154 debug_rcu_head_unqueue(list);
2155 if (__rcu_reclaim(rsp->name, list))
2158 /* Stop only if limit reached and CPU has something to do. */
2159 if (++count >= bl &&
2161 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2165 local_irq_save(flags);
2166 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2167 is_idle_task(current),
2168 rcu_is_callbacks_kthread());
2170 /* Update count, and requeue any remaining callbacks. */
2172 *tail = rdp->nxtlist;
2173 rdp->nxtlist = list;
2174 for (i = 0; i < RCU_NEXT_SIZE; i++)
2175 if (&rdp->nxtlist == rdp->nxttail[i])
2176 rdp->nxttail[i] = tail;
2180 smp_mb(); /* List handling before counting for rcu_barrier(). */
2181 rdp->qlen_lazy -= count_lazy;
2182 ACCESS_ONCE(rdp->qlen) -= count;
2183 rdp->n_cbs_invoked += count;
2185 /* Reinstate batch limit if we have worked down the excess. */
2186 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2187 rdp->blimit = blimit;
2189 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2190 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2191 rdp->qlen_last_fqs_check = 0;
2192 rdp->n_force_qs_snap = rsp->n_force_qs;
2193 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2194 rdp->qlen_last_fqs_check = rdp->qlen;
2195 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2197 local_irq_restore(flags);
2199 /* Re-invoke RCU core processing if there are callbacks remaining. */
2200 if (cpu_has_callbacks_ready_to_invoke(rdp))
2205 * Check to see if this CPU is in a non-context-switch quiescent state
2206 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2207 * Also schedule RCU core processing.
2209 * This function must be called from hardirq context. It is normally
2210 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2211 * false, there is no point in invoking rcu_check_callbacks().
2213 void rcu_check_callbacks(int cpu, int user)
2215 trace_rcu_utilization(TPS("Start scheduler-tick"));
2216 increment_cpu_stall_ticks();
2217 if (user || rcu_is_cpu_rrupt_from_idle()) {
2220 * Get here if this CPU took its interrupt from user
2221 * mode or from the idle loop, and if this is not a
2222 * nested interrupt. In this case, the CPU is in
2223 * a quiescent state, so note it.
2225 * No memory barrier is required here because both
2226 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2227 * variables that other CPUs neither access nor modify,
2228 * at least not while the corresponding CPU is online.
2234 } else if (!in_softirq()) {
2237 * Get here if this CPU did not take its interrupt from
2238 * softirq, in other words, if it is not interrupting
2239 * a rcu_bh read-side critical section. This is an _bh
2240 * critical section, so note it.
2245 rcu_preempt_check_callbacks(cpu);
2246 if (rcu_pending(cpu))
2248 trace_rcu_utilization(TPS("End scheduler-tick"));
2252 * Scan the leaf rcu_node structures, processing dyntick state for any that
2253 * have not yet encountered a quiescent state, using the function specified.
2254 * Also initiate boosting for any threads blocked on the root rcu_node.
2256 * The caller must have suppressed start of new grace periods.
2258 static void force_qs_rnp(struct rcu_state *rsp,
2259 int (*f)(struct rcu_data *rsp, bool *isidle,
2260 unsigned long *maxj),
2261 bool *isidle, unsigned long *maxj)
2265 unsigned long flags;
2267 struct rcu_node *rnp;
2269 rcu_for_each_leaf_node(rsp, rnp) {
2272 raw_spin_lock_irqsave(&rnp->lock, flags);
2273 smp_mb__after_unlock_lock();
2274 if (!rcu_gp_in_progress(rsp)) {
2275 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2278 if (rnp->qsmask == 0) {
2279 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2284 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2285 if ((rnp->qsmask & bit) != 0) {
2286 if ((rnp->qsmaskinit & bit) != 0)
2288 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2294 /* rcu_report_qs_rnp() releases rnp->lock. */
2295 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2298 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2300 rnp = rcu_get_root(rsp);
2301 if (rnp->qsmask == 0) {
2302 raw_spin_lock_irqsave(&rnp->lock, flags);
2303 smp_mb__after_unlock_lock();
2304 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2309 * Force quiescent states on reluctant CPUs, and also detect which
2310 * CPUs are in dyntick-idle mode.
2312 static void force_quiescent_state(struct rcu_state *rsp)
2314 unsigned long flags;
2316 struct rcu_node *rnp;
2317 struct rcu_node *rnp_old = NULL;
2319 /* Funnel through hierarchy to reduce memory contention. */
2320 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2321 for (; rnp != NULL; rnp = rnp->parent) {
2322 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2323 !raw_spin_trylock(&rnp->fqslock);
2324 if (rnp_old != NULL)
2325 raw_spin_unlock(&rnp_old->fqslock);
2327 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2332 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2334 /* Reached the root of the rcu_node tree, acquire lock. */
2335 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2336 smp_mb__after_unlock_lock();
2337 raw_spin_unlock(&rnp_old->fqslock);
2338 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2339 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2340 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2341 return; /* Someone beat us to it. */
2343 ACCESS_ONCE(rsp->gp_flags) |= RCU_GP_FLAG_FQS;
2344 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2345 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2349 * This does the RCU core processing work for the specified rcu_state
2350 * and rcu_data structures. This may be called only from the CPU to
2351 * whom the rdp belongs.
2354 __rcu_process_callbacks(struct rcu_state *rsp)
2356 unsigned long flags;
2357 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2359 WARN_ON_ONCE(rdp->beenonline == 0);
2361 /* Update RCU state based on any recent quiescent states. */
2362 rcu_check_quiescent_state(rsp, rdp);
2364 /* Does this CPU require a not-yet-started grace period? */
2365 local_irq_save(flags);
2366 if (cpu_needs_another_gp(rsp, rdp)) {
2367 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2369 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2371 local_irq_restore(flags);
2374 /* If there are callbacks ready, invoke them. */
2375 if (cpu_has_callbacks_ready_to_invoke(rdp))
2376 invoke_rcu_callbacks(rsp, rdp);
2378 /* Do any needed deferred wakeups of rcuo kthreads. */
2379 do_nocb_deferred_wakeup(rdp);
2383 * Do RCU core processing for the current CPU.
2385 static void rcu_process_callbacks(struct softirq_action *unused)
2387 struct rcu_state *rsp;
2389 if (cpu_is_offline(smp_processor_id()))
2391 trace_rcu_utilization(TPS("Start RCU core"));
2392 for_each_rcu_flavor(rsp)
2393 __rcu_process_callbacks(rsp);
2394 trace_rcu_utilization(TPS("End RCU core"));
2398 * Schedule RCU callback invocation. If the specified type of RCU
2399 * does not support RCU priority boosting, just do a direct call,
2400 * otherwise wake up the per-CPU kernel kthread. Note that because we
2401 * are running on the current CPU with interrupts disabled, the
2402 * rcu_cpu_kthread_task cannot disappear out from under us.
2404 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2406 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2408 if (likely(!rsp->boost)) {
2409 rcu_do_batch(rsp, rdp);
2412 invoke_rcu_callbacks_kthread();
2415 static void invoke_rcu_core(void)
2417 if (cpu_online(smp_processor_id()))
2418 raise_softirq(RCU_SOFTIRQ);
2422 * Handle any core-RCU processing required by a call_rcu() invocation.
2424 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2425 struct rcu_head *head, unsigned long flags)
2428 * If called from an extended quiescent state, invoke the RCU
2429 * core in order to force a re-evaluation of RCU's idleness.
2431 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2434 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2435 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2439 * Force the grace period if too many callbacks or too long waiting.
2440 * Enforce hysteresis, and don't invoke force_quiescent_state()
2441 * if some other CPU has recently done so. Also, don't bother
2442 * invoking force_quiescent_state() if the newly enqueued callback
2443 * is the only one waiting for a grace period to complete.
2445 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2447 /* Are we ignoring a completed grace period? */
2448 note_gp_changes(rsp, rdp);
2450 /* Start a new grace period if one not already started. */
2451 if (!rcu_gp_in_progress(rsp)) {
2452 struct rcu_node *rnp_root = rcu_get_root(rsp);
2454 raw_spin_lock(&rnp_root->lock);
2455 smp_mb__after_unlock_lock();
2457 raw_spin_unlock(&rnp_root->lock);
2459 /* Give the grace period a kick. */
2460 rdp->blimit = LONG_MAX;
2461 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2462 *rdp->nxttail[RCU_DONE_TAIL] != head)
2463 force_quiescent_state(rsp);
2464 rdp->n_force_qs_snap = rsp->n_force_qs;
2465 rdp->qlen_last_fqs_check = rdp->qlen;
2471 * RCU callback function to leak a callback.
2473 static void rcu_leak_callback(struct rcu_head *rhp)
2478 * Helper function for call_rcu() and friends. The cpu argument will
2479 * normally be -1, indicating "currently running CPU". It may specify
2480 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2481 * is expected to specify a CPU.
2484 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2485 struct rcu_state *rsp, int cpu, bool lazy)
2487 unsigned long flags;
2488 struct rcu_data *rdp;
2490 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2491 if (debug_rcu_head_queue(head)) {
2492 /* Probable double call_rcu(), so leak the callback. */
2493 ACCESS_ONCE(head->func) = rcu_leak_callback;
2494 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2501 * Opportunistically note grace-period endings and beginnings.
2502 * Note that we might see a beginning right after we see an
2503 * end, but never vice versa, since this CPU has to pass through
2504 * a quiescent state betweentimes.
2506 local_irq_save(flags);
2507 rdp = this_cpu_ptr(rsp->rda);
2509 /* Add the callback to our list. */
2510 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2514 rdp = per_cpu_ptr(rsp->rda, cpu);
2515 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2516 WARN_ON_ONCE(offline);
2517 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2518 local_irq_restore(flags);
2521 ACCESS_ONCE(rdp->qlen)++;
2525 rcu_idle_count_callbacks_posted();
2526 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2527 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2528 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2530 if (__is_kfree_rcu_offset((unsigned long)func))
2531 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2532 rdp->qlen_lazy, rdp->qlen);
2534 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2536 /* Go handle any RCU core processing required. */
2537 __call_rcu_core(rsp, rdp, head, flags);
2538 local_irq_restore(flags);
2542 * Queue an RCU-sched callback for invocation after a grace period.
2544 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2546 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2548 EXPORT_SYMBOL_GPL(call_rcu_sched);
2551 * Queue an RCU callback for invocation after a quicker grace period.
2553 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2555 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2557 EXPORT_SYMBOL_GPL(call_rcu_bh);
2560 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2561 * any blocking grace-period wait automatically implies a grace period
2562 * if there is only one CPU online at any point time during execution
2563 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2564 * occasionally incorrectly indicate that there are multiple CPUs online
2565 * when there was in fact only one the whole time, as this just adds
2566 * some overhead: RCU still operates correctly.
2568 static inline int rcu_blocking_is_gp(void)
2572 might_sleep(); /* Check for RCU read-side critical section. */
2574 ret = num_online_cpus() <= 1;
2580 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2582 * Control will return to the caller some time after a full rcu-sched
2583 * grace period has elapsed, in other words after all currently executing
2584 * rcu-sched read-side critical sections have completed. These read-side
2585 * critical sections are delimited by rcu_read_lock_sched() and
2586 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2587 * local_irq_disable(), and so on may be used in place of
2588 * rcu_read_lock_sched().
2590 * This means that all preempt_disable code sequences, including NMI and
2591 * non-threaded hardware-interrupt handlers, in progress on entry will
2592 * have completed before this primitive returns. However, this does not
2593 * guarantee that softirq handlers will have completed, since in some
2594 * kernels, these handlers can run in process context, and can block.
2596 * Note that this guarantee implies further memory-ordering guarantees.
2597 * On systems with more than one CPU, when synchronize_sched() returns,
2598 * each CPU is guaranteed to have executed a full memory barrier since the
2599 * end of its last RCU-sched read-side critical section whose beginning
2600 * preceded the call to synchronize_sched(). In addition, each CPU having
2601 * an RCU read-side critical section that extends beyond the return from
2602 * synchronize_sched() is guaranteed to have executed a full memory barrier
2603 * after the beginning of synchronize_sched() and before the beginning of
2604 * that RCU read-side critical section. Note that these guarantees include
2605 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2606 * that are executing in the kernel.
2608 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2609 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2610 * to have executed a full memory barrier during the execution of
2611 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2612 * again only if the system has more than one CPU).
2614 * This primitive provides the guarantees made by the (now removed)
2615 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2616 * guarantees that rcu_read_lock() sections will have completed.
2617 * In "classic RCU", these two guarantees happen to be one and
2618 * the same, but can differ in realtime RCU implementations.
2620 void synchronize_sched(void)
2622 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2623 !lock_is_held(&rcu_lock_map) &&
2624 !lock_is_held(&rcu_sched_lock_map),
2625 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2626 if (rcu_blocking_is_gp())
2629 synchronize_sched_expedited();
2631 wait_rcu_gp(call_rcu_sched);
2633 EXPORT_SYMBOL_GPL(synchronize_sched);
2636 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2638 * Control will return to the caller some time after a full rcu_bh grace
2639 * period has elapsed, in other words after all currently executing rcu_bh
2640 * read-side critical sections have completed. RCU read-side critical
2641 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2642 * and may be nested.
2644 * See the description of synchronize_sched() for more detailed information
2645 * on memory ordering guarantees.
2647 void synchronize_rcu_bh(void)
2649 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2650 !lock_is_held(&rcu_lock_map) &&
2651 !lock_is_held(&rcu_sched_lock_map),
2652 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2653 if (rcu_blocking_is_gp())
2656 synchronize_rcu_bh_expedited();
2658 wait_rcu_gp(call_rcu_bh);
2660 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2663 * get_state_synchronize_rcu - Snapshot current RCU state
2665 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2666 * to determine whether or not a full grace period has elapsed in the
2669 unsigned long get_state_synchronize_rcu(void)
2672 * Any prior manipulation of RCU-protected data must happen
2673 * before the load from ->gpnum.
2678 * Make sure this load happens before the purportedly
2679 * time-consuming work between get_state_synchronize_rcu()
2680 * and cond_synchronize_rcu().
2682 return smp_load_acquire(&rcu_state->gpnum);
2684 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2687 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2689 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2691 * If a full RCU grace period has elapsed since the earlier call to
2692 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2693 * synchronize_rcu() to wait for a full grace period.
2695 * Yes, this function does not take counter wrap into account. But
2696 * counter wrap is harmless. If the counter wraps, we have waited for
2697 * more than 2 billion grace periods (and way more on a 64-bit system!),
2698 * so waiting for one additional grace period should be just fine.
2700 void cond_synchronize_rcu(unsigned long oldstate)
2702 unsigned long newstate;
2705 * Ensure that this load happens before any RCU-destructive
2706 * actions the caller might carry out after we return.
2708 newstate = smp_load_acquire(&rcu_state->completed);
2709 if (ULONG_CMP_GE(oldstate, newstate))
2712 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2714 static int synchronize_sched_expedited_cpu_stop(void *data)
2717 * There must be a full memory barrier on each affected CPU
2718 * between the time that try_stop_cpus() is called and the
2719 * time that it returns.
2721 * In the current initial implementation of cpu_stop, the
2722 * above condition is already met when the control reaches
2723 * this point and the following smp_mb() is not strictly
2724 * necessary. Do smp_mb() anyway for documentation and
2725 * robustness against future implementation changes.
2727 smp_mb(); /* See above comment block. */
2732 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2734 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2735 * approach to force the grace period to end quickly. This consumes
2736 * significant time on all CPUs and is unfriendly to real-time workloads,
2737 * so is thus not recommended for any sort of common-case code. In fact,
2738 * if you are using synchronize_sched_expedited() in a loop, please
2739 * restructure your code to batch your updates, and then use a single
2740 * synchronize_sched() instead.
2742 * Note that it is illegal to call this function while holding any lock
2743 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2744 * to call this function from a CPU-hotplug notifier. Failing to observe
2745 * these restriction will result in deadlock.
2747 * This implementation can be thought of as an application of ticket
2748 * locking to RCU, with sync_sched_expedited_started and
2749 * sync_sched_expedited_done taking on the roles of the halves
2750 * of the ticket-lock word. Each task atomically increments
2751 * sync_sched_expedited_started upon entry, snapshotting the old value,
2752 * then attempts to stop all the CPUs. If this succeeds, then each
2753 * CPU will have executed a context switch, resulting in an RCU-sched
2754 * grace period. We are then done, so we use atomic_cmpxchg() to
2755 * update sync_sched_expedited_done to match our snapshot -- but
2756 * only if someone else has not already advanced past our snapshot.
2758 * On the other hand, if try_stop_cpus() fails, we check the value
2759 * of sync_sched_expedited_done. If it has advanced past our
2760 * initial snapshot, then someone else must have forced a grace period
2761 * some time after we took our snapshot. In this case, our work is
2762 * done for us, and we can simply return. Otherwise, we try again,
2763 * but keep our initial snapshot for purposes of checking for someone
2764 * doing our work for us.
2766 * If we fail too many times in a row, we fall back to synchronize_sched().
2768 void synchronize_sched_expedited(void)
2770 long firstsnap, s, snap;
2772 struct rcu_state *rsp = &rcu_sched_state;
2775 * If we are in danger of counter wrap, just do synchronize_sched().
2776 * By allowing sync_sched_expedited_started to advance no more than
2777 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2778 * that more than 3.5 billion CPUs would be required to force a
2779 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2780 * course be required on a 64-bit system.
2782 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2783 (ulong)atomic_long_read(&rsp->expedited_done) +
2785 synchronize_sched();
2786 atomic_long_inc(&rsp->expedited_wrap);
2791 * Take a ticket. Note that atomic_inc_return() implies a
2792 * full memory barrier.
2794 snap = atomic_long_inc_return(&rsp->expedited_start);
2797 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2800 * Each pass through the following loop attempts to force a
2801 * context switch on each CPU.
2803 while (try_stop_cpus(cpu_online_mask,
2804 synchronize_sched_expedited_cpu_stop,
2807 atomic_long_inc(&rsp->expedited_tryfail);
2809 /* Check to see if someone else did our work for us. */
2810 s = atomic_long_read(&rsp->expedited_done);
2811 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2812 /* ensure test happens before caller kfree */
2813 smp_mb__before_atomic_inc(); /* ^^^ */
2814 atomic_long_inc(&rsp->expedited_workdone1);
2818 /* No joy, try again later. Or just synchronize_sched(). */
2819 if (trycount++ < 10) {
2820 udelay(trycount * num_online_cpus());
2822 wait_rcu_gp(call_rcu_sched);
2823 atomic_long_inc(&rsp->expedited_normal);
2827 /* Recheck to see if someone else did our work for us. */
2828 s = atomic_long_read(&rsp->expedited_done);
2829 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2830 /* ensure test happens before caller kfree */
2831 smp_mb__before_atomic_inc(); /* ^^^ */
2832 atomic_long_inc(&rsp->expedited_workdone2);
2837 * Refetching sync_sched_expedited_started allows later
2838 * callers to piggyback on our grace period. We retry
2839 * after they started, so our grace period works for them,
2840 * and they started after our first try, so their grace
2841 * period works for us.
2844 snap = atomic_long_read(&rsp->expedited_start);
2845 smp_mb(); /* ensure read is before try_stop_cpus(). */
2847 atomic_long_inc(&rsp->expedited_stoppedcpus);
2850 * Everyone up to our most recent fetch is covered by our grace
2851 * period. Update the counter, but only if our work is still
2852 * relevant -- which it won't be if someone who started later
2853 * than we did already did their update.
2856 atomic_long_inc(&rsp->expedited_done_tries);
2857 s = atomic_long_read(&rsp->expedited_done);
2858 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2859 /* ensure test happens before caller kfree */
2860 smp_mb__before_atomic_inc(); /* ^^^ */
2861 atomic_long_inc(&rsp->expedited_done_lost);
2864 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2865 atomic_long_inc(&rsp->expedited_done_exit);
2869 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2872 * Check to see if there is any immediate RCU-related work to be done
2873 * by the current CPU, for the specified type of RCU, returning 1 if so.
2874 * The checks are in order of increasing expense: checks that can be
2875 * carried out against CPU-local state are performed first. However,
2876 * we must check for CPU stalls first, else we might not get a chance.
2878 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2880 struct rcu_node *rnp = rdp->mynode;
2882 rdp->n_rcu_pending++;
2884 /* Check for CPU stalls, if enabled. */
2885 check_cpu_stall(rsp, rdp);
2887 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2888 if (rcu_nohz_full_cpu(rsp))
2891 /* Is the RCU core waiting for a quiescent state from this CPU? */
2892 if (rcu_scheduler_fully_active &&
2893 rdp->qs_pending && !rdp->passed_quiesce) {
2894 rdp->n_rp_qs_pending++;
2895 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2896 rdp->n_rp_report_qs++;
2900 /* Does this CPU have callbacks ready to invoke? */
2901 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2902 rdp->n_rp_cb_ready++;
2906 /* Has RCU gone idle with this CPU needing another grace period? */
2907 if (cpu_needs_another_gp(rsp, rdp)) {
2908 rdp->n_rp_cpu_needs_gp++;
2912 /* Has another RCU grace period completed? */
2913 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2914 rdp->n_rp_gp_completed++;
2918 /* Has a new RCU grace period started? */
2919 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2920 rdp->n_rp_gp_started++;
2924 /* Does this CPU need a deferred NOCB wakeup? */
2925 if (rcu_nocb_need_deferred_wakeup(rdp)) {
2926 rdp->n_rp_nocb_defer_wakeup++;
2931 rdp->n_rp_need_nothing++;
2936 * Check to see if there is any immediate RCU-related work to be done
2937 * by the current CPU, returning 1 if so. This function is part of the
2938 * RCU implementation; it is -not- an exported member of the RCU API.
2940 static int rcu_pending(int cpu)
2942 struct rcu_state *rsp;
2944 for_each_rcu_flavor(rsp)
2945 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2951 * Return true if the specified CPU has any callback. If all_lazy is
2952 * non-NULL, store an indication of whether all callbacks are lazy.
2953 * (If there are no callbacks, all of them are deemed to be lazy.)
2955 static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2959 struct rcu_data *rdp;
2960 struct rcu_state *rsp;
2962 for_each_rcu_flavor(rsp) {
2963 rdp = per_cpu_ptr(rsp->rda, cpu);
2967 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2978 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2979 * the compiler is expected to optimize this away.
2981 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2982 int cpu, unsigned long done)
2984 trace_rcu_barrier(rsp->name, s, cpu,
2985 atomic_read(&rsp->barrier_cpu_count), done);
2989 * RCU callback function for _rcu_barrier(). If we are last, wake
2990 * up the task executing _rcu_barrier().
2992 static void rcu_barrier_callback(struct rcu_head *rhp)
2994 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2995 struct rcu_state *rsp = rdp->rsp;
2997 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2998 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2999 complete(&rsp->barrier_completion);
3001 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
3006 * Called with preemption disabled, and from cross-cpu IRQ context.
3008 static void rcu_barrier_func(void *type)
3010 struct rcu_state *rsp = type;
3011 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
3013 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
3014 atomic_inc(&rsp->barrier_cpu_count);
3015 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3019 * Orchestrate the specified type of RCU barrier, waiting for all
3020 * RCU callbacks of the specified type to complete.
3022 static void _rcu_barrier(struct rcu_state *rsp)
3025 struct rcu_data *rdp;
3026 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3027 unsigned long snap_done;
3029 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3031 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3032 mutex_lock(&rsp->barrier_mutex);
3035 * Ensure that all prior references, including to ->n_barrier_done,
3036 * are ordered before the _rcu_barrier() machinery.
3038 smp_mb(); /* See above block comment. */
3041 * Recheck ->n_barrier_done to see if others did our work for us.
3042 * This means checking ->n_barrier_done for an even-to-odd-to-even
3043 * transition. The "if" expression below therefore rounds the old
3044 * value up to the next even number and adds two before comparing.
3046 snap_done = rsp->n_barrier_done;
3047 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3050 * If the value in snap is odd, we needed to wait for the current
3051 * rcu_barrier() to complete, then wait for the next one, in other
3052 * words, we need the value of snap_done to be three larger than
3053 * the value of snap. On the other hand, if the value in snap is
3054 * even, we only had to wait for the next rcu_barrier() to complete,
3055 * in other words, we need the value of snap_done to be only two
3056 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3057 * this for us (thank you, Linus!).
3059 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3060 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3061 smp_mb(); /* caller's subsequent code after above check. */
3062 mutex_unlock(&rsp->barrier_mutex);
3067 * Increment ->n_barrier_done to avoid duplicate work. Use
3068 * ACCESS_ONCE() to prevent the compiler from speculating
3069 * the increment to precede the early-exit check.
3071 ACCESS_ONCE(rsp->n_barrier_done)++;
3072 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3073 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3074 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3077 * Initialize the count to one rather than to zero in order to
3078 * avoid a too-soon return to zero in case of a short grace period
3079 * (or preemption of this task). Exclude CPU-hotplug operations
3080 * to ensure that no offline CPU has callbacks queued.
3082 init_completion(&rsp->barrier_completion);
3083 atomic_set(&rsp->barrier_cpu_count, 1);
3087 * Force each CPU with callbacks to register a new callback.
3088 * When that callback is invoked, we will know that all of the
3089 * corresponding CPU's preceding callbacks have been invoked.
3091 for_each_possible_cpu(cpu) {
3092 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3094 rdp = per_cpu_ptr(rsp->rda, cpu);
3095 if (rcu_is_nocb_cpu(cpu)) {
3096 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3097 rsp->n_barrier_done);
3098 atomic_inc(&rsp->barrier_cpu_count);
3099 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3101 } else if (ACCESS_ONCE(rdp->qlen)) {
3102 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3103 rsp->n_barrier_done);
3104 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3106 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3107 rsp->n_barrier_done);
3113 * Now that we have an rcu_barrier_callback() callback on each
3114 * CPU, and thus each counted, remove the initial count.
3116 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3117 complete(&rsp->barrier_completion);
3119 /* Increment ->n_barrier_done to prevent duplicate work. */
3120 smp_mb(); /* Keep increment after above mechanism. */
3121 ACCESS_ONCE(rsp->n_barrier_done)++;
3122 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3123 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3124 smp_mb(); /* Keep increment before caller's subsequent code. */
3126 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3127 wait_for_completion(&rsp->barrier_completion);
3129 /* Other rcu_barrier() invocations can now safely proceed. */
3130 mutex_unlock(&rsp->barrier_mutex);
3134 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3136 void rcu_barrier_bh(void)
3138 _rcu_barrier(&rcu_bh_state);
3140 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3143 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3145 void rcu_barrier_sched(void)
3147 _rcu_barrier(&rcu_sched_state);
3149 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3152 * Do boot-time initialization of a CPU's per-CPU RCU data.
3155 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3157 unsigned long flags;
3158 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3159 struct rcu_node *rnp = rcu_get_root(rsp);
3161 /* Set up local state, ensuring consistent view of global state. */
3162 raw_spin_lock_irqsave(&rnp->lock, flags);
3163 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3164 init_callback_list(rdp);
3166 ACCESS_ONCE(rdp->qlen) = 0;
3167 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3168 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3169 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3172 rcu_boot_init_nocb_percpu_data(rdp);
3173 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3177 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3178 * offline event can be happening at a given time. Note also that we
3179 * can accept some slop in the rsp->completed access due to the fact
3180 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3183 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3185 unsigned long flags;
3187 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3188 struct rcu_node *rnp = rcu_get_root(rsp);
3190 /* Exclude new grace periods. */
3191 mutex_lock(&rsp->onoff_mutex);
3193 /* Set up local state, ensuring consistent view of global state. */
3194 raw_spin_lock_irqsave(&rnp->lock, flags);
3195 rdp->beenonline = 1; /* We have now been online. */
3196 rdp->preemptible = preemptible;
3197 rdp->qlen_last_fqs_check = 0;
3198 rdp->n_force_qs_snap = rsp->n_force_qs;
3199 rdp->blimit = blimit;
3200 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3201 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3202 rcu_sysidle_init_percpu_data(rdp->dynticks);
3203 atomic_set(&rdp->dynticks->dynticks,
3204 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3205 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3207 /* Add CPU to rcu_node bitmasks. */
3209 mask = rdp->grpmask;
3211 /* Exclude any attempts to start a new GP on small systems. */
3212 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3213 rnp->qsmaskinit |= mask;
3214 mask = rnp->grpmask;
3215 if (rnp == rdp->mynode) {
3217 * If there is a grace period in progress, we will
3218 * set up to wait for it next time we run the
3221 rdp->gpnum = rnp->completed;
3222 rdp->completed = rnp->completed;
3223 rdp->passed_quiesce = 0;
3224 rdp->qs_pending = 0;
3225 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3227 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3229 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3230 local_irq_restore(flags);
3232 mutex_unlock(&rsp->onoff_mutex);
3235 static void rcu_prepare_cpu(int cpu)
3237 struct rcu_state *rsp;
3239 for_each_rcu_flavor(rsp)
3240 rcu_init_percpu_data(cpu, rsp,
3241 strcmp(rsp->name, "rcu_preempt") == 0);
3245 * Handle CPU online/offline notification events.
3247 static int rcu_cpu_notify(struct notifier_block *self,
3248 unsigned long action, void *hcpu)
3250 long cpu = (long)hcpu;
3251 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3252 struct rcu_node *rnp = rdp->mynode;
3253 struct rcu_state *rsp;
3255 trace_rcu_utilization(TPS("Start CPU hotplug"));
3257 case CPU_UP_PREPARE:
3258 case CPU_UP_PREPARE_FROZEN:
3259 rcu_prepare_cpu(cpu);
3260 rcu_prepare_kthreads(cpu);
3263 case CPU_DOWN_FAILED:
3264 rcu_boost_kthread_setaffinity(rnp, -1);
3266 case CPU_DOWN_PREPARE:
3267 rcu_boost_kthread_setaffinity(rnp, cpu);
3270 case CPU_DYING_FROZEN:
3271 for_each_rcu_flavor(rsp)
3272 rcu_cleanup_dying_cpu(rsp);
3275 case CPU_DEAD_FROZEN:
3276 case CPU_UP_CANCELED:
3277 case CPU_UP_CANCELED_FROZEN:
3278 for_each_rcu_flavor(rsp)
3279 rcu_cleanup_dead_cpu(cpu, rsp);
3284 trace_rcu_utilization(TPS("End CPU hotplug"));
3288 static int rcu_pm_notify(struct notifier_block *self,
3289 unsigned long action, void *hcpu)
3292 case PM_HIBERNATION_PREPARE:
3293 case PM_SUSPEND_PREPARE:
3294 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3297 case PM_POST_HIBERNATION:
3298 case PM_POST_SUSPEND:
3308 * Spawn the kthread that handles this RCU flavor's grace periods.
3310 static int __init rcu_spawn_gp_kthread(void)
3312 unsigned long flags;
3313 struct rcu_node *rnp;
3314 struct rcu_state *rsp;
3315 struct task_struct *t;
3317 for_each_rcu_flavor(rsp) {
3318 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3320 rnp = rcu_get_root(rsp);
3321 raw_spin_lock_irqsave(&rnp->lock, flags);
3322 rsp->gp_kthread = t;
3323 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3324 rcu_spawn_nocb_kthreads(rsp);
3328 early_initcall(rcu_spawn_gp_kthread);
3331 * This function is invoked towards the end of the scheduler's initialization
3332 * process. Before this is called, the idle task might contain
3333 * RCU read-side critical sections (during which time, this idle
3334 * task is booting the system). After this function is called, the
3335 * idle tasks are prohibited from containing RCU read-side critical
3336 * sections. This function also enables RCU lockdep checking.
3338 void rcu_scheduler_starting(void)
3340 WARN_ON(num_online_cpus() != 1);
3341 WARN_ON(nr_context_switches() > 0);
3342 rcu_scheduler_active = 1;
3346 * Compute the per-level fanout, either using the exact fanout specified
3347 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3349 #ifdef CONFIG_RCU_FANOUT_EXACT
3350 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3354 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3355 for (i = rcu_num_lvls - 2; i >= 0; i--)
3356 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3358 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3359 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3366 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3367 ccur = rsp->levelcnt[i];
3368 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3372 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3375 * Helper function for rcu_init() that initializes one rcu_state structure.
3377 static void __init rcu_init_one(struct rcu_state *rsp,
3378 struct rcu_data __percpu *rda)
3380 static char *buf[] = { "rcu_node_0",
3383 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3384 static char *fqs[] = { "rcu_node_fqs_0",
3387 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3391 struct rcu_node *rnp;
3393 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3395 /* Silence gcc 4.8 warning about array index out of range. */
3396 if (rcu_num_lvls > RCU_NUM_LVLS)
3397 panic("rcu_init_one: rcu_num_lvls overflow");
3399 /* Initialize the level-tracking arrays. */
3401 for (i = 0; i < rcu_num_lvls; i++)
3402 rsp->levelcnt[i] = num_rcu_lvl[i];
3403 for (i = 1; i < rcu_num_lvls; i++)
3404 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3405 rcu_init_levelspread(rsp);
3407 /* Initialize the elements themselves, starting from the leaves. */
3409 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3410 cpustride *= rsp->levelspread[i];
3411 rnp = rsp->level[i];
3412 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3413 raw_spin_lock_init(&rnp->lock);
3414 lockdep_set_class_and_name(&rnp->lock,
3415 &rcu_node_class[i], buf[i]);
3416 raw_spin_lock_init(&rnp->fqslock);
3417 lockdep_set_class_and_name(&rnp->fqslock,
3418 &rcu_fqs_class[i], fqs[i]);
3419 rnp->gpnum = rsp->gpnum;
3420 rnp->completed = rsp->completed;
3422 rnp->qsmaskinit = 0;
3423 rnp->grplo = j * cpustride;
3424 rnp->grphi = (j + 1) * cpustride - 1;
3425 if (rnp->grphi >= NR_CPUS)
3426 rnp->grphi = NR_CPUS - 1;
3432 rnp->grpnum = j % rsp->levelspread[i - 1];
3433 rnp->grpmask = 1UL << rnp->grpnum;
3434 rnp->parent = rsp->level[i - 1] +
3435 j / rsp->levelspread[i - 1];
3438 INIT_LIST_HEAD(&rnp->blkd_tasks);
3439 rcu_init_one_nocb(rnp);
3444 init_waitqueue_head(&rsp->gp_wq);
3445 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3446 rnp = rsp->level[rcu_num_lvls - 1];
3447 for_each_possible_cpu(i) {
3448 while (i > rnp->grphi)
3450 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3451 rcu_boot_init_percpu_data(i, rsp);
3453 list_add(&rsp->flavors, &rcu_struct_flavors);
3457 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3458 * replace the definitions in tree.h because those are needed to size
3459 * the ->node array in the rcu_state structure.
3461 static void __init rcu_init_geometry(void)
3467 int rcu_capacity[MAX_RCU_LVLS + 1];
3470 * Initialize any unspecified boot parameters.
3471 * The default values of jiffies_till_first_fqs and
3472 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3473 * value, which is a function of HZ, then adding one for each
3474 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3476 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3477 if (jiffies_till_first_fqs == ULONG_MAX)
3478 jiffies_till_first_fqs = d;
3479 if (jiffies_till_next_fqs == ULONG_MAX)
3480 jiffies_till_next_fqs = d;
3482 /* If the compile-time values are accurate, just leave. */
3483 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3484 nr_cpu_ids == NR_CPUS)
3486 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3487 rcu_fanout_leaf, nr_cpu_ids);
3490 * Compute number of nodes that can be handled an rcu_node tree
3491 * with the given number of levels. Setting rcu_capacity[0] makes
3492 * some of the arithmetic easier.
3494 rcu_capacity[0] = 1;
3495 rcu_capacity[1] = rcu_fanout_leaf;
3496 for (i = 2; i <= MAX_RCU_LVLS; i++)
3497 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3500 * The boot-time rcu_fanout_leaf parameter is only permitted
3501 * to increase the leaf-level fanout, not decrease it. Of course,
3502 * the leaf-level fanout cannot exceed the number of bits in
3503 * the rcu_node masks. Finally, the tree must be able to accommodate
3504 * the configured number of CPUs. Complain and fall back to the
3505 * compile-time values if these limits are exceeded.
3507 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3508 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3509 n > rcu_capacity[MAX_RCU_LVLS]) {
3514 /* Calculate the number of rcu_nodes at each level of the tree. */
3515 for (i = 1; i <= MAX_RCU_LVLS; i++)
3516 if (n <= rcu_capacity[i]) {
3517 for (j = 0; j <= i; j++)
3519 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3521 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3526 /* Calculate the total number of rcu_node structures. */
3528 for (i = 0; i <= MAX_RCU_LVLS; i++)
3529 rcu_num_nodes += num_rcu_lvl[i];
3533 void __init rcu_init(void)
3537 rcu_bootup_announce();
3538 rcu_init_geometry();
3539 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3540 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3541 __rcu_init_preempt();
3542 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3545 * We don't need protection against CPU-hotplug here because
3546 * this is called early in boot, before either interrupts
3547 * or the scheduler are operational.
3549 cpu_notifier(rcu_cpu_notify, 0);
3550 pm_notifier(rcu_pm_notify, 0);
3551 for_each_online_cpu(cpu)
3552 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3555 #include "tree_plugin.h"