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_p;
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;
209 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
211 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
212 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
213 .dynticks = ATOMIC_INIT(1),
214 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
215 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
216 .dynticks_idle = ATOMIC_INIT(1),
217 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
221 * Let the RCU core know that this CPU has gone through the scheduler,
222 * which is a quiescent state. This is called when the need for a
223 * quiescent state is urgent, so we burn an atomic operation and full
224 * memory barriers to let the RCU core know about it, regardless of what
225 * this CPU might (or might not) do in the near future.
227 * We inform the RCU core by emulating a zero-duration dyntick-idle
228 * period, which we in turn do by incrementing the ->dynticks counter
231 static void rcu_momentary_dyntick_idle(void)
234 struct rcu_data *rdp;
235 struct rcu_dynticks *rdtp;
237 struct rcu_state *rsp;
239 local_irq_save(flags);
242 * Yes, we can lose flag-setting operations. This is OK, because
243 * the flag will be set again after some delay.
245 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
246 raw_cpu_write(rcu_sched_qs_mask, 0);
248 /* Find the flavor that needs a quiescent state. */
249 for_each_rcu_flavor(rsp) {
250 rdp = raw_cpu_ptr(rsp->rda);
251 if (!(resched_mask & rsp->flavor_mask))
253 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
254 if (ACCESS_ONCE(rdp->mynode->completed) !=
255 ACCESS_ONCE(rdp->cond_resched_completed))
259 * Pretend to be momentarily idle for the quiescent state.
260 * This allows the grace-period kthread to record the
261 * quiescent state, with no need for this CPU to do anything
264 rdtp = this_cpu_ptr(&rcu_dynticks);
265 smp_mb__before_atomic(); /* Earlier stuff before QS. */
266 atomic_add(2, &rdtp->dynticks); /* QS. */
267 smp_mb__after_atomic(); /* Later stuff after QS. */
270 local_irq_restore(flags);
274 * Note a context switch. This is a quiescent state for RCU-sched,
275 * and requires special handling for preemptible RCU.
276 * The caller must have disabled preemption.
278 void rcu_note_context_switch(int cpu)
280 trace_rcu_utilization(TPS("Start context switch"));
282 rcu_preempt_note_context_switch(cpu);
283 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
284 rcu_momentary_dyntick_idle();
285 trace_rcu_utilization(TPS("End context switch"));
287 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
289 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
290 static long qhimark = 10000; /* If this many pending, ignore blimit. */
291 static long qlowmark = 100; /* Once only this many pending, use blimit. */
293 module_param(blimit, long, 0444);
294 module_param(qhimark, long, 0444);
295 module_param(qlowmark, long, 0444);
297 static ulong jiffies_till_first_fqs = ULONG_MAX;
298 static ulong jiffies_till_next_fqs = ULONG_MAX;
300 module_param(jiffies_till_first_fqs, ulong, 0644);
301 module_param(jiffies_till_next_fqs, ulong, 0644);
304 * How long the grace period must be before we start recruiting
305 * quiescent-state help from rcu_note_context_switch().
307 static ulong jiffies_till_sched_qs = HZ / 20;
308 module_param(jiffies_till_sched_qs, ulong, 0644);
310 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
311 struct rcu_data *rdp);
312 static void force_qs_rnp(struct rcu_state *rsp,
313 int (*f)(struct rcu_data *rsp, bool *isidle,
314 unsigned long *maxj),
315 bool *isidle, unsigned long *maxj);
316 static void force_quiescent_state(struct rcu_state *rsp);
317 static int rcu_pending(int cpu);
320 * Return the number of RCU-sched batches processed thus far for debug & stats.
322 long rcu_batches_completed_sched(void)
324 return rcu_sched_state.completed;
326 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
329 * Return the number of RCU BH batches processed thus far for debug & stats.
331 long rcu_batches_completed_bh(void)
333 return rcu_bh_state.completed;
335 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
338 * Force a quiescent state.
340 void rcu_force_quiescent_state(void)
342 force_quiescent_state(rcu_state_p);
344 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
347 * Force a quiescent state for RCU BH.
349 void rcu_bh_force_quiescent_state(void)
351 force_quiescent_state(&rcu_bh_state);
353 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
356 * Show the state of the grace-period kthreads.
358 void show_rcu_gp_kthreads(void)
360 struct rcu_state *rsp;
362 for_each_rcu_flavor(rsp) {
363 pr_info("%s: wait state: %d ->state: %#lx\n",
364 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
365 /* sched_show_task(rsp->gp_kthread); */
368 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
371 * Record the number of times rcutorture tests have been initiated and
372 * terminated. This information allows the debugfs tracing stats to be
373 * correlated to the rcutorture messages, even when the rcutorture module
374 * is being repeatedly loaded and unloaded. In other words, we cannot
375 * store this state in rcutorture itself.
377 void rcutorture_record_test_transition(void)
379 rcutorture_testseq++;
380 rcutorture_vernum = 0;
382 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
385 * Send along grace-period-related data for rcutorture diagnostics.
387 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
388 unsigned long *gpnum, unsigned long *completed)
390 struct rcu_state *rsp = NULL;
399 case RCU_SCHED_FLAVOR:
400 rsp = &rcu_sched_state;
406 *flags = ACCESS_ONCE(rsp->gp_flags);
407 *gpnum = ACCESS_ONCE(rsp->gpnum);
408 *completed = ACCESS_ONCE(rsp->completed);
415 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
418 * Record the number of writer passes through the current rcutorture test.
419 * This is also used to correlate debugfs tracing stats with the rcutorture
422 void rcutorture_record_progress(unsigned long vernum)
426 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
429 * Force a quiescent state for RCU-sched.
431 void rcu_sched_force_quiescent_state(void)
433 force_quiescent_state(&rcu_sched_state);
435 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
438 * Does the CPU have callbacks ready to be invoked?
441 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
443 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
444 rdp->nxttail[RCU_DONE_TAIL] != NULL;
448 * Return the root node of the specified rcu_state structure.
450 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
452 return &rsp->node[0];
456 * Is there any need for future grace periods?
457 * Interrupts must be disabled. If the caller does not hold the root
458 * rnp_node structure's ->lock, the results are advisory only.
460 static int rcu_future_needs_gp(struct rcu_state *rsp)
462 struct rcu_node *rnp = rcu_get_root(rsp);
463 int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1;
464 int *fp = &rnp->need_future_gp[idx];
466 return ACCESS_ONCE(*fp);
470 * Does the current CPU require a not-yet-started grace period?
471 * The caller must have disabled interrupts to prevent races with
472 * normal callback registry.
475 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
479 if (rcu_gp_in_progress(rsp))
480 return 0; /* No, a grace period is already in progress. */
481 if (rcu_future_needs_gp(rsp))
482 return 1; /* Yes, a no-CBs CPU needs one. */
483 if (!rdp->nxttail[RCU_NEXT_TAIL])
484 return 0; /* No, this is a no-CBs (or offline) CPU. */
485 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
486 return 1; /* Yes, this CPU has newly registered callbacks. */
487 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
488 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
489 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
490 rdp->nxtcompleted[i]))
491 return 1; /* Yes, CBs for future grace period. */
492 return 0; /* No grace period needed. */
496 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
498 * If the new value of the ->dynticks_nesting counter now is zero,
499 * we really have entered idle, and must do the appropriate accounting.
500 * The caller must have disabled interrupts.
502 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
505 struct rcu_state *rsp;
506 struct rcu_data *rdp;
508 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
509 if (!user && !is_idle_task(current)) {
510 struct task_struct *idle __maybe_unused =
511 idle_task(smp_processor_id());
513 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
514 ftrace_dump(DUMP_ORIG);
515 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
516 current->pid, current->comm,
517 idle->pid, idle->comm); /* must be idle task! */
519 for_each_rcu_flavor(rsp) {
520 rdp = this_cpu_ptr(rsp->rda);
521 do_nocb_deferred_wakeup(rdp);
523 rcu_prepare_for_idle(smp_processor_id());
524 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
525 smp_mb__before_atomic(); /* See above. */
526 atomic_inc(&rdtp->dynticks);
527 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
528 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
531 * It is illegal to enter an extended quiescent state while
532 * in an RCU read-side critical section.
534 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
535 "Illegal idle entry in RCU read-side critical section.");
536 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
537 "Illegal idle entry in RCU-bh read-side critical section.");
538 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
539 "Illegal idle entry in RCU-sched read-side critical section.");
543 * Enter an RCU extended quiescent state, which can be either the
544 * idle loop or adaptive-tickless usermode execution.
546 static void rcu_eqs_enter(bool user)
549 struct rcu_dynticks *rdtp;
551 rdtp = this_cpu_ptr(&rcu_dynticks);
552 oldval = rdtp->dynticks_nesting;
553 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
554 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
555 rdtp->dynticks_nesting = 0;
556 rcu_eqs_enter_common(rdtp, oldval, user);
558 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
563 * rcu_idle_enter - inform RCU that current CPU is entering idle
565 * Enter idle mode, in other words, -leave- the mode in which RCU
566 * read-side critical sections can occur. (Though RCU read-side
567 * critical sections can occur in irq handlers in idle, a possibility
568 * handled by irq_enter() and irq_exit().)
570 * We crowbar the ->dynticks_nesting field to zero to allow for
571 * the possibility of usermode upcalls having messed up our count
572 * of interrupt nesting level during the prior busy period.
574 void rcu_idle_enter(void)
578 local_irq_save(flags);
579 rcu_eqs_enter(false);
580 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
581 local_irq_restore(flags);
583 EXPORT_SYMBOL_GPL(rcu_idle_enter);
585 #ifdef CONFIG_RCU_USER_QS
587 * rcu_user_enter - inform RCU that we are resuming userspace.
589 * Enter RCU idle mode right before resuming userspace. No use of RCU
590 * is permitted between this call and rcu_user_exit(). This way the
591 * CPU doesn't need to maintain the tick for RCU maintenance purposes
592 * when the CPU runs in userspace.
594 void rcu_user_enter(void)
598 #endif /* CONFIG_RCU_USER_QS */
601 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
603 * Exit from an interrupt handler, which might possibly result in entering
604 * idle mode, in other words, leaving the mode in which read-side critical
605 * sections can occur.
607 * This code assumes that the idle loop never does anything that might
608 * result in unbalanced calls to irq_enter() and irq_exit(). If your
609 * architecture violates this assumption, RCU will give you what you
610 * deserve, good and hard. But very infrequently and irreproducibly.
612 * Use things like work queues to work around this limitation.
614 * You have been warned.
616 void rcu_irq_exit(void)
620 struct rcu_dynticks *rdtp;
622 local_irq_save(flags);
623 rdtp = this_cpu_ptr(&rcu_dynticks);
624 oldval = rdtp->dynticks_nesting;
625 rdtp->dynticks_nesting--;
626 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
627 if (rdtp->dynticks_nesting)
628 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
630 rcu_eqs_enter_common(rdtp, oldval, true);
631 rcu_sysidle_enter(rdtp, 1);
632 local_irq_restore(flags);
636 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
638 * If the new value of the ->dynticks_nesting counter was previously zero,
639 * we really have exited idle, and must do the appropriate accounting.
640 * The caller must have disabled interrupts.
642 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
645 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
646 atomic_inc(&rdtp->dynticks);
647 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
648 smp_mb__after_atomic(); /* See above. */
649 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
650 rcu_cleanup_after_idle(smp_processor_id());
651 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
652 if (!user && !is_idle_task(current)) {
653 struct task_struct *idle __maybe_unused =
654 idle_task(smp_processor_id());
656 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
657 oldval, rdtp->dynticks_nesting);
658 ftrace_dump(DUMP_ORIG);
659 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
660 current->pid, current->comm,
661 idle->pid, idle->comm); /* must be idle task! */
666 * Exit an RCU extended quiescent state, which can be either the
667 * idle loop or adaptive-tickless usermode execution.
669 static void rcu_eqs_exit(bool user)
671 struct rcu_dynticks *rdtp;
674 rdtp = this_cpu_ptr(&rcu_dynticks);
675 oldval = rdtp->dynticks_nesting;
676 WARN_ON_ONCE(oldval < 0);
677 if (oldval & DYNTICK_TASK_NEST_MASK) {
678 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
680 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
681 rcu_eqs_exit_common(rdtp, oldval, user);
686 * rcu_idle_exit - inform RCU that current CPU is leaving idle
688 * Exit idle mode, in other words, -enter- the mode in which RCU
689 * read-side critical sections can occur.
691 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
692 * allow for the possibility of usermode upcalls messing up our count
693 * of interrupt nesting level during the busy period that is just
696 void rcu_idle_exit(void)
700 local_irq_save(flags);
702 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
703 local_irq_restore(flags);
705 EXPORT_SYMBOL_GPL(rcu_idle_exit);
707 #ifdef CONFIG_RCU_USER_QS
709 * rcu_user_exit - inform RCU that we are exiting userspace.
711 * Exit RCU idle mode while entering the kernel because it can
712 * run a RCU read side critical section anytime.
714 void rcu_user_exit(void)
718 #endif /* CONFIG_RCU_USER_QS */
721 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
723 * Enter an interrupt handler, which might possibly result in exiting
724 * idle mode, in other words, entering the mode in which read-side critical
725 * sections can occur.
727 * Note that the Linux kernel is fully capable of entering an interrupt
728 * handler that it never exits, for example when doing upcalls to
729 * user mode! This code assumes that the idle loop never does upcalls to
730 * user mode. If your architecture does do upcalls from the idle loop (or
731 * does anything else that results in unbalanced calls to the irq_enter()
732 * and irq_exit() functions), RCU will give you what you deserve, good
733 * and hard. But very infrequently and irreproducibly.
735 * Use things like work queues to work around this limitation.
737 * You have been warned.
739 void rcu_irq_enter(void)
742 struct rcu_dynticks *rdtp;
745 local_irq_save(flags);
746 rdtp = this_cpu_ptr(&rcu_dynticks);
747 oldval = rdtp->dynticks_nesting;
748 rdtp->dynticks_nesting++;
749 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
751 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
753 rcu_eqs_exit_common(rdtp, oldval, true);
754 rcu_sysidle_exit(rdtp, 1);
755 local_irq_restore(flags);
759 * rcu_nmi_enter - inform RCU of entry to NMI context
761 * If the CPU was idle with dynamic ticks active, and there is no
762 * irq handler running, this updates rdtp->dynticks_nmi to let the
763 * RCU grace-period handling know that the CPU is active.
765 void rcu_nmi_enter(void)
767 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
769 if (rdtp->dynticks_nmi_nesting == 0 &&
770 (atomic_read(&rdtp->dynticks) & 0x1))
772 rdtp->dynticks_nmi_nesting++;
773 smp_mb__before_atomic(); /* Force delay from prior write. */
774 atomic_inc(&rdtp->dynticks);
775 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
776 smp_mb__after_atomic(); /* See above. */
777 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
781 * rcu_nmi_exit - inform RCU of exit from NMI context
783 * If the CPU was idle with dynamic ticks active, and there is no
784 * irq handler running, this updates rdtp->dynticks_nmi to let the
785 * RCU grace-period handling know that the CPU is no longer active.
787 void rcu_nmi_exit(void)
789 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
791 if (rdtp->dynticks_nmi_nesting == 0 ||
792 --rdtp->dynticks_nmi_nesting != 0)
794 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
795 smp_mb__before_atomic(); /* See above. */
796 atomic_inc(&rdtp->dynticks);
797 smp_mb__after_atomic(); /* Force delay to next write. */
798 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
802 * __rcu_is_watching - are RCU read-side critical sections safe?
804 * Return true if RCU is watching the running CPU, which means that
805 * this CPU can safely enter RCU read-side critical sections. Unlike
806 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
807 * least disabled preemption.
809 bool notrace __rcu_is_watching(void)
811 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
815 * rcu_is_watching - see if RCU thinks that the current CPU is idle
817 * If the current CPU is in its idle loop and is neither in an interrupt
818 * or NMI handler, return true.
820 bool notrace rcu_is_watching(void)
825 ret = __rcu_is_watching();
829 EXPORT_SYMBOL_GPL(rcu_is_watching);
831 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
834 * Is the current CPU online? Disable preemption to avoid false positives
835 * that could otherwise happen due to the current CPU number being sampled,
836 * this task being preempted, its old CPU being taken offline, resuming
837 * on some other CPU, then determining that its old CPU is now offline.
838 * It is OK to use RCU on an offline processor during initial boot, hence
839 * the check for rcu_scheduler_fully_active. Note also that it is OK
840 * for a CPU coming online to use RCU for one jiffy prior to marking itself
841 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
842 * offline to continue to use RCU for one jiffy after marking itself
843 * offline in the cpu_online_mask. This leniency is necessary given the
844 * non-atomic nature of the online and offline processing, for example,
845 * the fact that a CPU enters the scheduler after completing the CPU_DYING
848 * This is also why RCU internally marks CPUs online during the
849 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
851 * Disable checking if in an NMI handler because we cannot safely report
852 * errors from NMI handlers anyway.
854 bool rcu_lockdep_current_cpu_online(void)
856 struct rcu_data *rdp;
857 struct rcu_node *rnp;
863 rdp = this_cpu_ptr(&rcu_sched_data);
865 ret = (rdp->grpmask & rnp->qsmaskinit) ||
866 !rcu_scheduler_fully_active;
870 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
872 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
875 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
877 * If the current CPU is idle or running at a first-level (not nested)
878 * interrupt from idle, return true. The caller must have at least
879 * disabled preemption.
881 static int rcu_is_cpu_rrupt_from_idle(void)
883 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
887 * Snapshot the specified CPU's dynticks counter so that we can later
888 * credit them with an implicit quiescent state. Return 1 if this CPU
889 * is in dynticks idle mode, which is an extended quiescent state.
891 static int dyntick_save_progress_counter(struct rcu_data *rdp,
892 bool *isidle, unsigned long *maxj)
894 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
895 rcu_sysidle_check_cpu(rdp, isidle, maxj);
896 if ((rdp->dynticks_snap & 0x1) == 0) {
897 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
905 * This function really isn't for public consumption, but RCU is special in
906 * that context switches can allow the state machine to make progress.
908 extern void resched_cpu(int cpu);
911 * Return true if the specified CPU has passed through a quiescent
912 * state by virtue of being in or having passed through an dynticks
913 * idle state since the last call to dyntick_save_progress_counter()
914 * for this same CPU, or by virtue of having been offline.
916 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
917 bool *isidle, unsigned long *maxj)
923 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
924 snap = (unsigned int)rdp->dynticks_snap;
927 * If the CPU passed through or entered a dynticks idle phase with
928 * no active irq/NMI handlers, then we can safely pretend that the CPU
929 * already acknowledged the request to pass through a quiescent
930 * state. Either way, that CPU cannot possibly be in an RCU
931 * read-side critical section that started before the beginning
932 * of the current RCU grace period.
934 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
935 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
941 * Check for the CPU being offline, but only if the grace period
942 * is old enough. We don't need to worry about the CPU changing
943 * state: If we see it offline even once, it has been through a
946 * The reason for insisting that the grace period be at least
947 * one jiffy old is that CPUs that are not quite online and that
948 * have just gone offline can still execute RCU read-side critical
951 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
952 return 0; /* Grace period is not old enough. */
954 if (cpu_is_offline(rdp->cpu)) {
955 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
961 * A CPU running for an extended time within the kernel can
962 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
963 * even context-switching back and forth between a pair of
964 * in-kernel CPU-bound tasks cannot advance grace periods.
965 * So if the grace period is old enough, make the CPU pay attention.
966 * Note that the unsynchronized assignments to the per-CPU
967 * rcu_sched_qs_mask variable are safe. Yes, setting of
968 * bits can be lost, but they will be set again on the next
969 * force-quiescent-state pass. So lost bit sets do not result
970 * in incorrect behavior, merely in a grace period lasting
971 * a few jiffies longer than it might otherwise. Because
972 * there are at most four threads involved, and because the
973 * updates are only once every few jiffies, the probability of
974 * lossage (and thus of slight grace-period extension) is
977 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
978 * is set too high, we override with half of the RCU CPU stall
981 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
982 if (ULONG_CMP_GE(jiffies,
983 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
984 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
985 if (!(ACCESS_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
986 ACCESS_ONCE(rdp->cond_resched_completed) =
987 ACCESS_ONCE(rdp->mynode->completed);
988 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
989 ACCESS_ONCE(*rcrmp) =
990 ACCESS_ONCE(*rcrmp) + rdp->rsp->flavor_mask;
991 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
992 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
993 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
994 /* Time to beat on that CPU again! */
995 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
996 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1003 static void record_gp_stall_check_time(struct rcu_state *rsp)
1005 unsigned long j = jiffies;
1009 smp_wmb(); /* Record start time before stall time. */
1010 j1 = rcu_jiffies_till_stall_check();
1011 ACCESS_ONCE(rsp->jiffies_stall) = j + j1;
1012 rsp->jiffies_resched = j + j1 / 2;
1016 * Dump stacks of all tasks running on stalled CPUs.
1018 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1021 unsigned long flags;
1022 struct rcu_node *rnp;
1024 rcu_for_each_leaf_node(rsp, rnp) {
1025 raw_spin_lock_irqsave(&rnp->lock, flags);
1026 if (rnp->qsmask != 0) {
1027 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1028 if (rnp->qsmask & (1UL << cpu))
1029 dump_cpu_task(rnp->grplo + cpu);
1031 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1035 static void print_other_cpu_stall(struct rcu_state *rsp)
1039 unsigned long flags;
1041 struct rcu_node *rnp = rcu_get_root(rsp);
1044 /* Only let one CPU complain about others per time interval. */
1046 raw_spin_lock_irqsave(&rnp->lock, flags);
1047 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
1048 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1049 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1052 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1053 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1056 * OK, time to rat on our buddy...
1057 * See Documentation/RCU/stallwarn.txt for info on how to debug
1058 * RCU CPU stall warnings.
1060 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1062 print_cpu_stall_info_begin();
1063 rcu_for_each_leaf_node(rsp, rnp) {
1064 raw_spin_lock_irqsave(&rnp->lock, flags);
1065 ndetected += rcu_print_task_stall(rnp);
1066 if (rnp->qsmask != 0) {
1067 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1068 if (rnp->qsmask & (1UL << cpu)) {
1069 print_cpu_stall_info(rsp,
1074 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1078 * Now rat on any tasks that got kicked up to the root rcu_node
1079 * due to CPU offlining.
1081 rnp = rcu_get_root(rsp);
1082 raw_spin_lock_irqsave(&rnp->lock, flags);
1083 ndetected += rcu_print_task_stall(rnp);
1084 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1086 print_cpu_stall_info_end();
1087 for_each_possible_cpu(cpu)
1088 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1089 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1090 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1091 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1093 pr_err("INFO: Stall ended before state dump start\n");
1095 rcu_dump_cpu_stacks(rsp);
1097 /* Complain about tasks blocking the grace period. */
1099 rcu_print_detail_task_stall(rsp);
1101 force_quiescent_state(rsp); /* Kick them all. */
1104 static void print_cpu_stall(struct rcu_state *rsp)
1107 unsigned long flags;
1108 struct rcu_node *rnp = rcu_get_root(rsp);
1112 * OK, time to rat on ourselves...
1113 * See Documentation/RCU/stallwarn.txt for info on how to debug
1114 * RCU CPU stall warnings.
1116 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1117 print_cpu_stall_info_begin();
1118 print_cpu_stall_info(rsp, smp_processor_id());
1119 print_cpu_stall_info_end();
1120 for_each_possible_cpu(cpu)
1121 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1122 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1123 jiffies - rsp->gp_start,
1124 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1125 rcu_dump_cpu_stacks(rsp);
1127 raw_spin_lock_irqsave(&rnp->lock, flags);
1128 if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall)))
1129 ACCESS_ONCE(rsp->jiffies_stall) = jiffies +
1130 3 * rcu_jiffies_till_stall_check() + 3;
1131 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1134 * Attempt to revive the RCU machinery by forcing a context switch.
1136 * A context switch would normally allow the RCU state machine to make
1137 * progress and it could be we're stuck in kernel space without context
1138 * switches for an entirely unreasonable amount of time.
1140 resched_cpu(smp_processor_id());
1143 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1145 unsigned long completed;
1146 unsigned long gpnum;
1150 struct rcu_node *rnp;
1152 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1157 * Lots of memory barriers to reject false positives.
1159 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1160 * then rsp->gp_start, and finally rsp->completed. These values
1161 * are updated in the opposite order with memory barriers (or
1162 * equivalent) during grace-period initialization and cleanup.
1163 * Now, a false positive can occur if we get an new value of
1164 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1165 * the memory barriers, the only way that this can happen is if one
1166 * grace period ends and another starts between these two fetches.
1167 * Detect this by comparing rsp->completed with the previous fetch
1170 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1171 * and rsp->gp_start suffice to forestall false positives.
1173 gpnum = ACCESS_ONCE(rsp->gpnum);
1174 smp_rmb(); /* Pick up ->gpnum first... */
1175 js = ACCESS_ONCE(rsp->jiffies_stall);
1176 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1177 gps = ACCESS_ONCE(rsp->gp_start);
1178 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1179 completed = ACCESS_ONCE(rsp->completed);
1180 if (ULONG_CMP_GE(completed, gpnum) ||
1181 ULONG_CMP_LT(j, js) ||
1182 ULONG_CMP_GE(gps, js))
1183 return; /* No stall or GP completed since entering function. */
1185 if (rcu_gp_in_progress(rsp) &&
1186 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1188 /* We haven't checked in, so go dump stack. */
1189 print_cpu_stall(rsp);
1191 } else if (rcu_gp_in_progress(rsp) &&
1192 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1194 /* They had a few time units to dump stack, so complain. */
1195 print_other_cpu_stall(rsp);
1200 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1202 * Set the stall-warning timeout way off into the future, thus preventing
1203 * any RCU CPU stall-warning messages from appearing in the current set of
1204 * RCU grace periods.
1206 * The caller must disable hard irqs.
1208 void rcu_cpu_stall_reset(void)
1210 struct rcu_state *rsp;
1212 for_each_rcu_flavor(rsp)
1213 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2;
1217 * Initialize the specified rcu_data structure's callback list to empty.
1219 static void init_callback_list(struct rcu_data *rdp)
1223 if (init_nocb_callback_list(rdp))
1225 rdp->nxtlist = NULL;
1226 for (i = 0; i < RCU_NEXT_SIZE; i++)
1227 rdp->nxttail[i] = &rdp->nxtlist;
1231 * Determine the value that ->completed will have at the end of the
1232 * next subsequent grace period. This is used to tag callbacks so that
1233 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1234 * been dyntick-idle for an extended period with callbacks under the
1235 * influence of RCU_FAST_NO_HZ.
1237 * The caller must hold rnp->lock with interrupts disabled.
1239 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1240 struct rcu_node *rnp)
1243 * If RCU is idle, we just wait for the next grace period.
1244 * But we can only be sure that RCU is idle if we are looking
1245 * at the root rcu_node structure -- otherwise, a new grace
1246 * period might have started, but just not yet gotten around
1247 * to initializing the current non-root rcu_node structure.
1249 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1250 return rnp->completed + 1;
1253 * Otherwise, wait for a possible partial grace period and
1254 * then the subsequent full grace period.
1256 return rnp->completed + 2;
1260 * Trace-event helper function for rcu_start_future_gp() and
1261 * rcu_nocb_wait_gp().
1263 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1264 unsigned long c, const char *s)
1266 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1267 rnp->completed, c, rnp->level,
1268 rnp->grplo, rnp->grphi, s);
1272 * Start some future grace period, as needed to handle newly arrived
1273 * callbacks. The required future grace periods are recorded in each
1274 * rcu_node structure's ->need_future_gp field. Returns true if there
1275 * is reason to awaken the grace-period kthread.
1277 * The caller must hold the specified rcu_node structure's ->lock.
1279 static bool __maybe_unused
1280 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1281 unsigned long *c_out)
1286 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1289 * Pick up grace-period number for new callbacks. If this
1290 * grace period is already marked as needed, return to the caller.
1292 c = rcu_cbs_completed(rdp->rsp, rnp);
1293 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1294 if (rnp->need_future_gp[c & 0x1]) {
1295 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1300 * If either this rcu_node structure or the root rcu_node structure
1301 * believe that a grace period is in progress, then we must wait
1302 * for the one following, which is in "c". Because our request
1303 * will be noticed at the end of the current grace period, we don't
1304 * need to explicitly start one. We only do the lockless check
1305 * of rnp_root's fields if the current rcu_node structure thinks
1306 * there is no grace period in flight, and because we hold rnp->lock,
1307 * the only possible change is when rnp_root's two fields are
1308 * equal, in which case rnp_root->gpnum might be concurrently
1309 * incremented. But that is OK, as it will just result in our
1310 * doing some extra useless work.
1312 if (rnp->gpnum != rnp->completed ||
1313 ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) {
1314 rnp->need_future_gp[c & 0x1]++;
1315 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1320 * There might be no grace period in progress. If we don't already
1321 * hold it, acquire the root rcu_node structure's lock in order to
1322 * start one (if needed).
1324 if (rnp != rnp_root) {
1325 raw_spin_lock(&rnp_root->lock);
1326 smp_mb__after_unlock_lock();
1330 * Get a new grace-period number. If there really is no grace
1331 * period in progress, it will be smaller than the one we obtained
1332 * earlier. Adjust callbacks as needed. Note that even no-CBs
1333 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1335 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1336 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1337 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1338 rdp->nxtcompleted[i] = c;
1341 * If the needed for the required grace period is already
1342 * recorded, trace and leave.
1344 if (rnp_root->need_future_gp[c & 0x1]) {
1345 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1349 /* Record the need for the future grace period. */
1350 rnp_root->need_future_gp[c & 0x1]++;
1352 /* If a grace period is not already in progress, start one. */
1353 if (rnp_root->gpnum != rnp_root->completed) {
1354 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1356 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1357 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1360 if (rnp != rnp_root)
1361 raw_spin_unlock(&rnp_root->lock);
1369 * Clean up any old requests for the just-ended grace period. Also return
1370 * whether any additional grace periods have been requested. Also invoke
1371 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1372 * waiting for this grace period to complete.
1374 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1376 int c = rnp->completed;
1378 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1380 rcu_nocb_gp_cleanup(rsp, rnp);
1381 rnp->need_future_gp[c & 0x1] = 0;
1382 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1383 trace_rcu_future_gp(rnp, rdp, c,
1384 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1389 * Awaken the grace-period kthread for the specified flavor of RCU.
1390 * Don't do a self-awaken, and don't bother awakening when there is
1391 * nothing for the grace-period kthread to do (as in several CPUs
1392 * raced to awaken, and we lost), and finally don't try to awaken
1393 * a kthread that has not yet been created.
1395 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1397 if (current == rsp->gp_kthread ||
1398 !ACCESS_ONCE(rsp->gp_flags) ||
1401 wake_up(&rsp->gp_wq);
1405 * If there is room, assign a ->completed number to any callbacks on
1406 * this CPU that have not already been assigned. Also accelerate any
1407 * callbacks that were previously assigned a ->completed number that has
1408 * since proven to be too conservative, which can happen if callbacks get
1409 * assigned a ->completed number while RCU is idle, but with reference to
1410 * a non-root rcu_node structure. This function is idempotent, so it does
1411 * not hurt to call it repeatedly. Returns an flag saying that we should
1412 * awaken the RCU grace-period kthread.
1414 * The caller must hold rnp->lock with interrupts disabled.
1416 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1417 struct rcu_data *rdp)
1423 /* If the CPU has no callbacks, nothing to do. */
1424 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1428 * Starting from the sublist containing the callbacks most
1429 * recently assigned a ->completed number and working down, find the
1430 * first sublist that is not assignable to an upcoming grace period.
1431 * Such a sublist has something in it (first two tests) and has
1432 * a ->completed number assigned that will complete sooner than
1433 * the ->completed number for newly arrived callbacks (last test).
1435 * The key point is that any later sublist can be assigned the
1436 * same ->completed number as the newly arrived callbacks, which
1437 * means that the callbacks in any of these later sublist can be
1438 * grouped into a single sublist, whether or not they have already
1439 * been assigned a ->completed number.
1441 c = rcu_cbs_completed(rsp, rnp);
1442 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1443 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1444 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1448 * If there are no sublist for unassigned callbacks, leave.
1449 * At the same time, advance "i" one sublist, so that "i" will
1450 * index into the sublist where all the remaining callbacks should
1453 if (++i >= RCU_NEXT_TAIL)
1457 * Assign all subsequent callbacks' ->completed number to the next
1458 * full grace period and group them all in the sublist initially
1461 for (; i <= RCU_NEXT_TAIL; i++) {
1462 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1463 rdp->nxtcompleted[i] = c;
1465 /* Record any needed additional grace periods. */
1466 ret = rcu_start_future_gp(rnp, rdp, NULL);
1468 /* Trace depending on how much we were able to accelerate. */
1469 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1470 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1472 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1477 * Move any callbacks whose grace period has completed to the
1478 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1479 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1480 * sublist. This function is idempotent, so it does not hurt to
1481 * invoke it repeatedly. As long as it is not invoked -too- often...
1482 * Returns true if the RCU grace-period kthread needs to be awakened.
1484 * The caller must hold rnp->lock with interrupts disabled.
1486 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1487 struct rcu_data *rdp)
1491 /* If the CPU has no callbacks, nothing to do. */
1492 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1496 * Find all callbacks whose ->completed numbers indicate that they
1497 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1499 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1500 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1502 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1504 /* Clean up any sublist tail pointers that were misordered above. */
1505 for (j = RCU_WAIT_TAIL; j < i; j++)
1506 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1508 /* Copy down callbacks to fill in empty sublists. */
1509 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1510 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1512 rdp->nxttail[j] = rdp->nxttail[i];
1513 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1516 /* Classify any remaining callbacks. */
1517 return rcu_accelerate_cbs(rsp, rnp, rdp);
1521 * Update CPU-local rcu_data state to record the beginnings and ends of
1522 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1523 * structure corresponding to the current CPU, and must have irqs disabled.
1524 * Returns true if the grace-period kthread needs to be awakened.
1526 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1527 struct rcu_data *rdp)
1531 /* Handle the ends of any preceding grace periods first. */
1532 if (rdp->completed == rnp->completed) {
1534 /* No grace period end, so just accelerate recent callbacks. */
1535 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1539 /* Advance callbacks. */
1540 ret = rcu_advance_cbs(rsp, rnp, rdp);
1542 /* Remember that we saw this grace-period completion. */
1543 rdp->completed = rnp->completed;
1544 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1547 if (rdp->gpnum != rnp->gpnum) {
1549 * If the current grace period is waiting for this CPU,
1550 * set up to detect a quiescent state, otherwise don't
1551 * go looking for one.
1553 rdp->gpnum = rnp->gpnum;
1554 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1555 rdp->passed_quiesce = 0;
1556 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1557 zero_cpu_stall_ticks(rdp);
1562 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1564 unsigned long flags;
1566 struct rcu_node *rnp;
1568 local_irq_save(flags);
1570 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1571 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1572 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1573 local_irq_restore(flags);
1576 smp_mb__after_unlock_lock();
1577 needwake = __note_gp_changes(rsp, rnp, rdp);
1578 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1580 rcu_gp_kthread_wake(rsp);
1584 * Initialize a new grace period. Return 0 if no grace period required.
1586 static int rcu_gp_init(struct rcu_state *rsp)
1588 struct rcu_data *rdp;
1589 struct rcu_node *rnp = rcu_get_root(rsp);
1591 rcu_bind_gp_kthread();
1592 raw_spin_lock_irq(&rnp->lock);
1593 smp_mb__after_unlock_lock();
1594 if (!ACCESS_ONCE(rsp->gp_flags)) {
1595 /* Spurious wakeup, tell caller to go back to sleep. */
1596 raw_spin_unlock_irq(&rnp->lock);
1599 ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */
1601 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1603 * Grace period already in progress, don't start another.
1604 * Not supposed to be able to happen.
1606 raw_spin_unlock_irq(&rnp->lock);
1610 /* Advance to a new grace period and initialize state. */
1611 record_gp_stall_check_time(rsp);
1612 /* Record GP times before starting GP, hence smp_store_release(). */
1613 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1614 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1615 raw_spin_unlock_irq(&rnp->lock);
1617 /* Exclude any concurrent CPU-hotplug operations. */
1618 mutex_lock(&rsp->onoff_mutex);
1619 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1622 * Set the quiescent-state-needed bits in all the rcu_node
1623 * structures for all currently online CPUs in breadth-first order,
1624 * starting from the root rcu_node structure, relying on the layout
1625 * of the tree within the rsp->node[] array. Note that other CPUs
1626 * will access only the leaves of the hierarchy, thus seeing that no
1627 * grace period is in progress, at least until the corresponding
1628 * leaf node has been initialized. In addition, we have excluded
1629 * CPU-hotplug operations.
1631 * The grace period cannot complete until the initialization
1632 * process finishes, because this kthread handles both.
1634 rcu_for_each_node_breadth_first(rsp, rnp) {
1635 raw_spin_lock_irq(&rnp->lock);
1636 smp_mb__after_unlock_lock();
1637 rdp = this_cpu_ptr(rsp->rda);
1638 rcu_preempt_check_blocked_tasks(rnp);
1639 rnp->qsmask = rnp->qsmaskinit;
1640 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1641 WARN_ON_ONCE(rnp->completed != rsp->completed);
1642 ACCESS_ONCE(rnp->completed) = rsp->completed;
1643 if (rnp == rdp->mynode)
1644 (void)__note_gp_changes(rsp, rnp, rdp);
1645 rcu_preempt_boost_start_gp(rnp);
1646 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1647 rnp->level, rnp->grplo,
1648 rnp->grphi, rnp->qsmask);
1649 raw_spin_unlock_irq(&rnp->lock);
1650 cond_resched_rcu_qs();
1653 mutex_unlock(&rsp->onoff_mutex);
1658 * Do one round of quiescent-state forcing.
1660 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1662 int fqs_state = fqs_state_in;
1663 bool isidle = false;
1665 struct rcu_node *rnp = rcu_get_root(rsp);
1668 if (fqs_state == RCU_SAVE_DYNTICK) {
1669 /* Collect dyntick-idle snapshots. */
1670 if (is_sysidle_rcu_state(rsp)) {
1672 maxj = jiffies - ULONG_MAX / 4;
1674 force_qs_rnp(rsp, dyntick_save_progress_counter,
1676 rcu_sysidle_report_gp(rsp, isidle, maxj);
1677 fqs_state = RCU_FORCE_QS;
1679 /* Handle dyntick-idle and offline CPUs. */
1681 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1683 /* Clear flag to prevent immediate re-entry. */
1684 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1685 raw_spin_lock_irq(&rnp->lock);
1686 smp_mb__after_unlock_lock();
1687 ACCESS_ONCE(rsp->gp_flags) &= ~RCU_GP_FLAG_FQS;
1688 raw_spin_unlock_irq(&rnp->lock);
1694 * Clean up after the old grace period.
1696 static void rcu_gp_cleanup(struct rcu_state *rsp)
1698 unsigned long gp_duration;
1699 bool needgp = false;
1701 struct rcu_data *rdp;
1702 struct rcu_node *rnp = rcu_get_root(rsp);
1704 raw_spin_lock_irq(&rnp->lock);
1705 smp_mb__after_unlock_lock();
1706 gp_duration = jiffies - rsp->gp_start;
1707 if (gp_duration > rsp->gp_max)
1708 rsp->gp_max = gp_duration;
1711 * We know the grace period is complete, but to everyone else
1712 * it appears to still be ongoing. But it is also the case
1713 * that to everyone else it looks like there is nothing that
1714 * they can do to advance the grace period. It is therefore
1715 * safe for us to drop the lock in order to mark the grace
1716 * period as completed in all of the rcu_node structures.
1718 raw_spin_unlock_irq(&rnp->lock);
1721 * Propagate new ->completed value to rcu_node structures so
1722 * that other CPUs don't have to wait until the start of the next
1723 * grace period to process their callbacks. This also avoids
1724 * some nasty RCU grace-period initialization races by forcing
1725 * the end of the current grace period to be completely recorded in
1726 * all of the rcu_node structures before the beginning of the next
1727 * grace period is recorded in any of the rcu_node structures.
1729 rcu_for_each_node_breadth_first(rsp, rnp) {
1730 raw_spin_lock_irq(&rnp->lock);
1731 smp_mb__after_unlock_lock();
1732 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1733 rdp = this_cpu_ptr(rsp->rda);
1734 if (rnp == rdp->mynode)
1735 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
1736 /* smp_mb() provided by prior unlock-lock pair. */
1737 nocb += rcu_future_gp_cleanup(rsp, rnp);
1738 raw_spin_unlock_irq(&rnp->lock);
1739 cond_resched_rcu_qs();
1741 rnp = rcu_get_root(rsp);
1742 raw_spin_lock_irq(&rnp->lock);
1743 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1744 rcu_nocb_gp_set(rnp, nocb);
1746 /* Declare grace period done. */
1747 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1748 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1749 rsp->fqs_state = RCU_GP_IDLE;
1750 rdp = this_cpu_ptr(rsp->rda);
1751 /* Advance CBs to reduce false positives below. */
1752 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
1753 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
1754 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1755 trace_rcu_grace_period(rsp->name,
1756 ACCESS_ONCE(rsp->gpnum),
1759 raw_spin_unlock_irq(&rnp->lock);
1763 * Body of kthread that handles grace periods.
1765 static int __noreturn rcu_gp_kthread(void *arg)
1771 struct rcu_state *rsp = arg;
1772 struct rcu_node *rnp = rcu_get_root(rsp);
1776 /* Handle grace-period start. */
1778 trace_rcu_grace_period(rsp->name,
1779 ACCESS_ONCE(rsp->gpnum),
1781 rsp->gp_state = RCU_GP_WAIT_GPS;
1782 wait_event_interruptible(rsp->gp_wq,
1783 ACCESS_ONCE(rsp->gp_flags) &
1785 /* Locking provides needed memory barrier. */
1786 if (rcu_gp_init(rsp))
1788 cond_resched_rcu_qs();
1789 flush_signals(current);
1790 trace_rcu_grace_period(rsp->name,
1791 ACCESS_ONCE(rsp->gpnum),
1795 /* Handle quiescent-state forcing. */
1796 fqs_state = RCU_SAVE_DYNTICK;
1797 j = jiffies_till_first_fqs;
1800 jiffies_till_first_fqs = HZ;
1805 rsp->jiffies_force_qs = jiffies + j;
1806 trace_rcu_grace_period(rsp->name,
1807 ACCESS_ONCE(rsp->gpnum),
1809 rsp->gp_state = RCU_GP_WAIT_FQS;
1810 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1811 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1813 (!ACCESS_ONCE(rnp->qsmask) &&
1814 !rcu_preempt_blocked_readers_cgp(rnp)),
1816 /* Locking provides needed memory barriers. */
1817 /* If grace period done, leave loop. */
1818 if (!ACCESS_ONCE(rnp->qsmask) &&
1819 !rcu_preempt_blocked_readers_cgp(rnp))
1821 /* If time for quiescent-state forcing, do it. */
1822 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1823 (gf & RCU_GP_FLAG_FQS)) {
1824 trace_rcu_grace_period(rsp->name,
1825 ACCESS_ONCE(rsp->gpnum),
1827 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1828 trace_rcu_grace_period(rsp->name,
1829 ACCESS_ONCE(rsp->gpnum),
1831 cond_resched_rcu_qs();
1833 /* Deal with stray signal. */
1834 cond_resched_rcu_qs();
1835 flush_signals(current);
1836 trace_rcu_grace_period(rsp->name,
1837 ACCESS_ONCE(rsp->gpnum),
1840 j = jiffies_till_next_fqs;
1843 jiffies_till_next_fqs = HZ;
1846 jiffies_till_next_fqs = 1;
1850 /* Handle grace-period end. */
1851 rcu_gp_cleanup(rsp);
1856 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1857 * in preparation for detecting the next grace period. The caller must hold
1858 * the root node's ->lock and hard irqs must be disabled.
1860 * Note that it is legal for a dying CPU (which is marked as offline) to
1861 * invoke this function. This can happen when the dying CPU reports its
1864 * Returns true if the grace-period kthread must be awakened.
1867 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1868 struct rcu_data *rdp)
1870 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1872 * Either we have not yet spawned the grace-period
1873 * task, this CPU does not need another grace period,
1874 * or a grace period is already in progress.
1875 * Either way, don't start a new grace period.
1879 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1880 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1884 * We can't do wakeups while holding the rnp->lock, as that
1885 * could cause possible deadlocks with the rq->lock. Defer
1886 * the wakeup to our caller.
1892 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1893 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1894 * is invoked indirectly from rcu_advance_cbs(), which would result in
1895 * endless recursion -- or would do so if it wasn't for the self-deadlock
1896 * that is encountered beforehand.
1898 * Returns true if the grace-period kthread needs to be awakened.
1900 static bool rcu_start_gp(struct rcu_state *rsp)
1902 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1903 struct rcu_node *rnp = rcu_get_root(rsp);
1907 * If there is no grace period in progress right now, any
1908 * callbacks we have up to this point will be satisfied by the
1909 * next grace period. Also, advancing the callbacks reduces the
1910 * probability of false positives from cpu_needs_another_gp()
1911 * resulting in pointless grace periods. So, advance callbacks
1912 * then start the grace period!
1914 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
1915 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
1920 * Report a full set of quiescent states to the specified rcu_state
1921 * data structure. This involves cleaning up after the prior grace
1922 * period and letting rcu_start_gp() start up the next grace period
1923 * if one is needed. Note that the caller must hold rnp->lock, which
1924 * is released before return.
1926 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1927 __releases(rcu_get_root(rsp)->lock)
1929 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1930 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1931 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1935 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1936 * Allows quiescent states for a group of CPUs to be reported at one go
1937 * to the specified rcu_node structure, though all the CPUs in the group
1938 * must be represented by the same rcu_node structure (which need not be
1939 * a leaf rcu_node structure, though it often will be). That structure's
1940 * lock must be held upon entry, and it is released before return.
1943 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1944 struct rcu_node *rnp, unsigned long flags)
1945 __releases(rnp->lock)
1947 struct rcu_node *rnp_c;
1949 /* Walk up the rcu_node hierarchy. */
1951 if (!(rnp->qsmask & mask)) {
1953 /* Our bit has already been cleared, so done. */
1954 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1957 rnp->qsmask &= ~mask;
1958 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1959 mask, rnp->qsmask, rnp->level,
1960 rnp->grplo, rnp->grphi,
1962 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1964 /* Other bits still set at this level, so done. */
1965 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1968 mask = rnp->grpmask;
1969 if (rnp->parent == NULL) {
1971 /* No more levels. Exit loop holding root lock. */
1975 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1978 raw_spin_lock_irqsave(&rnp->lock, flags);
1979 smp_mb__after_unlock_lock();
1980 WARN_ON_ONCE(rnp_c->qsmask);
1984 * Get here if we are the last CPU to pass through a quiescent
1985 * state for this grace period. Invoke rcu_report_qs_rsp()
1986 * to clean up and start the next grace period if one is needed.
1988 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1992 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1993 * structure. This must be either called from the specified CPU, or
1994 * called when the specified CPU is known to be offline (and when it is
1995 * also known that no other CPU is concurrently trying to help the offline
1996 * CPU). The lastcomp argument is used to make sure we are still in the
1997 * grace period of interest. We don't want to end the current grace period
1998 * based on quiescent states detected in an earlier grace period!
2001 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2003 unsigned long flags;
2006 struct rcu_node *rnp;
2009 raw_spin_lock_irqsave(&rnp->lock, flags);
2010 smp_mb__after_unlock_lock();
2011 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
2012 rnp->completed == rnp->gpnum) {
2015 * The grace period in which this quiescent state was
2016 * recorded has ended, so don't report it upwards.
2017 * We will instead need a new quiescent state that lies
2018 * within the current grace period.
2020 rdp->passed_quiesce = 0; /* need qs for new gp. */
2021 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2024 mask = rdp->grpmask;
2025 if ((rnp->qsmask & mask) == 0) {
2026 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2028 rdp->qs_pending = 0;
2031 * This GP can't end until cpu checks in, so all of our
2032 * callbacks can be processed during the next GP.
2034 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2036 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
2038 rcu_gp_kthread_wake(rsp);
2043 * Check to see if there is a new grace period of which this CPU
2044 * is not yet aware, and if so, set up local rcu_data state for it.
2045 * Otherwise, see if this CPU has just passed through its first
2046 * quiescent state for this grace period, and record that fact if so.
2049 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2051 /* Check for grace-period ends and beginnings. */
2052 note_gp_changes(rsp, rdp);
2055 * Does this CPU still need to do its part for current grace period?
2056 * If no, return and let the other CPUs do their part as well.
2058 if (!rdp->qs_pending)
2062 * Was there a quiescent state since the beginning of the grace
2063 * period? If no, then exit and wait for the next call.
2065 if (!rdp->passed_quiesce)
2069 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2072 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2075 #ifdef CONFIG_HOTPLUG_CPU
2078 * Send the specified CPU's RCU callbacks to the orphanage. The
2079 * specified CPU must be offline, and the caller must hold the
2083 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2084 struct rcu_node *rnp, struct rcu_data *rdp)
2086 /* No-CBs CPUs do not have orphanable callbacks. */
2087 if (rcu_is_nocb_cpu(rdp->cpu))
2091 * Orphan the callbacks. First adjust the counts. This is safe
2092 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2093 * cannot be running now. Thus no memory barrier is required.
2095 if (rdp->nxtlist != NULL) {
2096 rsp->qlen_lazy += rdp->qlen_lazy;
2097 rsp->qlen += rdp->qlen;
2098 rdp->n_cbs_orphaned += rdp->qlen;
2100 ACCESS_ONCE(rdp->qlen) = 0;
2104 * Next, move those callbacks still needing a grace period to
2105 * the orphanage, where some other CPU will pick them up.
2106 * Some of the callbacks might have gone partway through a grace
2107 * period, but that is too bad. They get to start over because we
2108 * cannot assume that grace periods are synchronized across CPUs.
2109 * We don't bother updating the ->nxttail[] array yet, instead
2110 * we just reset the whole thing later on.
2112 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2113 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2114 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2115 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2119 * Then move the ready-to-invoke callbacks to the orphanage,
2120 * where some other CPU will pick them up. These will not be
2121 * required to pass though another grace period: They are done.
2123 if (rdp->nxtlist != NULL) {
2124 *rsp->orphan_donetail = rdp->nxtlist;
2125 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2128 /* Finally, initialize the rcu_data structure's list to empty. */
2129 init_callback_list(rdp);
2133 * Adopt the RCU callbacks from the specified rcu_state structure's
2134 * orphanage. The caller must hold the ->orphan_lock.
2136 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2139 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2141 /* No-CBs CPUs are handled specially. */
2142 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2145 /* Do the accounting first. */
2146 rdp->qlen_lazy += rsp->qlen_lazy;
2147 rdp->qlen += rsp->qlen;
2148 rdp->n_cbs_adopted += rsp->qlen;
2149 if (rsp->qlen_lazy != rsp->qlen)
2150 rcu_idle_count_callbacks_posted();
2155 * We do not need a memory barrier here because the only way we
2156 * can get here if there is an rcu_barrier() in flight is if
2157 * we are the task doing the rcu_barrier().
2160 /* First adopt the ready-to-invoke callbacks. */
2161 if (rsp->orphan_donelist != NULL) {
2162 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2163 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2164 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2165 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2166 rdp->nxttail[i] = rsp->orphan_donetail;
2167 rsp->orphan_donelist = NULL;
2168 rsp->orphan_donetail = &rsp->orphan_donelist;
2171 /* And then adopt the callbacks that still need a grace period. */
2172 if (rsp->orphan_nxtlist != NULL) {
2173 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2174 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2175 rsp->orphan_nxtlist = NULL;
2176 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2181 * Trace the fact that this CPU is going offline.
2183 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2185 RCU_TRACE(unsigned long mask);
2186 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2187 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2189 RCU_TRACE(mask = rdp->grpmask);
2190 trace_rcu_grace_period(rsp->name,
2191 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2196 * The CPU has been completely removed, and some other CPU is reporting
2197 * this fact from process context. Do the remainder of the cleanup,
2198 * including orphaning the outgoing CPU's RCU callbacks, and also
2199 * adopting them. There can only be one CPU hotplug operation at a time,
2200 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2202 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2204 unsigned long flags;
2206 int need_report = 0;
2207 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2208 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2210 /* Adjust any no-longer-needed kthreads. */
2211 rcu_boost_kthread_setaffinity(rnp, -1);
2213 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2215 /* Exclude any attempts to start a new grace period. */
2216 mutex_lock(&rsp->onoff_mutex);
2217 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2219 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2220 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2221 rcu_adopt_orphan_cbs(rsp, flags);
2223 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2224 mask = rdp->grpmask; /* rnp->grplo is constant. */
2226 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2227 smp_mb__after_unlock_lock();
2228 rnp->qsmaskinit &= ~mask;
2229 if (rnp->qsmaskinit != 0) {
2230 if (rnp != rdp->mynode)
2231 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2234 if (rnp == rdp->mynode)
2235 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2237 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2238 mask = rnp->grpmask;
2240 } while (rnp != NULL);
2243 * We still hold the leaf rcu_node structure lock here, and
2244 * irqs are still disabled. The reason for this subterfuge is
2245 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2246 * held leads to deadlock.
2248 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2250 if (need_report & RCU_OFL_TASKS_NORM_GP)
2251 rcu_report_unblock_qs_rnp(rnp, flags);
2253 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2254 if (need_report & RCU_OFL_TASKS_EXP_GP)
2255 rcu_report_exp_rnp(rsp, rnp, true);
2256 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2257 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2258 cpu, rdp->qlen, rdp->nxtlist);
2259 init_callback_list(rdp);
2260 /* Disallow further callbacks on this CPU. */
2261 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2262 mutex_unlock(&rsp->onoff_mutex);
2265 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2267 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2271 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2275 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2278 * Invoke any RCU callbacks that have made it to the end of their grace
2279 * period. Thottle as specified by rdp->blimit.
2281 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2283 unsigned long flags;
2284 struct rcu_head *next, *list, **tail;
2285 long bl, count, count_lazy;
2288 /* If no callbacks are ready, just return. */
2289 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2290 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2291 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2292 need_resched(), is_idle_task(current),
2293 rcu_is_callbacks_kthread());
2298 * Extract the list of ready callbacks, disabling to prevent
2299 * races with call_rcu() from interrupt handlers.
2301 local_irq_save(flags);
2302 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2304 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2305 list = rdp->nxtlist;
2306 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2307 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2308 tail = rdp->nxttail[RCU_DONE_TAIL];
2309 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2310 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2311 rdp->nxttail[i] = &rdp->nxtlist;
2312 local_irq_restore(flags);
2314 /* Invoke callbacks. */
2315 count = count_lazy = 0;
2319 debug_rcu_head_unqueue(list);
2320 if (__rcu_reclaim(rsp->name, list))
2323 /* Stop only if limit reached and CPU has something to do. */
2324 if (++count >= bl &&
2326 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2330 local_irq_save(flags);
2331 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2332 is_idle_task(current),
2333 rcu_is_callbacks_kthread());
2335 /* Update count, and requeue any remaining callbacks. */
2337 *tail = rdp->nxtlist;
2338 rdp->nxtlist = list;
2339 for (i = 0; i < RCU_NEXT_SIZE; i++)
2340 if (&rdp->nxtlist == rdp->nxttail[i])
2341 rdp->nxttail[i] = tail;
2345 smp_mb(); /* List handling before counting for rcu_barrier(). */
2346 rdp->qlen_lazy -= count_lazy;
2347 ACCESS_ONCE(rdp->qlen) = rdp->qlen - count;
2348 rdp->n_cbs_invoked += count;
2350 /* Reinstate batch limit if we have worked down the excess. */
2351 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2352 rdp->blimit = blimit;
2354 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2355 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2356 rdp->qlen_last_fqs_check = 0;
2357 rdp->n_force_qs_snap = rsp->n_force_qs;
2358 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2359 rdp->qlen_last_fqs_check = rdp->qlen;
2360 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2362 local_irq_restore(flags);
2364 /* Re-invoke RCU core processing if there are callbacks remaining. */
2365 if (cpu_has_callbacks_ready_to_invoke(rdp))
2370 * Check to see if this CPU is in a non-context-switch quiescent state
2371 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2372 * Also schedule RCU core processing.
2374 * This function must be called from hardirq context. It is normally
2375 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2376 * false, there is no point in invoking rcu_check_callbacks().
2378 void rcu_check_callbacks(int cpu, int user)
2380 trace_rcu_utilization(TPS("Start scheduler-tick"));
2381 increment_cpu_stall_ticks();
2382 if (user || rcu_is_cpu_rrupt_from_idle()) {
2385 * Get here if this CPU took its interrupt from user
2386 * mode or from the idle loop, and if this is not a
2387 * nested interrupt. In this case, the CPU is in
2388 * a quiescent state, so note it.
2390 * No memory barrier is required here because both
2391 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2392 * variables that other CPUs neither access nor modify,
2393 * at least not while the corresponding CPU is online.
2399 } else if (!in_softirq()) {
2402 * Get here if this CPU did not take its interrupt from
2403 * softirq, in other words, if it is not interrupting
2404 * a rcu_bh read-side critical section. This is an _bh
2405 * critical section, so note it.
2410 rcu_preempt_check_callbacks(cpu);
2411 if (rcu_pending(cpu))
2414 rcu_note_voluntary_context_switch(current);
2415 trace_rcu_utilization(TPS("End scheduler-tick"));
2419 * Scan the leaf rcu_node structures, processing dyntick state for any that
2420 * have not yet encountered a quiescent state, using the function specified.
2421 * Also initiate boosting for any threads blocked on the root rcu_node.
2423 * The caller must have suppressed start of new grace periods.
2425 static void force_qs_rnp(struct rcu_state *rsp,
2426 int (*f)(struct rcu_data *rsp, bool *isidle,
2427 unsigned long *maxj),
2428 bool *isidle, unsigned long *maxj)
2432 unsigned long flags;
2434 struct rcu_node *rnp;
2436 rcu_for_each_leaf_node(rsp, rnp) {
2437 cond_resched_rcu_qs();
2439 raw_spin_lock_irqsave(&rnp->lock, flags);
2440 smp_mb__after_unlock_lock();
2441 if (!rcu_gp_in_progress(rsp)) {
2442 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2445 if (rnp->qsmask == 0) {
2446 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2451 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2452 if ((rnp->qsmask & bit) != 0) {
2453 if ((rnp->qsmaskinit & bit) != 0)
2455 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2461 /* rcu_report_qs_rnp() releases rnp->lock. */
2462 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2465 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2467 rnp = rcu_get_root(rsp);
2468 if (rnp->qsmask == 0) {
2469 raw_spin_lock_irqsave(&rnp->lock, flags);
2470 smp_mb__after_unlock_lock();
2471 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2476 * Force quiescent states on reluctant CPUs, and also detect which
2477 * CPUs are in dyntick-idle mode.
2479 static void force_quiescent_state(struct rcu_state *rsp)
2481 unsigned long flags;
2483 struct rcu_node *rnp;
2484 struct rcu_node *rnp_old = NULL;
2486 /* Funnel through hierarchy to reduce memory contention. */
2487 rnp = __this_cpu_read(rsp->rda->mynode);
2488 for (; rnp != NULL; rnp = rnp->parent) {
2489 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2490 !raw_spin_trylock(&rnp->fqslock);
2491 if (rnp_old != NULL)
2492 raw_spin_unlock(&rnp_old->fqslock);
2494 rsp->n_force_qs_lh++;
2499 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2501 /* Reached the root of the rcu_node tree, acquire lock. */
2502 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2503 smp_mb__after_unlock_lock();
2504 raw_spin_unlock(&rnp_old->fqslock);
2505 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2506 rsp->n_force_qs_lh++;
2507 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2508 return; /* Someone beat us to it. */
2510 ACCESS_ONCE(rsp->gp_flags) |= RCU_GP_FLAG_FQS;
2511 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2512 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2516 * This does the RCU core processing work for the specified rcu_state
2517 * and rcu_data structures. This may be called only from the CPU to
2518 * whom the rdp belongs.
2521 __rcu_process_callbacks(struct rcu_state *rsp)
2523 unsigned long flags;
2525 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2527 WARN_ON_ONCE(rdp->beenonline == 0);
2529 /* Update RCU state based on any recent quiescent states. */
2530 rcu_check_quiescent_state(rsp, rdp);
2532 /* Does this CPU require a not-yet-started grace period? */
2533 local_irq_save(flags);
2534 if (cpu_needs_another_gp(rsp, rdp)) {
2535 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2536 needwake = rcu_start_gp(rsp);
2537 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2539 rcu_gp_kthread_wake(rsp);
2541 local_irq_restore(flags);
2544 /* If there are callbacks ready, invoke them. */
2545 if (cpu_has_callbacks_ready_to_invoke(rdp))
2546 invoke_rcu_callbacks(rsp, rdp);
2548 /* Do any needed deferred wakeups of rcuo kthreads. */
2549 do_nocb_deferred_wakeup(rdp);
2553 * Do RCU core processing for the current CPU.
2555 static void rcu_process_callbacks(struct softirq_action *unused)
2557 struct rcu_state *rsp;
2559 if (cpu_is_offline(smp_processor_id()))
2561 trace_rcu_utilization(TPS("Start RCU core"));
2562 for_each_rcu_flavor(rsp)
2563 __rcu_process_callbacks(rsp);
2564 trace_rcu_utilization(TPS("End RCU core"));
2568 * Schedule RCU callback invocation. If the specified type of RCU
2569 * does not support RCU priority boosting, just do a direct call,
2570 * otherwise wake up the per-CPU kernel kthread. Note that because we
2571 * are running on the current CPU with interrupts disabled, the
2572 * rcu_cpu_kthread_task cannot disappear out from under us.
2574 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2576 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2578 if (likely(!rsp->boost)) {
2579 rcu_do_batch(rsp, rdp);
2582 invoke_rcu_callbacks_kthread();
2585 static void invoke_rcu_core(void)
2587 if (cpu_online(smp_processor_id()))
2588 raise_softirq(RCU_SOFTIRQ);
2592 * Handle any core-RCU processing required by a call_rcu() invocation.
2594 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2595 struct rcu_head *head, unsigned long flags)
2600 * If called from an extended quiescent state, invoke the RCU
2601 * core in order to force a re-evaluation of RCU's idleness.
2603 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2606 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2607 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2611 * Force the grace period if too many callbacks or too long waiting.
2612 * Enforce hysteresis, and don't invoke force_quiescent_state()
2613 * if some other CPU has recently done so. Also, don't bother
2614 * invoking force_quiescent_state() if the newly enqueued callback
2615 * is the only one waiting for a grace period to complete.
2617 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2619 /* Are we ignoring a completed grace period? */
2620 note_gp_changes(rsp, rdp);
2622 /* Start a new grace period if one not already started. */
2623 if (!rcu_gp_in_progress(rsp)) {
2624 struct rcu_node *rnp_root = rcu_get_root(rsp);
2626 raw_spin_lock(&rnp_root->lock);
2627 smp_mb__after_unlock_lock();
2628 needwake = rcu_start_gp(rsp);
2629 raw_spin_unlock(&rnp_root->lock);
2631 rcu_gp_kthread_wake(rsp);
2633 /* Give the grace period a kick. */
2634 rdp->blimit = LONG_MAX;
2635 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2636 *rdp->nxttail[RCU_DONE_TAIL] != head)
2637 force_quiescent_state(rsp);
2638 rdp->n_force_qs_snap = rsp->n_force_qs;
2639 rdp->qlen_last_fqs_check = rdp->qlen;
2645 * RCU callback function to leak a callback.
2647 static void rcu_leak_callback(struct rcu_head *rhp)
2652 * Helper function for call_rcu() and friends. The cpu argument will
2653 * normally be -1, indicating "currently running CPU". It may specify
2654 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2655 * is expected to specify a CPU.
2658 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2659 struct rcu_state *rsp, int cpu, bool lazy)
2661 unsigned long flags;
2662 struct rcu_data *rdp;
2664 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
2665 if (debug_rcu_head_queue(head)) {
2666 /* Probable double call_rcu(), so leak the callback. */
2667 ACCESS_ONCE(head->func) = rcu_leak_callback;
2668 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2675 * Opportunistically note grace-period endings and beginnings.
2676 * Note that we might see a beginning right after we see an
2677 * end, but never vice versa, since this CPU has to pass through
2678 * a quiescent state betweentimes.
2680 local_irq_save(flags);
2681 rdp = this_cpu_ptr(rsp->rda);
2683 /* Add the callback to our list. */
2684 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2688 rdp = per_cpu_ptr(rsp->rda, cpu);
2689 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2690 WARN_ON_ONCE(offline);
2691 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2692 local_irq_restore(flags);
2695 ACCESS_ONCE(rdp->qlen) = rdp->qlen + 1;
2699 rcu_idle_count_callbacks_posted();
2700 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2701 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2702 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2704 if (__is_kfree_rcu_offset((unsigned long)func))
2705 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2706 rdp->qlen_lazy, rdp->qlen);
2708 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2710 /* Go handle any RCU core processing required. */
2711 __call_rcu_core(rsp, rdp, head, flags);
2712 local_irq_restore(flags);
2716 * Queue an RCU-sched callback for invocation after a grace period.
2718 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2720 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2722 EXPORT_SYMBOL_GPL(call_rcu_sched);
2725 * Queue an RCU callback for invocation after a quicker grace period.
2727 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2729 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2731 EXPORT_SYMBOL_GPL(call_rcu_bh);
2734 * Queue an RCU callback for lazy invocation after a grace period.
2735 * This will likely be later named something like "call_rcu_lazy()",
2736 * but this change will require some way of tagging the lazy RCU
2737 * callbacks in the list of pending callbacks. Until then, this
2738 * function may only be called from __kfree_rcu().
2740 void kfree_call_rcu(struct rcu_head *head,
2741 void (*func)(struct rcu_head *rcu))
2743 __call_rcu(head, func, rcu_state_p, -1, 1);
2745 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2748 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2749 * any blocking grace-period wait automatically implies a grace period
2750 * if there is only one CPU online at any point time during execution
2751 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2752 * occasionally incorrectly indicate that there are multiple CPUs online
2753 * when there was in fact only one the whole time, as this just adds
2754 * some overhead: RCU still operates correctly.
2756 static inline int rcu_blocking_is_gp(void)
2760 might_sleep(); /* Check for RCU read-side critical section. */
2762 ret = num_online_cpus() <= 1;
2768 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2770 * Control will return to the caller some time after a full rcu-sched
2771 * grace period has elapsed, in other words after all currently executing
2772 * rcu-sched read-side critical sections have completed. These read-side
2773 * critical sections are delimited by rcu_read_lock_sched() and
2774 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2775 * local_irq_disable(), and so on may be used in place of
2776 * rcu_read_lock_sched().
2778 * This means that all preempt_disable code sequences, including NMI and
2779 * non-threaded hardware-interrupt handlers, in progress on entry will
2780 * have completed before this primitive returns. However, this does not
2781 * guarantee that softirq handlers will have completed, since in some
2782 * kernels, these handlers can run in process context, and can block.
2784 * Note that this guarantee implies further memory-ordering guarantees.
2785 * On systems with more than one CPU, when synchronize_sched() returns,
2786 * each CPU is guaranteed to have executed a full memory barrier since the
2787 * end of its last RCU-sched read-side critical section whose beginning
2788 * preceded the call to synchronize_sched(). In addition, each CPU having
2789 * an RCU read-side critical section that extends beyond the return from
2790 * synchronize_sched() is guaranteed to have executed a full memory barrier
2791 * after the beginning of synchronize_sched() and before the beginning of
2792 * that RCU read-side critical section. Note that these guarantees include
2793 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2794 * that are executing in the kernel.
2796 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2797 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2798 * to have executed a full memory barrier during the execution of
2799 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2800 * again only if the system has more than one CPU).
2802 * This primitive provides the guarantees made by the (now removed)
2803 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2804 * guarantees that rcu_read_lock() sections will have completed.
2805 * In "classic RCU", these two guarantees happen to be one and
2806 * the same, but can differ in realtime RCU implementations.
2808 void synchronize_sched(void)
2810 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2811 !lock_is_held(&rcu_lock_map) &&
2812 !lock_is_held(&rcu_sched_lock_map),
2813 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2814 if (rcu_blocking_is_gp())
2817 synchronize_sched_expedited();
2819 wait_rcu_gp(call_rcu_sched);
2821 EXPORT_SYMBOL_GPL(synchronize_sched);
2824 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2826 * Control will return to the caller some time after a full rcu_bh grace
2827 * period has elapsed, in other words after all currently executing rcu_bh
2828 * read-side critical sections have completed. RCU read-side critical
2829 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2830 * and may be nested.
2832 * See the description of synchronize_sched() for more detailed information
2833 * on memory ordering guarantees.
2835 void synchronize_rcu_bh(void)
2837 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2838 !lock_is_held(&rcu_lock_map) &&
2839 !lock_is_held(&rcu_sched_lock_map),
2840 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2841 if (rcu_blocking_is_gp())
2844 synchronize_rcu_bh_expedited();
2846 wait_rcu_gp(call_rcu_bh);
2848 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2851 * get_state_synchronize_rcu - Snapshot current RCU state
2853 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2854 * to determine whether or not a full grace period has elapsed in the
2857 unsigned long get_state_synchronize_rcu(void)
2860 * Any prior manipulation of RCU-protected data must happen
2861 * before the load from ->gpnum.
2866 * Make sure this load happens before the purportedly
2867 * time-consuming work between get_state_synchronize_rcu()
2868 * and cond_synchronize_rcu().
2870 return smp_load_acquire(&rcu_state_p->gpnum);
2872 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2875 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2877 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2879 * If a full RCU grace period has elapsed since the earlier call to
2880 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2881 * synchronize_rcu() to wait for a full grace period.
2883 * Yes, this function does not take counter wrap into account. But
2884 * counter wrap is harmless. If the counter wraps, we have waited for
2885 * more than 2 billion grace periods (and way more on a 64-bit system!),
2886 * so waiting for one additional grace period should be just fine.
2888 void cond_synchronize_rcu(unsigned long oldstate)
2890 unsigned long newstate;
2893 * Ensure that this load happens before any RCU-destructive
2894 * actions the caller might carry out after we return.
2896 newstate = smp_load_acquire(&rcu_state_p->completed);
2897 if (ULONG_CMP_GE(oldstate, newstate))
2900 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2902 static int synchronize_sched_expedited_cpu_stop(void *data)
2905 * There must be a full memory barrier on each affected CPU
2906 * between the time that try_stop_cpus() is called and the
2907 * time that it returns.
2909 * In the current initial implementation of cpu_stop, the
2910 * above condition is already met when the control reaches
2911 * this point and the following smp_mb() is not strictly
2912 * necessary. Do smp_mb() anyway for documentation and
2913 * robustness against future implementation changes.
2915 smp_mb(); /* See above comment block. */
2920 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2922 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2923 * approach to force the grace period to end quickly. This consumes
2924 * significant time on all CPUs and is unfriendly to real-time workloads,
2925 * so is thus not recommended for any sort of common-case code. In fact,
2926 * if you are using synchronize_sched_expedited() in a loop, please
2927 * restructure your code to batch your updates, and then use a single
2928 * synchronize_sched() instead.
2930 * Note that it is illegal to call this function while holding any lock
2931 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2932 * to call this function from a CPU-hotplug notifier. Failing to observe
2933 * these restriction will result in deadlock.
2935 * This implementation can be thought of as an application of ticket
2936 * locking to RCU, with sync_sched_expedited_started and
2937 * sync_sched_expedited_done taking on the roles of the halves
2938 * of the ticket-lock word. Each task atomically increments
2939 * sync_sched_expedited_started upon entry, snapshotting the old value,
2940 * then attempts to stop all the CPUs. If this succeeds, then each
2941 * CPU will have executed a context switch, resulting in an RCU-sched
2942 * grace period. We are then done, so we use atomic_cmpxchg() to
2943 * update sync_sched_expedited_done to match our snapshot -- but
2944 * only if someone else has not already advanced past our snapshot.
2946 * On the other hand, if try_stop_cpus() fails, we check the value
2947 * of sync_sched_expedited_done. If it has advanced past our
2948 * initial snapshot, then someone else must have forced a grace period
2949 * some time after we took our snapshot. In this case, our work is
2950 * done for us, and we can simply return. Otherwise, we try again,
2951 * but keep our initial snapshot for purposes of checking for someone
2952 * doing our work for us.
2954 * If we fail too many times in a row, we fall back to synchronize_sched().
2956 void synchronize_sched_expedited(void)
2958 long firstsnap, s, snap;
2960 struct rcu_state *rsp = &rcu_sched_state;
2963 * If we are in danger of counter wrap, just do synchronize_sched().
2964 * By allowing sync_sched_expedited_started to advance no more than
2965 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2966 * that more than 3.5 billion CPUs would be required to force a
2967 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2968 * course be required on a 64-bit system.
2970 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2971 (ulong)atomic_long_read(&rsp->expedited_done) +
2973 synchronize_sched();
2974 atomic_long_inc(&rsp->expedited_wrap);
2979 * Take a ticket. Note that atomic_inc_return() implies a
2980 * full memory barrier.
2982 snap = atomic_long_inc_return(&rsp->expedited_start);
2985 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2988 * Each pass through the following loop attempts to force a
2989 * context switch on each CPU.
2991 while (try_stop_cpus(cpu_online_mask,
2992 synchronize_sched_expedited_cpu_stop,
2995 atomic_long_inc(&rsp->expedited_tryfail);
2997 /* Check to see if someone else did our work for us. */
2998 s = atomic_long_read(&rsp->expedited_done);
2999 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3000 /* ensure test happens before caller kfree */
3001 smp_mb__before_atomic(); /* ^^^ */
3002 atomic_long_inc(&rsp->expedited_workdone1);
3006 /* No joy, try again later. Or just synchronize_sched(). */
3007 if (trycount++ < 10) {
3008 udelay(trycount * num_online_cpus());
3010 wait_rcu_gp(call_rcu_sched);
3011 atomic_long_inc(&rsp->expedited_normal);
3015 /* Recheck to see if someone else did our work for us. */
3016 s = atomic_long_read(&rsp->expedited_done);
3017 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3018 /* ensure test happens before caller kfree */
3019 smp_mb__before_atomic(); /* ^^^ */
3020 atomic_long_inc(&rsp->expedited_workdone2);
3025 * Refetching sync_sched_expedited_started allows later
3026 * callers to piggyback on our grace period. We retry
3027 * after they started, so our grace period works for them,
3028 * and they started after our first try, so their grace
3029 * period works for us.
3032 snap = atomic_long_read(&rsp->expedited_start);
3033 smp_mb(); /* ensure read is before try_stop_cpus(). */
3035 atomic_long_inc(&rsp->expedited_stoppedcpus);
3038 * Everyone up to our most recent fetch is covered by our grace
3039 * period. Update the counter, but only if our work is still
3040 * relevant -- which it won't be if someone who started later
3041 * than we did already did their update.
3044 atomic_long_inc(&rsp->expedited_done_tries);
3045 s = atomic_long_read(&rsp->expedited_done);
3046 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
3047 /* ensure test happens before caller kfree */
3048 smp_mb__before_atomic(); /* ^^^ */
3049 atomic_long_inc(&rsp->expedited_done_lost);
3052 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
3053 atomic_long_inc(&rsp->expedited_done_exit);
3057 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3060 * Check to see if there is any immediate RCU-related work to be done
3061 * by the current CPU, for the specified type of RCU, returning 1 if so.
3062 * The checks are in order of increasing expense: checks that can be
3063 * carried out against CPU-local state are performed first. However,
3064 * we must check for CPU stalls first, else we might not get a chance.
3066 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3068 struct rcu_node *rnp = rdp->mynode;
3070 rdp->n_rcu_pending++;
3072 /* Check for CPU stalls, if enabled. */
3073 check_cpu_stall(rsp, rdp);
3075 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3076 if (rcu_nohz_full_cpu(rsp))
3079 /* Is the RCU core waiting for a quiescent state from this CPU? */
3080 if (rcu_scheduler_fully_active &&
3081 rdp->qs_pending && !rdp->passed_quiesce) {
3082 rdp->n_rp_qs_pending++;
3083 } else if (rdp->qs_pending && rdp->passed_quiesce) {
3084 rdp->n_rp_report_qs++;
3088 /* Does this CPU have callbacks ready to invoke? */
3089 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3090 rdp->n_rp_cb_ready++;
3094 /* Has RCU gone idle with this CPU needing another grace period? */
3095 if (cpu_needs_another_gp(rsp, rdp)) {
3096 rdp->n_rp_cpu_needs_gp++;
3100 /* Has another RCU grace period completed? */
3101 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3102 rdp->n_rp_gp_completed++;
3106 /* Has a new RCU grace period started? */
3107 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
3108 rdp->n_rp_gp_started++;
3112 /* Does this CPU need a deferred NOCB wakeup? */
3113 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3114 rdp->n_rp_nocb_defer_wakeup++;
3119 rdp->n_rp_need_nothing++;
3124 * Check to see if there is any immediate RCU-related work to be done
3125 * by the current CPU, returning 1 if so. This function is part of the
3126 * RCU implementation; it is -not- an exported member of the RCU API.
3128 static int rcu_pending(int cpu)
3130 struct rcu_state *rsp;
3132 for_each_rcu_flavor(rsp)
3133 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
3139 * Return true if the specified CPU has any callback. If all_lazy is
3140 * non-NULL, store an indication of whether all callbacks are lazy.
3141 * (If there are no callbacks, all of them are deemed to be lazy.)
3143 static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
3147 struct rcu_data *rdp;
3148 struct rcu_state *rsp;
3150 for_each_rcu_flavor(rsp) {
3151 rdp = per_cpu_ptr(rsp->rda, cpu);
3155 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3166 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3167 * the compiler is expected to optimize this away.
3169 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3170 int cpu, unsigned long done)
3172 trace_rcu_barrier(rsp->name, s, cpu,
3173 atomic_read(&rsp->barrier_cpu_count), done);
3177 * RCU callback function for _rcu_barrier(). If we are last, wake
3178 * up the task executing _rcu_barrier().
3180 static void rcu_barrier_callback(struct rcu_head *rhp)
3182 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3183 struct rcu_state *rsp = rdp->rsp;
3185 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3186 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
3187 complete(&rsp->barrier_completion);
3189 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
3194 * Called with preemption disabled, and from cross-cpu IRQ context.
3196 static void rcu_barrier_func(void *type)
3198 struct rcu_state *rsp = type;
3199 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3201 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
3202 atomic_inc(&rsp->barrier_cpu_count);
3203 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3207 * Orchestrate the specified type of RCU barrier, waiting for all
3208 * RCU callbacks of the specified type to complete.
3210 static void _rcu_barrier(struct rcu_state *rsp)
3213 struct rcu_data *rdp;
3214 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3215 unsigned long snap_done;
3217 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3219 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3220 mutex_lock(&rsp->barrier_mutex);
3223 * Ensure that all prior references, including to ->n_barrier_done,
3224 * are ordered before the _rcu_barrier() machinery.
3226 smp_mb(); /* See above block comment. */
3229 * Recheck ->n_barrier_done to see if others did our work for us.
3230 * This means checking ->n_barrier_done for an even-to-odd-to-even
3231 * transition. The "if" expression below therefore rounds the old
3232 * value up to the next even number and adds two before comparing.
3234 snap_done = rsp->n_barrier_done;
3235 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3238 * If the value in snap is odd, we needed to wait for the current
3239 * rcu_barrier() to complete, then wait for the next one, in other
3240 * words, we need the value of snap_done to be three larger than
3241 * the value of snap. On the other hand, if the value in snap is
3242 * even, we only had to wait for the next rcu_barrier() to complete,
3243 * in other words, we need the value of snap_done to be only two
3244 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3245 * this for us (thank you, Linus!).
3247 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3248 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3249 smp_mb(); /* caller's subsequent code after above check. */
3250 mutex_unlock(&rsp->barrier_mutex);
3255 * Increment ->n_barrier_done to avoid duplicate work. Use
3256 * ACCESS_ONCE() to prevent the compiler from speculating
3257 * the increment to precede the early-exit check.
3259 ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1;
3260 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3261 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3262 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3265 * Initialize the count to one rather than to zero in order to
3266 * avoid a too-soon return to zero in case of a short grace period
3267 * (or preemption of this task). Exclude CPU-hotplug operations
3268 * to ensure that no offline CPU has callbacks queued.
3270 init_completion(&rsp->barrier_completion);
3271 atomic_set(&rsp->barrier_cpu_count, 1);
3275 * Force each CPU with callbacks to register a new callback.
3276 * When that callback is invoked, we will know that all of the
3277 * corresponding CPU's preceding callbacks have been invoked.
3279 for_each_possible_cpu(cpu) {
3280 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3282 rdp = per_cpu_ptr(rsp->rda, cpu);
3283 if (rcu_is_nocb_cpu(cpu)) {
3284 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3285 rsp->n_barrier_done);
3286 atomic_inc(&rsp->barrier_cpu_count);
3287 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3289 } else if (ACCESS_ONCE(rdp->qlen)) {
3290 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3291 rsp->n_barrier_done);
3292 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3294 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3295 rsp->n_barrier_done);
3301 * Now that we have an rcu_barrier_callback() callback on each
3302 * CPU, and thus each counted, remove the initial count.
3304 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3305 complete(&rsp->barrier_completion);
3307 /* Increment ->n_barrier_done to prevent duplicate work. */
3308 smp_mb(); /* Keep increment after above mechanism. */
3309 ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1;
3310 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3311 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3312 smp_mb(); /* Keep increment before caller's subsequent code. */
3314 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3315 wait_for_completion(&rsp->barrier_completion);
3317 /* Other rcu_barrier() invocations can now safely proceed. */
3318 mutex_unlock(&rsp->barrier_mutex);
3322 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3324 void rcu_barrier_bh(void)
3326 _rcu_barrier(&rcu_bh_state);
3328 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3331 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3333 void rcu_barrier_sched(void)
3335 _rcu_barrier(&rcu_sched_state);
3337 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3340 * Do boot-time initialization of a CPU's per-CPU RCU data.
3343 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3345 unsigned long flags;
3346 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3347 struct rcu_node *rnp = rcu_get_root(rsp);
3349 /* Set up local state, ensuring consistent view of global state. */
3350 raw_spin_lock_irqsave(&rnp->lock, flags);
3351 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3352 init_callback_list(rdp);
3354 ACCESS_ONCE(rdp->qlen) = 0;
3355 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3356 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3357 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3360 rcu_boot_init_nocb_percpu_data(rdp);
3361 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3365 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3366 * offline event can be happening at a given time. Note also that we
3367 * can accept some slop in the rsp->completed access due to the fact
3368 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3371 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3373 unsigned long flags;
3375 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3376 struct rcu_node *rnp = rcu_get_root(rsp);
3378 /* Exclude new grace periods. */
3379 mutex_lock(&rsp->onoff_mutex);
3381 /* Set up local state, ensuring consistent view of global state. */
3382 raw_spin_lock_irqsave(&rnp->lock, flags);
3383 rdp->beenonline = 1; /* We have now been online. */
3384 rdp->qlen_last_fqs_check = 0;
3385 rdp->n_force_qs_snap = rsp->n_force_qs;
3386 rdp->blimit = blimit;
3387 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3388 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3389 rcu_sysidle_init_percpu_data(rdp->dynticks);
3390 atomic_set(&rdp->dynticks->dynticks,
3391 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3392 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3394 /* Add CPU to rcu_node bitmasks. */
3396 mask = rdp->grpmask;
3398 /* Exclude any attempts to start a new GP on small systems. */
3399 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3400 rnp->qsmaskinit |= mask;
3401 mask = rnp->grpmask;
3402 if (rnp == rdp->mynode) {
3404 * If there is a grace period in progress, we will
3405 * set up to wait for it next time we run the
3408 rdp->gpnum = rnp->completed;
3409 rdp->completed = rnp->completed;
3410 rdp->passed_quiesce = 0;
3411 rdp->qs_pending = 0;
3412 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3414 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3416 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3417 local_irq_restore(flags);
3419 mutex_unlock(&rsp->onoff_mutex);
3422 static void rcu_prepare_cpu(int cpu)
3424 struct rcu_state *rsp;
3426 for_each_rcu_flavor(rsp)
3427 rcu_init_percpu_data(cpu, rsp);
3431 * Handle CPU online/offline notification events.
3433 static int rcu_cpu_notify(struct notifier_block *self,
3434 unsigned long action, void *hcpu)
3436 long cpu = (long)hcpu;
3437 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3438 struct rcu_node *rnp = rdp->mynode;
3439 struct rcu_state *rsp;
3441 trace_rcu_utilization(TPS("Start CPU hotplug"));
3443 case CPU_UP_PREPARE:
3444 case CPU_UP_PREPARE_FROZEN:
3445 rcu_prepare_cpu(cpu);
3446 rcu_prepare_kthreads(cpu);
3449 case CPU_DOWN_FAILED:
3450 rcu_boost_kthread_setaffinity(rnp, -1);
3452 case CPU_DOWN_PREPARE:
3453 rcu_boost_kthread_setaffinity(rnp, cpu);
3456 case CPU_DYING_FROZEN:
3457 for_each_rcu_flavor(rsp)
3458 rcu_cleanup_dying_cpu(rsp);
3461 case CPU_DEAD_FROZEN:
3462 case CPU_UP_CANCELED:
3463 case CPU_UP_CANCELED_FROZEN:
3464 for_each_rcu_flavor(rsp)
3465 rcu_cleanup_dead_cpu(cpu, rsp);
3470 trace_rcu_utilization(TPS("End CPU hotplug"));
3474 static int rcu_pm_notify(struct notifier_block *self,
3475 unsigned long action, void *hcpu)
3478 case PM_HIBERNATION_PREPARE:
3479 case PM_SUSPEND_PREPARE:
3480 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3483 case PM_POST_HIBERNATION:
3484 case PM_POST_SUSPEND:
3494 * Spawn the kthread that handles this RCU flavor's grace periods.
3496 static int __init rcu_spawn_gp_kthread(void)
3498 unsigned long flags;
3499 struct rcu_node *rnp;
3500 struct rcu_state *rsp;
3501 struct task_struct *t;
3503 for_each_rcu_flavor(rsp) {
3504 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3506 rnp = rcu_get_root(rsp);
3507 raw_spin_lock_irqsave(&rnp->lock, flags);
3508 rsp->gp_kthread = t;
3509 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3510 rcu_spawn_nocb_kthreads(rsp);
3514 early_initcall(rcu_spawn_gp_kthread);
3517 * This function is invoked towards the end of the scheduler's initialization
3518 * process. Before this is called, the idle task might contain
3519 * RCU read-side critical sections (during which time, this idle
3520 * task is booting the system). After this function is called, the
3521 * idle tasks are prohibited from containing RCU read-side critical
3522 * sections. This function also enables RCU lockdep checking.
3524 void rcu_scheduler_starting(void)
3526 WARN_ON(num_online_cpus() != 1);
3527 WARN_ON(nr_context_switches() > 0);
3528 rcu_scheduler_active = 1;
3532 * Compute the per-level fanout, either using the exact fanout specified
3533 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3535 #ifdef CONFIG_RCU_FANOUT_EXACT
3536 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3540 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3541 for (i = rcu_num_lvls - 2; i >= 0; i--)
3542 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3544 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3545 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3552 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3553 ccur = rsp->levelcnt[i];
3554 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3558 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3561 * Helper function for rcu_init() that initializes one rcu_state structure.
3563 static void __init rcu_init_one(struct rcu_state *rsp,
3564 struct rcu_data __percpu *rda)
3566 static const char * const buf[] = {
3570 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3571 static const char * const fqs[] = {
3575 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3576 static u8 fl_mask = 0x1;
3580 struct rcu_node *rnp;
3582 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3584 /* Silence gcc 4.8 warning about array index out of range. */
3585 if (rcu_num_lvls > RCU_NUM_LVLS)
3586 panic("rcu_init_one: rcu_num_lvls overflow");
3588 /* Initialize the level-tracking arrays. */
3590 for (i = 0; i < rcu_num_lvls; i++)
3591 rsp->levelcnt[i] = num_rcu_lvl[i];
3592 for (i = 1; i < rcu_num_lvls; i++)
3593 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3594 rcu_init_levelspread(rsp);
3595 rsp->flavor_mask = fl_mask;
3598 /* Initialize the elements themselves, starting from the leaves. */
3600 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3601 cpustride *= rsp->levelspread[i];
3602 rnp = rsp->level[i];
3603 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3604 raw_spin_lock_init(&rnp->lock);
3605 lockdep_set_class_and_name(&rnp->lock,
3606 &rcu_node_class[i], buf[i]);
3607 raw_spin_lock_init(&rnp->fqslock);
3608 lockdep_set_class_and_name(&rnp->fqslock,
3609 &rcu_fqs_class[i], fqs[i]);
3610 rnp->gpnum = rsp->gpnum;
3611 rnp->completed = rsp->completed;
3613 rnp->qsmaskinit = 0;
3614 rnp->grplo = j * cpustride;
3615 rnp->grphi = (j + 1) * cpustride - 1;
3616 if (rnp->grphi >= nr_cpu_ids)
3617 rnp->grphi = nr_cpu_ids - 1;
3623 rnp->grpnum = j % rsp->levelspread[i - 1];
3624 rnp->grpmask = 1UL << rnp->grpnum;
3625 rnp->parent = rsp->level[i - 1] +
3626 j / rsp->levelspread[i - 1];
3629 INIT_LIST_HEAD(&rnp->blkd_tasks);
3630 rcu_init_one_nocb(rnp);
3635 init_waitqueue_head(&rsp->gp_wq);
3636 rnp = rsp->level[rcu_num_lvls - 1];
3637 for_each_possible_cpu(i) {
3638 while (i > rnp->grphi)
3640 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3641 rcu_boot_init_percpu_data(i, rsp);
3643 list_add(&rsp->flavors, &rcu_struct_flavors);
3647 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3648 * replace the definitions in tree.h because those are needed to size
3649 * the ->node array in the rcu_state structure.
3651 static void __init rcu_init_geometry(void)
3657 int rcu_capacity[MAX_RCU_LVLS + 1];
3660 * Initialize any unspecified boot parameters.
3661 * The default values of jiffies_till_first_fqs and
3662 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3663 * value, which is a function of HZ, then adding one for each
3664 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3666 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3667 if (jiffies_till_first_fqs == ULONG_MAX)
3668 jiffies_till_first_fqs = d;
3669 if (jiffies_till_next_fqs == ULONG_MAX)
3670 jiffies_till_next_fqs = d;
3672 /* If the compile-time values are accurate, just leave. */
3673 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3674 nr_cpu_ids == NR_CPUS)
3676 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3677 rcu_fanout_leaf, nr_cpu_ids);
3680 * Compute number of nodes that can be handled an rcu_node tree
3681 * with the given number of levels. Setting rcu_capacity[0] makes
3682 * some of the arithmetic easier.
3684 rcu_capacity[0] = 1;
3685 rcu_capacity[1] = rcu_fanout_leaf;
3686 for (i = 2; i <= MAX_RCU_LVLS; i++)
3687 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3690 * The boot-time rcu_fanout_leaf parameter is only permitted
3691 * to increase the leaf-level fanout, not decrease it. Of course,
3692 * the leaf-level fanout cannot exceed the number of bits in
3693 * the rcu_node masks. Finally, the tree must be able to accommodate
3694 * the configured number of CPUs. Complain and fall back to the
3695 * compile-time values if these limits are exceeded.
3697 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3698 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3699 n > rcu_capacity[MAX_RCU_LVLS]) {
3704 /* Calculate the number of rcu_nodes at each level of the tree. */
3705 for (i = 1; i <= MAX_RCU_LVLS; i++)
3706 if (n <= rcu_capacity[i]) {
3707 for (j = 0; j <= i; j++)
3709 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3711 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3716 /* Calculate the total number of rcu_node structures. */
3718 for (i = 0; i <= MAX_RCU_LVLS; i++)
3719 rcu_num_nodes += num_rcu_lvl[i];
3723 void __init rcu_init(void)
3727 rcu_bootup_announce();
3728 rcu_init_geometry();
3729 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3730 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3731 __rcu_init_preempt();
3732 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3735 * We don't need protection against CPU-hotplug here because
3736 * this is called early in boot, before either interrupts
3737 * or the scheduler are operational.
3739 cpu_notifier(rcu_cpu_notify, 0);
3740 pm_notifier(rcu_pm_notify, 0);
3741 for_each_online_cpu(cpu)
3742 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3745 #include "tree_plugin.h"