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.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 struct rcu_state sname##_state = { \
95 .level = { &sname##_state.node[0] }, \
97 .fqs_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
101 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
102 .orphan_donetail = &sname##_state.orphan_donelist, \
103 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
104 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
105 .name = RCU_STATE_NAME(sname), \
108 DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data)
110 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
111 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
113 static struct rcu_state *rcu_state_p;
114 LIST_HEAD(rcu_struct_flavors);
116 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
117 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
118 module_param(rcu_fanout_leaf, int, 0444);
119 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
120 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
127 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
130 * The rcu_scheduler_active variable transitions from zero to one just
131 * before the first task is spawned. So when this variable is zero, RCU
132 * can assume that there is but one task, allowing RCU to (for example)
133 * optimize synchronize_sched() to a simple barrier(). When this variable
134 * is one, RCU must actually do all the hard work required to detect real
135 * grace periods. This variable is also used to suppress boot-time false
136 * positives from lockdep-RCU error checking.
138 int rcu_scheduler_active __read_mostly;
139 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
142 * The rcu_scheduler_fully_active variable transitions from zero to one
143 * during the early_initcall() processing, which is after the scheduler
144 * is capable of creating new tasks. So RCU processing (for example,
145 * creating tasks for RCU priority boosting) must be delayed until after
146 * rcu_scheduler_fully_active transitions from zero to one. We also
147 * currently delay invocation of any RCU callbacks until after this point.
149 * It might later prove better for people registering RCU callbacks during
150 * early boot to take responsibility for these callbacks, but one step at
153 static int rcu_scheduler_fully_active __read_mostly;
155 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
156 static void invoke_rcu_core(void);
157 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
160 * Track the rcutorture test sequence number and the update version
161 * number within a given test. The rcutorture_testseq is incremented
162 * on every rcutorture module load and unload, so has an odd value
163 * when a test is running. The rcutorture_vernum is set to zero
164 * when rcutorture starts and is incremented on each rcutorture update.
165 * These variables enable correlating rcutorture output with the
166 * RCU tracing information.
168 unsigned long rcutorture_testseq;
169 unsigned long rcutorture_vernum;
172 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
173 * permit this function to be invoked without holding the root rcu_node
174 * structure's ->lock, but of course results can be subject to change.
176 static int rcu_gp_in_progress(struct rcu_state *rsp)
178 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
182 * Note a quiescent state. Because we do not need to know
183 * how many quiescent states passed, just if there was at least
184 * one since the start of the grace period, this just sets a flag.
185 * The caller must have disabled preemption.
187 void rcu_sched_qs(void)
189 if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) {
190 trace_rcu_grace_period(TPS("rcu_sched"),
191 __this_cpu_read(rcu_sched_data.gpnum),
193 __this_cpu_write(rcu_sched_data.passed_quiesce, 1);
199 if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) {
200 trace_rcu_grace_period(TPS("rcu_bh"),
201 __this_cpu_read(rcu_bh_data.gpnum),
203 __this_cpu_write(rcu_bh_data.passed_quiesce, 1);
207 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
209 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
210 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
211 .dynticks = ATOMIC_INIT(1),
212 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
213 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
214 .dynticks_idle = ATOMIC_INIT(1),
215 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
219 * Let the RCU core know that this CPU has gone through the scheduler,
220 * which is a quiescent state. This is called when the need for a
221 * quiescent state is urgent, so we burn an atomic operation and full
222 * memory barriers to let the RCU core know about it, regardless of what
223 * this CPU might (or might not) do in the near future.
225 * We inform the RCU core by emulating a zero-duration dyntick-idle
226 * period, which we in turn do by incrementing the ->dynticks counter
229 static void rcu_momentary_dyntick_idle(void)
232 struct rcu_data *rdp;
233 struct rcu_dynticks *rdtp;
235 struct rcu_state *rsp;
237 local_irq_save(flags);
240 * Yes, we can lose flag-setting operations. This is OK, because
241 * the flag will be set again after some delay.
243 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
244 raw_cpu_write(rcu_sched_qs_mask, 0);
246 /* Find the flavor that needs a quiescent state. */
247 for_each_rcu_flavor(rsp) {
248 rdp = raw_cpu_ptr(rsp->rda);
249 if (!(resched_mask & rsp->flavor_mask))
251 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
252 if (ACCESS_ONCE(rdp->mynode->completed) !=
253 ACCESS_ONCE(rdp->cond_resched_completed))
257 * Pretend to be momentarily idle for the quiescent state.
258 * This allows the grace-period kthread to record the
259 * quiescent state, with no need for this CPU to do anything
262 rdtp = this_cpu_ptr(&rcu_dynticks);
263 smp_mb__before_atomic(); /* Earlier stuff before QS. */
264 atomic_add(2, &rdtp->dynticks); /* QS. */
265 smp_mb__after_atomic(); /* Later stuff after QS. */
268 local_irq_restore(flags);
272 * Note a context switch. This is a quiescent state for RCU-sched,
273 * and requires special handling for preemptible RCU.
274 * The caller must have disabled preemption.
276 void rcu_note_context_switch(void)
278 trace_rcu_utilization(TPS("Start context switch"));
280 rcu_preempt_note_context_switch();
281 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
282 rcu_momentary_dyntick_idle();
283 trace_rcu_utilization(TPS("End context switch"));
285 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
287 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
288 static long qhimark = 10000; /* If this many pending, ignore blimit. */
289 static long qlowmark = 100; /* Once only this many pending, use blimit. */
291 module_param(blimit, long, 0444);
292 module_param(qhimark, long, 0444);
293 module_param(qlowmark, long, 0444);
295 static ulong jiffies_till_first_fqs = ULONG_MAX;
296 static ulong jiffies_till_next_fqs = ULONG_MAX;
298 module_param(jiffies_till_first_fqs, ulong, 0644);
299 module_param(jiffies_till_next_fqs, ulong, 0644);
302 * How long the grace period must be before we start recruiting
303 * quiescent-state help from rcu_note_context_switch().
305 static ulong jiffies_till_sched_qs = HZ / 20;
306 module_param(jiffies_till_sched_qs, ulong, 0644);
308 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
309 struct rcu_data *rdp);
310 static void force_qs_rnp(struct rcu_state *rsp,
311 int (*f)(struct rcu_data *rsp, bool *isidle,
312 unsigned long *maxj),
313 bool *isidle, unsigned long *maxj);
314 static void force_quiescent_state(struct rcu_state *rsp);
315 static int rcu_pending(void);
318 * Return the number of RCU-sched batches processed thus far for debug & stats.
320 long rcu_batches_completed_sched(void)
322 return rcu_sched_state.completed;
324 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
327 * Return the number of RCU BH batches processed thus far for debug & stats.
329 long rcu_batches_completed_bh(void)
331 return rcu_bh_state.completed;
333 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
336 * Force a quiescent state.
338 void rcu_force_quiescent_state(void)
340 force_quiescent_state(rcu_state_p);
342 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
345 * Force a quiescent state for RCU BH.
347 void rcu_bh_force_quiescent_state(void)
349 force_quiescent_state(&rcu_bh_state);
351 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
354 * Show the state of the grace-period kthreads.
356 void show_rcu_gp_kthreads(void)
358 struct rcu_state *rsp;
360 for_each_rcu_flavor(rsp) {
361 pr_info("%s: wait state: %d ->state: %#lx\n",
362 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
363 /* sched_show_task(rsp->gp_kthread); */
366 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
369 * Record the number of times rcutorture tests have been initiated and
370 * terminated. This information allows the debugfs tracing stats to be
371 * correlated to the rcutorture messages, even when the rcutorture module
372 * is being repeatedly loaded and unloaded. In other words, we cannot
373 * store this state in rcutorture itself.
375 void rcutorture_record_test_transition(void)
377 rcutorture_testseq++;
378 rcutorture_vernum = 0;
380 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
383 * Send along grace-period-related data for rcutorture diagnostics.
385 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
386 unsigned long *gpnum, unsigned long *completed)
388 struct rcu_state *rsp = NULL;
397 case RCU_SCHED_FLAVOR:
398 rsp = &rcu_sched_state;
404 *flags = ACCESS_ONCE(rsp->gp_flags);
405 *gpnum = ACCESS_ONCE(rsp->gpnum);
406 *completed = ACCESS_ONCE(rsp->completed);
413 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
416 * Record the number of writer passes through the current rcutorture test.
417 * This is also used to correlate debugfs tracing stats with the rcutorture
420 void rcutorture_record_progress(unsigned long vernum)
424 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
427 * Force a quiescent state for RCU-sched.
429 void rcu_sched_force_quiescent_state(void)
431 force_quiescent_state(&rcu_sched_state);
433 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
436 * Does the CPU have callbacks ready to be invoked?
439 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
441 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
442 rdp->nxttail[RCU_DONE_TAIL] != NULL;
446 * Return the root node of the specified rcu_state structure.
448 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
450 return &rsp->node[0];
454 * Is there any need for future grace periods?
455 * Interrupts must be disabled. If the caller does not hold the root
456 * rnp_node structure's ->lock, the results are advisory only.
458 static int rcu_future_needs_gp(struct rcu_state *rsp)
460 struct rcu_node *rnp = rcu_get_root(rsp);
461 int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1;
462 int *fp = &rnp->need_future_gp[idx];
464 return ACCESS_ONCE(*fp);
468 * Does the current CPU require a not-yet-started grace period?
469 * The caller must have disabled interrupts to prevent races with
470 * normal callback registry.
473 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
477 if (rcu_gp_in_progress(rsp))
478 return 0; /* No, a grace period is already in progress. */
479 if (rcu_future_needs_gp(rsp))
480 return 1; /* Yes, a no-CBs CPU needs one. */
481 if (!rdp->nxttail[RCU_NEXT_TAIL])
482 return 0; /* No, this is a no-CBs (or offline) CPU. */
483 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
484 return 1; /* Yes, this CPU has newly registered callbacks. */
485 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
486 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
487 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
488 rdp->nxtcompleted[i]))
489 return 1; /* Yes, CBs for future grace period. */
490 return 0; /* No grace period needed. */
494 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
496 * If the new value of the ->dynticks_nesting counter now is zero,
497 * we really have entered idle, and must do the appropriate accounting.
498 * The caller must have disabled interrupts.
500 static void rcu_eqs_enter_common(long long oldval, bool user)
502 struct rcu_state *rsp;
503 struct rcu_data *rdp;
504 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
506 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
507 if (!user && !is_idle_task(current)) {
508 struct task_struct *idle __maybe_unused =
509 idle_task(smp_processor_id());
511 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
512 ftrace_dump(DUMP_ORIG);
513 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
514 current->pid, current->comm,
515 idle->pid, idle->comm); /* must be idle task! */
517 for_each_rcu_flavor(rsp) {
518 rdp = this_cpu_ptr(rsp->rda);
519 do_nocb_deferred_wakeup(rdp);
521 rcu_prepare_for_idle();
522 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
523 smp_mb__before_atomic(); /* See above. */
524 atomic_inc(&rdtp->dynticks);
525 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
526 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
527 rcu_dynticks_task_enter();
530 * It is illegal to enter an extended quiescent state while
531 * in an RCU read-side critical section.
533 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
534 "Illegal idle entry in RCU read-side critical section.");
535 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
536 "Illegal idle entry in RCU-bh read-side critical section.");
537 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
538 "Illegal idle entry in RCU-sched read-side critical section.");
542 * Enter an RCU extended quiescent state, which can be either the
543 * idle loop or adaptive-tickless usermode execution.
545 static void rcu_eqs_enter(bool user)
548 struct rcu_dynticks *rdtp;
550 rdtp = this_cpu_ptr(&rcu_dynticks);
551 oldval = rdtp->dynticks_nesting;
552 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
553 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
554 rdtp->dynticks_nesting = 0;
555 rcu_eqs_enter_common(oldval, user);
557 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
562 * rcu_idle_enter - inform RCU that current CPU is entering idle
564 * Enter idle mode, in other words, -leave- the mode in which RCU
565 * read-side critical sections can occur. (Though RCU read-side
566 * critical sections can occur in irq handlers in idle, a possibility
567 * handled by irq_enter() and irq_exit().)
569 * We crowbar the ->dynticks_nesting field to zero to allow for
570 * the possibility of usermode upcalls having messed up our count
571 * of interrupt nesting level during the prior busy period.
573 void rcu_idle_enter(void)
577 local_irq_save(flags);
578 rcu_eqs_enter(false);
579 rcu_sysidle_enter(0);
580 local_irq_restore(flags);
582 EXPORT_SYMBOL_GPL(rcu_idle_enter);
584 #ifdef CONFIG_RCU_USER_QS
586 * rcu_user_enter - inform RCU that we are resuming userspace.
588 * Enter RCU idle mode right before resuming userspace. No use of RCU
589 * is permitted between this call and rcu_user_exit(). This way the
590 * CPU doesn't need to maintain the tick for RCU maintenance purposes
591 * when the CPU runs in userspace.
593 void rcu_user_enter(void)
597 #endif /* CONFIG_RCU_USER_QS */
600 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
602 * Exit from an interrupt handler, which might possibly result in entering
603 * idle mode, in other words, leaving the mode in which read-side critical
604 * sections can occur.
606 * This code assumes that the idle loop never does anything that might
607 * result in unbalanced calls to irq_enter() and irq_exit(). If your
608 * architecture violates this assumption, RCU will give you what you
609 * deserve, good and hard. But very infrequently and irreproducibly.
611 * Use things like work queues to work around this limitation.
613 * You have been warned.
615 void rcu_irq_exit(void)
619 struct rcu_dynticks *rdtp;
621 local_irq_save(flags);
622 rdtp = this_cpu_ptr(&rcu_dynticks);
623 oldval = rdtp->dynticks_nesting;
624 rdtp->dynticks_nesting--;
625 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
626 if (rdtp->dynticks_nesting)
627 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
629 rcu_eqs_enter_common(oldval, true);
630 rcu_sysidle_enter(1);
631 local_irq_restore(flags);
635 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
637 * If the new value of the ->dynticks_nesting counter was previously zero,
638 * we really have exited idle, and must do the appropriate accounting.
639 * The caller must have disabled interrupts.
641 static void rcu_eqs_exit_common(long long oldval, int user)
643 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
645 rcu_dynticks_task_exit();
646 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
647 atomic_inc(&rdtp->dynticks);
648 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
649 smp_mb__after_atomic(); /* See above. */
650 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
651 rcu_cleanup_after_idle();
652 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
653 if (!user && !is_idle_task(current)) {
654 struct task_struct *idle __maybe_unused =
655 idle_task(smp_processor_id());
657 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
658 oldval, rdtp->dynticks_nesting);
659 ftrace_dump(DUMP_ORIG);
660 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
661 current->pid, current->comm,
662 idle->pid, idle->comm); /* must be idle task! */
667 * Exit an RCU extended quiescent state, which can be either the
668 * idle loop or adaptive-tickless usermode execution.
670 static void rcu_eqs_exit(bool user)
672 struct rcu_dynticks *rdtp;
675 rdtp = this_cpu_ptr(&rcu_dynticks);
676 oldval = rdtp->dynticks_nesting;
677 WARN_ON_ONCE(oldval < 0);
678 if (oldval & DYNTICK_TASK_NEST_MASK) {
679 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
681 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
682 rcu_eqs_exit_common(oldval, user);
687 * rcu_idle_exit - inform RCU that current CPU is leaving idle
689 * Exit idle mode, in other words, -enter- the mode in which RCU
690 * read-side critical sections can occur.
692 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
693 * allow for the possibility of usermode upcalls messing up our count
694 * of interrupt nesting level during the busy period that is just
697 void rcu_idle_exit(void)
701 local_irq_save(flags);
704 local_irq_restore(flags);
706 EXPORT_SYMBOL_GPL(rcu_idle_exit);
708 #ifdef CONFIG_RCU_USER_QS
710 * rcu_user_exit - inform RCU that we are exiting userspace.
712 * Exit RCU idle mode while entering the kernel because it can
713 * run a RCU read side critical section anytime.
715 void rcu_user_exit(void)
719 #endif /* CONFIG_RCU_USER_QS */
722 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
724 * Enter an interrupt handler, which might possibly result in exiting
725 * idle mode, in other words, entering the mode in which read-side critical
726 * sections can occur.
728 * Note that the Linux kernel is fully capable of entering an interrupt
729 * handler that it never exits, for example when doing upcalls to
730 * user mode! This code assumes that the idle loop never does upcalls to
731 * user mode. If your architecture does do upcalls from the idle loop (or
732 * does anything else that results in unbalanced calls to the irq_enter()
733 * and irq_exit() functions), RCU will give you what you deserve, good
734 * and hard. But very infrequently and irreproducibly.
736 * Use things like work queues to work around this limitation.
738 * You have been warned.
740 void rcu_irq_enter(void)
743 struct rcu_dynticks *rdtp;
746 local_irq_save(flags);
747 rdtp = this_cpu_ptr(&rcu_dynticks);
748 oldval = rdtp->dynticks_nesting;
749 rdtp->dynticks_nesting++;
750 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
752 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
754 rcu_eqs_exit_common(oldval, true);
756 local_irq_restore(flags);
760 * rcu_nmi_enter - inform RCU of entry to NMI context
762 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
763 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
764 * that the CPU is active. This implementation permits nested NMIs, as
765 * long as the nesting level does not overflow an int. (You will probably
766 * run out of stack space first.)
768 void rcu_nmi_enter(void)
770 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
773 /* Complain about underflow. */
774 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
777 * If idle from RCU viewpoint, atomically increment ->dynticks
778 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
779 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
780 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
781 * to be in the outermost NMI handler that interrupted an RCU-idle
782 * period (observation due to Andy Lutomirski).
784 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
785 smp_mb__before_atomic(); /* Force delay from prior write. */
786 atomic_inc(&rdtp->dynticks);
787 /* atomic_inc() before later RCU read-side crit sects */
788 smp_mb__after_atomic(); /* See above. */
789 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
792 rdtp->dynticks_nmi_nesting += incby;
797 * rcu_nmi_exit - inform RCU of exit from NMI context
799 * If we are returning from the outermost NMI handler that interrupted an
800 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
801 * to let the RCU grace-period handling know that the CPU is back to
804 void rcu_nmi_exit(void)
806 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
809 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
810 * (We are exiting an NMI handler, so RCU better be paying attention
813 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
814 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
817 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
818 * leave it in non-RCU-idle state.
820 if (rdtp->dynticks_nmi_nesting != 1) {
821 rdtp->dynticks_nmi_nesting -= 2;
825 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
826 rdtp->dynticks_nmi_nesting = 0;
827 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
828 smp_mb__before_atomic(); /* See above. */
829 atomic_inc(&rdtp->dynticks);
830 smp_mb__after_atomic(); /* Force delay to next write. */
831 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
835 * __rcu_is_watching - are RCU read-side critical sections safe?
837 * Return true if RCU is watching the running CPU, which means that
838 * this CPU can safely enter RCU read-side critical sections. Unlike
839 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
840 * least disabled preemption.
842 bool notrace __rcu_is_watching(void)
844 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
848 * rcu_is_watching - see if RCU thinks that the current CPU is idle
850 * If the current CPU is in its idle loop and is neither in an interrupt
851 * or NMI handler, return true.
853 bool notrace rcu_is_watching(void)
858 ret = __rcu_is_watching();
862 EXPORT_SYMBOL_GPL(rcu_is_watching);
864 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
867 * Is the current CPU online? Disable preemption to avoid false positives
868 * that could otherwise happen due to the current CPU number being sampled,
869 * this task being preempted, its old CPU being taken offline, resuming
870 * on some other CPU, then determining that its old CPU is now offline.
871 * It is OK to use RCU on an offline processor during initial boot, hence
872 * the check for rcu_scheduler_fully_active. Note also that it is OK
873 * for a CPU coming online to use RCU for one jiffy prior to marking itself
874 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
875 * offline to continue to use RCU for one jiffy after marking itself
876 * offline in the cpu_online_mask. This leniency is necessary given the
877 * non-atomic nature of the online and offline processing, for example,
878 * the fact that a CPU enters the scheduler after completing the CPU_DYING
881 * This is also why RCU internally marks CPUs online during the
882 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
884 * Disable checking if in an NMI handler because we cannot safely report
885 * errors from NMI handlers anyway.
887 bool rcu_lockdep_current_cpu_online(void)
889 struct rcu_data *rdp;
890 struct rcu_node *rnp;
896 rdp = this_cpu_ptr(&rcu_sched_data);
898 ret = (rdp->grpmask & rnp->qsmaskinit) ||
899 !rcu_scheduler_fully_active;
903 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
905 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
908 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
910 * If the current CPU is idle or running at a first-level (not nested)
911 * interrupt from idle, return true. The caller must have at least
912 * disabled preemption.
914 static int rcu_is_cpu_rrupt_from_idle(void)
916 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
920 * Snapshot the specified CPU's dynticks counter so that we can later
921 * credit them with an implicit quiescent state. Return 1 if this CPU
922 * is in dynticks idle mode, which is an extended quiescent state.
924 static int dyntick_save_progress_counter(struct rcu_data *rdp,
925 bool *isidle, unsigned long *maxj)
927 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
928 rcu_sysidle_check_cpu(rdp, isidle, maxj);
929 if ((rdp->dynticks_snap & 0x1) == 0) {
930 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
933 if (ULONG_CMP_LT(ACCESS_ONCE(rdp->gpnum) + ULONG_MAX / 4,
935 ACCESS_ONCE(rdp->gpwrap) = true;
941 * This function really isn't for public consumption, but RCU is special in
942 * that context switches can allow the state machine to make progress.
944 extern void resched_cpu(int cpu);
947 * Return true if the specified CPU has passed through a quiescent
948 * state by virtue of being in or having passed through an dynticks
949 * idle state since the last call to dyntick_save_progress_counter()
950 * for this same CPU, or by virtue of having been offline.
952 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
953 bool *isidle, unsigned long *maxj)
959 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
960 snap = (unsigned int)rdp->dynticks_snap;
963 * If the CPU passed through or entered a dynticks idle phase with
964 * no active irq/NMI handlers, then we can safely pretend that the CPU
965 * already acknowledged the request to pass through a quiescent
966 * state. Either way, that CPU cannot possibly be in an RCU
967 * read-side critical section that started before the beginning
968 * of the current RCU grace period.
970 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
971 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
977 * Check for the CPU being offline, but only if the grace period
978 * is old enough. We don't need to worry about the CPU changing
979 * state: If we see it offline even once, it has been through a
982 * The reason for insisting that the grace period be at least
983 * one jiffy old is that CPUs that are not quite online and that
984 * have just gone offline can still execute RCU read-side critical
987 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
988 return 0; /* Grace period is not old enough. */
990 if (cpu_is_offline(rdp->cpu)) {
991 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
997 * A CPU running for an extended time within the kernel can
998 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
999 * even context-switching back and forth between a pair of
1000 * in-kernel CPU-bound tasks cannot advance grace periods.
1001 * So if the grace period is old enough, make the CPU pay attention.
1002 * Note that the unsynchronized assignments to the per-CPU
1003 * rcu_sched_qs_mask variable are safe. Yes, setting of
1004 * bits can be lost, but they will be set again on the next
1005 * force-quiescent-state pass. So lost bit sets do not result
1006 * in incorrect behavior, merely in a grace period lasting
1007 * a few jiffies longer than it might otherwise. Because
1008 * there are at most four threads involved, and because the
1009 * updates are only once every few jiffies, the probability of
1010 * lossage (and thus of slight grace-period extension) is
1013 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1014 * is set too high, we override with half of the RCU CPU stall
1017 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1018 if (ULONG_CMP_GE(jiffies,
1019 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1020 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1021 if (!(ACCESS_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1022 ACCESS_ONCE(rdp->cond_resched_completed) =
1023 ACCESS_ONCE(rdp->mynode->completed);
1024 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1025 ACCESS_ONCE(*rcrmp) =
1026 ACCESS_ONCE(*rcrmp) + rdp->rsp->flavor_mask;
1027 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1028 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
1029 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1030 /* Time to beat on that CPU again! */
1031 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1032 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1039 static void record_gp_stall_check_time(struct rcu_state *rsp)
1041 unsigned long j = jiffies;
1045 smp_wmb(); /* Record start time before stall time. */
1046 j1 = rcu_jiffies_till_stall_check();
1047 ACCESS_ONCE(rsp->jiffies_stall) = j + j1;
1048 rsp->jiffies_resched = j + j1 / 2;
1049 rsp->n_force_qs_gpstart = ACCESS_ONCE(rsp->n_force_qs);
1053 * Dump stacks of all tasks running on stalled CPUs.
1055 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1058 unsigned long flags;
1059 struct rcu_node *rnp;
1061 rcu_for_each_leaf_node(rsp, rnp) {
1062 raw_spin_lock_irqsave(&rnp->lock, flags);
1063 if (rnp->qsmask != 0) {
1064 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1065 if (rnp->qsmask & (1UL << cpu))
1066 dump_cpu_task(rnp->grplo + cpu);
1068 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1072 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1076 unsigned long flags;
1080 struct rcu_node *rnp = rcu_get_root(rsp);
1083 /* Only let one CPU complain about others per time interval. */
1085 raw_spin_lock_irqsave(&rnp->lock, flags);
1086 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
1087 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1088 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1091 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1092 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1095 * OK, time to rat on our buddy...
1096 * See Documentation/RCU/stallwarn.txt for info on how to debug
1097 * RCU CPU stall warnings.
1099 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1101 print_cpu_stall_info_begin();
1102 rcu_for_each_leaf_node(rsp, rnp) {
1103 raw_spin_lock_irqsave(&rnp->lock, flags);
1104 ndetected += rcu_print_task_stall(rnp);
1105 if (rnp->qsmask != 0) {
1106 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1107 if (rnp->qsmask & (1UL << cpu)) {
1108 print_cpu_stall_info(rsp,
1113 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1117 * Now rat on any tasks that got kicked up to the root rcu_node
1118 * due to CPU offlining.
1120 rnp = rcu_get_root(rsp);
1121 raw_spin_lock_irqsave(&rnp->lock, flags);
1122 ndetected += rcu_print_task_stall(rnp);
1123 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1125 print_cpu_stall_info_end();
1126 for_each_possible_cpu(cpu)
1127 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1128 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1129 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1130 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1132 rcu_dump_cpu_stacks(rsp);
1134 if (ACCESS_ONCE(rsp->gpnum) != gpnum ||
1135 ACCESS_ONCE(rsp->completed) == gpnum) {
1136 pr_err("INFO: Stall ended before state dump start\n");
1139 gpa = ACCESS_ONCE(rsp->gp_activity);
1140 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld\n",
1141 rsp->name, j - gpa, j, gpa,
1142 jiffies_till_next_fqs);
1143 /* In this case, the current CPU might be at fault. */
1144 sched_show_task(current);
1148 /* Complain about tasks blocking the grace period. */
1150 rcu_print_detail_task_stall(rsp);
1152 force_quiescent_state(rsp); /* Kick them all. */
1155 static void print_cpu_stall(struct rcu_state *rsp)
1158 unsigned long flags;
1159 struct rcu_node *rnp = rcu_get_root(rsp);
1163 * OK, time to rat on ourselves...
1164 * See Documentation/RCU/stallwarn.txt for info on how to debug
1165 * RCU CPU stall warnings.
1167 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1168 print_cpu_stall_info_begin();
1169 print_cpu_stall_info(rsp, smp_processor_id());
1170 print_cpu_stall_info_end();
1171 for_each_possible_cpu(cpu)
1172 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1173 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1174 jiffies - rsp->gp_start,
1175 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1176 rcu_dump_cpu_stacks(rsp);
1178 raw_spin_lock_irqsave(&rnp->lock, flags);
1179 if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall)))
1180 ACCESS_ONCE(rsp->jiffies_stall) = jiffies +
1181 3 * rcu_jiffies_till_stall_check() + 3;
1182 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1185 * Attempt to revive the RCU machinery by forcing a context switch.
1187 * A context switch would normally allow the RCU state machine to make
1188 * progress and it could be we're stuck in kernel space without context
1189 * switches for an entirely unreasonable amount of time.
1191 resched_cpu(smp_processor_id());
1194 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1196 unsigned long completed;
1197 unsigned long gpnum;
1201 struct rcu_node *rnp;
1203 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1208 * Lots of memory barriers to reject false positives.
1210 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1211 * then rsp->gp_start, and finally rsp->completed. These values
1212 * are updated in the opposite order with memory barriers (or
1213 * equivalent) during grace-period initialization and cleanup.
1214 * Now, a false positive can occur if we get an new value of
1215 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1216 * the memory barriers, the only way that this can happen is if one
1217 * grace period ends and another starts between these two fetches.
1218 * Detect this by comparing rsp->completed with the previous fetch
1221 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1222 * and rsp->gp_start suffice to forestall false positives.
1224 gpnum = ACCESS_ONCE(rsp->gpnum);
1225 smp_rmb(); /* Pick up ->gpnum first... */
1226 js = ACCESS_ONCE(rsp->jiffies_stall);
1227 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1228 gps = ACCESS_ONCE(rsp->gp_start);
1229 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1230 completed = ACCESS_ONCE(rsp->completed);
1231 if (ULONG_CMP_GE(completed, gpnum) ||
1232 ULONG_CMP_LT(j, js) ||
1233 ULONG_CMP_GE(gps, js))
1234 return; /* No stall or GP completed since entering function. */
1236 if (rcu_gp_in_progress(rsp) &&
1237 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1239 /* We haven't checked in, so go dump stack. */
1240 print_cpu_stall(rsp);
1242 } else if (rcu_gp_in_progress(rsp) &&
1243 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1245 /* They had a few time units to dump stack, so complain. */
1246 print_other_cpu_stall(rsp, gpnum);
1251 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1253 * Set the stall-warning timeout way off into the future, thus preventing
1254 * any RCU CPU stall-warning messages from appearing in the current set of
1255 * RCU grace periods.
1257 * The caller must disable hard irqs.
1259 void rcu_cpu_stall_reset(void)
1261 struct rcu_state *rsp;
1263 for_each_rcu_flavor(rsp)
1264 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2;
1268 * Initialize the specified rcu_data structure's callback list to empty.
1270 static void init_callback_list(struct rcu_data *rdp)
1274 if (init_nocb_callback_list(rdp))
1276 rdp->nxtlist = NULL;
1277 for (i = 0; i < RCU_NEXT_SIZE; i++)
1278 rdp->nxttail[i] = &rdp->nxtlist;
1282 * Determine the value that ->completed will have at the end of the
1283 * next subsequent grace period. This is used to tag callbacks so that
1284 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1285 * been dyntick-idle for an extended period with callbacks under the
1286 * influence of RCU_FAST_NO_HZ.
1288 * The caller must hold rnp->lock with interrupts disabled.
1290 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1291 struct rcu_node *rnp)
1294 * If RCU is idle, we just wait for the next grace period.
1295 * But we can only be sure that RCU is idle if we are looking
1296 * at the root rcu_node structure -- otherwise, a new grace
1297 * period might have started, but just not yet gotten around
1298 * to initializing the current non-root rcu_node structure.
1300 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1301 return rnp->completed + 1;
1304 * Otherwise, wait for a possible partial grace period and
1305 * then the subsequent full grace period.
1307 return rnp->completed + 2;
1311 * Trace-event helper function for rcu_start_future_gp() and
1312 * rcu_nocb_wait_gp().
1314 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1315 unsigned long c, const char *s)
1317 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1318 rnp->completed, c, rnp->level,
1319 rnp->grplo, rnp->grphi, s);
1323 * Start some future grace period, as needed to handle newly arrived
1324 * callbacks. The required future grace periods are recorded in each
1325 * rcu_node structure's ->need_future_gp field. Returns true if there
1326 * is reason to awaken the grace-period kthread.
1328 * The caller must hold the specified rcu_node structure's ->lock.
1330 static bool __maybe_unused
1331 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1332 unsigned long *c_out)
1337 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1340 * Pick up grace-period number for new callbacks. If this
1341 * grace period is already marked as needed, return to the caller.
1343 c = rcu_cbs_completed(rdp->rsp, rnp);
1344 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1345 if (rnp->need_future_gp[c & 0x1]) {
1346 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1351 * If either this rcu_node structure or the root rcu_node structure
1352 * believe that a grace period is in progress, then we must wait
1353 * for the one following, which is in "c". Because our request
1354 * will be noticed at the end of the current grace period, we don't
1355 * need to explicitly start one. We only do the lockless check
1356 * of rnp_root's fields if the current rcu_node structure thinks
1357 * there is no grace period in flight, and because we hold rnp->lock,
1358 * the only possible change is when rnp_root's two fields are
1359 * equal, in which case rnp_root->gpnum might be concurrently
1360 * incremented. But that is OK, as it will just result in our
1361 * doing some extra useless work.
1363 if (rnp->gpnum != rnp->completed ||
1364 ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) {
1365 rnp->need_future_gp[c & 0x1]++;
1366 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1371 * There might be no grace period in progress. If we don't already
1372 * hold it, acquire the root rcu_node structure's lock in order to
1373 * start one (if needed).
1375 if (rnp != rnp_root) {
1376 raw_spin_lock(&rnp_root->lock);
1377 smp_mb__after_unlock_lock();
1381 * Get a new grace-period number. If there really is no grace
1382 * period in progress, it will be smaller than the one we obtained
1383 * earlier. Adjust callbacks as needed. Note that even no-CBs
1384 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1386 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1387 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1388 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1389 rdp->nxtcompleted[i] = c;
1392 * If the needed for the required grace period is already
1393 * recorded, trace and leave.
1395 if (rnp_root->need_future_gp[c & 0x1]) {
1396 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1400 /* Record the need for the future grace period. */
1401 rnp_root->need_future_gp[c & 0x1]++;
1403 /* If a grace period is not already in progress, start one. */
1404 if (rnp_root->gpnum != rnp_root->completed) {
1405 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1407 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1408 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1411 if (rnp != rnp_root)
1412 raw_spin_unlock(&rnp_root->lock);
1420 * Clean up any old requests for the just-ended grace period. Also return
1421 * whether any additional grace periods have been requested. Also invoke
1422 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1423 * waiting for this grace period to complete.
1425 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1427 int c = rnp->completed;
1429 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1431 rcu_nocb_gp_cleanup(rsp, rnp);
1432 rnp->need_future_gp[c & 0x1] = 0;
1433 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1434 trace_rcu_future_gp(rnp, rdp, c,
1435 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1440 * Awaken the grace-period kthread for the specified flavor of RCU.
1441 * Don't do a self-awaken, and don't bother awakening when there is
1442 * nothing for the grace-period kthread to do (as in several CPUs
1443 * raced to awaken, and we lost), and finally don't try to awaken
1444 * a kthread that has not yet been created.
1446 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1448 if (current == rsp->gp_kthread ||
1449 !ACCESS_ONCE(rsp->gp_flags) ||
1452 wake_up(&rsp->gp_wq);
1456 * If there is room, assign a ->completed number to any callbacks on
1457 * this CPU that have not already been assigned. Also accelerate any
1458 * callbacks that were previously assigned a ->completed number that has
1459 * since proven to be too conservative, which can happen if callbacks get
1460 * assigned a ->completed number while RCU is idle, but with reference to
1461 * a non-root rcu_node structure. This function is idempotent, so it does
1462 * not hurt to call it repeatedly. Returns an flag saying that we should
1463 * awaken the RCU grace-period kthread.
1465 * The caller must hold rnp->lock with interrupts disabled.
1467 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1468 struct rcu_data *rdp)
1474 /* If the CPU has no callbacks, nothing to do. */
1475 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1479 * Starting from the sublist containing the callbacks most
1480 * recently assigned a ->completed number and working down, find the
1481 * first sublist that is not assignable to an upcoming grace period.
1482 * Such a sublist has something in it (first two tests) and has
1483 * a ->completed number assigned that will complete sooner than
1484 * the ->completed number for newly arrived callbacks (last test).
1486 * The key point is that any later sublist can be assigned the
1487 * same ->completed number as the newly arrived callbacks, which
1488 * means that the callbacks in any of these later sublist can be
1489 * grouped into a single sublist, whether or not they have already
1490 * been assigned a ->completed number.
1492 c = rcu_cbs_completed(rsp, rnp);
1493 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1494 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1495 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1499 * If there are no sublist for unassigned callbacks, leave.
1500 * At the same time, advance "i" one sublist, so that "i" will
1501 * index into the sublist where all the remaining callbacks should
1504 if (++i >= RCU_NEXT_TAIL)
1508 * Assign all subsequent callbacks' ->completed number to the next
1509 * full grace period and group them all in the sublist initially
1512 for (; i <= RCU_NEXT_TAIL; i++) {
1513 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1514 rdp->nxtcompleted[i] = c;
1516 /* Record any needed additional grace periods. */
1517 ret = rcu_start_future_gp(rnp, rdp, NULL);
1519 /* Trace depending on how much we were able to accelerate. */
1520 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1521 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1523 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1528 * Move any callbacks whose grace period has completed to the
1529 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1530 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1531 * sublist. This function is idempotent, so it does not hurt to
1532 * invoke it repeatedly. As long as it is not invoked -too- often...
1533 * Returns true if the RCU grace-period kthread needs to be awakened.
1535 * The caller must hold rnp->lock with interrupts disabled.
1537 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1538 struct rcu_data *rdp)
1542 /* If the CPU has no callbacks, nothing to do. */
1543 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1547 * Find all callbacks whose ->completed numbers indicate that they
1548 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1550 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1551 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1553 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1555 /* Clean up any sublist tail pointers that were misordered above. */
1556 for (j = RCU_WAIT_TAIL; j < i; j++)
1557 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1559 /* Copy down callbacks to fill in empty sublists. */
1560 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1561 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1563 rdp->nxttail[j] = rdp->nxttail[i];
1564 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1567 /* Classify any remaining callbacks. */
1568 return rcu_accelerate_cbs(rsp, rnp, rdp);
1572 * Update CPU-local rcu_data state to record the beginnings and ends of
1573 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1574 * structure corresponding to the current CPU, and must have irqs disabled.
1575 * Returns true if the grace-period kthread needs to be awakened.
1577 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1578 struct rcu_data *rdp)
1582 /* Handle the ends of any preceding grace periods first. */
1583 if (rdp->completed == rnp->completed &&
1584 !unlikely(ACCESS_ONCE(rdp->gpwrap))) {
1586 /* No grace period end, so just accelerate recent callbacks. */
1587 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1591 /* Advance callbacks. */
1592 ret = rcu_advance_cbs(rsp, rnp, rdp);
1594 /* Remember that we saw this grace-period completion. */
1595 rdp->completed = rnp->completed;
1596 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1599 if (rdp->gpnum != rnp->gpnum || unlikely(ACCESS_ONCE(rdp->gpwrap))) {
1601 * If the current grace period is waiting for this CPU,
1602 * set up to detect a quiescent state, otherwise don't
1603 * go looking for one.
1605 rdp->gpnum = rnp->gpnum;
1606 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1607 rdp->passed_quiesce = 0;
1608 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1609 zero_cpu_stall_ticks(rdp);
1610 ACCESS_ONCE(rdp->gpwrap) = false;
1615 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1617 unsigned long flags;
1619 struct rcu_node *rnp;
1621 local_irq_save(flags);
1623 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1624 rdp->completed == ACCESS_ONCE(rnp->completed) &&
1625 !unlikely(ACCESS_ONCE(rdp->gpwrap))) || /* w/out lock. */
1626 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1627 local_irq_restore(flags);
1630 smp_mb__after_unlock_lock();
1631 needwake = __note_gp_changes(rsp, rnp, rdp);
1632 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1634 rcu_gp_kthread_wake(rsp);
1638 * Initialize a new grace period. Return 0 if no grace period required.
1640 static int rcu_gp_init(struct rcu_state *rsp)
1642 struct rcu_data *rdp;
1643 struct rcu_node *rnp = rcu_get_root(rsp);
1645 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1646 rcu_bind_gp_kthread();
1647 raw_spin_lock_irq(&rnp->lock);
1648 smp_mb__after_unlock_lock();
1649 if (!ACCESS_ONCE(rsp->gp_flags)) {
1650 /* Spurious wakeup, tell caller to go back to sleep. */
1651 raw_spin_unlock_irq(&rnp->lock);
1654 ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */
1656 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1658 * Grace period already in progress, don't start another.
1659 * Not supposed to be able to happen.
1661 raw_spin_unlock_irq(&rnp->lock);
1665 /* Advance to a new grace period and initialize state. */
1666 record_gp_stall_check_time(rsp);
1667 /* Record GP times before starting GP, hence smp_store_release(). */
1668 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1669 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1670 raw_spin_unlock_irq(&rnp->lock);
1672 /* Exclude any concurrent CPU-hotplug operations. */
1673 mutex_lock(&rsp->onoff_mutex);
1674 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1677 * Set the quiescent-state-needed bits in all the rcu_node
1678 * structures for all currently online CPUs in breadth-first order,
1679 * starting from the root rcu_node structure, relying on the layout
1680 * of the tree within the rsp->node[] array. Note that other CPUs
1681 * will access only the leaves of the hierarchy, thus seeing that no
1682 * grace period is in progress, at least until the corresponding
1683 * leaf node has been initialized. In addition, we have excluded
1684 * CPU-hotplug operations.
1686 * The grace period cannot complete until the initialization
1687 * process finishes, because this kthread handles both.
1689 rcu_for_each_node_breadth_first(rsp, rnp) {
1690 raw_spin_lock_irq(&rnp->lock);
1691 smp_mb__after_unlock_lock();
1692 rdp = this_cpu_ptr(rsp->rda);
1693 rcu_preempt_check_blocked_tasks(rnp);
1694 rnp->qsmask = rnp->qsmaskinit;
1695 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1696 WARN_ON_ONCE(rnp->completed != rsp->completed);
1697 ACCESS_ONCE(rnp->completed) = rsp->completed;
1698 if (rnp == rdp->mynode)
1699 (void)__note_gp_changes(rsp, rnp, rdp);
1700 rcu_preempt_boost_start_gp(rnp);
1701 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1702 rnp->level, rnp->grplo,
1703 rnp->grphi, rnp->qsmask);
1704 raw_spin_unlock_irq(&rnp->lock);
1705 cond_resched_rcu_qs();
1706 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1709 mutex_unlock(&rsp->onoff_mutex);
1714 * Do one round of quiescent-state forcing.
1716 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1718 int fqs_state = fqs_state_in;
1719 bool isidle = false;
1721 struct rcu_node *rnp = rcu_get_root(rsp);
1723 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1725 if (fqs_state == RCU_SAVE_DYNTICK) {
1726 /* Collect dyntick-idle snapshots. */
1727 if (is_sysidle_rcu_state(rsp)) {
1729 maxj = jiffies - ULONG_MAX / 4;
1731 force_qs_rnp(rsp, dyntick_save_progress_counter,
1733 rcu_sysidle_report_gp(rsp, isidle, maxj);
1734 fqs_state = RCU_FORCE_QS;
1736 /* Handle dyntick-idle and offline CPUs. */
1738 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1740 /* Clear flag to prevent immediate re-entry. */
1741 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1742 raw_spin_lock_irq(&rnp->lock);
1743 smp_mb__after_unlock_lock();
1744 ACCESS_ONCE(rsp->gp_flags) =
1745 ACCESS_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS;
1746 raw_spin_unlock_irq(&rnp->lock);
1752 * Clean up after the old grace period.
1754 static void rcu_gp_cleanup(struct rcu_state *rsp)
1756 unsigned long gp_duration;
1757 bool needgp = false;
1759 struct rcu_data *rdp;
1760 struct rcu_node *rnp = rcu_get_root(rsp);
1762 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1763 raw_spin_lock_irq(&rnp->lock);
1764 smp_mb__after_unlock_lock();
1765 gp_duration = jiffies - rsp->gp_start;
1766 if (gp_duration > rsp->gp_max)
1767 rsp->gp_max = gp_duration;
1770 * We know the grace period is complete, but to everyone else
1771 * it appears to still be ongoing. But it is also the case
1772 * that to everyone else it looks like there is nothing that
1773 * they can do to advance the grace period. It is therefore
1774 * safe for us to drop the lock in order to mark the grace
1775 * period as completed in all of the rcu_node structures.
1777 raw_spin_unlock_irq(&rnp->lock);
1780 * Propagate new ->completed value to rcu_node structures so
1781 * that other CPUs don't have to wait until the start of the next
1782 * grace period to process their callbacks. This also avoids
1783 * some nasty RCU grace-period initialization races by forcing
1784 * the end of the current grace period to be completely recorded in
1785 * all of the rcu_node structures before the beginning of the next
1786 * grace period is recorded in any of the rcu_node structures.
1788 rcu_for_each_node_breadth_first(rsp, rnp) {
1789 raw_spin_lock_irq(&rnp->lock);
1790 smp_mb__after_unlock_lock();
1791 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1792 rdp = this_cpu_ptr(rsp->rda);
1793 if (rnp == rdp->mynode)
1794 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
1795 /* smp_mb() provided by prior unlock-lock pair. */
1796 nocb += rcu_future_gp_cleanup(rsp, rnp);
1797 raw_spin_unlock_irq(&rnp->lock);
1798 cond_resched_rcu_qs();
1799 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1801 rnp = rcu_get_root(rsp);
1802 raw_spin_lock_irq(&rnp->lock);
1803 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1804 rcu_nocb_gp_set(rnp, nocb);
1806 /* Declare grace period done. */
1807 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1808 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1809 rsp->fqs_state = RCU_GP_IDLE;
1810 rdp = this_cpu_ptr(rsp->rda);
1811 /* Advance CBs to reduce false positives below. */
1812 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
1813 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
1814 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1815 trace_rcu_grace_period(rsp->name,
1816 ACCESS_ONCE(rsp->gpnum),
1819 raw_spin_unlock_irq(&rnp->lock);
1823 * Body of kthread that handles grace periods.
1825 static int __noreturn rcu_gp_kthread(void *arg)
1831 struct rcu_state *rsp = arg;
1832 struct rcu_node *rnp = rcu_get_root(rsp);
1836 /* Handle grace-period start. */
1838 trace_rcu_grace_period(rsp->name,
1839 ACCESS_ONCE(rsp->gpnum),
1841 rsp->gp_state = RCU_GP_WAIT_GPS;
1842 wait_event_interruptible(rsp->gp_wq,
1843 ACCESS_ONCE(rsp->gp_flags) &
1845 /* Locking provides needed memory barrier. */
1846 if (rcu_gp_init(rsp))
1848 cond_resched_rcu_qs();
1849 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1850 WARN_ON(signal_pending(current));
1851 trace_rcu_grace_period(rsp->name,
1852 ACCESS_ONCE(rsp->gpnum),
1856 /* Handle quiescent-state forcing. */
1857 fqs_state = RCU_SAVE_DYNTICK;
1858 j = jiffies_till_first_fqs;
1861 jiffies_till_first_fqs = HZ;
1866 rsp->jiffies_force_qs = jiffies + j;
1867 trace_rcu_grace_period(rsp->name,
1868 ACCESS_ONCE(rsp->gpnum),
1870 rsp->gp_state = RCU_GP_WAIT_FQS;
1871 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1872 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1874 (!ACCESS_ONCE(rnp->qsmask) &&
1875 !rcu_preempt_blocked_readers_cgp(rnp)),
1877 /* Locking provides needed memory barriers. */
1878 /* If grace period done, leave loop. */
1879 if (!ACCESS_ONCE(rnp->qsmask) &&
1880 !rcu_preempt_blocked_readers_cgp(rnp))
1882 /* If time for quiescent-state forcing, do it. */
1883 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1884 (gf & RCU_GP_FLAG_FQS)) {
1885 trace_rcu_grace_period(rsp->name,
1886 ACCESS_ONCE(rsp->gpnum),
1888 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1889 trace_rcu_grace_period(rsp->name,
1890 ACCESS_ONCE(rsp->gpnum),
1892 cond_resched_rcu_qs();
1893 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1895 /* Deal with stray signal. */
1896 cond_resched_rcu_qs();
1897 ACCESS_ONCE(rsp->gp_activity) = jiffies;
1898 WARN_ON(signal_pending(current));
1899 trace_rcu_grace_period(rsp->name,
1900 ACCESS_ONCE(rsp->gpnum),
1903 j = jiffies_till_next_fqs;
1906 jiffies_till_next_fqs = HZ;
1909 jiffies_till_next_fqs = 1;
1913 /* Handle grace-period end. */
1914 rcu_gp_cleanup(rsp);
1919 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1920 * in preparation for detecting the next grace period. The caller must hold
1921 * the root node's ->lock and hard irqs must be disabled.
1923 * Note that it is legal for a dying CPU (which is marked as offline) to
1924 * invoke this function. This can happen when the dying CPU reports its
1927 * Returns true if the grace-period kthread must be awakened.
1930 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1931 struct rcu_data *rdp)
1933 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1935 * Either we have not yet spawned the grace-period
1936 * task, this CPU does not need another grace period,
1937 * or a grace period is already in progress.
1938 * Either way, don't start a new grace period.
1942 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1943 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1947 * We can't do wakeups while holding the rnp->lock, as that
1948 * could cause possible deadlocks with the rq->lock. Defer
1949 * the wakeup to our caller.
1955 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1956 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1957 * is invoked indirectly from rcu_advance_cbs(), which would result in
1958 * endless recursion -- or would do so if it wasn't for the self-deadlock
1959 * that is encountered beforehand.
1961 * Returns true if the grace-period kthread needs to be awakened.
1963 static bool rcu_start_gp(struct rcu_state *rsp)
1965 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1966 struct rcu_node *rnp = rcu_get_root(rsp);
1970 * If there is no grace period in progress right now, any
1971 * callbacks we have up to this point will be satisfied by the
1972 * next grace period. Also, advancing the callbacks reduces the
1973 * probability of false positives from cpu_needs_another_gp()
1974 * resulting in pointless grace periods. So, advance callbacks
1975 * then start the grace period!
1977 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
1978 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
1983 * Report a full set of quiescent states to the specified rcu_state
1984 * data structure. This involves cleaning up after the prior grace
1985 * period and letting rcu_start_gp() start up the next grace period
1986 * if one is needed. Note that the caller must hold rnp->lock, which
1987 * is released before return.
1989 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1990 __releases(rcu_get_root(rsp)->lock)
1992 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1993 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1994 rcu_gp_kthread_wake(rsp);
1998 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1999 * Allows quiescent states for a group of CPUs to be reported at one go
2000 * to the specified rcu_node structure, though all the CPUs in the group
2001 * must be represented by the same rcu_node structure (which need not be
2002 * a leaf rcu_node structure, though it often will be). That structure's
2003 * lock must be held upon entry, and it is released before return.
2006 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2007 struct rcu_node *rnp, unsigned long flags)
2008 __releases(rnp->lock)
2010 struct rcu_node *rnp_c;
2012 /* Walk up the rcu_node hierarchy. */
2014 if (!(rnp->qsmask & mask)) {
2016 /* Our bit has already been cleared, so done. */
2017 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2020 rnp->qsmask &= ~mask;
2021 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2022 mask, rnp->qsmask, rnp->level,
2023 rnp->grplo, rnp->grphi,
2025 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2027 /* Other bits still set at this level, so done. */
2028 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2031 mask = rnp->grpmask;
2032 if (rnp->parent == NULL) {
2034 /* No more levels. Exit loop holding root lock. */
2038 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2041 raw_spin_lock_irqsave(&rnp->lock, flags);
2042 smp_mb__after_unlock_lock();
2043 WARN_ON_ONCE(rnp_c->qsmask);
2047 * Get here if we are the last CPU to pass through a quiescent
2048 * state for this grace period. Invoke rcu_report_qs_rsp()
2049 * to clean up and start the next grace period if one is needed.
2051 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2055 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2056 * structure. This must be either called from the specified CPU, or
2057 * called when the specified CPU is known to be offline (and when it is
2058 * also known that no other CPU is concurrently trying to help the offline
2059 * CPU). The lastcomp argument is used to make sure we are still in the
2060 * grace period of interest. We don't want to end the current grace period
2061 * based on quiescent states detected in an earlier grace period!
2064 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2066 unsigned long flags;
2069 struct rcu_node *rnp;
2072 raw_spin_lock_irqsave(&rnp->lock, flags);
2073 smp_mb__after_unlock_lock();
2074 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
2075 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2078 * The grace period in which this quiescent state was
2079 * recorded has ended, so don't report it upwards.
2080 * We will instead need a new quiescent state that lies
2081 * within the current grace period.
2083 rdp->passed_quiesce = 0; /* need qs for new gp. */
2084 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2087 mask = rdp->grpmask;
2088 if ((rnp->qsmask & mask) == 0) {
2089 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2091 rdp->qs_pending = 0;
2094 * This GP can't end until cpu checks in, so all of our
2095 * callbacks can be processed during the next GP.
2097 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2099 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
2101 rcu_gp_kthread_wake(rsp);
2106 * Check to see if there is a new grace period of which this CPU
2107 * is not yet aware, and if so, set up local rcu_data state for it.
2108 * Otherwise, see if this CPU has just passed through its first
2109 * quiescent state for this grace period, and record that fact if so.
2112 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2114 /* Check for grace-period ends and beginnings. */
2115 note_gp_changes(rsp, rdp);
2118 * Does this CPU still need to do its part for current grace period?
2119 * If no, return and let the other CPUs do their part as well.
2121 if (!rdp->qs_pending)
2125 * Was there a quiescent state since the beginning of the grace
2126 * period? If no, then exit and wait for the next call.
2128 if (!rdp->passed_quiesce)
2132 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2135 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2138 #ifdef CONFIG_HOTPLUG_CPU
2141 * Send the specified CPU's RCU callbacks to the orphanage. The
2142 * specified CPU must be offline, and the caller must hold the
2146 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2147 struct rcu_node *rnp, struct rcu_data *rdp)
2149 /* No-CBs CPUs do not have orphanable callbacks. */
2150 if (rcu_is_nocb_cpu(rdp->cpu))
2154 * Orphan the callbacks. First adjust the counts. This is safe
2155 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2156 * cannot be running now. Thus no memory barrier is required.
2158 if (rdp->nxtlist != NULL) {
2159 rsp->qlen_lazy += rdp->qlen_lazy;
2160 rsp->qlen += rdp->qlen;
2161 rdp->n_cbs_orphaned += rdp->qlen;
2163 ACCESS_ONCE(rdp->qlen) = 0;
2167 * Next, move those callbacks still needing a grace period to
2168 * the orphanage, where some other CPU will pick them up.
2169 * Some of the callbacks might have gone partway through a grace
2170 * period, but that is too bad. They get to start over because we
2171 * cannot assume that grace periods are synchronized across CPUs.
2172 * We don't bother updating the ->nxttail[] array yet, instead
2173 * we just reset the whole thing later on.
2175 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2176 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2177 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2178 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2182 * Then move the ready-to-invoke callbacks to the orphanage,
2183 * where some other CPU will pick them up. These will not be
2184 * required to pass though another grace period: They are done.
2186 if (rdp->nxtlist != NULL) {
2187 *rsp->orphan_donetail = rdp->nxtlist;
2188 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2191 /* Finally, initialize the rcu_data structure's list to empty. */
2192 init_callback_list(rdp);
2196 * Adopt the RCU callbacks from the specified rcu_state structure's
2197 * orphanage. The caller must hold the ->orphan_lock.
2199 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2202 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2204 /* No-CBs CPUs are handled specially. */
2205 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2208 /* Do the accounting first. */
2209 rdp->qlen_lazy += rsp->qlen_lazy;
2210 rdp->qlen += rsp->qlen;
2211 rdp->n_cbs_adopted += rsp->qlen;
2212 if (rsp->qlen_lazy != rsp->qlen)
2213 rcu_idle_count_callbacks_posted();
2218 * We do not need a memory barrier here because the only way we
2219 * can get here if there is an rcu_barrier() in flight is if
2220 * we are the task doing the rcu_barrier().
2223 /* First adopt the ready-to-invoke callbacks. */
2224 if (rsp->orphan_donelist != NULL) {
2225 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2226 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2227 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2228 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2229 rdp->nxttail[i] = rsp->orphan_donetail;
2230 rsp->orphan_donelist = NULL;
2231 rsp->orphan_donetail = &rsp->orphan_donelist;
2234 /* And then adopt the callbacks that still need a grace period. */
2235 if (rsp->orphan_nxtlist != NULL) {
2236 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2237 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2238 rsp->orphan_nxtlist = NULL;
2239 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2244 * Trace the fact that this CPU is going offline.
2246 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2248 RCU_TRACE(unsigned long mask);
2249 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2250 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2252 RCU_TRACE(mask = rdp->grpmask);
2253 trace_rcu_grace_period(rsp->name,
2254 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2259 * The CPU has been completely removed, and some other CPU is reporting
2260 * this fact from process context. Do the remainder of the cleanup,
2261 * including orphaning the outgoing CPU's RCU callbacks, and also
2262 * adopting them. There can only be one CPU hotplug operation at a time,
2263 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2265 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2267 unsigned long flags;
2269 int need_report = 0;
2270 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2271 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2273 /* Adjust any no-longer-needed kthreads. */
2274 rcu_boost_kthread_setaffinity(rnp, -1);
2276 /* Exclude any attempts to start a new grace period. */
2277 mutex_lock(&rsp->onoff_mutex);
2278 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2280 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2281 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2282 rcu_adopt_orphan_cbs(rsp, flags);
2284 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2285 mask = rdp->grpmask; /* rnp->grplo is constant. */
2287 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2288 smp_mb__after_unlock_lock();
2289 rnp->qsmaskinit &= ~mask;
2290 if (rnp->qsmaskinit != 0) {
2291 if (rnp != rdp->mynode)
2292 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2295 if (rnp == rdp->mynode)
2296 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2298 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2299 mask = rnp->grpmask;
2301 } while (rnp != NULL);
2304 * We still hold the leaf rcu_node structure lock here, and
2305 * irqs are still disabled. The reason for this subterfuge is
2306 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2307 * held leads to deadlock.
2309 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2311 if (need_report & RCU_OFL_TASKS_NORM_GP)
2312 rcu_report_unblock_qs_rnp(rnp, flags);
2314 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2315 if (need_report & RCU_OFL_TASKS_EXP_GP)
2316 rcu_report_exp_rnp(rsp, rnp, true);
2317 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2318 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2319 cpu, rdp->qlen, rdp->nxtlist);
2320 init_callback_list(rdp);
2321 /* Disallow further callbacks on this CPU. */
2322 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2323 mutex_unlock(&rsp->onoff_mutex);
2326 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2328 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2332 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2336 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2339 * Invoke any RCU callbacks that have made it to the end of their grace
2340 * period. Thottle as specified by rdp->blimit.
2342 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2344 unsigned long flags;
2345 struct rcu_head *next, *list, **tail;
2346 long bl, count, count_lazy;
2349 /* If no callbacks are ready, just return. */
2350 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2351 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2352 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2353 need_resched(), is_idle_task(current),
2354 rcu_is_callbacks_kthread());
2359 * Extract the list of ready callbacks, disabling to prevent
2360 * races with call_rcu() from interrupt handlers.
2362 local_irq_save(flags);
2363 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2365 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2366 list = rdp->nxtlist;
2367 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2368 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2369 tail = rdp->nxttail[RCU_DONE_TAIL];
2370 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2371 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2372 rdp->nxttail[i] = &rdp->nxtlist;
2373 local_irq_restore(flags);
2375 /* Invoke callbacks. */
2376 count = count_lazy = 0;
2380 debug_rcu_head_unqueue(list);
2381 if (__rcu_reclaim(rsp->name, list))
2384 /* Stop only if limit reached and CPU has something to do. */
2385 if (++count >= bl &&
2387 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2391 local_irq_save(flags);
2392 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2393 is_idle_task(current),
2394 rcu_is_callbacks_kthread());
2396 /* Update count, and requeue any remaining callbacks. */
2398 *tail = rdp->nxtlist;
2399 rdp->nxtlist = list;
2400 for (i = 0; i < RCU_NEXT_SIZE; i++)
2401 if (&rdp->nxtlist == rdp->nxttail[i])
2402 rdp->nxttail[i] = tail;
2406 smp_mb(); /* List handling before counting for rcu_barrier(). */
2407 rdp->qlen_lazy -= count_lazy;
2408 ACCESS_ONCE(rdp->qlen) = rdp->qlen - count;
2409 rdp->n_cbs_invoked += count;
2411 /* Reinstate batch limit if we have worked down the excess. */
2412 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2413 rdp->blimit = blimit;
2415 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2416 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2417 rdp->qlen_last_fqs_check = 0;
2418 rdp->n_force_qs_snap = rsp->n_force_qs;
2419 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2420 rdp->qlen_last_fqs_check = rdp->qlen;
2421 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2423 local_irq_restore(flags);
2425 /* Re-invoke RCU core processing if there are callbacks remaining. */
2426 if (cpu_has_callbacks_ready_to_invoke(rdp))
2431 * Check to see if this CPU is in a non-context-switch quiescent state
2432 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2433 * Also schedule RCU core processing.
2435 * This function must be called from hardirq context. It is normally
2436 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2437 * false, there is no point in invoking rcu_check_callbacks().
2439 void rcu_check_callbacks(int user)
2441 trace_rcu_utilization(TPS("Start scheduler-tick"));
2442 increment_cpu_stall_ticks();
2443 if (user || rcu_is_cpu_rrupt_from_idle()) {
2446 * Get here if this CPU took its interrupt from user
2447 * mode or from the idle loop, and if this is not a
2448 * nested interrupt. In this case, the CPU is in
2449 * a quiescent state, so note it.
2451 * No memory barrier is required here because both
2452 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2453 * variables that other CPUs neither access nor modify,
2454 * at least not while the corresponding CPU is online.
2460 } else if (!in_softirq()) {
2463 * Get here if this CPU did not take its interrupt from
2464 * softirq, in other words, if it is not interrupting
2465 * a rcu_bh read-side critical section. This is an _bh
2466 * critical section, so note it.
2471 rcu_preempt_check_callbacks();
2475 rcu_note_voluntary_context_switch(current);
2476 trace_rcu_utilization(TPS("End scheduler-tick"));
2480 * Scan the leaf rcu_node structures, processing dyntick state for any that
2481 * have not yet encountered a quiescent state, using the function specified.
2482 * Also initiate boosting for any threads blocked on the root rcu_node.
2484 * The caller must have suppressed start of new grace periods.
2486 static void force_qs_rnp(struct rcu_state *rsp,
2487 int (*f)(struct rcu_data *rsp, bool *isidle,
2488 unsigned long *maxj),
2489 bool *isidle, unsigned long *maxj)
2493 unsigned long flags;
2495 struct rcu_node *rnp;
2497 rcu_for_each_leaf_node(rsp, rnp) {
2498 cond_resched_rcu_qs();
2500 raw_spin_lock_irqsave(&rnp->lock, flags);
2501 smp_mb__after_unlock_lock();
2502 if (!rcu_gp_in_progress(rsp)) {
2503 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2506 if (rnp->qsmask == 0) {
2507 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2512 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2513 if ((rnp->qsmask & bit) != 0) {
2514 if ((rnp->qsmaskinit & bit) != 0)
2516 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2522 /* rcu_report_qs_rnp() releases rnp->lock. */
2523 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2526 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2528 rnp = rcu_get_root(rsp);
2529 if (rnp->qsmask == 0) {
2530 raw_spin_lock_irqsave(&rnp->lock, flags);
2531 smp_mb__after_unlock_lock();
2532 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2537 * Force quiescent states on reluctant CPUs, and also detect which
2538 * CPUs are in dyntick-idle mode.
2540 static void force_quiescent_state(struct rcu_state *rsp)
2542 unsigned long flags;
2544 struct rcu_node *rnp;
2545 struct rcu_node *rnp_old = NULL;
2547 /* Funnel through hierarchy to reduce memory contention. */
2548 rnp = __this_cpu_read(rsp->rda->mynode);
2549 for (; rnp != NULL; rnp = rnp->parent) {
2550 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2551 !raw_spin_trylock(&rnp->fqslock);
2552 if (rnp_old != NULL)
2553 raw_spin_unlock(&rnp_old->fqslock);
2555 rsp->n_force_qs_lh++;
2560 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2562 /* Reached the root of the rcu_node tree, acquire lock. */
2563 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2564 smp_mb__after_unlock_lock();
2565 raw_spin_unlock(&rnp_old->fqslock);
2566 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2567 rsp->n_force_qs_lh++;
2568 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2569 return; /* Someone beat us to it. */
2571 ACCESS_ONCE(rsp->gp_flags) =
2572 ACCESS_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS;
2573 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2574 rcu_gp_kthread_wake(rsp);
2578 * This does the RCU core processing work for the specified rcu_state
2579 * and rcu_data structures. This may be called only from the CPU to
2580 * whom the rdp belongs.
2583 __rcu_process_callbacks(struct rcu_state *rsp)
2585 unsigned long flags;
2587 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2589 WARN_ON_ONCE(rdp->beenonline == 0);
2591 /* Update RCU state based on any recent quiescent states. */
2592 rcu_check_quiescent_state(rsp, rdp);
2594 /* Does this CPU require a not-yet-started grace period? */
2595 local_irq_save(flags);
2596 if (cpu_needs_another_gp(rsp, rdp)) {
2597 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2598 needwake = rcu_start_gp(rsp);
2599 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2601 rcu_gp_kthread_wake(rsp);
2603 local_irq_restore(flags);
2606 /* If there are callbacks ready, invoke them. */
2607 if (cpu_has_callbacks_ready_to_invoke(rdp))
2608 invoke_rcu_callbacks(rsp, rdp);
2610 /* Do any needed deferred wakeups of rcuo kthreads. */
2611 do_nocb_deferred_wakeup(rdp);
2615 * Do RCU core processing for the current CPU.
2617 static void rcu_process_callbacks(struct softirq_action *unused)
2619 struct rcu_state *rsp;
2621 if (cpu_is_offline(smp_processor_id()))
2623 trace_rcu_utilization(TPS("Start RCU core"));
2624 for_each_rcu_flavor(rsp)
2625 __rcu_process_callbacks(rsp);
2626 trace_rcu_utilization(TPS("End RCU core"));
2630 * Schedule RCU callback invocation. If the specified type of RCU
2631 * does not support RCU priority boosting, just do a direct call,
2632 * otherwise wake up the per-CPU kernel kthread. Note that because we
2633 * are running on the current CPU with interrupts disabled, the
2634 * rcu_cpu_kthread_task cannot disappear out from under us.
2636 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2638 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2640 if (likely(!rsp->boost)) {
2641 rcu_do_batch(rsp, rdp);
2644 invoke_rcu_callbacks_kthread();
2647 static void invoke_rcu_core(void)
2649 if (cpu_online(smp_processor_id()))
2650 raise_softirq(RCU_SOFTIRQ);
2654 * Handle any core-RCU processing required by a call_rcu() invocation.
2656 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2657 struct rcu_head *head, unsigned long flags)
2662 * If called from an extended quiescent state, invoke the RCU
2663 * core in order to force a re-evaluation of RCU's idleness.
2665 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2668 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2669 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2673 * Force the grace period if too many callbacks or too long waiting.
2674 * Enforce hysteresis, and don't invoke force_quiescent_state()
2675 * if some other CPU has recently done so. Also, don't bother
2676 * invoking force_quiescent_state() if the newly enqueued callback
2677 * is the only one waiting for a grace period to complete.
2679 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2681 /* Are we ignoring a completed grace period? */
2682 note_gp_changes(rsp, rdp);
2684 /* Start a new grace period if one not already started. */
2685 if (!rcu_gp_in_progress(rsp)) {
2686 struct rcu_node *rnp_root = rcu_get_root(rsp);
2688 raw_spin_lock(&rnp_root->lock);
2689 smp_mb__after_unlock_lock();
2690 needwake = rcu_start_gp(rsp);
2691 raw_spin_unlock(&rnp_root->lock);
2693 rcu_gp_kthread_wake(rsp);
2695 /* Give the grace period a kick. */
2696 rdp->blimit = LONG_MAX;
2697 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2698 *rdp->nxttail[RCU_DONE_TAIL] != head)
2699 force_quiescent_state(rsp);
2700 rdp->n_force_qs_snap = rsp->n_force_qs;
2701 rdp->qlen_last_fqs_check = rdp->qlen;
2707 * RCU callback function to leak a callback.
2709 static void rcu_leak_callback(struct rcu_head *rhp)
2714 * Helper function for call_rcu() and friends. The cpu argument will
2715 * normally be -1, indicating "currently running CPU". It may specify
2716 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2717 * is expected to specify a CPU.
2720 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2721 struct rcu_state *rsp, int cpu, bool lazy)
2723 unsigned long flags;
2724 struct rcu_data *rdp;
2726 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
2727 if (debug_rcu_head_queue(head)) {
2728 /* Probable double call_rcu(), so leak the callback. */
2729 ACCESS_ONCE(head->func) = rcu_leak_callback;
2730 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2737 * Opportunistically note grace-period endings and beginnings.
2738 * Note that we might see a beginning right after we see an
2739 * end, but never vice versa, since this CPU has to pass through
2740 * a quiescent state betweentimes.
2742 local_irq_save(flags);
2743 rdp = this_cpu_ptr(rsp->rda);
2745 /* Add the callback to our list. */
2746 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2750 rdp = per_cpu_ptr(rsp->rda, cpu);
2751 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2752 WARN_ON_ONCE(offline);
2753 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2754 local_irq_restore(flags);
2757 ACCESS_ONCE(rdp->qlen) = rdp->qlen + 1;
2761 rcu_idle_count_callbacks_posted();
2762 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2763 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2764 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2766 if (__is_kfree_rcu_offset((unsigned long)func))
2767 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2768 rdp->qlen_lazy, rdp->qlen);
2770 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2772 /* Go handle any RCU core processing required. */
2773 __call_rcu_core(rsp, rdp, head, flags);
2774 local_irq_restore(flags);
2778 * Queue an RCU-sched callback for invocation after a grace period.
2780 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2782 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2784 EXPORT_SYMBOL_GPL(call_rcu_sched);
2787 * Queue an RCU callback for invocation after a quicker grace period.
2789 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2791 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2793 EXPORT_SYMBOL_GPL(call_rcu_bh);
2796 * Queue an RCU callback for lazy invocation after a grace period.
2797 * This will likely be later named something like "call_rcu_lazy()",
2798 * but this change will require some way of tagging the lazy RCU
2799 * callbacks in the list of pending callbacks. Until then, this
2800 * function may only be called from __kfree_rcu().
2802 void kfree_call_rcu(struct rcu_head *head,
2803 void (*func)(struct rcu_head *rcu))
2805 __call_rcu(head, func, rcu_state_p, -1, 1);
2807 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2810 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2811 * any blocking grace-period wait automatically implies a grace period
2812 * if there is only one CPU online at any point time during execution
2813 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2814 * occasionally incorrectly indicate that there are multiple CPUs online
2815 * when there was in fact only one the whole time, as this just adds
2816 * some overhead: RCU still operates correctly.
2818 static inline int rcu_blocking_is_gp(void)
2822 might_sleep(); /* Check for RCU read-side critical section. */
2824 ret = num_online_cpus() <= 1;
2830 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2832 * Control will return to the caller some time after a full rcu-sched
2833 * grace period has elapsed, in other words after all currently executing
2834 * rcu-sched read-side critical sections have completed. These read-side
2835 * critical sections are delimited by rcu_read_lock_sched() and
2836 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2837 * local_irq_disable(), and so on may be used in place of
2838 * rcu_read_lock_sched().
2840 * This means that all preempt_disable code sequences, including NMI and
2841 * non-threaded hardware-interrupt handlers, in progress on entry will
2842 * have completed before this primitive returns. However, this does not
2843 * guarantee that softirq handlers will have completed, since in some
2844 * kernels, these handlers can run in process context, and can block.
2846 * Note that this guarantee implies further memory-ordering guarantees.
2847 * On systems with more than one CPU, when synchronize_sched() returns,
2848 * each CPU is guaranteed to have executed a full memory barrier since the
2849 * end of its last RCU-sched read-side critical section whose beginning
2850 * preceded the call to synchronize_sched(). In addition, each CPU having
2851 * an RCU read-side critical section that extends beyond the return from
2852 * synchronize_sched() is guaranteed to have executed a full memory barrier
2853 * after the beginning of synchronize_sched() and before the beginning of
2854 * that RCU read-side critical section. Note that these guarantees include
2855 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2856 * that are executing in the kernel.
2858 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2859 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2860 * to have executed a full memory barrier during the execution of
2861 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2862 * again only if the system has more than one CPU).
2864 * This primitive provides the guarantees made by the (now removed)
2865 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2866 * guarantees that rcu_read_lock() sections will have completed.
2867 * In "classic RCU", these two guarantees happen to be one and
2868 * the same, but can differ in realtime RCU implementations.
2870 void synchronize_sched(void)
2872 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2873 !lock_is_held(&rcu_lock_map) &&
2874 !lock_is_held(&rcu_sched_lock_map),
2875 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2876 if (rcu_blocking_is_gp())
2879 synchronize_sched_expedited();
2881 wait_rcu_gp(call_rcu_sched);
2883 EXPORT_SYMBOL_GPL(synchronize_sched);
2886 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2888 * Control will return to the caller some time after a full rcu_bh grace
2889 * period has elapsed, in other words after all currently executing rcu_bh
2890 * read-side critical sections have completed. RCU read-side critical
2891 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2892 * and may be nested.
2894 * See the description of synchronize_sched() for more detailed information
2895 * on memory ordering guarantees.
2897 void synchronize_rcu_bh(void)
2899 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2900 !lock_is_held(&rcu_lock_map) &&
2901 !lock_is_held(&rcu_sched_lock_map),
2902 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2903 if (rcu_blocking_is_gp())
2906 synchronize_rcu_bh_expedited();
2908 wait_rcu_gp(call_rcu_bh);
2910 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2913 * get_state_synchronize_rcu - Snapshot current RCU state
2915 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2916 * to determine whether or not a full grace period has elapsed in the
2919 unsigned long get_state_synchronize_rcu(void)
2922 * Any prior manipulation of RCU-protected data must happen
2923 * before the load from ->gpnum.
2928 * Make sure this load happens before the purportedly
2929 * time-consuming work between get_state_synchronize_rcu()
2930 * and cond_synchronize_rcu().
2932 return smp_load_acquire(&rcu_state_p->gpnum);
2934 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2937 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2939 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2941 * If a full RCU grace period has elapsed since the earlier call to
2942 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2943 * synchronize_rcu() to wait for a full grace period.
2945 * Yes, this function does not take counter wrap into account. But
2946 * counter wrap is harmless. If the counter wraps, we have waited for
2947 * more than 2 billion grace periods (and way more on a 64-bit system!),
2948 * so waiting for one additional grace period should be just fine.
2950 void cond_synchronize_rcu(unsigned long oldstate)
2952 unsigned long newstate;
2955 * Ensure that this load happens before any RCU-destructive
2956 * actions the caller might carry out after we return.
2958 newstate = smp_load_acquire(&rcu_state_p->completed);
2959 if (ULONG_CMP_GE(oldstate, newstate))
2962 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2964 static int synchronize_sched_expedited_cpu_stop(void *data)
2967 * There must be a full memory barrier on each affected CPU
2968 * between the time that try_stop_cpus() is called and the
2969 * time that it returns.
2971 * In the current initial implementation of cpu_stop, the
2972 * above condition is already met when the control reaches
2973 * this point and the following smp_mb() is not strictly
2974 * necessary. Do smp_mb() anyway for documentation and
2975 * robustness against future implementation changes.
2977 smp_mb(); /* See above comment block. */
2982 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2984 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2985 * approach to force the grace period to end quickly. This consumes
2986 * significant time on all CPUs and is unfriendly to real-time workloads,
2987 * so is thus not recommended for any sort of common-case code. In fact,
2988 * if you are using synchronize_sched_expedited() in a loop, please
2989 * restructure your code to batch your updates, and then use a single
2990 * synchronize_sched() instead.
2992 * This implementation can be thought of as an application of ticket
2993 * locking to RCU, with sync_sched_expedited_started and
2994 * sync_sched_expedited_done taking on the roles of the halves
2995 * of the ticket-lock word. Each task atomically increments
2996 * sync_sched_expedited_started upon entry, snapshotting the old value,
2997 * then attempts to stop all the CPUs. If this succeeds, then each
2998 * CPU will have executed a context switch, resulting in an RCU-sched
2999 * grace period. We are then done, so we use atomic_cmpxchg() to
3000 * update sync_sched_expedited_done to match our snapshot -- but
3001 * only if someone else has not already advanced past our snapshot.
3003 * On the other hand, if try_stop_cpus() fails, we check the value
3004 * of sync_sched_expedited_done. If it has advanced past our
3005 * initial snapshot, then someone else must have forced a grace period
3006 * some time after we took our snapshot. In this case, our work is
3007 * done for us, and we can simply return. Otherwise, we try again,
3008 * but keep our initial snapshot for purposes of checking for someone
3009 * doing our work for us.
3011 * If we fail too many times in a row, we fall back to synchronize_sched().
3013 void synchronize_sched_expedited(void)
3018 long firstsnap, s, snap;
3020 struct rcu_state *rsp = &rcu_sched_state;
3023 * If we are in danger of counter wrap, just do synchronize_sched().
3024 * By allowing sync_sched_expedited_started to advance no more than
3025 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3026 * that more than 3.5 billion CPUs would be required to force a
3027 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3028 * course be required on a 64-bit system.
3030 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
3031 (ulong)atomic_long_read(&rsp->expedited_done) +
3033 synchronize_sched();
3034 atomic_long_inc(&rsp->expedited_wrap);
3039 * Take a ticket. Note that atomic_inc_return() implies a
3040 * full memory barrier.
3042 snap = atomic_long_inc_return(&rsp->expedited_start);
3044 if (!try_get_online_cpus()) {
3045 /* CPU hotplug operation in flight, fall back to normal GP. */
3046 wait_rcu_gp(call_rcu_sched);
3047 atomic_long_inc(&rsp->expedited_normal);
3050 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3052 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3053 cma = zalloc_cpumask_var(&cm, GFP_KERNEL);
3055 cpumask_copy(cm, cpu_online_mask);
3056 cpumask_clear_cpu(raw_smp_processor_id(), cm);
3057 for_each_cpu(cpu, cm) {
3058 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3060 if (!(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3061 cpumask_clear_cpu(cpu, cm);
3063 if (cpumask_weight(cm) == 0)
3068 * Each pass through the following loop attempts to force a
3069 * context switch on each CPU.
3071 while (try_stop_cpus(cma ? cm : cpu_online_mask,
3072 synchronize_sched_expedited_cpu_stop,
3075 atomic_long_inc(&rsp->expedited_tryfail);
3077 /* Check to see if someone else did our work for us. */
3078 s = atomic_long_read(&rsp->expedited_done);
3079 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3080 /* ensure test happens before caller kfree */
3081 smp_mb__before_atomic(); /* ^^^ */
3082 atomic_long_inc(&rsp->expedited_workdone1);
3083 free_cpumask_var(cm);
3087 /* No joy, try again later. Or just synchronize_sched(). */
3088 if (trycount++ < 10) {
3089 udelay(trycount * num_online_cpus());
3091 wait_rcu_gp(call_rcu_sched);
3092 atomic_long_inc(&rsp->expedited_normal);
3093 free_cpumask_var(cm);
3097 /* Recheck to see if someone else did our work for us. */
3098 s = atomic_long_read(&rsp->expedited_done);
3099 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3100 /* ensure test happens before caller kfree */
3101 smp_mb__before_atomic(); /* ^^^ */
3102 atomic_long_inc(&rsp->expedited_workdone2);
3103 free_cpumask_var(cm);
3108 * Refetching sync_sched_expedited_started allows later
3109 * callers to piggyback on our grace period. We retry
3110 * after they started, so our grace period works for them,
3111 * and they started after our first try, so their grace
3112 * period works for us.
3114 if (!try_get_online_cpus()) {
3115 /* CPU hotplug operation in flight, use normal GP. */
3116 wait_rcu_gp(call_rcu_sched);
3117 atomic_long_inc(&rsp->expedited_normal);
3118 free_cpumask_var(cm);
3121 snap = atomic_long_read(&rsp->expedited_start);
3122 smp_mb(); /* ensure read is before try_stop_cpus(). */
3124 atomic_long_inc(&rsp->expedited_stoppedcpus);
3127 free_cpumask_var(cm);
3130 * Everyone up to our most recent fetch is covered by our grace
3131 * period. Update the counter, but only if our work is still
3132 * relevant -- which it won't be if someone who started later
3133 * than we did already did their update.
3136 atomic_long_inc(&rsp->expedited_done_tries);
3137 s = atomic_long_read(&rsp->expedited_done);
3138 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
3139 /* ensure test happens before caller kfree */
3140 smp_mb__before_atomic(); /* ^^^ */
3141 atomic_long_inc(&rsp->expedited_done_lost);
3144 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
3145 atomic_long_inc(&rsp->expedited_done_exit);
3149 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3152 * Check to see if there is any immediate RCU-related work to be done
3153 * by the current CPU, for the specified type of RCU, returning 1 if so.
3154 * The checks are in order of increasing expense: checks that can be
3155 * carried out against CPU-local state are performed first. However,
3156 * we must check for CPU stalls first, else we might not get a chance.
3158 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3160 struct rcu_node *rnp = rdp->mynode;
3162 rdp->n_rcu_pending++;
3164 /* Check for CPU stalls, if enabled. */
3165 check_cpu_stall(rsp, rdp);
3167 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3168 if (rcu_nohz_full_cpu(rsp))
3171 /* Is the RCU core waiting for a quiescent state from this CPU? */
3172 if (rcu_scheduler_fully_active &&
3173 rdp->qs_pending && !rdp->passed_quiesce) {
3174 rdp->n_rp_qs_pending++;
3175 } else if (rdp->qs_pending && rdp->passed_quiesce) {
3176 rdp->n_rp_report_qs++;
3180 /* Does this CPU have callbacks ready to invoke? */
3181 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3182 rdp->n_rp_cb_ready++;
3186 /* Has RCU gone idle with this CPU needing another grace period? */
3187 if (cpu_needs_another_gp(rsp, rdp)) {
3188 rdp->n_rp_cpu_needs_gp++;
3192 /* Has another RCU grace period completed? */
3193 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3194 rdp->n_rp_gp_completed++;
3198 /* Has a new RCU grace period started? */
3199 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum ||
3200 unlikely(ACCESS_ONCE(rdp->gpwrap))) { /* outside lock */
3201 rdp->n_rp_gp_started++;
3205 /* Does this CPU need a deferred NOCB wakeup? */
3206 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3207 rdp->n_rp_nocb_defer_wakeup++;
3212 rdp->n_rp_need_nothing++;
3217 * Check to see if there is any immediate RCU-related work to be done
3218 * by the current CPU, returning 1 if so. This function is part of the
3219 * RCU implementation; it is -not- an exported member of the RCU API.
3221 static int rcu_pending(void)
3223 struct rcu_state *rsp;
3225 for_each_rcu_flavor(rsp)
3226 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3232 * Return true if the specified CPU has any callback. If all_lazy is
3233 * non-NULL, store an indication of whether all callbacks are lazy.
3234 * (If there are no callbacks, all of them are deemed to be lazy.)
3236 static int __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3240 struct rcu_data *rdp;
3241 struct rcu_state *rsp;
3243 for_each_rcu_flavor(rsp) {
3244 rdp = this_cpu_ptr(rsp->rda);
3248 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3259 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3260 * the compiler is expected to optimize this away.
3262 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3263 int cpu, unsigned long done)
3265 trace_rcu_barrier(rsp->name, s, cpu,
3266 atomic_read(&rsp->barrier_cpu_count), done);
3270 * RCU callback function for _rcu_barrier(). If we are last, wake
3271 * up the task executing _rcu_barrier().
3273 static void rcu_barrier_callback(struct rcu_head *rhp)
3275 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3276 struct rcu_state *rsp = rdp->rsp;
3278 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3279 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
3280 complete(&rsp->barrier_completion);
3282 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
3287 * Called with preemption disabled, and from cross-cpu IRQ context.
3289 static void rcu_barrier_func(void *type)
3291 struct rcu_state *rsp = type;
3292 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3294 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
3295 atomic_inc(&rsp->barrier_cpu_count);
3296 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3300 * Orchestrate the specified type of RCU barrier, waiting for all
3301 * RCU callbacks of the specified type to complete.
3303 static void _rcu_barrier(struct rcu_state *rsp)
3306 struct rcu_data *rdp;
3307 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3308 unsigned long snap_done;
3310 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3312 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3313 mutex_lock(&rsp->barrier_mutex);
3316 * Ensure that all prior references, including to ->n_barrier_done,
3317 * are ordered before the _rcu_barrier() machinery.
3319 smp_mb(); /* See above block comment. */
3322 * Recheck ->n_barrier_done to see if others did our work for us.
3323 * This means checking ->n_barrier_done for an even-to-odd-to-even
3324 * transition. The "if" expression below therefore rounds the old
3325 * value up to the next even number and adds two before comparing.
3327 snap_done = rsp->n_barrier_done;
3328 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3331 * If the value in snap is odd, we needed to wait for the current
3332 * rcu_barrier() to complete, then wait for the next one, in other
3333 * words, we need the value of snap_done to be three larger than
3334 * the value of snap. On the other hand, if the value in snap is
3335 * even, we only had to wait for the next rcu_barrier() to complete,
3336 * in other words, we need the value of snap_done to be only two
3337 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3338 * this for us (thank you, Linus!).
3340 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3341 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3342 smp_mb(); /* caller's subsequent code after above check. */
3343 mutex_unlock(&rsp->barrier_mutex);
3348 * Increment ->n_barrier_done to avoid duplicate work. Use
3349 * ACCESS_ONCE() to prevent the compiler from speculating
3350 * the increment to precede the early-exit check.
3352 ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1;
3353 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3354 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3355 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3358 * Initialize the count to one rather than to zero in order to
3359 * avoid a too-soon return to zero in case of a short grace period
3360 * (or preemption of this task). Exclude CPU-hotplug operations
3361 * to ensure that no offline CPU has callbacks queued.
3363 init_completion(&rsp->barrier_completion);
3364 atomic_set(&rsp->barrier_cpu_count, 1);
3368 * Force each CPU with callbacks to register a new callback.
3369 * When that callback is invoked, we will know that all of the
3370 * corresponding CPU's preceding callbacks have been invoked.
3372 for_each_possible_cpu(cpu) {
3373 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3375 rdp = per_cpu_ptr(rsp->rda, cpu);
3376 if (rcu_is_nocb_cpu(cpu)) {
3377 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3378 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3379 rsp->n_barrier_done);
3381 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3382 rsp->n_barrier_done);
3383 atomic_inc(&rsp->barrier_cpu_count);
3384 __call_rcu(&rdp->barrier_head,
3385 rcu_barrier_callback, rsp, cpu, 0);
3387 } else if (ACCESS_ONCE(rdp->qlen)) {
3388 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3389 rsp->n_barrier_done);
3390 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3392 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3393 rsp->n_barrier_done);
3399 * Now that we have an rcu_barrier_callback() callback on each
3400 * CPU, and thus each counted, remove the initial count.
3402 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3403 complete(&rsp->barrier_completion);
3405 /* Increment ->n_barrier_done to prevent duplicate work. */
3406 smp_mb(); /* Keep increment after above mechanism. */
3407 ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1;
3408 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3409 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3410 smp_mb(); /* Keep increment before caller's subsequent code. */
3412 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3413 wait_for_completion(&rsp->barrier_completion);
3415 /* Other rcu_barrier() invocations can now safely proceed. */
3416 mutex_unlock(&rsp->barrier_mutex);
3420 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3422 void rcu_barrier_bh(void)
3424 _rcu_barrier(&rcu_bh_state);
3426 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3429 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3431 void rcu_barrier_sched(void)
3433 _rcu_barrier(&rcu_sched_state);
3435 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3438 * Do boot-time initialization of a CPU's per-CPU RCU data.
3441 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3443 unsigned long flags;
3444 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3445 struct rcu_node *rnp = rcu_get_root(rsp);
3447 /* Set up local state, ensuring consistent view of global state. */
3448 raw_spin_lock_irqsave(&rnp->lock, flags);
3449 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3450 init_callback_list(rdp);
3452 ACCESS_ONCE(rdp->qlen) = 0;
3453 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3454 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3455 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3458 rcu_boot_init_nocb_percpu_data(rdp);
3459 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3463 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3464 * offline event can be happening at a given time. Note also that we
3465 * can accept some slop in the rsp->completed access due to the fact
3466 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3469 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3471 unsigned long flags;
3473 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3474 struct rcu_node *rnp = rcu_get_root(rsp);
3476 /* Exclude new grace periods. */
3477 mutex_lock(&rsp->onoff_mutex);
3479 /* Set up local state, ensuring consistent view of global state. */
3480 raw_spin_lock_irqsave(&rnp->lock, flags);
3481 rdp->beenonline = 1; /* We have now been online. */
3482 rdp->qlen_last_fqs_check = 0;
3483 rdp->n_force_qs_snap = rsp->n_force_qs;
3484 rdp->blimit = blimit;
3485 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3486 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3487 rcu_sysidle_init_percpu_data(rdp->dynticks);
3488 atomic_set(&rdp->dynticks->dynticks,
3489 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3490 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3492 /* Add CPU to rcu_node bitmasks. */
3494 mask = rdp->grpmask;
3496 /* Exclude any attempts to start a new GP on small systems. */
3497 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3498 rnp->qsmaskinit |= mask;
3499 mask = rnp->grpmask;
3500 if (rnp == rdp->mynode) {
3502 * If there is a grace period in progress, we will
3503 * set up to wait for it next time we run the
3506 rdp->gpnum = rnp->completed;
3507 rdp->completed = rnp->completed;
3508 rdp->passed_quiesce = 0;
3509 rdp->qs_pending = 0;
3510 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3512 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3514 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3515 local_irq_restore(flags);
3517 mutex_unlock(&rsp->onoff_mutex);
3520 static void rcu_prepare_cpu(int cpu)
3522 struct rcu_state *rsp;
3524 for_each_rcu_flavor(rsp)
3525 rcu_init_percpu_data(cpu, rsp);
3529 * Handle CPU online/offline notification events.
3531 static int rcu_cpu_notify(struct notifier_block *self,
3532 unsigned long action, void *hcpu)
3534 long cpu = (long)hcpu;
3535 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3536 struct rcu_node *rnp = rdp->mynode;
3537 struct rcu_state *rsp;
3539 trace_rcu_utilization(TPS("Start CPU hotplug"));
3541 case CPU_UP_PREPARE:
3542 case CPU_UP_PREPARE_FROZEN:
3543 rcu_prepare_cpu(cpu);
3544 rcu_prepare_kthreads(cpu);
3545 rcu_spawn_all_nocb_kthreads(cpu);
3548 case CPU_DOWN_FAILED:
3549 rcu_boost_kthread_setaffinity(rnp, -1);
3551 case CPU_DOWN_PREPARE:
3552 rcu_boost_kthread_setaffinity(rnp, cpu);
3555 case CPU_DYING_FROZEN:
3556 for_each_rcu_flavor(rsp)
3557 rcu_cleanup_dying_cpu(rsp);
3560 case CPU_DEAD_FROZEN:
3561 case CPU_UP_CANCELED:
3562 case CPU_UP_CANCELED_FROZEN:
3563 for_each_rcu_flavor(rsp) {
3564 rcu_cleanup_dead_cpu(cpu, rsp);
3565 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3571 trace_rcu_utilization(TPS("End CPU hotplug"));
3575 static int rcu_pm_notify(struct notifier_block *self,
3576 unsigned long action, void *hcpu)
3579 case PM_HIBERNATION_PREPARE:
3580 case PM_SUSPEND_PREPARE:
3581 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3584 case PM_POST_HIBERNATION:
3585 case PM_POST_SUSPEND:
3595 * Spawn the kthreads that handle each RCU flavor's grace periods.
3597 static int __init rcu_spawn_gp_kthread(void)
3599 unsigned long flags;
3600 struct rcu_node *rnp;
3601 struct rcu_state *rsp;
3602 struct task_struct *t;
3604 rcu_scheduler_fully_active = 1;
3605 for_each_rcu_flavor(rsp) {
3606 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3608 rnp = rcu_get_root(rsp);
3609 raw_spin_lock_irqsave(&rnp->lock, flags);
3610 rsp->gp_kthread = t;
3611 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3613 rcu_spawn_nocb_kthreads();
3614 rcu_spawn_boost_kthreads();
3617 early_initcall(rcu_spawn_gp_kthread);
3620 * This function is invoked towards the end of the scheduler's initialization
3621 * process. Before this is called, the idle task might contain
3622 * RCU read-side critical sections (during which time, this idle
3623 * task is booting the system). After this function is called, the
3624 * idle tasks are prohibited from containing RCU read-side critical
3625 * sections. This function also enables RCU lockdep checking.
3627 void rcu_scheduler_starting(void)
3629 WARN_ON(num_online_cpus() != 1);
3630 WARN_ON(nr_context_switches() > 0);
3631 rcu_scheduler_active = 1;
3635 * Compute the per-level fanout, either using the exact fanout specified
3636 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3638 #ifdef CONFIG_RCU_FANOUT_EXACT
3639 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3643 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3644 for (i = rcu_num_lvls - 2; i >= 0; i--)
3645 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3647 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3648 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3655 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3656 ccur = rsp->levelcnt[i];
3657 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3661 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3664 * Helper function for rcu_init() that initializes one rcu_state structure.
3666 static void __init rcu_init_one(struct rcu_state *rsp,
3667 struct rcu_data __percpu *rda)
3669 static const char * const buf[] = {
3673 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3674 static const char * const fqs[] = {
3678 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3679 static u8 fl_mask = 0x1;
3683 struct rcu_node *rnp;
3685 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3687 /* Silence gcc 4.8 warning about array index out of range. */
3688 if (rcu_num_lvls > RCU_NUM_LVLS)
3689 panic("rcu_init_one: rcu_num_lvls overflow");
3691 /* Initialize the level-tracking arrays. */
3693 for (i = 0; i < rcu_num_lvls; i++)
3694 rsp->levelcnt[i] = num_rcu_lvl[i];
3695 for (i = 1; i < rcu_num_lvls; i++)
3696 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3697 rcu_init_levelspread(rsp);
3698 rsp->flavor_mask = fl_mask;
3701 /* Initialize the elements themselves, starting from the leaves. */
3703 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3704 cpustride *= rsp->levelspread[i];
3705 rnp = rsp->level[i];
3706 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3707 raw_spin_lock_init(&rnp->lock);
3708 lockdep_set_class_and_name(&rnp->lock,
3709 &rcu_node_class[i], buf[i]);
3710 raw_spin_lock_init(&rnp->fqslock);
3711 lockdep_set_class_and_name(&rnp->fqslock,
3712 &rcu_fqs_class[i], fqs[i]);
3713 rnp->gpnum = rsp->gpnum;
3714 rnp->completed = rsp->completed;
3716 rnp->qsmaskinit = 0;
3717 rnp->grplo = j * cpustride;
3718 rnp->grphi = (j + 1) * cpustride - 1;
3719 if (rnp->grphi >= nr_cpu_ids)
3720 rnp->grphi = nr_cpu_ids - 1;
3726 rnp->grpnum = j % rsp->levelspread[i - 1];
3727 rnp->grpmask = 1UL << rnp->grpnum;
3728 rnp->parent = rsp->level[i - 1] +
3729 j / rsp->levelspread[i - 1];
3732 INIT_LIST_HEAD(&rnp->blkd_tasks);
3733 rcu_init_one_nocb(rnp);
3738 init_waitqueue_head(&rsp->gp_wq);
3739 rnp = rsp->level[rcu_num_lvls - 1];
3740 for_each_possible_cpu(i) {
3741 while (i > rnp->grphi)
3743 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3744 rcu_boot_init_percpu_data(i, rsp);
3746 list_add(&rsp->flavors, &rcu_struct_flavors);
3750 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3751 * replace the definitions in tree.h because those are needed to size
3752 * the ->node array in the rcu_state structure.
3754 static void __init rcu_init_geometry(void)
3760 int rcu_capacity[MAX_RCU_LVLS + 1];
3763 * Initialize any unspecified boot parameters.
3764 * The default values of jiffies_till_first_fqs and
3765 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3766 * value, which is a function of HZ, then adding one for each
3767 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3769 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3770 if (jiffies_till_first_fqs == ULONG_MAX)
3771 jiffies_till_first_fqs = d;
3772 if (jiffies_till_next_fqs == ULONG_MAX)
3773 jiffies_till_next_fqs = d;
3775 /* If the compile-time values are accurate, just leave. */
3776 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3777 nr_cpu_ids == NR_CPUS)
3779 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3780 rcu_fanout_leaf, nr_cpu_ids);
3783 * Compute number of nodes that can be handled an rcu_node tree
3784 * with the given number of levels. Setting rcu_capacity[0] makes
3785 * some of the arithmetic easier.
3787 rcu_capacity[0] = 1;
3788 rcu_capacity[1] = rcu_fanout_leaf;
3789 for (i = 2; i <= MAX_RCU_LVLS; i++)
3790 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3793 * The boot-time rcu_fanout_leaf parameter is only permitted
3794 * to increase the leaf-level fanout, not decrease it. Of course,
3795 * the leaf-level fanout cannot exceed the number of bits in
3796 * the rcu_node masks. Finally, the tree must be able to accommodate
3797 * the configured number of CPUs. Complain and fall back to the
3798 * compile-time values if these limits are exceeded.
3800 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3801 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3802 n > rcu_capacity[MAX_RCU_LVLS]) {
3807 /* Calculate the number of rcu_nodes at each level of the tree. */
3808 for (i = 1; i <= MAX_RCU_LVLS; i++)
3809 if (n <= rcu_capacity[i]) {
3810 for (j = 0; j <= i; j++)
3812 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3814 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3819 /* Calculate the total number of rcu_node structures. */
3821 for (i = 0; i <= MAX_RCU_LVLS; i++)
3822 rcu_num_nodes += num_rcu_lvl[i];
3826 void __init rcu_init(void)
3830 rcu_bootup_announce();
3831 rcu_init_geometry();
3832 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3833 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3834 __rcu_init_preempt();
3835 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3838 * We don't need protection against CPU-hotplug here because
3839 * this is called early in boot, before either interrupts
3840 * or the scheduler are operational.
3842 cpu_notifier(rcu_cpu_notify, 0);
3843 pm_notifier(rcu_pm_notify, 0);
3844 for_each_online_cpu(cpu)
3845 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3847 rcu_early_boot_tests();
3850 #include "tree_plugin.h"