2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/suspend.h>
60 #include <trace/events/rcu.h>
65 * Strings used in tracepoints need to be exported via the
66 * tracing system such that tools like perf and trace-cmd can
67 * translate the string address pointers to actual text.
69 #define TPS(x) tracepoint_string(x)
71 /* Data structures. */
73 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
74 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
77 * In order to export the rcu_state name to the tracing tools, it
78 * needs to be added in the __tracepoint_string section.
79 * This requires defining a separate variable tp_<sname>_varname
80 * that points to the string being used, and this will allow
81 * the tracing userspace tools to be able to decipher the string
82 * address to the matching string.
84 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
87 struct rcu_state sname##_state = { \
88 .level = { &sname##_state.node[0] }, \
90 .fqs_state = RCU_GP_IDLE, \
91 .gpnum = 0UL - 300UL, \
92 .completed = 0UL - 300UL, \
93 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
94 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
95 .orphan_donetail = &sname##_state.orphan_donelist, \
96 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
97 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
98 .name = sname##_varname, \
101 DEFINE_PER_CPU(struct rcu_data, sname##_data)
103 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
104 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
106 static struct rcu_state *rcu_state;
107 LIST_HEAD(rcu_struct_flavors);
109 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
110 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
111 module_param(rcu_fanout_leaf, int, 0444);
112 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
113 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
120 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
123 * The rcu_scheduler_active variable transitions from zero to one just
124 * before the first task is spawned. So when this variable is zero, RCU
125 * can assume that there is but one task, allowing RCU to (for example)
126 * optimize synchronize_sched() to a simple barrier(). When this variable
127 * is one, RCU must actually do all the hard work required to detect real
128 * grace periods. This variable is also used to suppress boot-time false
129 * positives from lockdep-RCU error checking.
131 int rcu_scheduler_active __read_mostly;
132 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
135 * The rcu_scheduler_fully_active variable transitions from zero to one
136 * during the early_initcall() processing, which is after the scheduler
137 * is capable of creating new tasks. So RCU processing (for example,
138 * creating tasks for RCU priority boosting) must be delayed until after
139 * rcu_scheduler_fully_active transitions from zero to one. We also
140 * currently delay invocation of any RCU callbacks until after this point.
142 * It might later prove better for people registering RCU callbacks during
143 * early boot to take responsibility for these callbacks, but one step at
146 static int rcu_scheduler_fully_active __read_mostly;
148 #ifdef CONFIG_RCU_BOOST
151 * Control variables for per-CPU and per-rcu_node kthreads. These
152 * handle all flavors of RCU.
154 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
155 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
156 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
157 DEFINE_PER_CPU(char, rcu_cpu_has_work);
159 #endif /* #ifdef CONFIG_RCU_BOOST */
161 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
162 static void invoke_rcu_core(void);
163 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
166 * Track the rcutorture test sequence number and the update version
167 * number within a given test. The rcutorture_testseq is incremented
168 * on every rcutorture module load and unload, so has an odd value
169 * when a test is running. The rcutorture_vernum is set to zero
170 * when rcutorture starts and is incremented on each rcutorture update.
171 * These variables enable correlating rcutorture output with the
172 * RCU tracing information.
174 unsigned long rcutorture_testseq;
175 unsigned long rcutorture_vernum;
178 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
179 * permit this function to be invoked without holding the root rcu_node
180 * structure's ->lock, but of course results can be subject to change.
182 static int rcu_gp_in_progress(struct rcu_state *rsp)
184 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
188 * Note a quiescent state. Because we do not need to know
189 * how many quiescent states passed, just if there was at least
190 * one since the start of the grace period, this just sets a flag.
191 * The caller must have disabled preemption.
193 void rcu_sched_qs(int cpu)
195 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
197 if (rdp->passed_quiesce == 0)
198 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
199 rdp->passed_quiesce = 1;
202 void rcu_bh_qs(int cpu)
204 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
206 if (rdp->passed_quiesce == 0)
207 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
208 rdp->passed_quiesce = 1;
212 * Note a context switch. This is a quiescent state for RCU-sched,
213 * and requires special handling for preemptible RCU.
214 * The caller must have disabled preemption.
216 void rcu_note_context_switch(int cpu)
218 trace_rcu_utilization(TPS("Start context switch"));
220 rcu_preempt_note_context_switch(cpu);
221 trace_rcu_utilization(TPS("End context switch"));
223 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
225 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
226 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
227 .dynticks = ATOMIC_INIT(1),
230 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
231 static long qhimark = 10000; /* If this many pending, ignore blimit. */
232 static long qlowmark = 100; /* Once only this many pending, use blimit. */
234 module_param(blimit, long, 0444);
235 module_param(qhimark, long, 0444);
236 module_param(qlowmark, long, 0444);
238 static ulong jiffies_till_first_fqs = ULONG_MAX;
239 static ulong jiffies_till_next_fqs = ULONG_MAX;
241 module_param(jiffies_till_first_fqs, ulong, 0644);
242 module_param(jiffies_till_next_fqs, ulong, 0644);
244 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
245 struct rcu_data *rdp);
246 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
247 static void force_quiescent_state(struct rcu_state *rsp);
248 static int rcu_pending(int cpu);
251 * Return the number of RCU-sched batches processed thus far for debug & stats.
253 long rcu_batches_completed_sched(void)
255 return rcu_sched_state.completed;
257 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
260 * Return the number of RCU BH batches processed thus far for debug & stats.
262 long rcu_batches_completed_bh(void)
264 return rcu_bh_state.completed;
266 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
269 * Force a quiescent state for RCU BH.
271 void rcu_bh_force_quiescent_state(void)
273 force_quiescent_state(&rcu_bh_state);
275 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
278 * Record the number of times rcutorture tests have been initiated and
279 * terminated. This information allows the debugfs tracing stats to be
280 * correlated to the rcutorture messages, even when the rcutorture module
281 * is being repeatedly loaded and unloaded. In other words, we cannot
282 * store this state in rcutorture itself.
284 void rcutorture_record_test_transition(void)
286 rcutorture_testseq++;
287 rcutorture_vernum = 0;
289 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
292 * Record the number of writer passes through the current rcutorture test.
293 * This is also used to correlate debugfs tracing stats with the rcutorture
296 void rcutorture_record_progress(unsigned long vernum)
300 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
303 * Force a quiescent state for RCU-sched.
305 void rcu_sched_force_quiescent_state(void)
307 force_quiescent_state(&rcu_sched_state);
309 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
312 * Does the CPU have callbacks ready to be invoked?
315 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
317 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
318 rdp->nxttail[RCU_DONE_TAIL] != NULL;
322 * Does the current CPU require a not-yet-started grace period?
323 * The caller must have disabled interrupts to prevent races with
324 * normal callback registry.
327 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
331 if (rcu_gp_in_progress(rsp))
332 return 0; /* No, a grace period is already in progress. */
333 if (rcu_nocb_needs_gp(rsp))
334 return 1; /* Yes, a no-CBs CPU needs one. */
335 if (!rdp->nxttail[RCU_NEXT_TAIL])
336 return 0; /* No, this is a no-CBs (or offline) CPU. */
337 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
338 return 1; /* Yes, this CPU has newly registered callbacks. */
339 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
340 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
341 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
342 rdp->nxtcompleted[i]))
343 return 1; /* Yes, CBs for future grace period. */
344 return 0; /* No grace period needed. */
348 * Return the root node of the specified rcu_state structure.
350 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
352 return &rsp->node[0];
356 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
358 * If the new value of the ->dynticks_nesting counter now is zero,
359 * we really have entered idle, and must do the appropriate accounting.
360 * The caller must have disabled interrupts.
362 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
365 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
366 if (!user && !is_idle_task(current)) {
367 struct task_struct *idle = idle_task(smp_processor_id());
369 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
370 ftrace_dump(DUMP_ORIG);
371 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
372 current->pid, current->comm,
373 idle->pid, idle->comm); /* must be idle task! */
375 rcu_prepare_for_idle(smp_processor_id());
376 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
377 smp_mb__before_atomic_inc(); /* See above. */
378 atomic_inc(&rdtp->dynticks);
379 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
380 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
383 * It is illegal to enter an extended quiescent state while
384 * in an RCU read-side critical section.
386 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
387 "Illegal idle entry in RCU read-side critical section.");
388 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
389 "Illegal idle entry in RCU-bh read-side critical section.");
390 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
391 "Illegal idle entry in RCU-sched read-side critical section.");
395 * Enter an RCU extended quiescent state, which can be either the
396 * idle loop or adaptive-tickless usermode execution.
398 static void rcu_eqs_enter(bool user)
401 struct rcu_dynticks *rdtp;
403 rdtp = &__get_cpu_var(rcu_dynticks);
404 oldval = rdtp->dynticks_nesting;
405 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
406 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
407 rdtp->dynticks_nesting = 0;
409 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
410 rcu_eqs_enter_common(rdtp, oldval, user);
414 * rcu_idle_enter - inform RCU that current CPU is entering idle
416 * Enter idle mode, in other words, -leave- the mode in which RCU
417 * read-side critical sections can occur. (Though RCU read-side
418 * critical sections can occur in irq handlers in idle, a possibility
419 * handled by irq_enter() and irq_exit().)
421 * We crowbar the ->dynticks_nesting field to zero to allow for
422 * the possibility of usermode upcalls having messed up our count
423 * of interrupt nesting level during the prior busy period.
425 void rcu_idle_enter(void)
429 local_irq_save(flags);
430 rcu_eqs_enter(false);
431 local_irq_restore(flags);
433 EXPORT_SYMBOL_GPL(rcu_idle_enter);
435 #ifdef CONFIG_RCU_USER_QS
437 * rcu_user_enter - inform RCU that we are resuming userspace.
439 * Enter RCU idle mode right before resuming userspace. No use of RCU
440 * is permitted between this call and rcu_user_exit(). This way the
441 * CPU doesn't need to maintain the tick for RCU maintenance purposes
442 * when the CPU runs in userspace.
444 void rcu_user_enter(void)
450 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
451 * after the current irq returns.
453 * This is similar to rcu_user_enter() but in the context of a non-nesting
454 * irq. After this call, RCU enters into idle mode when the interrupt
457 void rcu_user_enter_after_irq(void)
460 struct rcu_dynticks *rdtp;
462 local_irq_save(flags);
463 rdtp = &__get_cpu_var(rcu_dynticks);
464 /* Ensure this irq is interrupting a non-idle RCU state. */
465 WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
466 rdtp->dynticks_nesting = 1;
467 local_irq_restore(flags);
469 #endif /* CONFIG_RCU_USER_QS */
472 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
474 * Exit from an interrupt handler, which might possibly result in entering
475 * idle mode, in other words, leaving the mode in which read-side critical
476 * sections can occur.
478 * This code assumes that the idle loop never does anything that might
479 * result in unbalanced calls to irq_enter() and irq_exit(). If your
480 * architecture violates this assumption, RCU will give you what you
481 * deserve, good and hard. But very infrequently and irreproducibly.
483 * Use things like work queues to work around this limitation.
485 * You have been warned.
487 void rcu_irq_exit(void)
491 struct rcu_dynticks *rdtp;
493 local_irq_save(flags);
494 rdtp = &__get_cpu_var(rcu_dynticks);
495 oldval = rdtp->dynticks_nesting;
496 rdtp->dynticks_nesting--;
497 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
498 if (rdtp->dynticks_nesting)
499 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
501 rcu_eqs_enter_common(rdtp, oldval, true);
502 local_irq_restore(flags);
506 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
508 * If the new value of the ->dynticks_nesting counter was previously zero,
509 * we really have exited idle, and must do the appropriate accounting.
510 * The caller must have disabled interrupts.
512 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
515 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
516 atomic_inc(&rdtp->dynticks);
517 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
518 smp_mb__after_atomic_inc(); /* See above. */
519 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
520 rcu_cleanup_after_idle(smp_processor_id());
521 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
522 if (!user && !is_idle_task(current)) {
523 struct task_struct *idle = idle_task(smp_processor_id());
525 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
526 oldval, rdtp->dynticks_nesting);
527 ftrace_dump(DUMP_ORIG);
528 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
529 current->pid, current->comm,
530 idle->pid, idle->comm); /* must be idle task! */
535 * Exit an RCU extended quiescent state, which can be either the
536 * idle loop or adaptive-tickless usermode execution.
538 static void rcu_eqs_exit(bool user)
540 struct rcu_dynticks *rdtp;
543 rdtp = &__get_cpu_var(rcu_dynticks);
544 oldval = rdtp->dynticks_nesting;
545 WARN_ON_ONCE(oldval < 0);
546 if (oldval & DYNTICK_TASK_NEST_MASK)
547 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
549 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
550 rcu_eqs_exit_common(rdtp, oldval, user);
554 * rcu_idle_exit - inform RCU that current CPU is leaving idle
556 * Exit idle mode, in other words, -enter- the mode in which RCU
557 * read-side critical sections can occur.
559 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
560 * allow for the possibility of usermode upcalls messing up our count
561 * of interrupt nesting level during the busy period that is just
564 void rcu_idle_exit(void)
568 local_irq_save(flags);
570 local_irq_restore(flags);
572 EXPORT_SYMBOL_GPL(rcu_idle_exit);
574 #ifdef CONFIG_RCU_USER_QS
576 * rcu_user_exit - inform RCU that we are exiting userspace.
578 * Exit RCU idle mode while entering the kernel because it can
579 * run a RCU read side critical section anytime.
581 void rcu_user_exit(void)
587 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
588 * idle mode after the current non-nesting irq returns.
590 * This is similar to rcu_user_exit() but in the context of an irq.
591 * This is called when the irq has interrupted a userspace RCU idle mode
592 * context. When the current non-nesting interrupt returns after this call,
593 * the CPU won't restore the RCU idle mode.
595 void rcu_user_exit_after_irq(void)
598 struct rcu_dynticks *rdtp;
600 local_irq_save(flags);
601 rdtp = &__get_cpu_var(rcu_dynticks);
602 /* Ensure we are interrupting an RCU idle mode. */
603 WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
604 rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
605 local_irq_restore(flags);
607 #endif /* CONFIG_RCU_USER_QS */
610 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
612 * Enter an interrupt handler, which might possibly result in exiting
613 * idle mode, in other words, entering the mode in which read-side critical
614 * sections can occur.
616 * Note that the Linux kernel is fully capable of entering an interrupt
617 * handler that it never exits, for example when doing upcalls to
618 * user mode! This code assumes that the idle loop never does upcalls to
619 * user mode. If your architecture does do upcalls from the idle loop (or
620 * does anything else that results in unbalanced calls to the irq_enter()
621 * and irq_exit() functions), RCU will give you what you deserve, good
622 * and hard. But very infrequently and irreproducibly.
624 * Use things like work queues to work around this limitation.
626 * You have been warned.
628 void rcu_irq_enter(void)
631 struct rcu_dynticks *rdtp;
634 local_irq_save(flags);
635 rdtp = &__get_cpu_var(rcu_dynticks);
636 oldval = rdtp->dynticks_nesting;
637 rdtp->dynticks_nesting++;
638 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
640 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
642 rcu_eqs_exit_common(rdtp, oldval, true);
643 local_irq_restore(flags);
647 * rcu_nmi_enter - inform RCU of entry to NMI context
649 * If the CPU was idle with dynamic ticks active, and there is no
650 * irq handler running, this updates rdtp->dynticks_nmi to let the
651 * RCU grace-period handling know that the CPU is active.
653 void rcu_nmi_enter(void)
655 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
657 if (rdtp->dynticks_nmi_nesting == 0 &&
658 (atomic_read(&rdtp->dynticks) & 0x1))
660 rdtp->dynticks_nmi_nesting++;
661 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
662 atomic_inc(&rdtp->dynticks);
663 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
664 smp_mb__after_atomic_inc(); /* See above. */
665 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
669 * rcu_nmi_exit - inform RCU of exit from NMI context
671 * If the CPU was idle with dynamic ticks active, and there is no
672 * irq handler running, this updates rdtp->dynticks_nmi to let the
673 * RCU grace-period handling know that the CPU is no longer active.
675 void rcu_nmi_exit(void)
677 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
679 if (rdtp->dynticks_nmi_nesting == 0 ||
680 --rdtp->dynticks_nmi_nesting != 0)
682 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
683 smp_mb__before_atomic_inc(); /* See above. */
684 atomic_inc(&rdtp->dynticks);
685 smp_mb__after_atomic_inc(); /* Force delay to next write. */
686 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
690 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
692 * If the current CPU is in its idle loop and is neither in an interrupt
693 * or NMI handler, return true.
695 int rcu_is_cpu_idle(void)
700 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
704 EXPORT_SYMBOL(rcu_is_cpu_idle);
706 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
709 * Is the current CPU online? Disable preemption to avoid false positives
710 * that could otherwise happen due to the current CPU number being sampled,
711 * this task being preempted, its old CPU being taken offline, resuming
712 * on some other CPU, then determining that its old CPU is now offline.
713 * It is OK to use RCU on an offline processor during initial boot, hence
714 * the check for rcu_scheduler_fully_active. Note also that it is OK
715 * for a CPU coming online to use RCU for one jiffy prior to marking itself
716 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
717 * offline to continue to use RCU for one jiffy after marking itself
718 * offline in the cpu_online_mask. This leniency is necessary given the
719 * non-atomic nature of the online and offline processing, for example,
720 * the fact that a CPU enters the scheduler after completing the CPU_DYING
723 * This is also why RCU internally marks CPUs online during the
724 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
726 * Disable checking if in an NMI handler because we cannot safely report
727 * errors from NMI handlers anyway.
729 bool rcu_lockdep_current_cpu_online(void)
731 struct rcu_data *rdp;
732 struct rcu_node *rnp;
738 rdp = &__get_cpu_var(rcu_sched_data);
740 ret = (rdp->grpmask & rnp->qsmaskinit) ||
741 !rcu_scheduler_fully_active;
745 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
747 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
750 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
752 * If the current CPU is idle or running at a first-level (not nested)
753 * interrupt from idle, return true. The caller must have at least
754 * disabled preemption.
756 static int rcu_is_cpu_rrupt_from_idle(void)
758 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
762 * Snapshot the specified CPU's dynticks counter so that we can later
763 * credit them with an implicit quiescent state. Return 1 if this CPU
764 * is in dynticks idle mode, which is an extended quiescent state.
766 static int dyntick_save_progress_counter(struct rcu_data *rdp)
768 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
769 return (rdp->dynticks_snap & 0x1) == 0;
773 * Return true if the specified CPU has passed through a quiescent
774 * state by virtue of being in or having passed through an dynticks
775 * idle state since the last call to dyntick_save_progress_counter()
776 * for this same CPU, or by virtue of having been offline.
778 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
783 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
784 snap = (unsigned int)rdp->dynticks_snap;
787 * If the CPU passed through or entered a dynticks idle phase with
788 * no active irq/NMI handlers, then we can safely pretend that the CPU
789 * already acknowledged the request to pass through a quiescent
790 * state. Either way, that CPU cannot possibly be in an RCU
791 * read-side critical section that started before the beginning
792 * of the current RCU grace period.
794 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
795 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
801 * Check for the CPU being offline, but only if the grace period
802 * is old enough. We don't need to worry about the CPU changing
803 * state: If we see it offline even once, it has been through a
806 * The reason for insisting that the grace period be at least
807 * one jiffy old is that CPUs that are not quite online and that
808 * have just gone offline can still execute RCU read-side critical
811 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
812 return 0; /* Grace period is not old enough. */
814 if (cpu_is_offline(rdp->cpu)) {
815 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
821 * There is a possibility that a CPU in adaptive-ticks state
822 * might run in the kernel with the scheduling-clock tick disabled
823 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
824 * force the CPU to restart the scheduling-clock tick in this
825 * CPU is in this state.
827 rcu_kick_nohz_cpu(rdp->cpu);
832 static void record_gp_stall_check_time(struct rcu_state *rsp)
834 rsp->gp_start = jiffies;
835 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
839 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
840 * for architectures that do not implement trigger_all_cpu_backtrace().
841 * The NMI-triggered stack traces are more accurate because they are
842 * printed by the target CPU.
844 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
848 struct rcu_node *rnp;
850 rcu_for_each_leaf_node(rsp, rnp) {
851 raw_spin_lock_irqsave(&rnp->lock, flags);
852 if (rnp->qsmask != 0) {
853 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
854 if (rnp->qsmask & (1UL << cpu))
855 dump_cpu_task(rnp->grplo + cpu);
857 raw_spin_unlock_irqrestore(&rnp->lock, flags);
861 static void print_other_cpu_stall(struct rcu_state *rsp)
867 struct rcu_node *rnp = rcu_get_root(rsp);
870 /* Only let one CPU complain about others per time interval. */
872 raw_spin_lock_irqsave(&rnp->lock, flags);
873 delta = jiffies - rsp->jiffies_stall;
874 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
875 raw_spin_unlock_irqrestore(&rnp->lock, flags);
878 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
879 raw_spin_unlock_irqrestore(&rnp->lock, flags);
882 * OK, time to rat on our buddy...
883 * See Documentation/RCU/stallwarn.txt for info on how to debug
884 * RCU CPU stall warnings.
886 pr_err("INFO: %s detected stalls on CPUs/tasks:",
888 print_cpu_stall_info_begin();
889 rcu_for_each_leaf_node(rsp, rnp) {
890 raw_spin_lock_irqsave(&rnp->lock, flags);
891 ndetected += rcu_print_task_stall(rnp);
892 if (rnp->qsmask != 0) {
893 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
894 if (rnp->qsmask & (1UL << cpu)) {
895 print_cpu_stall_info(rsp,
900 raw_spin_unlock_irqrestore(&rnp->lock, flags);
904 * Now rat on any tasks that got kicked up to the root rcu_node
905 * due to CPU offlining.
907 rnp = rcu_get_root(rsp);
908 raw_spin_lock_irqsave(&rnp->lock, flags);
909 ndetected += rcu_print_task_stall(rnp);
910 raw_spin_unlock_irqrestore(&rnp->lock, flags);
912 print_cpu_stall_info_end();
913 for_each_possible_cpu(cpu)
914 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
915 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
916 smp_processor_id(), (long)(jiffies - rsp->gp_start),
917 rsp->gpnum, rsp->completed, totqlen);
919 pr_err("INFO: Stall ended before state dump start\n");
920 else if (!trigger_all_cpu_backtrace())
921 rcu_dump_cpu_stacks(rsp);
923 /* Complain about tasks blocking the grace period. */
925 rcu_print_detail_task_stall(rsp);
927 force_quiescent_state(rsp); /* Kick them all. */
930 static void print_cpu_stall(struct rcu_state *rsp)
934 struct rcu_node *rnp = rcu_get_root(rsp);
938 * OK, time to rat on ourselves...
939 * See Documentation/RCU/stallwarn.txt for info on how to debug
940 * RCU CPU stall warnings.
942 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
943 print_cpu_stall_info_begin();
944 print_cpu_stall_info(rsp, smp_processor_id());
945 print_cpu_stall_info_end();
946 for_each_possible_cpu(cpu)
947 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
948 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
949 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
950 if (!trigger_all_cpu_backtrace())
953 raw_spin_lock_irqsave(&rnp->lock, flags);
954 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
955 rsp->jiffies_stall = jiffies +
956 3 * rcu_jiffies_till_stall_check() + 3;
957 raw_spin_unlock_irqrestore(&rnp->lock, flags);
959 set_need_resched(); /* kick ourselves to get things going. */
962 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
966 struct rcu_node *rnp;
968 if (rcu_cpu_stall_suppress)
970 j = ACCESS_ONCE(jiffies);
971 js = ACCESS_ONCE(rsp->jiffies_stall);
973 if (rcu_gp_in_progress(rsp) &&
974 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
976 /* We haven't checked in, so go dump stack. */
977 print_cpu_stall(rsp);
979 } else if (rcu_gp_in_progress(rsp) &&
980 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
982 /* They had a few time units to dump stack, so complain. */
983 print_other_cpu_stall(rsp);
988 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
990 * Set the stall-warning timeout way off into the future, thus preventing
991 * any RCU CPU stall-warning messages from appearing in the current set of
994 * The caller must disable hard irqs.
996 void rcu_cpu_stall_reset(void)
998 struct rcu_state *rsp;
1000 for_each_rcu_flavor(rsp)
1001 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1005 * Initialize the specified rcu_data structure's callback list to empty.
1007 static void init_callback_list(struct rcu_data *rdp)
1011 if (init_nocb_callback_list(rdp))
1013 rdp->nxtlist = NULL;
1014 for (i = 0; i < RCU_NEXT_SIZE; i++)
1015 rdp->nxttail[i] = &rdp->nxtlist;
1019 * Determine the value that ->completed will have at the end of the
1020 * next subsequent grace period. This is used to tag callbacks so that
1021 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1022 * been dyntick-idle for an extended period with callbacks under the
1023 * influence of RCU_FAST_NO_HZ.
1025 * The caller must hold rnp->lock with interrupts disabled.
1027 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1028 struct rcu_node *rnp)
1031 * If RCU is idle, we just wait for the next grace period.
1032 * But we can only be sure that RCU is idle if we are looking
1033 * at the root rcu_node structure -- otherwise, a new grace
1034 * period might have started, but just not yet gotten around
1035 * to initializing the current non-root rcu_node structure.
1037 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1038 return rnp->completed + 1;
1041 * Otherwise, wait for a possible partial grace period and
1042 * then the subsequent full grace period.
1044 return rnp->completed + 2;
1048 * Trace-event helper function for rcu_start_future_gp() and
1049 * rcu_nocb_wait_gp().
1051 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1052 unsigned long c, const char *s)
1054 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1055 rnp->completed, c, rnp->level,
1056 rnp->grplo, rnp->grphi, s);
1060 * Start some future grace period, as needed to handle newly arrived
1061 * callbacks. The required future grace periods are recorded in each
1062 * rcu_node structure's ->need_future_gp field.
1064 * The caller must hold the specified rcu_node structure's ->lock.
1066 static unsigned long __maybe_unused
1067 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1071 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1074 * Pick up grace-period number for new callbacks. If this
1075 * grace period is already marked as needed, return to the caller.
1077 c = rcu_cbs_completed(rdp->rsp, rnp);
1078 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1079 if (rnp->need_future_gp[c & 0x1]) {
1080 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1085 * If either this rcu_node structure or the root rcu_node structure
1086 * believe that a grace period is in progress, then we must wait
1087 * for the one following, which is in "c". Because our request
1088 * will be noticed at the end of the current grace period, we don't
1089 * need to explicitly start one.
1091 if (rnp->gpnum != rnp->completed ||
1092 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1093 rnp->need_future_gp[c & 0x1]++;
1094 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1099 * There might be no grace period in progress. If we don't already
1100 * hold it, acquire the root rcu_node structure's lock in order to
1101 * start one (if needed).
1103 if (rnp != rnp_root)
1104 raw_spin_lock(&rnp_root->lock);
1107 * Get a new grace-period number. If there really is no grace
1108 * period in progress, it will be smaller than the one we obtained
1109 * earlier. Adjust callbacks as needed. Note that even no-CBs
1110 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1112 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1113 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1114 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1115 rdp->nxtcompleted[i] = c;
1118 * If the needed for the required grace period is already
1119 * recorded, trace and leave.
1121 if (rnp_root->need_future_gp[c & 0x1]) {
1122 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1126 /* Record the need for the future grace period. */
1127 rnp_root->need_future_gp[c & 0x1]++;
1129 /* If a grace period is not already in progress, start one. */
1130 if (rnp_root->gpnum != rnp_root->completed) {
1131 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1133 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1134 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1137 if (rnp != rnp_root)
1138 raw_spin_unlock(&rnp_root->lock);
1143 * Clean up any old requests for the just-ended grace period. Also return
1144 * whether any additional grace periods have been requested. Also invoke
1145 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1146 * waiting for this grace period to complete.
1148 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1150 int c = rnp->completed;
1152 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1154 rcu_nocb_gp_cleanup(rsp, rnp);
1155 rnp->need_future_gp[c & 0x1] = 0;
1156 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1157 trace_rcu_future_gp(rnp, rdp, c,
1158 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1163 * If there is room, assign a ->completed number to any callbacks on
1164 * this CPU that have not already been assigned. Also accelerate any
1165 * callbacks that were previously assigned a ->completed number that has
1166 * since proven to be too conservative, which can happen if callbacks get
1167 * assigned a ->completed number while RCU is idle, but with reference to
1168 * a non-root rcu_node structure. This function is idempotent, so it does
1169 * not hurt to call it repeatedly.
1171 * The caller must hold rnp->lock with interrupts disabled.
1173 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1174 struct rcu_data *rdp)
1179 /* If the CPU has no callbacks, nothing to do. */
1180 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1184 * Starting from the sublist containing the callbacks most
1185 * recently assigned a ->completed number and working down, find the
1186 * first sublist that is not assignable to an upcoming grace period.
1187 * Such a sublist has something in it (first two tests) and has
1188 * a ->completed number assigned that will complete sooner than
1189 * the ->completed number for newly arrived callbacks (last test).
1191 * The key point is that any later sublist can be assigned the
1192 * same ->completed number as the newly arrived callbacks, which
1193 * means that the callbacks in any of these later sublist can be
1194 * grouped into a single sublist, whether or not they have already
1195 * been assigned a ->completed number.
1197 c = rcu_cbs_completed(rsp, rnp);
1198 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1199 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1200 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1204 * If there are no sublist for unassigned callbacks, leave.
1205 * At the same time, advance "i" one sublist, so that "i" will
1206 * index into the sublist where all the remaining callbacks should
1209 if (++i >= RCU_NEXT_TAIL)
1213 * Assign all subsequent callbacks' ->completed number to the next
1214 * full grace period and group them all in the sublist initially
1217 for (; i <= RCU_NEXT_TAIL; i++) {
1218 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1219 rdp->nxtcompleted[i] = c;
1221 /* Record any needed additional grace periods. */
1222 rcu_start_future_gp(rnp, rdp);
1224 /* Trace depending on how much we were able to accelerate. */
1225 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1226 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1228 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1232 * Move any callbacks whose grace period has completed to the
1233 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1234 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1235 * sublist. This function is idempotent, so it does not hurt to
1236 * invoke it repeatedly. As long as it is not invoked -too- often...
1238 * The caller must hold rnp->lock with interrupts disabled.
1240 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1241 struct rcu_data *rdp)
1245 /* If the CPU has no callbacks, nothing to do. */
1246 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1250 * Find all callbacks whose ->completed numbers indicate that they
1251 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1253 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1254 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1256 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1258 /* Clean up any sublist tail pointers that were misordered above. */
1259 for (j = RCU_WAIT_TAIL; j < i; j++)
1260 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1262 /* Copy down callbacks to fill in empty sublists. */
1263 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1264 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1266 rdp->nxttail[j] = rdp->nxttail[i];
1267 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1270 /* Classify any remaining callbacks. */
1271 rcu_accelerate_cbs(rsp, rnp, rdp);
1275 * Update CPU-local rcu_data state to record the beginnings and ends of
1276 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1277 * structure corresponding to the current CPU, and must have irqs disabled.
1279 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1281 /* Handle the ends of any preceding grace periods first. */
1282 if (rdp->completed == rnp->completed) {
1284 /* No grace period end, so just accelerate recent callbacks. */
1285 rcu_accelerate_cbs(rsp, rnp, rdp);
1289 /* Advance callbacks. */
1290 rcu_advance_cbs(rsp, rnp, rdp);
1292 /* Remember that we saw this grace-period completion. */
1293 rdp->completed = rnp->completed;
1294 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1297 if (rdp->gpnum != rnp->gpnum) {
1299 * If the current grace period is waiting for this CPU,
1300 * set up to detect a quiescent state, otherwise don't
1301 * go looking for one.
1303 rdp->gpnum = rnp->gpnum;
1304 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1305 rdp->passed_quiesce = 0;
1306 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1307 zero_cpu_stall_ticks(rdp);
1311 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1313 unsigned long flags;
1314 struct rcu_node *rnp;
1316 local_irq_save(flags);
1318 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1319 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1320 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1321 local_irq_restore(flags);
1324 __note_gp_changes(rsp, rnp, rdp);
1325 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1329 * Initialize a new grace period.
1331 static int rcu_gp_init(struct rcu_state *rsp)
1333 struct rcu_data *rdp;
1334 struct rcu_node *rnp = rcu_get_root(rsp);
1336 raw_spin_lock_irq(&rnp->lock);
1337 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1339 if (rcu_gp_in_progress(rsp)) {
1340 /* Grace period already in progress, don't start another. */
1341 raw_spin_unlock_irq(&rnp->lock);
1345 /* Advance to a new grace period and initialize state. */
1347 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1348 record_gp_stall_check_time(rsp);
1349 raw_spin_unlock_irq(&rnp->lock);
1351 /* Exclude any concurrent CPU-hotplug operations. */
1352 mutex_lock(&rsp->onoff_mutex);
1355 * Set the quiescent-state-needed bits in all the rcu_node
1356 * structures for all currently online CPUs in breadth-first order,
1357 * starting from the root rcu_node structure, relying on the layout
1358 * of the tree within the rsp->node[] array. Note that other CPUs
1359 * will access only the leaves of the hierarchy, thus seeing that no
1360 * grace period is in progress, at least until the corresponding
1361 * leaf node has been initialized. In addition, we have excluded
1362 * CPU-hotplug operations.
1364 * The grace period cannot complete until the initialization
1365 * process finishes, because this kthread handles both.
1367 rcu_for_each_node_breadth_first(rsp, rnp) {
1368 raw_spin_lock_irq(&rnp->lock);
1369 rdp = this_cpu_ptr(rsp->rda);
1370 rcu_preempt_check_blocked_tasks(rnp);
1371 rnp->qsmask = rnp->qsmaskinit;
1372 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1373 WARN_ON_ONCE(rnp->completed != rsp->completed);
1374 ACCESS_ONCE(rnp->completed) = rsp->completed;
1375 if (rnp == rdp->mynode)
1376 __note_gp_changes(rsp, rnp, rdp);
1377 rcu_preempt_boost_start_gp(rnp);
1378 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1379 rnp->level, rnp->grplo,
1380 rnp->grphi, rnp->qsmask);
1381 raw_spin_unlock_irq(&rnp->lock);
1382 #ifdef CONFIG_PROVE_RCU_DELAY
1383 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1384 system_state == SYSTEM_RUNNING)
1386 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1390 mutex_unlock(&rsp->onoff_mutex);
1395 * Do one round of quiescent-state forcing.
1397 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1399 int fqs_state = fqs_state_in;
1400 struct rcu_node *rnp = rcu_get_root(rsp);
1403 if (fqs_state == RCU_SAVE_DYNTICK) {
1404 /* Collect dyntick-idle snapshots. */
1405 force_qs_rnp(rsp, dyntick_save_progress_counter);
1406 fqs_state = RCU_FORCE_QS;
1408 /* Handle dyntick-idle and offline CPUs. */
1409 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1411 /* Clear flag to prevent immediate re-entry. */
1412 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1413 raw_spin_lock_irq(&rnp->lock);
1414 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1415 raw_spin_unlock_irq(&rnp->lock);
1421 * Clean up after the old grace period.
1423 static void rcu_gp_cleanup(struct rcu_state *rsp)
1425 unsigned long gp_duration;
1427 struct rcu_data *rdp;
1428 struct rcu_node *rnp = rcu_get_root(rsp);
1430 raw_spin_lock_irq(&rnp->lock);
1431 gp_duration = jiffies - rsp->gp_start;
1432 if (gp_duration > rsp->gp_max)
1433 rsp->gp_max = gp_duration;
1436 * We know the grace period is complete, but to everyone else
1437 * it appears to still be ongoing. But it is also the case
1438 * that to everyone else it looks like there is nothing that
1439 * they can do to advance the grace period. It is therefore
1440 * safe for us to drop the lock in order to mark the grace
1441 * period as completed in all of the rcu_node structures.
1443 raw_spin_unlock_irq(&rnp->lock);
1446 * Propagate new ->completed value to rcu_node structures so
1447 * that other CPUs don't have to wait until the start of the next
1448 * grace period to process their callbacks. This also avoids
1449 * some nasty RCU grace-period initialization races by forcing
1450 * the end of the current grace period to be completely recorded in
1451 * all of the rcu_node structures before the beginning of the next
1452 * grace period is recorded in any of the rcu_node structures.
1454 rcu_for_each_node_breadth_first(rsp, rnp) {
1455 raw_spin_lock_irq(&rnp->lock);
1456 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1457 rdp = this_cpu_ptr(rsp->rda);
1458 if (rnp == rdp->mynode)
1459 __note_gp_changes(rsp, rnp, rdp);
1460 nocb += rcu_future_gp_cleanup(rsp, rnp);
1461 raw_spin_unlock_irq(&rnp->lock);
1464 rnp = rcu_get_root(rsp);
1465 raw_spin_lock_irq(&rnp->lock);
1466 rcu_nocb_gp_set(rnp, nocb);
1468 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1469 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1470 rsp->fqs_state = RCU_GP_IDLE;
1471 rdp = this_cpu_ptr(rsp->rda);
1472 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1473 if (cpu_needs_another_gp(rsp, rdp))
1475 raw_spin_unlock_irq(&rnp->lock);
1479 * Body of kthread that handles grace periods.
1481 static int __noreturn rcu_gp_kthread(void *arg)
1486 struct rcu_state *rsp = arg;
1487 struct rcu_node *rnp = rcu_get_root(rsp);
1491 /* Handle grace-period start. */
1493 wait_event_interruptible(rsp->gp_wq,
1496 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1500 flush_signals(current);
1503 /* Handle quiescent-state forcing. */
1504 fqs_state = RCU_SAVE_DYNTICK;
1505 j = jiffies_till_first_fqs;
1508 jiffies_till_first_fqs = HZ;
1511 rsp->jiffies_force_qs = jiffies + j;
1512 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1513 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1514 (!ACCESS_ONCE(rnp->qsmask) &&
1515 !rcu_preempt_blocked_readers_cgp(rnp)),
1517 /* If grace period done, leave loop. */
1518 if (!ACCESS_ONCE(rnp->qsmask) &&
1519 !rcu_preempt_blocked_readers_cgp(rnp))
1521 /* If time for quiescent-state forcing, do it. */
1522 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1523 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1526 /* Deal with stray signal. */
1528 flush_signals(current);
1530 j = jiffies_till_next_fqs;
1533 jiffies_till_next_fqs = HZ;
1536 jiffies_till_next_fqs = 1;
1540 /* Handle grace-period end. */
1541 rcu_gp_cleanup(rsp);
1545 static void rsp_wakeup(struct irq_work *work)
1547 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1549 /* Wake up rcu_gp_kthread() to start the grace period. */
1550 wake_up(&rsp->gp_wq);
1554 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1555 * in preparation for detecting the next grace period. The caller must hold
1556 * the root node's ->lock and hard irqs must be disabled.
1558 * Note that it is legal for a dying CPU (which is marked as offline) to
1559 * invoke this function. This can happen when the dying CPU reports its
1563 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1564 struct rcu_data *rdp)
1566 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1568 * Either we have not yet spawned the grace-period
1569 * task, this CPU does not need another grace period,
1570 * or a grace period is already in progress.
1571 * Either way, don't start a new grace period.
1575 rsp->gp_flags = RCU_GP_FLAG_INIT;
1578 * We can't do wakeups while holding the rnp->lock, as that
1579 * could cause possible deadlocks with the rq->lock. Defer
1580 * the wakeup to interrupt context. And don't bother waking
1581 * up the running kthread.
1583 if (current != rsp->gp_kthread)
1584 irq_work_queue(&rsp->wakeup_work);
1588 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1589 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1590 * is invoked indirectly from rcu_advance_cbs(), which would result in
1591 * endless recursion -- or would do so if it wasn't for the self-deadlock
1592 * that is encountered beforehand.
1595 rcu_start_gp(struct rcu_state *rsp)
1597 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1598 struct rcu_node *rnp = rcu_get_root(rsp);
1601 * If there is no grace period in progress right now, any
1602 * callbacks we have up to this point will be satisfied by the
1603 * next grace period. Also, advancing the callbacks reduces the
1604 * probability of false positives from cpu_needs_another_gp()
1605 * resulting in pointless grace periods. So, advance callbacks
1606 * then start the grace period!
1608 rcu_advance_cbs(rsp, rnp, rdp);
1609 rcu_start_gp_advanced(rsp, rnp, rdp);
1613 * Report a full set of quiescent states to the specified rcu_state
1614 * data structure. This involves cleaning up after the prior grace
1615 * period and letting rcu_start_gp() start up the next grace period
1616 * if one is needed. Note that the caller must hold rnp->lock, which
1617 * is released before return.
1619 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1620 __releases(rcu_get_root(rsp)->lock)
1622 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1623 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1624 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1628 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1629 * Allows quiescent states for a group of CPUs to be reported at one go
1630 * to the specified rcu_node structure, though all the CPUs in the group
1631 * must be represented by the same rcu_node structure (which need not be
1632 * a leaf rcu_node structure, though it often will be). That structure's
1633 * lock must be held upon entry, and it is released before return.
1636 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1637 struct rcu_node *rnp, unsigned long flags)
1638 __releases(rnp->lock)
1640 struct rcu_node *rnp_c;
1642 /* Walk up the rcu_node hierarchy. */
1644 if (!(rnp->qsmask & mask)) {
1646 /* Our bit has already been cleared, so done. */
1647 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1650 rnp->qsmask &= ~mask;
1651 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1652 mask, rnp->qsmask, rnp->level,
1653 rnp->grplo, rnp->grphi,
1655 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1657 /* Other bits still set at this level, so done. */
1658 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1661 mask = rnp->grpmask;
1662 if (rnp->parent == NULL) {
1664 /* No more levels. Exit loop holding root lock. */
1668 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1671 raw_spin_lock_irqsave(&rnp->lock, flags);
1672 WARN_ON_ONCE(rnp_c->qsmask);
1676 * Get here if we are the last CPU to pass through a quiescent
1677 * state for this grace period. Invoke rcu_report_qs_rsp()
1678 * to clean up and start the next grace period if one is needed.
1680 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1684 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1685 * structure. This must be either called from the specified CPU, or
1686 * called when the specified CPU is known to be offline (and when it is
1687 * also known that no other CPU is concurrently trying to help the offline
1688 * CPU). The lastcomp argument is used to make sure we are still in the
1689 * grace period of interest. We don't want to end the current grace period
1690 * based on quiescent states detected in an earlier grace period!
1693 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1695 unsigned long flags;
1697 struct rcu_node *rnp;
1700 raw_spin_lock_irqsave(&rnp->lock, flags);
1701 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1702 rnp->completed == rnp->gpnum) {
1705 * The grace period in which this quiescent state was
1706 * recorded has ended, so don't report it upwards.
1707 * We will instead need a new quiescent state that lies
1708 * within the current grace period.
1710 rdp->passed_quiesce = 0; /* need qs for new gp. */
1711 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1714 mask = rdp->grpmask;
1715 if ((rnp->qsmask & mask) == 0) {
1716 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1718 rdp->qs_pending = 0;
1721 * This GP can't end until cpu checks in, so all of our
1722 * callbacks can be processed during the next GP.
1724 rcu_accelerate_cbs(rsp, rnp, rdp);
1726 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1731 * Check to see if there is a new grace period of which this CPU
1732 * is not yet aware, and if so, set up local rcu_data state for it.
1733 * Otherwise, see if this CPU has just passed through its first
1734 * quiescent state for this grace period, and record that fact if so.
1737 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1739 /* Check for grace-period ends and beginnings. */
1740 note_gp_changes(rsp, rdp);
1743 * Does this CPU still need to do its part for current grace period?
1744 * If no, return and let the other CPUs do their part as well.
1746 if (!rdp->qs_pending)
1750 * Was there a quiescent state since the beginning of the grace
1751 * period? If no, then exit and wait for the next call.
1753 if (!rdp->passed_quiesce)
1757 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1760 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1763 #ifdef CONFIG_HOTPLUG_CPU
1766 * Send the specified CPU's RCU callbacks to the orphanage. The
1767 * specified CPU must be offline, and the caller must hold the
1771 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1772 struct rcu_node *rnp, struct rcu_data *rdp)
1774 /* No-CBs CPUs do not have orphanable callbacks. */
1775 if (rcu_is_nocb_cpu(rdp->cpu))
1779 * Orphan the callbacks. First adjust the counts. This is safe
1780 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1781 * cannot be running now. Thus no memory barrier is required.
1783 if (rdp->nxtlist != NULL) {
1784 rsp->qlen_lazy += rdp->qlen_lazy;
1785 rsp->qlen += rdp->qlen;
1786 rdp->n_cbs_orphaned += rdp->qlen;
1788 ACCESS_ONCE(rdp->qlen) = 0;
1792 * Next, move those callbacks still needing a grace period to
1793 * the orphanage, where some other CPU will pick them up.
1794 * Some of the callbacks might have gone partway through a grace
1795 * period, but that is too bad. They get to start over because we
1796 * cannot assume that grace periods are synchronized across CPUs.
1797 * We don't bother updating the ->nxttail[] array yet, instead
1798 * we just reset the whole thing later on.
1800 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1801 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1802 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1803 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1807 * Then move the ready-to-invoke callbacks to the orphanage,
1808 * where some other CPU will pick them up. These will not be
1809 * required to pass though another grace period: They are done.
1811 if (rdp->nxtlist != NULL) {
1812 *rsp->orphan_donetail = rdp->nxtlist;
1813 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1816 /* Finally, initialize the rcu_data structure's list to empty. */
1817 init_callback_list(rdp);
1821 * Adopt the RCU callbacks from the specified rcu_state structure's
1822 * orphanage. The caller must hold the ->orphan_lock.
1824 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1827 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1829 /* No-CBs CPUs are handled specially. */
1830 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1833 /* Do the accounting first. */
1834 rdp->qlen_lazy += rsp->qlen_lazy;
1835 rdp->qlen += rsp->qlen;
1836 rdp->n_cbs_adopted += rsp->qlen;
1837 if (rsp->qlen_lazy != rsp->qlen)
1838 rcu_idle_count_callbacks_posted();
1843 * We do not need a memory barrier here because the only way we
1844 * can get here if there is an rcu_barrier() in flight is if
1845 * we are the task doing the rcu_barrier().
1848 /* First adopt the ready-to-invoke callbacks. */
1849 if (rsp->orphan_donelist != NULL) {
1850 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1851 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1852 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1853 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1854 rdp->nxttail[i] = rsp->orphan_donetail;
1855 rsp->orphan_donelist = NULL;
1856 rsp->orphan_donetail = &rsp->orphan_donelist;
1859 /* And then adopt the callbacks that still need a grace period. */
1860 if (rsp->orphan_nxtlist != NULL) {
1861 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1862 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1863 rsp->orphan_nxtlist = NULL;
1864 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1869 * Trace the fact that this CPU is going offline.
1871 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1873 RCU_TRACE(unsigned long mask);
1874 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1875 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1877 RCU_TRACE(mask = rdp->grpmask);
1878 trace_rcu_grace_period(rsp->name,
1879 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1884 * The CPU has been completely removed, and some other CPU is reporting
1885 * this fact from process context. Do the remainder of the cleanup,
1886 * including orphaning the outgoing CPU's RCU callbacks, and also
1887 * adopting them. There can only be one CPU hotplug operation at a time,
1888 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1890 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1892 unsigned long flags;
1894 int need_report = 0;
1895 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1896 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1898 /* Adjust any no-longer-needed kthreads. */
1899 rcu_boost_kthread_setaffinity(rnp, -1);
1901 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1903 /* Exclude any attempts to start a new grace period. */
1904 mutex_lock(&rsp->onoff_mutex);
1905 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1907 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1908 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1909 rcu_adopt_orphan_cbs(rsp);
1911 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1912 mask = rdp->grpmask; /* rnp->grplo is constant. */
1914 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1915 rnp->qsmaskinit &= ~mask;
1916 if (rnp->qsmaskinit != 0) {
1917 if (rnp != rdp->mynode)
1918 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1921 if (rnp == rdp->mynode)
1922 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1924 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1925 mask = rnp->grpmask;
1927 } while (rnp != NULL);
1930 * We still hold the leaf rcu_node structure lock here, and
1931 * irqs are still disabled. The reason for this subterfuge is
1932 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1933 * held leads to deadlock.
1935 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1937 if (need_report & RCU_OFL_TASKS_NORM_GP)
1938 rcu_report_unblock_qs_rnp(rnp, flags);
1940 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1941 if (need_report & RCU_OFL_TASKS_EXP_GP)
1942 rcu_report_exp_rnp(rsp, rnp, true);
1943 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1944 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1945 cpu, rdp->qlen, rdp->nxtlist);
1946 init_callback_list(rdp);
1947 /* Disallow further callbacks on this CPU. */
1948 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1949 mutex_unlock(&rsp->onoff_mutex);
1952 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1954 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1958 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1962 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1965 * Invoke any RCU callbacks that have made it to the end of their grace
1966 * period. Thottle as specified by rdp->blimit.
1968 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1970 unsigned long flags;
1971 struct rcu_head *next, *list, **tail;
1972 long bl, count, count_lazy;
1975 /* If no callbacks are ready, just return. */
1976 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1977 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1978 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1979 need_resched(), is_idle_task(current),
1980 rcu_is_callbacks_kthread());
1985 * Extract the list of ready callbacks, disabling to prevent
1986 * races with call_rcu() from interrupt handlers.
1988 local_irq_save(flags);
1989 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1991 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1992 list = rdp->nxtlist;
1993 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1994 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1995 tail = rdp->nxttail[RCU_DONE_TAIL];
1996 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1997 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1998 rdp->nxttail[i] = &rdp->nxtlist;
1999 local_irq_restore(flags);
2001 /* Invoke callbacks. */
2002 count = count_lazy = 0;
2006 debug_rcu_head_unqueue(list);
2007 if (__rcu_reclaim(rsp->name, list))
2010 /* Stop only if limit reached and CPU has something to do. */
2011 if (++count >= bl &&
2013 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2017 local_irq_save(flags);
2018 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2019 is_idle_task(current),
2020 rcu_is_callbacks_kthread());
2022 /* Update count, and requeue any remaining callbacks. */
2024 *tail = rdp->nxtlist;
2025 rdp->nxtlist = list;
2026 for (i = 0; i < RCU_NEXT_SIZE; i++)
2027 if (&rdp->nxtlist == rdp->nxttail[i])
2028 rdp->nxttail[i] = tail;
2032 smp_mb(); /* List handling before counting for rcu_barrier(). */
2033 rdp->qlen_lazy -= count_lazy;
2034 ACCESS_ONCE(rdp->qlen) -= count;
2035 rdp->n_cbs_invoked += count;
2037 /* Reinstate batch limit if we have worked down the excess. */
2038 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2039 rdp->blimit = blimit;
2041 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2042 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2043 rdp->qlen_last_fqs_check = 0;
2044 rdp->n_force_qs_snap = rsp->n_force_qs;
2045 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2046 rdp->qlen_last_fqs_check = rdp->qlen;
2047 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2049 local_irq_restore(flags);
2051 /* Re-invoke RCU core processing if there are callbacks remaining. */
2052 if (cpu_has_callbacks_ready_to_invoke(rdp))
2057 * Check to see if this CPU is in a non-context-switch quiescent state
2058 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2059 * Also schedule RCU core processing.
2061 * This function must be called from hardirq context. It is normally
2062 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2063 * false, there is no point in invoking rcu_check_callbacks().
2065 void rcu_check_callbacks(int cpu, int user)
2067 trace_rcu_utilization(TPS("Start scheduler-tick"));
2068 increment_cpu_stall_ticks();
2069 if (user || rcu_is_cpu_rrupt_from_idle()) {
2072 * Get here if this CPU took its interrupt from user
2073 * mode or from the idle loop, and if this is not a
2074 * nested interrupt. In this case, the CPU is in
2075 * a quiescent state, so note it.
2077 * No memory barrier is required here because both
2078 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2079 * variables that other CPUs neither access nor modify,
2080 * at least not while the corresponding CPU is online.
2086 } else if (!in_softirq()) {
2089 * Get here if this CPU did not take its interrupt from
2090 * softirq, in other words, if it is not interrupting
2091 * a rcu_bh read-side critical section. This is an _bh
2092 * critical section, so note it.
2097 rcu_preempt_check_callbacks(cpu);
2098 if (rcu_pending(cpu))
2100 trace_rcu_utilization(TPS("End scheduler-tick"));
2104 * Scan the leaf rcu_node structures, processing dyntick state for any that
2105 * have not yet encountered a quiescent state, using the function specified.
2106 * Also initiate boosting for any threads blocked on the root rcu_node.
2108 * The caller must have suppressed start of new grace periods.
2110 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
2114 unsigned long flags;
2116 struct rcu_node *rnp;
2118 rcu_for_each_leaf_node(rsp, rnp) {
2121 raw_spin_lock_irqsave(&rnp->lock, flags);
2122 if (!rcu_gp_in_progress(rsp)) {
2123 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2126 if (rnp->qsmask == 0) {
2127 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2132 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2133 if ((rnp->qsmask & bit) != 0 &&
2134 f(per_cpu_ptr(rsp->rda, cpu)))
2139 /* rcu_report_qs_rnp() releases rnp->lock. */
2140 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2143 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2145 rnp = rcu_get_root(rsp);
2146 if (rnp->qsmask == 0) {
2147 raw_spin_lock_irqsave(&rnp->lock, flags);
2148 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2153 * Force quiescent states on reluctant CPUs, and also detect which
2154 * CPUs are in dyntick-idle mode.
2156 static void force_quiescent_state(struct rcu_state *rsp)
2158 unsigned long flags;
2160 struct rcu_node *rnp;
2161 struct rcu_node *rnp_old = NULL;
2163 /* Funnel through hierarchy to reduce memory contention. */
2164 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2165 for (; rnp != NULL; rnp = rnp->parent) {
2166 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2167 !raw_spin_trylock(&rnp->fqslock);
2168 if (rnp_old != NULL)
2169 raw_spin_unlock(&rnp_old->fqslock);
2171 rsp->n_force_qs_lh++;
2176 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2178 /* Reached the root of the rcu_node tree, acquire lock. */
2179 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2180 raw_spin_unlock(&rnp_old->fqslock);
2181 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2182 rsp->n_force_qs_lh++;
2183 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2184 return; /* Someone beat us to it. */
2186 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2187 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2188 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2192 * This does the RCU core processing work for the specified rcu_state
2193 * and rcu_data structures. This may be called only from the CPU to
2194 * whom the rdp belongs.
2197 __rcu_process_callbacks(struct rcu_state *rsp)
2199 unsigned long flags;
2200 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2202 WARN_ON_ONCE(rdp->beenonline == 0);
2204 /* Update RCU state based on any recent quiescent states. */
2205 rcu_check_quiescent_state(rsp, rdp);
2207 /* Does this CPU require a not-yet-started grace period? */
2208 local_irq_save(flags);
2209 if (cpu_needs_another_gp(rsp, rdp)) {
2210 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2212 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2214 local_irq_restore(flags);
2217 /* If there are callbacks ready, invoke them. */
2218 if (cpu_has_callbacks_ready_to_invoke(rdp))
2219 invoke_rcu_callbacks(rsp, rdp);
2223 * Do RCU core processing for the current CPU.
2225 static void rcu_process_callbacks(struct softirq_action *unused)
2227 struct rcu_state *rsp;
2229 if (cpu_is_offline(smp_processor_id()))
2231 trace_rcu_utilization(TPS("Start RCU core"));
2232 for_each_rcu_flavor(rsp)
2233 __rcu_process_callbacks(rsp);
2234 trace_rcu_utilization(TPS("End RCU core"));
2238 * Schedule RCU callback invocation. If the specified type of RCU
2239 * does not support RCU priority boosting, just do a direct call,
2240 * otherwise wake up the per-CPU kernel kthread. Note that because we
2241 * are running on the current CPU with interrupts disabled, the
2242 * rcu_cpu_kthread_task cannot disappear out from under us.
2244 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2246 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2248 if (likely(!rsp->boost)) {
2249 rcu_do_batch(rsp, rdp);
2252 invoke_rcu_callbacks_kthread();
2255 static void invoke_rcu_core(void)
2257 if (cpu_online(smp_processor_id()))
2258 raise_softirq(RCU_SOFTIRQ);
2262 * Handle any core-RCU processing required by a call_rcu() invocation.
2264 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2265 struct rcu_head *head, unsigned long flags)
2268 * If called from an extended quiescent state, invoke the RCU
2269 * core in order to force a re-evaluation of RCU's idleness.
2271 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2274 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2275 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2279 * Force the grace period if too many callbacks or too long waiting.
2280 * Enforce hysteresis, and don't invoke force_quiescent_state()
2281 * if some other CPU has recently done so. Also, don't bother
2282 * invoking force_quiescent_state() if the newly enqueued callback
2283 * is the only one waiting for a grace period to complete.
2285 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2287 /* Are we ignoring a completed grace period? */
2288 note_gp_changes(rsp, rdp);
2290 /* Start a new grace period if one not already started. */
2291 if (!rcu_gp_in_progress(rsp)) {
2292 struct rcu_node *rnp_root = rcu_get_root(rsp);
2294 raw_spin_lock(&rnp_root->lock);
2296 raw_spin_unlock(&rnp_root->lock);
2298 /* Give the grace period a kick. */
2299 rdp->blimit = LONG_MAX;
2300 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2301 *rdp->nxttail[RCU_DONE_TAIL] != head)
2302 force_quiescent_state(rsp);
2303 rdp->n_force_qs_snap = rsp->n_force_qs;
2304 rdp->qlen_last_fqs_check = rdp->qlen;
2310 * RCU callback function to leak a callback.
2312 static void rcu_leak_callback(struct rcu_head *rhp)
2317 * Helper function for call_rcu() and friends. The cpu argument will
2318 * normally be -1, indicating "currently running CPU". It may specify
2319 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2320 * is expected to specify a CPU.
2323 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2324 struct rcu_state *rsp, int cpu, bool lazy)
2326 unsigned long flags;
2327 struct rcu_data *rdp;
2329 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2330 if (debug_rcu_head_queue(head)) {
2331 /* Probable double call_rcu(), so leak the callback. */
2332 ACCESS_ONCE(head->func) = rcu_leak_callback;
2333 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2340 * Opportunistically note grace-period endings and beginnings.
2341 * Note that we might see a beginning right after we see an
2342 * end, but never vice versa, since this CPU has to pass through
2343 * a quiescent state betweentimes.
2345 local_irq_save(flags);
2346 rdp = this_cpu_ptr(rsp->rda);
2348 /* Add the callback to our list. */
2349 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2353 rdp = per_cpu_ptr(rsp->rda, cpu);
2354 offline = !__call_rcu_nocb(rdp, head, lazy);
2355 WARN_ON_ONCE(offline);
2356 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2357 local_irq_restore(flags);
2360 ACCESS_ONCE(rdp->qlen)++;
2364 rcu_idle_count_callbacks_posted();
2365 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2366 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2367 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2369 if (__is_kfree_rcu_offset((unsigned long)func))
2370 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2371 rdp->qlen_lazy, rdp->qlen);
2373 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2375 /* Go handle any RCU core processing required. */
2376 __call_rcu_core(rsp, rdp, head, flags);
2377 local_irq_restore(flags);
2381 * Queue an RCU-sched callback for invocation after a grace period.
2383 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2385 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2387 EXPORT_SYMBOL_GPL(call_rcu_sched);
2390 * Queue an RCU callback for invocation after a quicker grace period.
2392 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2394 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2396 EXPORT_SYMBOL_GPL(call_rcu_bh);
2399 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2400 * any blocking grace-period wait automatically implies a grace period
2401 * if there is only one CPU online at any point time during execution
2402 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2403 * occasionally incorrectly indicate that there are multiple CPUs online
2404 * when there was in fact only one the whole time, as this just adds
2405 * some overhead: RCU still operates correctly.
2407 static inline int rcu_blocking_is_gp(void)
2411 might_sleep(); /* Check for RCU read-side critical section. */
2413 ret = num_online_cpus() <= 1;
2419 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2421 * Control will return to the caller some time after a full rcu-sched
2422 * grace period has elapsed, in other words after all currently executing
2423 * rcu-sched read-side critical sections have completed. These read-side
2424 * critical sections are delimited by rcu_read_lock_sched() and
2425 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2426 * local_irq_disable(), and so on may be used in place of
2427 * rcu_read_lock_sched().
2429 * This means that all preempt_disable code sequences, including NMI and
2430 * non-threaded hardware-interrupt handlers, in progress on entry will
2431 * have completed before this primitive returns. However, this does not
2432 * guarantee that softirq handlers will have completed, since in some
2433 * kernels, these handlers can run in process context, and can block.
2435 * Note that this guarantee implies further memory-ordering guarantees.
2436 * On systems with more than one CPU, when synchronize_sched() returns,
2437 * each CPU is guaranteed to have executed a full memory barrier since the
2438 * end of its last RCU-sched read-side critical section whose beginning
2439 * preceded the call to synchronize_sched(). In addition, each CPU having
2440 * an RCU read-side critical section that extends beyond the return from
2441 * synchronize_sched() is guaranteed to have executed a full memory barrier
2442 * after the beginning of synchronize_sched() and before the beginning of
2443 * that RCU read-side critical section. Note that these guarantees include
2444 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2445 * that are executing in the kernel.
2447 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2448 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2449 * to have executed a full memory barrier during the execution of
2450 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2451 * again only if the system has more than one CPU).
2453 * This primitive provides the guarantees made by the (now removed)
2454 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2455 * guarantees that rcu_read_lock() sections will have completed.
2456 * In "classic RCU", these two guarantees happen to be one and
2457 * the same, but can differ in realtime RCU implementations.
2459 void synchronize_sched(void)
2461 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2462 !lock_is_held(&rcu_lock_map) &&
2463 !lock_is_held(&rcu_sched_lock_map),
2464 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2465 if (rcu_blocking_is_gp())
2468 synchronize_sched_expedited();
2470 wait_rcu_gp(call_rcu_sched);
2472 EXPORT_SYMBOL_GPL(synchronize_sched);
2475 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2477 * Control will return to the caller some time after a full rcu_bh grace
2478 * period has elapsed, in other words after all currently executing rcu_bh
2479 * read-side critical sections have completed. RCU read-side critical
2480 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2481 * and may be nested.
2483 * See the description of synchronize_sched() for more detailed information
2484 * on memory ordering guarantees.
2486 void synchronize_rcu_bh(void)
2488 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2489 !lock_is_held(&rcu_lock_map) &&
2490 !lock_is_held(&rcu_sched_lock_map),
2491 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2492 if (rcu_blocking_is_gp())
2495 synchronize_rcu_bh_expedited();
2497 wait_rcu_gp(call_rcu_bh);
2499 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2501 static int synchronize_sched_expedited_cpu_stop(void *data)
2504 * There must be a full memory barrier on each affected CPU
2505 * between the time that try_stop_cpus() is called and the
2506 * time that it returns.
2508 * In the current initial implementation of cpu_stop, the
2509 * above condition is already met when the control reaches
2510 * this point and the following smp_mb() is not strictly
2511 * necessary. Do smp_mb() anyway for documentation and
2512 * robustness against future implementation changes.
2514 smp_mb(); /* See above comment block. */
2519 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2521 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2522 * approach to force the grace period to end quickly. This consumes
2523 * significant time on all CPUs and is unfriendly to real-time workloads,
2524 * so is thus not recommended for any sort of common-case code. In fact,
2525 * if you are using synchronize_sched_expedited() in a loop, please
2526 * restructure your code to batch your updates, and then use a single
2527 * synchronize_sched() instead.
2529 * Note that it is illegal to call this function while holding any lock
2530 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2531 * to call this function from a CPU-hotplug notifier. Failing to observe
2532 * these restriction will result in deadlock.
2534 * This implementation can be thought of as an application of ticket
2535 * locking to RCU, with sync_sched_expedited_started and
2536 * sync_sched_expedited_done taking on the roles of the halves
2537 * of the ticket-lock word. Each task atomically increments
2538 * sync_sched_expedited_started upon entry, snapshotting the old value,
2539 * then attempts to stop all the CPUs. If this succeeds, then each
2540 * CPU will have executed a context switch, resulting in an RCU-sched
2541 * grace period. We are then done, so we use atomic_cmpxchg() to
2542 * update sync_sched_expedited_done to match our snapshot -- but
2543 * only if someone else has not already advanced past our snapshot.
2545 * On the other hand, if try_stop_cpus() fails, we check the value
2546 * of sync_sched_expedited_done. If it has advanced past our
2547 * initial snapshot, then someone else must have forced a grace period
2548 * some time after we took our snapshot. In this case, our work is
2549 * done for us, and we can simply return. Otherwise, we try again,
2550 * but keep our initial snapshot for purposes of checking for someone
2551 * doing our work for us.
2553 * If we fail too many times in a row, we fall back to synchronize_sched().
2555 void synchronize_sched_expedited(void)
2557 long firstsnap, s, snap;
2559 struct rcu_state *rsp = &rcu_sched_state;
2562 * If we are in danger of counter wrap, just do synchronize_sched().
2563 * By allowing sync_sched_expedited_started to advance no more than
2564 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2565 * that more than 3.5 billion CPUs would be required to force a
2566 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2567 * course be required on a 64-bit system.
2569 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2570 (ulong)atomic_long_read(&rsp->expedited_done) +
2572 synchronize_sched();
2573 atomic_long_inc(&rsp->expedited_wrap);
2578 * Take a ticket. Note that atomic_inc_return() implies a
2579 * full memory barrier.
2581 snap = atomic_long_inc_return(&rsp->expedited_start);
2584 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2587 * Each pass through the following loop attempts to force a
2588 * context switch on each CPU.
2590 while (try_stop_cpus(cpu_online_mask,
2591 synchronize_sched_expedited_cpu_stop,
2594 atomic_long_inc(&rsp->expedited_tryfail);
2596 /* Check to see if someone else did our work for us. */
2597 s = atomic_long_read(&rsp->expedited_done);
2598 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2599 /* ensure test happens before caller kfree */
2600 smp_mb__before_atomic_inc(); /* ^^^ */
2601 atomic_long_inc(&rsp->expedited_workdone1);
2605 /* No joy, try again later. Or just synchronize_sched(). */
2606 if (trycount++ < 10) {
2607 udelay(trycount * num_online_cpus());
2609 wait_rcu_gp(call_rcu_sched);
2610 atomic_long_inc(&rsp->expedited_normal);
2614 /* Recheck to see if someone else did our work for us. */
2615 s = atomic_long_read(&rsp->expedited_done);
2616 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2617 /* ensure test happens before caller kfree */
2618 smp_mb__before_atomic_inc(); /* ^^^ */
2619 atomic_long_inc(&rsp->expedited_workdone2);
2624 * Refetching sync_sched_expedited_started allows later
2625 * callers to piggyback on our grace period. We retry
2626 * after they started, so our grace period works for them,
2627 * and they started after our first try, so their grace
2628 * period works for us.
2631 snap = atomic_long_read(&rsp->expedited_start);
2632 smp_mb(); /* ensure read is before try_stop_cpus(). */
2634 atomic_long_inc(&rsp->expedited_stoppedcpus);
2637 * Everyone up to our most recent fetch is covered by our grace
2638 * period. Update the counter, but only if our work is still
2639 * relevant -- which it won't be if someone who started later
2640 * than we did already did their update.
2643 atomic_long_inc(&rsp->expedited_done_tries);
2644 s = atomic_long_read(&rsp->expedited_done);
2645 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2646 /* ensure test happens before caller kfree */
2647 smp_mb__before_atomic_inc(); /* ^^^ */
2648 atomic_long_inc(&rsp->expedited_done_lost);
2651 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2652 atomic_long_inc(&rsp->expedited_done_exit);
2656 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2659 * Check to see if there is any immediate RCU-related work to be done
2660 * by the current CPU, for the specified type of RCU, returning 1 if so.
2661 * The checks are in order of increasing expense: checks that can be
2662 * carried out against CPU-local state are performed first. However,
2663 * we must check for CPU stalls first, else we might not get a chance.
2665 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2667 struct rcu_node *rnp = rdp->mynode;
2669 rdp->n_rcu_pending++;
2671 /* Check for CPU stalls, if enabled. */
2672 check_cpu_stall(rsp, rdp);
2674 /* Is the RCU core waiting for a quiescent state from this CPU? */
2675 if (rcu_scheduler_fully_active &&
2676 rdp->qs_pending && !rdp->passed_quiesce) {
2677 rdp->n_rp_qs_pending++;
2678 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2679 rdp->n_rp_report_qs++;
2683 /* Does this CPU have callbacks ready to invoke? */
2684 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2685 rdp->n_rp_cb_ready++;
2689 /* Has RCU gone idle with this CPU needing another grace period? */
2690 if (cpu_needs_another_gp(rsp, rdp)) {
2691 rdp->n_rp_cpu_needs_gp++;
2695 /* Has another RCU grace period completed? */
2696 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2697 rdp->n_rp_gp_completed++;
2701 /* Has a new RCU grace period started? */
2702 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2703 rdp->n_rp_gp_started++;
2708 rdp->n_rp_need_nothing++;
2713 * Check to see if there is any immediate RCU-related work to be done
2714 * by the current CPU, returning 1 if so. This function is part of the
2715 * RCU implementation; it is -not- an exported member of the RCU API.
2717 static int rcu_pending(int cpu)
2719 struct rcu_state *rsp;
2721 for_each_rcu_flavor(rsp)
2722 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2728 * Return true if the specified CPU has any callback. If all_lazy is
2729 * non-NULL, store an indication of whether all callbacks are lazy.
2730 * (If there are no callbacks, all of them are deemed to be lazy.)
2732 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2736 struct rcu_data *rdp;
2737 struct rcu_state *rsp;
2739 for_each_rcu_flavor(rsp) {
2740 rdp = per_cpu_ptr(rsp->rda, cpu);
2741 if (rdp->qlen != rdp->qlen_lazy)
2752 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2753 * the compiler is expected to optimize this away.
2755 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2756 int cpu, unsigned long done)
2758 trace_rcu_barrier(rsp->name, s, cpu,
2759 atomic_read(&rsp->barrier_cpu_count), done);
2763 * RCU callback function for _rcu_barrier(). If we are last, wake
2764 * up the task executing _rcu_barrier().
2766 static void rcu_barrier_callback(struct rcu_head *rhp)
2768 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2769 struct rcu_state *rsp = rdp->rsp;
2771 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2772 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2773 complete(&rsp->barrier_completion);
2775 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2780 * Called with preemption disabled, and from cross-cpu IRQ context.
2782 static void rcu_barrier_func(void *type)
2784 struct rcu_state *rsp = type;
2785 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2787 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2788 atomic_inc(&rsp->barrier_cpu_count);
2789 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2793 * Orchestrate the specified type of RCU barrier, waiting for all
2794 * RCU callbacks of the specified type to complete.
2796 static void _rcu_barrier(struct rcu_state *rsp)
2799 struct rcu_data *rdp;
2800 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2801 unsigned long snap_done;
2803 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2805 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2806 mutex_lock(&rsp->barrier_mutex);
2809 * Ensure that all prior references, including to ->n_barrier_done,
2810 * are ordered before the _rcu_barrier() machinery.
2812 smp_mb(); /* See above block comment. */
2815 * Recheck ->n_barrier_done to see if others did our work for us.
2816 * This means checking ->n_barrier_done for an even-to-odd-to-even
2817 * transition. The "if" expression below therefore rounds the old
2818 * value up to the next even number and adds two before comparing.
2820 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2821 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2822 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2823 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2824 smp_mb(); /* caller's subsequent code after above check. */
2825 mutex_unlock(&rsp->barrier_mutex);
2830 * Increment ->n_barrier_done to avoid duplicate work. Use
2831 * ACCESS_ONCE() to prevent the compiler from speculating
2832 * the increment to precede the early-exit check.
2834 ACCESS_ONCE(rsp->n_barrier_done)++;
2835 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2836 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2837 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2840 * Initialize the count to one rather than to zero in order to
2841 * avoid a too-soon return to zero in case of a short grace period
2842 * (or preemption of this task). Exclude CPU-hotplug operations
2843 * to ensure that no offline CPU has callbacks queued.
2845 init_completion(&rsp->barrier_completion);
2846 atomic_set(&rsp->barrier_cpu_count, 1);
2850 * Force each CPU with callbacks to register a new callback.
2851 * When that callback is invoked, we will know that all of the
2852 * corresponding CPU's preceding callbacks have been invoked.
2854 for_each_possible_cpu(cpu) {
2855 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2857 rdp = per_cpu_ptr(rsp->rda, cpu);
2858 if (rcu_is_nocb_cpu(cpu)) {
2859 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2860 rsp->n_barrier_done);
2861 atomic_inc(&rsp->barrier_cpu_count);
2862 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2864 } else if (ACCESS_ONCE(rdp->qlen)) {
2865 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2866 rsp->n_barrier_done);
2867 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2869 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2870 rsp->n_barrier_done);
2876 * Now that we have an rcu_barrier_callback() callback on each
2877 * CPU, and thus each counted, remove the initial count.
2879 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2880 complete(&rsp->barrier_completion);
2882 /* Increment ->n_barrier_done to prevent duplicate work. */
2883 smp_mb(); /* Keep increment after above mechanism. */
2884 ACCESS_ONCE(rsp->n_barrier_done)++;
2885 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2886 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2887 smp_mb(); /* Keep increment before caller's subsequent code. */
2889 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2890 wait_for_completion(&rsp->barrier_completion);
2892 /* Other rcu_barrier() invocations can now safely proceed. */
2893 mutex_unlock(&rsp->barrier_mutex);
2897 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2899 void rcu_barrier_bh(void)
2901 _rcu_barrier(&rcu_bh_state);
2903 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2906 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2908 void rcu_barrier_sched(void)
2910 _rcu_barrier(&rcu_sched_state);
2912 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2915 * Do boot-time initialization of a CPU's per-CPU RCU data.
2918 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2920 unsigned long flags;
2921 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2922 struct rcu_node *rnp = rcu_get_root(rsp);
2924 /* Set up local state, ensuring consistent view of global state. */
2925 raw_spin_lock_irqsave(&rnp->lock, flags);
2926 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2927 init_callback_list(rdp);
2929 ACCESS_ONCE(rdp->qlen) = 0;
2930 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2931 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2932 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2935 rcu_boot_init_nocb_percpu_data(rdp);
2936 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2940 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2941 * offline event can be happening at a given time. Note also that we
2942 * can accept some slop in the rsp->completed access due to the fact
2943 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2946 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2948 unsigned long flags;
2950 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2951 struct rcu_node *rnp = rcu_get_root(rsp);
2953 /* Exclude new grace periods. */
2954 mutex_lock(&rsp->onoff_mutex);
2956 /* Set up local state, ensuring consistent view of global state. */
2957 raw_spin_lock_irqsave(&rnp->lock, flags);
2958 rdp->beenonline = 1; /* We have now been online. */
2959 rdp->preemptible = preemptible;
2960 rdp->qlen_last_fqs_check = 0;
2961 rdp->n_force_qs_snap = rsp->n_force_qs;
2962 rdp->blimit = blimit;
2963 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2964 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2965 atomic_set(&rdp->dynticks->dynticks,
2966 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2967 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2969 /* Add CPU to rcu_node bitmasks. */
2971 mask = rdp->grpmask;
2973 /* Exclude any attempts to start a new GP on small systems. */
2974 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2975 rnp->qsmaskinit |= mask;
2976 mask = rnp->grpmask;
2977 if (rnp == rdp->mynode) {
2979 * If there is a grace period in progress, we will
2980 * set up to wait for it next time we run the
2983 rdp->gpnum = rnp->completed;
2984 rdp->completed = rnp->completed;
2985 rdp->passed_quiesce = 0;
2986 rdp->qs_pending = 0;
2987 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
2989 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2991 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2992 local_irq_restore(flags);
2994 mutex_unlock(&rsp->onoff_mutex);
2997 static void rcu_prepare_cpu(int cpu)
2999 struct rcu_state *rsp;
3001 for_each_rcu_flavor(rsp)
3002 rcu_init_percpu_data(cpu, rsp,
3003 strcmp(rsp->name, "rcu_preempt") == 0);
3007 * Handle CPU online/offline notification events.
3009 static int rcu_cpu_notify(struct notifier_block *self,
3010 unsigned long action, void *hcpu)
3012 long cpu = (long)hcpu;
3013 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3014 struct rcu_node *rnp = rdp->mynode;
3015 struct rcu_state *rsp;
3017 trace_rcu_utilization(TPS("Start CPU hotplug"));
3019 case CPU_UP_PREPARE:
3020 case CPU_UP_PREPARE_FROZEN:
3021 rcu_prepare_cpu(cpu);
3022 rcu_prepare_kthreads(cpu);
3025 case CPU_DOWN_FAILED:
3026 rcu_boost_kthread_setaffinity(rnp, -1);
3028 case CPU_DOWN_PREPARE:
3029 rcu_boost_kthread_setaffinity(rnp, cpu);
3032 case CPU_DYING_FROZEN:
3033 for_each_rcu_flavor(rsp)
3034 rcu_cleanup_dying_cpu(rsp);
3037 case CPU_DEAD_FROZEN:
3038 case CPU_UP_CANCELED:
3039 case CPU_UP_CANCELED_FROZEN:
3040 for_each_rcu_flavor(rsp)
3041 rcu_cleanup_dead_cpu(cpu, rsp);
3046 trace_rcu_utilization(TPS("End CPU hotplug"));
3050 static int rcu_pm_notify(struct notifier_block *self,
3051 unsigned long action, void *hcpu)
3054 case PM_HIBERNATION_PREPARE:
3055 case PM_SUSPEND_PREPARE:
3056 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3059 case PM_POST_HIBERNATION:
3060 case PM_POST_SUSPEND:
3070 * Spawn the kthread that handles this RCU flavor's grace periods.
3072 static int __init rcu_spawn_gp_kthread(void)
3074 unsigned long flags;
3075 struct rcu_node *rnp;
3076 struct rcu_state *rsp;
3077 struct task_struct *t;
3079 for_each_rcu_flavor(rsp) {
3080 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3082 rnp = rcu_get_root(rsp);
3083 raw_spin_lock_irqsave(&rnp->lock, flags);
3084 rsp->gp_kthread = t;
3085 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3086 rcu_spawn_nocb_kthreads(rsp);
3090 early_initcall(rcu_spawn_gp_kthread);
3093 * This function is invoked towards the end of the scheduler's initialization
3094 * process. Before this is called, the idle task might contain
3095 * RCU read-side critical sections (during which time, this idle
3096 * task is booting the system). After this function is called, the
3097 * idle tasks are prohibited from containing RCU read-side critical
3098 * sections. This function also enables RCU lockdep checking.
3100 void rcu_scheduler_starting(void)
3102 WARN_ON(num_online_cpus() != 1);
3103 WARN_ON(nr_context_switches() > 0);
3104 rcu_scheduler_active = 1;
3108 * Compute the per-level fanout, either using the exact fanout specified
3109 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3111 #ifdef CONFIG_RCU_FANOUT_EXACT
3112 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3116 for (i = rcu_num_lvls - 1; i > 0; i--)
3117 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3118 rsp->levelspread[0] = rcu_fanout_leaf;
3120 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3121 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3128 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3129 ccur = rsp->levelcnt[i];
3130 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3134 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3137 * Helper function for rcu_init() that initializes one rcu_state structure.
3139 static void __init rcu_init_one(struct rcu_state *rsp,
3140 struct rcu_data __percpu *rda)
3142 static char *buf[] = { "rcu_node_0",
3145 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3146 static char *fqs[] = { "rcu_node_fqs_0",
3149 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3153 struct rcu_node *rnp;
3155 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3157 /* Silence gcc 4.8 warning about array index out of range. */
3158 if (rcu_num_lvls > RCU_NUM_LVLS)
3159 panic("rcu_init_one: rcu_num_lvls overflow");
3161 /* Initialize the level-tracking arrays. */
3163 for (i = 0; i < rcu_num_lvls; i++)
3164 rsp->levelcnt[i] = num_rcu_lvl[i];
3165 for (i = 1; i < rcu_num_lvls; i++)
3166 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3167 rcu_init_levelspread(rsp);
3169 /* Initialize the elements themselves, starting from the leaves. */
3171 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3172 cpustride *= rsp->levelspread[i];
3173 rnp = rsp->level[i];
3174 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3175 raw_spin_lock_init(&rnp->lock);
3176 lockdep_set_class_and_name(&rnp->lock,
3177 &rcu_node_class[i], buf[i]);
3178 raw_spin_lock_init(&rnp->fqslock);
3179 lockdep_set_class_and_name(&rnp->fqslock,
3180 &rcu_fqs_class[i], fqs[i]);
3181 rnp->gpnum = rsp->gpnum;
3182 rnp->completed = rsp->completed;
3184 rnp->qsmaskinit = 0;
3185 rnp->grplo = j * cpustride;
3186 rnp->grphi = (j + 1) * cpustride - 1;
3187 if (rnp->grphi >= NR_CPUS)
3188 rnp->grphi = NR_CPUS - 1;
3194 rnp->grpnum = j % rsp->levelspread[i - 1];
3195 rnp->grpmask = 1UL << rnp->grpnum;
3196 rnp->parent = rsp->level[i - 1] +
3197 j / rsp->levelspread[i - 1];
3200 INIT_LIST_HEAD(&rnp->blkd_tasks);
3201 rcu_init_one_nocb(rnp);
3206 init_waitqueue_head(&rsp->gp_wq);
3207 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3208 rnp = rsp->level[rcu_num_lvls - 1];
3209 for_each_possible_cpu(i) {
3210 while (i > rnp->grphi)
3212 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3213 rcu_boot_init_percpu_data(i, rsp);
3215 list_add(&rsp->flavors, &rcu_struct_flavors);
3219 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3220 * replace the definitions in rcutree.h because those are needed to size
3221 * the ->node array in the rcu_state structure.
3223 static void __init rcu_init_geometry(void)
3229 int rcu_capacity[MAX_RCU_LVLS + 1];
3232 * Initialize any unspecified boot parameters.
3233 * The default values of jiffies_till_first_fqs and
3234 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3235 * value, which is a function of HZ, then adding one for each
3236 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3238 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3239 if (jiffies_till_first_fqs == ULONG_MAX)
3240 jiffies_till_first_fqs = d;
3241 if (jiffies_till_next_fqs == ULONG_MAX)
3242 jiffies_till_next_fqs = d;
3244 /* If the compile-time values are accurate, just leave. */
3245 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3246 nr_cpu_ids == NR_CPUS)
3250 * Compute number of nodes that can be handled an rcu_node tree
3251 * with the given number of levels. Setting rcu_capacity[0] makes
3252 * some of the arithmetic easier.
3254 rcu_capacity[0] = 1;
3255 rcu_capacity[1] = rcu_fanout_leaf;
3256 for (i = 2; i <= MAX_RCU_LVLS; i++)
3257 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3260 * The boot-time rcu_fanout_leaf parameter is only permitted
3261 * to increase the leaf-level fanout, not decrease it. Of course,
3262 * the leaf-level fanout cannot exceed the number of bits in
3263 * the rcu_node masks. Finally, the tree must be able to accommodate
3264 * the configured number of CPUs. Complain and fall back to the
3265 * compile-time values if these limits are exceeded.
3267 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3268 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3269 n > rcu_capacity[MAX_RCU_LVLS]) {
3274 /* Calculate the number of rcu_nodes at each level of the tree. */
3275 for (i = 1; i <= MAX_RCU_LVLS; i++)
3276 if (n <= rcu_capacity[i]) {
3277 for (j = 0; j <= i; j++)
3279 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3281 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3286 /* Calculate the total number of rcu_node structures. */
3288 for (i = 0; i <= MAX_RCU_LVLS; i++)
3289 rcu_num_nodes += num_rcu_lvl[i];
3293 void __init rcu_init(void)
3297 rcu_bootup_announce();
3298 rcu_init_geometry();
3299 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3300 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3301 __rcu_init_preempt();
3302 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3305 * We don't need protection against CPU-hotplug here because
3306 * this is called early in boot, before either interrupts
3307 * or the scheduler are operational.
3309 cpu_notifier(rcu_cpu_notify, 0);
3310 pm_notifier(rcu_pm_notify, 0);
3311 for_each_online_cpu(cpu)
3312 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3315 #include "rcutree_plugin.h"