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>
59 #include <trace/events/rcu.h>
64 * Strings used in tracepoints need to be exported via the
65 * tracing system such that tools like perf and trace-cmd can
66 * translate the string address pointers to actual text.
68 #define TPS(x) tracepoint_string(x)
70 /* Data structures. */
72 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
73 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
76 * In order to export the rcu_state name to the tracing tools, it
77 * needs to be added in the __tracepoint_string section.
78 * This requires defining a separate variable tp_<sname>_varname
79 * that points to the string being used, and this will allow
80 * the tracing userspace tools to be able to decipher the string
81 * address to the matching string.
83 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
86 struct rcu_state sname##_state = { \
87 .level = { &sname##_state.node[0] }, \
89 .fqs_state = RCU_GP_IDLE, \
90 .gpnum = 0UL - 300UL, \
91 .completed = 0UL - 300UL, \
92 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
93 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
94 .orphan_donetail = &sname##_state.orphan_donelist, \
95 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
96 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
97 .name = sname##_varname, \
100 DEFINE_PER_CPU(struct rcu_data, sname##_data)
102 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
103 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
105 static struct rcu_state *rcu_state;
106 LIST_HEAD(rcu_struct_flavors);
108 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
109 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
110 module_param(rcu_fanout_leaf, int, 0444);
111 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
112 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
119 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
122 * The rcu_scheduler_active variable transitions from zero to one just
123 * before the first task is spawned. So when this variable is zero, RCU
124 * can assume that there is but one task, allowing RCU to (for example)
125 * optimize synchronize_sched() to a simple barrier(). When this variable
126 * is one, RCU must actually do all the hard work required to detect real
127 * grace periods. This variable is also used to suppress boot-time false
128 * positives from lockdep-RCU error checking.
130 int rcu_scheduler_active __read_mostly;
131 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
134 * The rcu_scheduler_fully_active variable transitions from zero to one
135 * during the early_initcall() processing, which is after the scheduler
136 * is capable of creating new tasks. So RCU processing (for example,
137 * creating tasks for RCU priority boosting) must be delayed until after
138 * rcu_scheduler_fully_active transitions from zero to one. We also
139 * currently delay invocation of any RCU callbacks until after this point.
141 * It might later prove better for people registering RCU callbacks during
142 * early boot to take responsibility for these callbacks, but one step at
145 static int rcu_scheduler_fully_active __read_mostly;
147 #ifdef CONFIG_RCU_BOOST
150 * Control variables for per-CPU and per-rcu_node kthreads. These
151 * handle all flavors of RCU.
153 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
155 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
156 DEFINE_PER_CPU(char, rcu_cpu_has_work);
158 #endif /* #ifdef CONFIG_RCU_BOOST */
160 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
165 * Track the rcutorture test sequence number and the update version
166 * number within a given test. The rcutorture_testseq is incremented
167 * on every rcutorture module load and unload, so has an odd value
168 * when a test is running. The rcutorture_vernum is set to zero
169 * when rcutorture starts and is incremented on each rcutorture update.
170 * These variables enable correlating rcutorture output with the
171 * RCU tracing information.
173 unsigned long rcutorture_testseq;
174 unsigned long rcutorture_vernum;
177 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
178 * permit this function to be invoked without holding the root rcu_node
179 * structure's ->lock, but of course results can be subject to change.
181 static int rcu_gp_in_progress(struct rcu_state *rsp)
183 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
187 * Note a quiescent state. Because we do not need to know
188 * how many quiescent states passed, just if there was at least
189 * one since the start of the grace period, this just sets a flag.
190 * The caller must have disabled preemption.
192 void rcu_sched_qs(int cpu)
194 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
196 if (rdp->passed_quiesce == 0)
197 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
198 rdp->passed_quiesce = 1;
201 void rcu_bh_qs(int cpu)
203 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
205 if (rdp->passed_quiesce == 0)
206 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
207 rdp->passed_quiesce = 1;
211 * Note a context switch. This is a quiescent state for RCU-sched,
212 * and requires special handling for preemptible RCU.
213 * The caller must have disabled preemption.
215 void rcu_note_context_switch(int cpu)
217 trace_rcu_utilization(TPS("Start context switch"));
219 rcu_preempt_note_context_switch(cpu);
220 trace_rcu_utilization(TPS("End context switch"));
222 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
224 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
225 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
226 .dynticks = ATOMIC_INIT(1),
227 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
228 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
229 .dynticks_idle = ATOMIC_INIT(1),
230 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
233 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
234 static long qhimark = 10000; /* If this many pending, ignore blimit. */
235 static long qlowmark = 100; /* Once only this many pending, use blimit. */
237 module_param(blimit, long, 0444);
238 module_param(qhimark, long, 0444);
239 module_param(qlowmark, long, 0444);
241 static ulong jiffies_till_first_fqs = ULONG_MAX;
242 static ulong jiffies_till_next_fqs = ULONG_MAX;
244 module_param(jiffies_till_first_fqs, ulong, 0644);
245 module_param(jiffies_till_next_fqs, ulong, 0644);
247 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
248 struct rcu_data *rdp);
249 static void force_qs_rnp(struct rcu_state *rsp,
250 int (*f)(struct rcu_data *rsp, bool *isidle,
251 unsigned long *maxj),
252 bool *isidle, unsigned long *maxj);
253 static void force_quiescent_state(struct rcu_state *rsp);
254 static int rcu_pending(int cpu);
257 * Return the number of RCU-sched batches processed thus far for debug & stats.
259 long rcu_batches_completed_sched(void)
261 return rcu_sched_state.completed;
263 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
266 * Return the number of RCU BH batches processed thus far for debug & stats.
268 long rcu_batches_completed_bh(void)
270 return rcu_bh_state.completed;
272 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
275 * Force a quiescent state for RCU BH.
277 void rcu_bh_force_quiescent_state(void)
279 force_quiescent_state(&rcu_bh_state);
281 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
284 * Record the number of times rcutorture tests have been initiated and
285 * terminated. This information allows the debugfs tracing stats to be
286 * correlated to the rcutorture messages, even when the rcutorture module
287 * is being repeatedly loaded and unloaded. In other words, we cannot
288 * store this state in rcutorture itself.
290 void rcutorture_record_test_transition(void)
292 rcutorture_testseq++;
293 rcutorture_vernum = 0;
295 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
298 * Record the number of writer passes through the current rcutorture test.
299 * This is also used to correlate debugfs tracing stats with the rcutorture
302 void rcutorture_record_progress(unsigned long vernum)
306 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
309 * Force a quiescent state for RCU-sched.
311 void rcu_sched_force_quiescent_state(void)
313 force_quiescent_state(&rcu_sched_state);
315 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
318 * Does the CPU have callbacks ready to be invoked?
321 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
323 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
324 rdp->nxttail[RCU_DONE_TAIL] != NULL;
328 * Does the current CPU require a not-yet-started grace period?
329 * The caller must have disabled interrupts to prevent races with
330 * normal callback registry.
333 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
337 if (rcu_gp_in_progress(rsp))
338 return 0; /* No, a grace period is already in progress. */
339 if (rcu_nocb_needs_gp(rsp))
340 return 1; /* Yes, a no-CBs CPU needs one. */
341 if (!rdp->nxttail[RCU_NEXT_TAIL])
342 return 0; /* No, this is a no-CBs (or offline) CPU. */
343 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
344 return 1; /* Yes, this CPU has newly registered callbacks. */
345 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
346 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
347 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
348 rdp->nxtcompleted[i]))
349 return 1; /* Yes, CBs for future grace period. */
350 return 0; /* No grace period needed. */
354 * Return the root node of the specified rcu_state structure.
356 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
358 return &rsp->node[0];
362 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
364 * If the new value of the ->dynticks_nesting counter now is zero,
365 * we really have entered idle, and must do the appropriate accounting.
366 * The caller must have disabled interrupts.
368 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
371 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
372 if (!user && !is_idle_task(current)) {
373 struct task_struct *idle = idle_task(smp_processor_id());
375 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
376 ftrace_dump(DUMP_ORIG);
377 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
378 current->pid, current->comm,
379 idle->pid, idle->comm); /* must be idle task! */
381 rcu_prepare_for_idle(smp_processor_id());
382 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
383 smp_mb__before_atomic_inc(); /* See above. */
384 atomic_inc(&rdtp->dynticks);
385 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
386 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
389 * It is illegal to enter an extended quiescent state while
390 * in an RCU read-side critical section.
392 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
393 "Illegal idle entry in RCU read-side critical section.");
394 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
395 "Illegal idle entry in RCU-bh read-side critical section.");
396 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
397 "Illegal idle entry in RCU-sched read-side critical section.");
401 * Enter an RCU extended quiescent state, which can be either the
402 * idle loop or adaptive-tickless usermode execution.
404 static void rcu_eqs_enter(bool user)
407 struct rcu_dynticks *rdtp;
409 rdtp = &__get_cpu_var(rcu_dynticks);
410 oldval = rdtp->dynticks_nesting;
411 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
412 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
413 rdtp->dynticks_nesting = 0;
415 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
416 rcu_eqs_enter_common(rdtp, oldval, user);
420 * rcu_idle_enter - inform RCU that current CPU is entering idle
422 * Enter idle mode, in other words, -leave- the mode in which RCU
423 * read-side critical sections can occur. (Though RCU read-side
424 * critical sections can occur in irq handlers in idle, a possibility
425 * handled by irq_enter() and irq_exit().)
427 * We crowbar the ->dynticks_nesting field to zero to allow for
428 * the possibility of usermode upcalls having messed up our count
429 * of interrupt nesting level during the prior busy period.
431 void rcu_idle_enter(void)
435 local_irq_save(flags);
436 rcu_eqs_enter(false);
437 rcu_sysidle_enter(&__get_cpu_var(rcu_dynticks), 0);
438 local_irq_restore(flags);
440 EXPORT_SYMBOL_GPL(rcu_idle_enter);
442 #ifdef CONFIG_RCU_USER_QS
444 * rcu_user_enter - inform RCU that we are resuming userspace.
446 * Enter RCU idle mode right before resuming userspace. No use of RCU
447 * is permitted between this call and rcu_user_exit(). This way the
448 * CPU doesn't need to maintain the tick for RCU maintenance purposes
449 * when the CPU runs in userspace.
451 void rcu_user_enter(void)
455 #endif /* CONFIG_RCU_USER_QS */
458 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
460 * Exit from an interrupt handler, which might possibly result in entering
461 * idle mode, in other words, leaving the mode in which read-side critical
462 * sections can occur.
464 * This code assumes that the idle loop never does anything that might
465 * result in unbalanced calls to irq_enter() and irq_exit(). If your
466 * architecture violates this assumption, RCU will give you what you
467 * deserve, good and hard. But very infrequently and irreproducibly.
469 * Use things like work queues to work around this limitation.
471 * You have been warned.
473 void rcu_irq_exit(void)
477 struct rcu_dynticks *rdtp;
479 local_irq_save(flags);
480 rdtp = &__get_cpu_var(rcu_dynticks);
481 oldval = rdtp->dynticks_nesting;
482 rdtp->dynticks_nesting--;
483 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
484 if (rdtp->dynticks_nesting)
485 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
487 rcu_eqs_enter_common(rdtp, oldval, true);
488 rcu_sysidle_enter(rdtp, 1);
489 local_irq_restore(flags);
493 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
495 * If the new value of the ->dynticks_nesting counter was previously zero,
496 * we really have exited idle, and must do the appropriate accounting.
497 * The caller must have disabled interrupts.
499 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
502 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
503 atomic_inc(&rdtp->dynticks);
504 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
505 smp_mb__after_atomic_inc(); /* See above. */
506 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
507 rcu_cleanup_after_idle(smp_processor_id());
508 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
509 if (!user && !is_idle_task(current)) {
510 struct task_struct *idle = idle_task(smp_processor_id());
512 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
513 oldval, rdtp->dynticks_nesting);
514 ftrace_dump(DUMP_ORIG);
515 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
516 current->pid, current->comm,
517 idle->pid, idle->comm); /* must be idle task! */
522 * Exit an RCU extended quiescent state, which can be either the
523 * idle loop or adaptive-tickless usermode execution.
525 static void rcu_eqs_exit(bool user)
527 struct rcu_dynticks *rdtp;
530 rdtp = &__get_cpu_var(rcu_dynticks);
531 oldval = rdtp->dynticks_nesting;
532 WARN_ON_ONCE(oldval < 0);
533 if (oldval & DYNTICK_TASK_NEST_MASK)
534 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
536 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
537 rcu_eqs_exit_common(rdtp, oldval, user);
541 * rcu_idle_exit - inform RCU that current CPU is leaving idle
543 * Exit idle mode, in other words, -enter- the mode in which RCU
544 * read-side critical sections can occur.
546 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
547 * allow for the possibility of usermode upcalls messing up our count
548 * of interrupt nesting level during the busy period that is just
551 void rcu_idle_exit(void)
555 local_irq_save(flags);
557 rcu_sysidle_exit(&__get_cpu_var(rcu_dynticks), 0);
558 local_irq_restore(flags);
560 EXPORT_SYMBOL_GPL(rcu_idle_exit);
562 #ifdef CONFIG_RCU_USER_QS
564 * rcu_user_exit - inform RCU that we are exiting userspace.
566 * Exit RCU idle mode while entering the kernel because it can
567 * run a RCU read side critical section anytime.
569 void rcu_user_exit(void)
573 #endif /* CONFIG_RCU_USER_QS */
576 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
578 * Enter an interrupt handler, which might possibly result in exiting
579 * idle mode, in other words, entering the mode in which read-side critical
580 * sections can occur.
582 * Note that the Linux kernel is fully capable of entering an interrupt
583 * handler that it never exits, for example when doing upcalls to
584 * user mode! This code assumes that the idle loop never does upcalls to
585 * user mode. If your architecture does do upcalls from the idle loop (or
586 * does anything else that results in unbalanced calls to the irq_enter()
587 * and irq_exit() functions), RCU will give you what you deserve, good
588 * and hard. But very infrequently and irreproducibly.
590 * Use things like work queues to work around this limitation.
592 * You have been warned.
594 void rcu_irq_enter(void)
597 struct rcu_dynticks *rdtp;
600 local_irq_save(flags);
601 rdtp = &__get_cpu_var(rcu_dynticks);
602 oldval = rdtp->dynticks_nesting;
603 rdtp->dynticks_nesting++;
604 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
606 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
608 rcu_eqs_exit_common(rdtp, oldval, true);
609 rcu_sysidle_exit(rdtp, 1);
610 local_irq_restore(flags);
614 * rcu_nmi_enter - inform RCU of entry to NMI context
616 * If the CPU was idle with dynamic ticks active, and there is no
617 * irq handler running, this updates rdtp->dynticks_nmi to let the
618 * RCU grace-period handling know that the CPU is active.
620 void rcu_nmi_enter(void)
622 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
624 if (rdtp->dynticks_nmi_nesting == 0 &&
625 (atomic_read(&rdtp->dynticks) & 0x1))
627 rdtp->dynticks_nmi_nesting++;
628 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
629 atomic_inc(&rdtp->dynticks);
630 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
631 smp_mb__after_atomic_inc(); /* See above. */
632 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
636 * rcu_nmi_exit - inform RCU of exit from NMI context
638 * If the CPU was idle with dynamic ticks active, and there is no
639 * irq handler running, this updates rdtp->dynticks_nmi to let the
640 * RCU grace-period handling know that the CPU is no longer active.
642 void rcu_nmi_exit(void)
644 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
646 if (rdtp->dynticks_nmi_nesting == 0 ||
647 --rdtp->dynticks_nmi_nesting != 0)
649 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
650 smp_mb__before_atomic_inc(); /* See above. */
651 atomic_inc(&rdtp->dynticks);
652 smp_mb__after_atomic_inc(); /* Force delay to next write. */
653 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
657 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
659 * If the current CPU is in its idle loop and is neither in an interrupt
660 * or NMI handler, return true.
662 int rcu_is_cpu_idle(void)
667 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
671 EXPORT_SYMBOL(rcu_is_cpu_idle);
673 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
676 * Is the current CPU online? Disable preemption to avoid false positives
677 * that could otherwise happen due to the current CPU number being sampled,
678 * this task being preempted, its old CPU being taken offline, resuming
679 * on some other CPU, then determining that its old CPU is now offline.
680 * It is OK to use RCU on an offline processor during initial boot, hence
681 * the check for rcu_scheduler_fully_active. Note also that it is OK
682 * for a CPU coming online to use RCU for one jiffy prior to marking itself
683 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
684 * offline to continue to use RCU for one jiffy after marking itself
685 * offline in the cpu_online_mask. This leniency is necessary given the
686 * non-atomic nature of the online and offline processing, for example,
687 * the fact that a CPU enters the scheduler after completing the CPU_DYING
690 * This is also why RCU internally marks CPUs online during the
691 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
693 * Disable checking if in an NMI handler because we cannot safely report
694 * errors from NMI handlers anyway.
696 bool rcu_lockdep_current_cpu_online(void)
698 struct rcu_data *rdp;
699 struct rcu_node *rnp;
705 rdp = &__get_cpu_var(rcu_sched_data);
707 ret = (rdp->grpmask & rnp->qsmaskinit) ||
708 !rcu_scheduler_fully_active;
712 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
714 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
717 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
719 * If the current CPU is idle or running at a first-level (not nested)
720 * interrupt from idle, return true. The caller must have at least
721 * disabled preemption.
723 static int rcu_is_cpu_rrupt_from_idle(void)
725 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
729 * Snapshot the specified CPU's dynticks counter so that we can later
730 * credit them with an implicit quiescent state. Return 1 if this CPU
731 * is in dynticks idle mode, which is an extended quiescent state.
733 static int dyntick_save_progress_counter(struct rcu_data *rdp,
734 bool *isidle, unsigned long *maxj)
736 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
737 rcu_sysidle_check_cpu(rdp, isidle, maxj);
738 return (rdp->dynticks_snap & 0x1) == 0;
742 * Return true if the specified CPU has passed through a quiescent
743 * state by virtue of being in or having passed through an dynticks
744 * idle state since the last call to dyntick_save_progress_counter()
745 * for this same CPU, or by virtue of having been offline.
747 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
748 bool *isidle, unsigned long *maxj)
753 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
754 snap = (unsigned int)rdp->dynticks_snap;
757 * If the CPU passed through or entered a dynticks idle phase with
758 * no active irq/NMI handlers, then we can safely pretend that the CPU
759 * already acknowledged the request to pass through a quiescent
760 * state. Either way, that CPU cannot possibly be in an RCU
761 * read-side critical section that started before the beginning
762 * of the current RCU grace period.
764 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
765 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
771 * Check for the CPU being offline, but only if the grace period
772 * is old enough. We don't need to worry about the CPU changing
773 * state: If we see it offline even once, it has been through a
776 * The reason for insisting that the grace period be at least
777 * one jiffy old is that CPUs that are not quite online and that
778 * have just gone offline can still execute RCU read-side critical
781 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
782 return 0; /* Grace period is not old enough. */
784 if (cpu_is_offline(rdp->cpu)) {
785 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
791 * There is a possibility that a CPU in adaptive-ticks state
792 * might run in the kernel with the scheduling-clock tick disabled
793 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
794 * force the CPU to restart the scheduling-clock tick in this
795 * CPU is in this state.
797 rcu_kick_nohz_cpu(rdp->cpu);
802 static void record_gp_stall_check_time(struct rcu_state *rsp)
804 rsp->gp_start = jiffies;
805 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
809 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
810 * for architectures that do not implement trigger_all_cpu_backtrace().
811 * The NMI-triggered stack traces are more accurate because they are
812 * printed by the target CPU.
814 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
818 struct rcu_node *rnp;
820 rcu_for_each_leaf_node(rsp, rnp) {
821 raw_spin_lock_irqsave(&rnp->lock, flags);
822 if (rnp->qsmask != 0) {
823 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
824 if (rnp->qsmask & (1UL << cpu))
825 dump_cpu_task(rnp->grplo + cpu);
827 raw_spin_unlock_irqrestore(&rnp->lock, flags);
831 static void print_other_cpu_stall(struct rcu_state *rsp)
837 struct rcu_node *rnp = rcu_get_root(rsp);
840 /* Only let one CPU complain about others per time interval. */
842 raw_spin_lock_irqsave(&rnp->lock, flags);
843 delta = jiffies - rsp->jiffies_stall;
844 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
845 raw_spin_unlock_irqrestore(&rnp->lock, flags);
848 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
849 raw_spin_unlock_irqrestore(&rnp->lock, flags);
852 * OK, time to rat on our buddy...
853 * See Documentation/RCU/stallwarn.txt for info on how to debug
854 * RCU CPU stall warnings.
856 pr_err("INFO: %s detected stalls on CPUs/tasks:",
858 print_cpu_stall_info_begin();
859 rcu_for_each_leaf_node(rsp, rnp) {
860 raw_spin_lock_irqsave(&rnp->lock, flags);
861 ndetected += rcu_print_task_stall(rnp);
862 if (rnp->qsmask != 0) {
863 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
864 if (rnp->qsmask & (1UL << cpu)) {
865 print_cpu_stall_info(rsp,
870 raw_spin_unlock_irqrestore(&rnp->lock, flags);
874 * Now rat on any tasks that got kicked up to the root rcu_node
875 * due to CPU offlining.
877 rnp = rcu_get_root(rsp);
878 raw_spin_lock_irqsave(&rnp->lock, flags);
879 ndetected += rcu_print_task_stall(rnp);
880 raw_spin_unlock_irqrestore(&rnp->lock, flags);
882 print_cpu_stall_info_end();
883 for_each_possible_cpu(cpu)
884 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
885 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
886 smp_processor_id(), (long)(jiffies - rsp->gp_start),
887 rsp->gpnum, rsp->completed, totqlen);
889 pr_err("INFO: Stall ended before state dump start\n");
890 else if (!trigger_all_cpu_backtrace())
891 rcu_dump_cpu_stacks(rsp);
893 /* Complain about tasks blocking the grace period. */
895 rcu_print_detail_task_stall(rsp);
897 force_quiescent_state(rsp); /* Kick them all. */
900 static void print_cpu_stall(struct rcu_state *rsp)
904 struct rcu_node *rnp = rcu_get_root(rsp);
908 * OK, time to rat on ourselves...
909 * See Documentation/RCU/stallwarn.txt for info on how to debug
910 * RCU CPU stall warnings.
912 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
913 print_cpu_stall_info_begin();
914 print_cpu_stall_info(rsp, smp_processor_id());
915 print_cpu_stall_info_end();
916 for_each_possible_cpu(cpu)
917 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
918 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
919 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
920 if (!trigger_all_cpu_backtrace())
923 raw_spin_lock_irqsave(&rnp->lock, flags);
924 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
925 rsp->jiffies_stall = jiffies +
926 3 * rcu_jiffies_till_stall_check() + 3;
927 raw_spin_unlock_irqrestore(&rnp->lock, flags);
929 set_need_resched(); /* kick ourselves to get things going. */
932 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
936 struct rcu_node *rnp;
938 if (rcu_cpu_stall_suppress)
940 j = ACCESS_ONCE(jiffies);
941 js = ACCESS_ONCE(rsp->jiffies_stall);
943 if (rcu_gp_in_progress(rsp) &&
944 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
946 /* We haven't checked in, so go dump stack. */
947 print_cpu_stall(rsp);
949 } else if (rcu_gp_in_progress(rsp) &&
950 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
952 /* They had a few time units to dump stack, so complain. */
953 print_other_cpu_stall(rsp);
958 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
960 * Set the stall-warning timeout way off into the future, thus preventing
961 * any RCU CPU stall-warning messages from appearing in the current set of
964 * The caller must disable hard irqs.
966 void rcu_cpu_stall_reset(void)
968 struct rcu_state *rsp;
970 for_each_rcu_flavor(rsp)
971 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
975 * Initialize the specified rcu_data structure's callback list to empty.
977 static void init_callback_list(struct rcu_data *rdp)
981 if (init_nocb_callback_list(rdp))
984 for (i = 0; i < RCU_NEXT_SIZE; i++)
985 rdp->nxttail[i] = &rdp->nxtlist;
989 * Determine the value that ->completed will have at the end of the
990 * next subsequent grace period. This is used to tag callbacks so that
991 * a CPU can invoke callbacks in a timely fashion even if that CPU has
992 * been dyntick-idle for an extended period with callbacks under the
993 * influence of RCU_FAST_NO_HZ.
995 * The caller must hold rnp->lock with interrupts disabled.
997 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
998 struct rcu_node *rnp)
1001 * If RCU is idle, we just wait for the next grace period.
1002 * But we can only be sure that RCU is idle if we are looking
1003 * at the root rcu_node structure -- otherwise, a new grace
1004 * period might have started, but just not yet gotten around
1005 * to initializing the current non-root rcu_node structure.
1007 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1008 return rnp->completed + 1;
1011 * Otherwise, wait for a possible partial grace period and
1012 * then the subsequent full grace period.
1014 return rnp->completed + 2;
1018 * Trace-event helper function for rcu_start_future_gp() and
1019 * rcu_nocb_wait_gp().
1021 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1022 unsigned long c, const char *s)
1024 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1025 rnp->completed, c, rnp->level,
1026 rnp->grplo, rnp->grphi, s);
1030 * Start some future grace period, as needed to handle newly arrived
1031 * callbacks. The required future grace periods are recorded in each
1032 * rcu_node structure's ->need_future_gp field.
1034 * The caller must hold the specified rcu_node structure's ->lock.
1036 static unsigned long __maybe_unused
1037 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1041 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1044 * Pick up grace-period number for new callbacks. If this
1045 * grace period is already marked as needed, return to the caller.
1047 c = rcu_cbs_completed(rdp->rsp, rnp);
1048 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1049 if (rnp->need_future_gp[c & 0x1]) {
1050 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1055 * If either this rcu_node structure or the root rcu_node structure
1056 * believe that a grace period is in progress, then we must wait
1057 * for the one following, which is in "c". Because our request
1058 * will be noticed at the end of the current grace period, we don't
1059 * need to explicitly start one.
1061 if (rnp->gpnum != rnp->completed ||
1062 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1063 rnp->need_future_gp[c & 0x1]++;
1064 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1069 * There might be no grace period in progress. If we don't already
1070 * hold it, acquire the root rcu_node structure's lock in order to
1071 * start one (if needed).
1073 if (rnp != rnp_root)
1074 raw_spin_lock(&rnp_root->lock);
1077 * Get a new grace-period number. If there really is no grace
1078 * period in progress, it will be smaller than the one we obtained
1079 * earlier. Adjust callbacks as needed. Note that even no-CBs
1080 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1082 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1083 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1084 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1085 rdp->nxtcompleted[i] = c;
1088 * If the needed for the required grace period is already
1089 * recorded, trace and leave.
1091 if (rnp_root->need_future_gp[c & 0x1]) {
1092 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1096 /* Record the need for the future grace period. */
1097 rnp_root->need_future_gp[c & 0x1]++;
1099 /* If a grace period is not already in progress, start one. */
1100 if (rnp_root->gpnum != rnp_root->completed) {
1101 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1103 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1104 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1107 if (rnp != rnp_root)
1108 raw_spin_unlock(&rnp_root->lock);
1113 * Clean up any old requests for the just-ended grace period. Also return
1114 * whether any additional grace periods have been requested. Also invoke
1115 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1116 * waiting for this grace period to complete.
1118 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1120 int c = rnp->completed;
1122 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1124 rcu_nocb_gp_cleanup(rsp, rnp);
1125 rnp->need_future_gp[c & 0x1] = 0;
1126 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1127 trace_rcu_future_gp(rnp, rdp, c,
1128 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1133 * If there is room, assign a ->completed number to any callbacks on
1134 * this CPU that have not already been assigned. Also accelerate any
1135 * callbacks that were previously assigned a ->completed number that has
1136 * since proven to be too conservative, which can happen if callbacks get
1137 * assigned a ->completed number while RCU is idle, but with reference to
1138 * a non-root rcu_node structure. This function is idempotent, so it does
1139 * not hurt to call it repeatedly.
1141 * The caller must hold rnp->lock with interrupts disabled.
1143 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1144 struct rcu_data *rdp)
1149 /* If the CPU has no callbacks, nothing to do. */
1150 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1154 * Starting from the sublist containing the callbacks most
1155 * recently assigned a ->completed number and working down, find the
1156 * first sublist that is not assignable to an upcoming grace period.
1157 * Such a sublist has something in it (first two tests) and has
1158 * a ->completed number assigned that will complete sooner than
1159 * the ->completed number for newly arrived callbacks (last test).
1161 * The key point is that any later sublist can be assigned the
1162 * same ->completed number as the newly arrived callbacks, which
1163 * means that the callbacks in any of these later sublist can be
1164 * grouped into a single sublist, whether or not they have already
1165 * been assigned a ->completed number.
1167 c = rcu_cbs_completed(rsp, rnp);
1168 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1169 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1170 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1174 * If there are no sublist for unassigned callbacks, leave.
1175 * At the same time, advance "i" one sublist, so that "i" will
1176 * index into the sublist where all the remaining callbacks should
1179 if (++i >= RCU_NEXT_TAIL)
1183 * Assign all subsequent callbacks' ->completed number to the next
1184 * full grace period and group them all in the sublist initially
1187 for (; i <= RCU_NEXT_TAIL; i++) {
1188 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1189 rdp->nxtcompleted[i] = c;
1191 /* Record any needed additional grace periods. */
1192 rcu_start_future_gp(rnp, rdp);
1194 /* Trace depending on how much we were able to accelerate. */
1195 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1196 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1198 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1202 * Move any callbacks whose grace period has completed to the
1203 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1204 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1205 * sublist. This function is idempotent, so it does not hurt to
1206 * invoke it repeatedly. As long as it is not invoked -too- often...
1208 * The caller must hold rnp->lock with interrupts disabled.
1210 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1211 struct rcu_data *rdp)
1215 /* If the CPU has no callbacks, nothing to do. */
1216 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1220 * Find all callbacks whose ->completed numbers indicate that they
1221 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1223 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1224 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1226 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1228 /* Clean up any sublist tail pointers that were misordered above. */
1229 for (j = RCU_WAIT_TAIL; j < i; j++)
1230 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1232 /* Copy down callbacks to fill in empty sublists. */
1233 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1234 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1236 rdp->nxttail[j] = rdp->nxttail[i];
1237 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1240 /* Classify any remaining callbacks. */
1241 rcu_accelerate_cbs(rsp, rnp, rdp);
1245 * Update CPU-local rcu_data state to record the beginnings and ends of
1246 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1247 * structure corresponding to the current CPU, and must have irqs disabled.
1249 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1251 /* Handle the ends of any preceding grace periods first. */
1252 if (rdp->completed == rnp->completed) {
1254 /* No grace period end, so just accelerate recent callbacks. */
1255 rcu_accelerate_cbs(rsp, rnp, rdp);
1259 /* Advance callbacks. */
1260 rcu_advance_cbs(rsp, rnp, rdp);
1262 /* Remember that we saw this grace-period completion. */
1263 rdp->completed = rnp->completed;
1264 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1267 if (rdp->gpnum != rnp->gpnum) {
1269 * If the current grace period is waiting for this CPU,
1270 * set up to detect a quiescent state, otherwise don't
1271 * go looking for one.
1273 rdp->gpnum = rnp->gpnum;
1274 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1275 rdp->passed_quiesce = 0;
1276 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1277 zero_cpu_stall_ticks(rdp);
1281 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1283 unsigned long flags;
1284 struct rcu_node *rnp;
1286 local_irq_save(flags);
1288 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1289 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1290 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1291 local_irq_restore(flags);
1294 __note_gp_changes(rsp, rnp, rdp);
1295 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1299 * Initialize a new grace period.
1301 static int rcu_gp_init(struct rcu_state *rsp)
1303 struct rcu_data *rdp;
1304 struct rcu_node *rnp = rcu_get_root(rsp);
1306 rcu_bind_gp_kthread();
1307 raw_spin_lock_irq(&rnp->lock);
1308 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1310 if (rcu_gp_in_progress(rsp)) {
1311 /* Grace period already in progress, don't start another. */
1312 raw_spin_unlock_irq(&rnp->lock);
1316 /* Advance to a new grace period and initialize state. */
1318 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1319 record_gp_stall_check_time(rsp);
1320 raw_spin_unlock_irq(&rnp->lock);
1322 /* Exclude any concurrent CPU-hotplug operations. */
1323 mutex_lock(&rsp->onoff_mutex);
1326 * Set the quiescent-state-needed bits in all the rcu_node
1327 * structures for all currently online CPUs in breadth-first order,
1328 * starting from the root rcu_node structure, relying on the layout
1329 * of the tree within the rsp->node[] array. Note that other CPUs
1330 * will access only the leaves of the hierarchy, thus seeing that no
1331 * grace period is in progress, at least until the corresponding
1332 * leaf node has been initialized. In addition, we have excluded
1333 * CPU-hotplug operations.
1335 * The grace period cannot complete until the initialization
1336 * process finishes, because this kthread handles both.
1338 rcu_for_each_node_breadth_first(rsp, rnp) {
1339 raw_spin_lock_irq(&rnp->lock);
1340 rdp = this_cpu_ptr(rsp->rda);
1341 rcu_preempt_check_blocked_tasks(rnp);
1342 rnp->qsmask = rnp->qsmaskinit;
1343 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1344 WARN_ON_ONCE(rnp->completed != rsp->completed);
1345 ACCESS_ONCE(rnp->completed) = rsp->completed;
1346 if (rnp == rdp->mynode)
1347 __note_gp_changes(rsp, rnp, rdp);
1348 rcu_preempt_boost_start_gp(rnp);
1349 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1350 rnp->level, rnp->grplo,
1351 rnp->grphi, rnp->qsmask);
1352 raw_spin_unlock_irq(&rnp->lock);
1353 #ifdef CONFIG_PROVE_RCU_DELAY
1354 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1355 system_state == SYSTEM_RUNNING)
1357 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1361 mutex_unlock(&rsp->onoff_mutex);
1366 * Do one round of quiescent-state forcing.
1368 int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1370 int fqs_state = fqs_state_in;
1371 bool isidle = false;
1373 struct rcu_node *rnp = rcu_get_root(rsp);
1376 if (fqs_state == RCU_SAVE_DYNTICK) {
1377 /* Collect dyntick-idle snapshots. */
1378 if (is_sysidle_rcu_state(rsp)) {
1380 maxj = jiffies - ULONG_MAX / 4;
1382 force_qs_rnp(rsp, dyntick_save_progress_counter,
1384 rcu_sysidle_report_gp(rsp, isidle, maxj);
1385 fqs_state = RCU_FORCE_QS;
1387 /* Handle dyntick-idle and offline CPUs. */
1389 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1391 /* Clear flag to prevent immediate re-entry. */
1392 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1393 raw_spin_lock_irq(&rnp->lock);
1394 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1395 raw_spin_unlock_irq(&rnp->lock);
1401 * Clean up after the old grace period.
1403 static void rcu_gp_cleanup(struct rcu_state *rsp)
1405 unsigned long gp_duration;
1407 struct rcu_data *rdp;
1408 struct rcu_node *rnp = rcu_get_root(rsp);
1410 raw_spin_lock_irq(&rnp->lock);
1411 gp_duration = jiffies - rsp->gp_start;
1412 if (gp_duration > rsp->gp_max)
1413 rsp->gp_max = gp_duration;
1416 * We know the grace period is complete, but to everyone else
1417 * it appears to still be ongoing. But it is also the case
1418 * that to everyone else it looks like there is nothing that
1419 * they can do to advance the grace period. It is therefore
1420 * safe for us to drop the lock in order to mark the grace
1421 * period as completed in all of the rcu_node structures.
1423 raw_spin_unlock_irq(&rnp->lock);
1426 * Propagate new ->completed value to rcu_node structures so
1427 * that other CPUs don't have to wait until the start of the next
1428 * grace period to process their callbacks. This also avoids
1429 * some nasty RCU grace-period initialization races by forcing
1430 * the end of the current grace period to be completely recorded in
1431 * all of the rcu_node structures before the beginning of the next
1432 * grace period is recorded in any of the rcu_node structures.
1434 rcu_for_each_node_breadth_first(rsp, rnp) {
1435 raw_spin_lock_irq(&rnp->lock);
1436 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1437 rdp = this_cpu_ptr(rsp->rda);
1438 if (rnp == rdp->mynode)
1439 __note_gp_changes(rsp, rnp, rdp);
1440 nocb += rcu_future_gp_cleanup(rsp, rnp);
1441 raw_spin_unlock_irq(&rnp->lock);
1444 rnp = rcu_get_root(rsp);
1445 raw_spin_lock_irq(&rnp->lock);
1446 rcu_nocb_gp_set(rnp, nocb);
1448 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1449 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1450 rsp->fqs_state = RCU_GP_IDLE;
1451 rdp = this_cpu_ptr(rsp->rda);
1452 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1453 if (cpu_needs_another_gp(rsp, rdp))
1455 raw_spin_unlock_irq(&rnp->lock);
1459 * Body of kthread that handles grace periods.
1461 static int __noreturn rcu_gp_kthread(void *arg)
1466 struct rcu_state *rsp = arg;
1467 struct rcu_node *rnp = rcu_get_root(rsp);
1471 /* Handle grace-period start. */
1473 wait_event_interruptible(rsp->gp_wq,
1476 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1480 flush_signals(current);
1483 /* Handle quiescent-state forcing. */
1484 fqs_state = RCU_SAVE_DYNTICK;
1485 j = jiffies_till_first_fqs;
1488 jiffies_till_first_fqs = HZ;
1491 rsp->jiffies_force_qs = jiffies + j;
1492 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1493 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1494 (!ACCESS_ONCE(rnp->qsmask) &&
1495 !rcu_preempt_blocked_readers_cgp(rnp)),
1497 /* If grace period done, leave loop. */
1498 if (!ACCESS_ONCE(rnp->qsmask) &&
1499 !rcu_preempt_blocked_readers_cgp(rnp))
1501 /* If time for quiescent-state forcing, do it. */
1502 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1503 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1506 /* Deal with stray signal. */
1508 flush_signals(current);
1510 j = jiffies_till_next_fqs;
1513 jiffies_till_next_fqs = HZ;
1516 jiffies_till_next_fqs = 1;
1520 /* Handle grace-period end. */
1521 rcu_gp_cleanup(rsp);
1525 static void rsp_wakeup(struct irq_work *work)
1527 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1529 /* Wake up rcu_gp_kthread() to start the grace period. */
1530 wake_up(&rsp->gp_wq);
1534 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1535 * in preparation for detecting the next grace period. The caller must hold
1536 * the root node's ->lock and hard irqs must be disabled.
1538 * Note that it is legal for a dying CPU (which is marked as offline) to
1539 * invoke this function. This can happen when the dying CPU reports its
1543 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1544 struct rcu_data *rdp)
1546 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1548 * Either we have not yet spawned the grace-period
1549 * task, this CPU does not need another grace period,
1550 * or a grace period is already in progress.
1551 * Either way, don't start a new grace period.
1555 rsp->gp_flags = RCU_GP_FLAG_INIT;
1558 * We can't do wakeups while holding the rnp->lock, as that
1559 * could cause possible deadlocks with the rq->lock. Deter
1560 * the wakeup to interrupt context.
1562 irq_work_queue(&rsp->wakeup_work);
1566 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1567 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1568 * is invoked indirectly from rcu_advance_cbs(), which would result in
1569 * endless recursion -- or would do so if it wasn't for the self-deadlock
1570 * that is encountered beforehand.
1573 rcu_start_gp(struct rcu_state *rsp)
1575 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1576 struct rcu_node *rnp = rcu_get_root(rsp);
1579 * If there is no grace period in progress right now, any
1580 * callbacks we have up to this point will be satisfied by the
1581 * next grace period. Also, advancing the callbacks reduces the
1582 * probability of false positives from cpu_needs_another_gp()
1583 * resulting in pointless grace periods. So, advance callbacks
1584 * then start the grace period!
1586 rcu_advance_cbs(rsp, rnp, rdp);
1587 rcu_start_gp_advanced(rsp, rnp, rdp);
1591 * Report a full set of quiescent states to the specified rcu_state
1592 * data structure. This involves cleaning up after the prior grace
1593 * period and letting rcu_start_gp() start up the next grace period
1594 * if one is needed. Note that the caller must hold rnp->lock, which
1595 * is released before return.
1597 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1598 __releases(rcu_get_root(rsp)->lock)
1600 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1601 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1602 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1606 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1607 * Allows quiescent states for a group of CPUs to be reported at one go
1608 * to the specified rcu_node structure, though all the CPUs in the group
1609 * must be represented by the same rcu_node structure (which need not be
1610 * a leaf rcu_node structure, though it often will be). That structure's
1611 * lock must be held upon entry, and it is released before return.
1614 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1615 struct rcu_node *rnp, unsigned long flags)
1616 __releases(rnp->lock)
1618 struct rcu_node *rnp_c;
1620 /* Walk up the rcu_node hierarchy. */
1622 if (!(rnp->qsmask & mask)) {
1624 /* Our bit has already been cleared, so done. */
1625 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1628 rnp->qsmask &= ~mask;
1629 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1630 mask, rnp->qsmask, rnp->level,
1631 rnp->grplo, rnp->grphi,
1633 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1635 /* Other bits still set at this level, so done. */
1636 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1639 mask = rnp->grpmask;
1640 if (rnp->parent == NULL) {
1642 /* No more levels. Exit loop holding root lock. */
1646 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1649 raw_spin_lock_irqsave(&rnp->lock, flags);
1650 WARN_ON_ONCE(rnp_c->qsmask);
1654 * Get here if we are the last CPU to pass through a quiescent
1655 * state for this grace period. Invoke rcu_report_qs_rsp()
1656 * to clean up and start the next grace period if one is needed.
1658 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1662 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1663 * structure. This must be either called from the specified CPU, or
1664 * called when the specified CPU is known to be offline (and when it is
1665 * also known that no other CPU is concurrently trying to help the offline
1666 * CPU). The lastcomp argument is used to make sure we are still in the
1667 * grace period of interest. We don't want to end the current grace period
1668 * based on quiescent states detected in an earlier grace period!
1671 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1673 unsigned long flags;
1675 struct rcu_node *rnp;
1678 raw_spin_lock_irqsave(&rnp->lock, flags);
1679 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1680 rnp->completed == rnp->gpnum) {
1683 * The grace period in which this quiescent state was
1684 * recorded has ended, so don't report it upwards.
1685 * We will instead need a new quiescent state that lies
1686 * within the current grace period.
1688 rdp->passed_quiesce = 0; /* need qs for new gp. */
1689 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1692 mask = rdp->grpmask;
1693 if ((rnp->qsmask & mask) == 0) {
1694 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1696 rdp->qs_pending = 0;
1699 * This GP can't end until cpu checks in, so all of our
1700 * callbacks can be processed during the next GP.
1702 rcu_accelerate_cbs(rsp, rnp, rdp);
1704 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1709 * Check to see if there is a new grace period of which this CPU
1710 * is not yet aware, and if so, set up local rcu_data state for it.
1711 * Otherwise, see if this CPU has just passed through its first
1712 * quiescent state for this grace period, and record that fact if so.
1715 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1717 /* Check for grace-period ends and beginnings. */
1718 note_gp_changes(rsp, rdp);
1721 * Does this CPU still need to do its part for current grace period?
1722 * If no, return and let the other CPUs do their part as well.
1724 if (!rdp->qs_pending)
1728 * Was there a quiescent state since the beginning of the grace
1729 * period? If no, then exit and wait for the next call.
1731 if (!rdp->passed_quiesce)
1735 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1738 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1741 #ifdef CONFIG_HOTPLUG_CPU
1744 * Send the specified CPU's RCU callbacks to the orphanage. The
1745 * specified CPU must be offline, and the caller must hold the
1749 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1750 struct rcu_node *rnp, struct rcu_data *rdp)
1752 /* No-CBs CPUs do not have orphanable callbacks. */
1753 if (rcu_is_nocb_cpu(rdp->cpu))
1757 * Orphan the callbacks. First adjust the counts. This is safe
1758 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1759 * cannot be running now. Thus no memory barrier is required.
1761 if (rdp->nxtlist != NULL) {
1762 rsp->qlen_lazy += rdp->qlen_lazy;
1763 rsp->qlen += rdp->qlen;
1764 rdp->n_cbs_orphaned += rdp->qlen;
1766 ACCESS_ONCE(rdp->qlen) = 0;
1770 * Next, move those callbacks still needing a grace period to
1771 * the orphanage, where some other CPU will pick them up.
1772 * Some of the callbacks might have gone partway through a grace
1773 * period, but that is too bad. They get to start over because we
1774 * cannot assume that grace periods are synchronized across CPUs.
1775 * We don't bother updating the ->nxttail[] array yet, instead
1776 * we just reset the whole thing later on.
1778 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1779 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1780 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1781 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1785 * Then move the ready-to-invoke callbacks to the orphanage,
1786 * where some other CPU will pick them up. These will not be
1787 * required to pass though another grace period: They are done.
1789 if (rdp->nxtlist != NULL) {
1790 *rsp->orphan_donetail = rdp->nxtlist;
1791 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1794 /* Finally, initialize the rcu_data structure's list to empty. */
1795 init_callback_list(rdp);
1799 * Adopt the RCU callbacks from the specified rcu_state structure's
1800 * orphanage. The caller must hold the ->orphan_lock.
1802 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1805 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1807 /* No-CBs CPUs are handled specially. */
1808 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1811 /* Do the accounting first. */
1812 rdp->qlen_lazy += rsp->qlen_lazy;
1813 rdp->qlen += rsp->qlen;
1814 rdp->n_cbs_adopted += rsp->qlen;
1815 if (rsp->qlen_lazy != rsp->qlen)
1816 rcu_idle_count_callbacks_posted();
1821 * We do not need a memory barrier here because the only way we
1822 * can get here if there is an rcu_barrier() in flight is if
1823 * we are the task doing the rcu_barrier().
1826 /* First adopt the ready-to-invoke callbacks. */
1827 if (rsp->orphan_donelist != NULL) {
1828 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1829 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1830 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1831 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1832 rdp->nxttail[i] = rsp->orphan_donetail;
1833 rsp->orphan_donelist = NULL;
1834 rsp->orphan_donetail = &rsp->orphan_donelist;
1837 /* And then adopt the callbacks that still need a grace period. */
1838 if (rsp->orphan_nxtlist != NULL) {
1839 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1840 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1841 rsp->orphan_nxtlist = NULL;
1842 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1847 * Trace the fact that this CPU is going offline.
1849 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1851 RCU_TRACE(unsigned long mask);
1852 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1853 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1855 RCU_TRACE(mask = rdp->grpmask);
1856 trace_rcu_grace_period(rsp->name,
1857 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1862 * The CPU has been completely removed, and some other CPU is reporting
1863 * this fact from process context. Do the remainder of the cleanup,
1864 * including orphaning the outgoing CPU's RCU callbacks, and also
1865 * adopting them. There can only be one CPU hotplug operation at a time,
1866 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1868 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1870 unsigned long flags;
1872 int need_report = 0;
1873 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1874 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1876 /* Adjust any no-longer-needed kthreads. */
1877 rcu_boost_kthread_setaffinity(rnp, -1);
1879 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1881 /* Exclude any attempts to start a new grace period. */
1882 mutex_lock(&rsp->onoff_mutex);
1883 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1885 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1886 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1887 rcu_adopt_orphan_cbs(rsp);
1889 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1890 mask = rdp->grpmask; /* rnp->grplo is constant. */
1892 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1893 rnp->qsmaskinit &= ~mask;
1894 if (rnp->qsmaskinit != 0) {
1895 if (rnp != rdp->mynode)
1896 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1899 if (rnp == rdp->mynode)
1900 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1902 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1903 mask = rnp->grpmask;
1905 } while (rnp != NULL);
1908 * We still hold the leaf rcu_node structure lock here, and
1909 * irqs are still disabled. The reason for this subterfuge is
1910 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1911 * held leads to deadlock.
1913 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1915 if (need_report & RCU_OFL_TASKS_NORM_GP)
1916 rcu_report_unblock_qs_rnp(rnp, flags);
1918 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1919 if (need_report & RCU_OFL_TASKS_EXP_GP)
1920 rcu_report_exp_rnp(rsp, rnp, true);
1921 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1922 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1923 cpu, rdp->qlen, rdp->nxtlist);
1924 init_callback_list(rdp);
1925 /* Disallow further callbacks on this CPU. */
1926 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1927 mutex_unlock(&rsp->onoff_mutex);
1930 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1932 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1936 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1940 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1943 * Invoke any RCU callbacks that have made it to the end of their grace
1944 * period. Thottle as specified by rdp->blimit.
1946 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1948 unsigned long flags;
1949 struct rcu_head *next, *list, **tail;
1950 long bl, count, count_lazy;
1953 /* If no callbacks are ready, just return. */
1954 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1955 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1956 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1957 need_resched(), is_idle_task(current),
1958 rcu_is_callbacks_kthread());
1963 * Extract the list of ready callbacks, disabling to prevent
1964 * races with call_rcu() from interrupt handlers.
1966 local_irq_save(flags);
1967 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1969 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1970 list = rdp->nxtlist;
1971 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1972 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1973 tail = rdp->nxttail[RCU_DONE_TAIL];
1974 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1975 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1976 rdp->nxttail[i] = &rdp->nxtlist;
1977 local_irq_restore(flags);
1979 /* Invoke callbacks. */
1980 count = count_lazy = 0;
1984 debug_rcu_head_unqueue(list);
1985 if (__rcu_reclaim(rsp->name, list))
1988 /* Stop only if limit reached and CPU has something to do. */
1989 if (++count >= bl &&
1991 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1995 local_irq_save(flags);
1996 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1997 is_idle_task(current),
1998 rcu_is_callbacks_kthread());
2000 /* Update count, and requeue any remaining callbacks. */
2002 *tail = rdp->nxtlist;
2003 rdp->nxtlist = list;
2004 for (i = 0; i < RCU_NEXT_SIZE; i++)
2005 if (&rdp->nxtlist == rdp->nxttail[i])
2006 rdp->nxttail[i] = tail;
2010 smp_mb(); /* List handling before counting for rcu_barrier(). */
2011 rdp->qlen_lazy -= count_lazy;
2012 ACCESS_ONCE(rdp->qlen) -= count;
2013 rdp->n_cbs_invoked += count;
2015 /* Reinstate batch limit if we have worked down the excess. */
2016 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2017 rdp->blimit = blimit;
2019 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2020 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2021 rdp->qlen_last_fqs_check = 0;
2022 rdp->n_force_qs_snap = rsp->n_force_qs;
2023 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2024 rdp->qlen_last_fqs_check = rdp->qlen;
2025 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2027 local_irq_restore(flags);
2029 /* Re-invoke RCU core processing if there are callbacks remaining. */
2030 if (cpu_has_callbacks_ready_to_invoke(rdp))
2035 * Check to see if this CPU is in a non-context-switch quiescent state
2036 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2037 * Also schedule RCU core processing.
2039 * This function must be called from hardirq context. It is normally
2040 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2041 * false, there is no point in invoking rcu_check_callbacks().
2043 void rcu_check_callbacks(int cpu, int user)
2045 trace_rcu_utilization(TPS("Start scheduler-tick"));
2046 increment_cpu_stall_ticks();
2047 if (user || rcu_is_cpu_rrupt_from_idle()) {
2050 * Get here if this CPU took its interrupt from user
2051 * mode or from the idle loop, and if this is not a
2052 * nested interrupt. In this case, the CPU is in
2053 * a quiescent state, so note it.
2055 * No memory barrier is required here because both
2056 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2057 * variables that other CPUs neither access nor modify,
2058 * at least not while the corresponding CPU is online.
2064 } else if (!in_softirq()) {
2067 * Get here if this CPU did not take its interrupt from
2068 * softirq, in other words, if it is not interrupting
2069 * a rcu_bh read-side critical section. This is an _bh
2070 * critical section, so note it.
2075 rcu_preempt_check_callbacks(cpu);
2076 if (rcu_pending(cpu))
2078 trace_rcu_utilization(TPS("End scheduler-tick"));
2082 * Scan the leaf rcu_node structures, processing dyntick state for any that
2083 * have not yet encountered a quiescent state, using the function specified.
2084 * Also initiate boosting for any threads blocked on the root rcu_node.
2086 * The caller must have suppressed start of new grace periods.
2088 static void force_qs_rnp(struct rcu_state *rsp,
2089 int (*f)(struct rcu_data *rsp, bool *isidle,
2090 unsigned long *maxj),
2091 bool *isidle, unsigned long *maxj)
2095 unsigned long flags;
2097 struct rcu_node *rnp;
2099 rcu_for_each_leaf_node(rsp, rnp) {
2102 raw_spin_lock_irqsave(&rnp->lock, flags);
2103 if (!rcu_gp_in_progress(rsp)) {
2104 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2107 if (rnp->qsmask == 0) {
2108 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2113 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2114 if ((rnp->qsmask & bit) != 0) {
2115 if ((rnp->qsmaskinit & bit) != 0)
2117 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2123 /* rcu_report_qs_rnp() releases rnp->lock. */
2124 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2127 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2129 rnp = rcu_get_root(rsp);
2130 if (rnp->qsmask == 0) {
2131 raw_spin_lock_irqsave(&rnp->lock, flags);
2132 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2137 * Force quiescent states on reluctant CPUs, and also detect which
2138 * CPUs are in dyntick-idle mode.
2140 static void force_quiescent_state(struct rcu_state *rsp)
2142 unsigned long flags;
2144 struct rcu_node *rnp;
2145 struct rcu_node *rnp_old = NULL;
2147 /* Funnel through hierarchy to reduce memory contention. */
2148 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2149 for (; rnp != NULL; rnp = rnp->parent) {
2150 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2151 !raw_spin_trylock(&rnp->fqslock);
2152 if (rnp_old != NULL)
2153 raw_spin_unlock(&rnp_old->fqslock);
2155 rsp->n_force_qs_lh++;
2160 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2162 /* Reached the root of the rcu_node tree, acquire lock. */
2163 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2164 raw_spin_unlock(&rnp_old->fqslock);
2165 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2166 rsp->n_force_qs_lh++;
2167 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2168 return; /* Someone beat us to it. */
2170 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2171 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2172 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2176 * This does the RCU core processing work for the specified rcu_state
2177 * and rcu_data structures. This may be called only from the CPU to
2178 * whom the rdp belongs.
2181 __rcu_process_callbacks(struct rcu_state *rsp)
2183 unsigned long flags;
2184 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2186 WARN_ON_ONCE(rdp->beenonline == 0);
2188 /* Update RCU state based on any recent quiescent states. */
2189 rcu_check_quiescent_state(rsp, rdp);
2191 /* Does this CPU require a not-yet-started grace period? */
2192 local_irq_save(flags);
2193 if (cpu_needs_another_gp(rsp, rdp)) {
2194 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2196 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2198 local_irq_restore(flags);
2201 /* If there are callbacks ready, invoke them. */
2202 if (cpu_has_callbacks_ready_to_invoke(rdp))
2203 invoke_rcu_callbacks(rsp, rdp);
2207 * Do RCU core processing for the current CPU.
2209 static void rcu_process_callbacks(struct softirq_action *unused)
2211 struct rcu_state *rsp;
2213 if (cpu_is_offline(smp_processor_id()))
2215 trace_rcu_utilization(TPS("Start RCU core"));
2216 for_each_rcu_flavor(rsp)
2217 __rcu_process_callbacks(rsp);
2218 trace_rcu_utilization(TPS("End RCU core"));
2222 * Schedule RCU callback invocation. If the specified type of RCU
2223 * does not support RCU priority boosting, just do a direct call,
2224 * otherwise wake up the per-CPU kernel kthread. Note that because we
2225 * are running on the current CPU with interrupts disabled, the
2226 * rcu_cpu_kthread_task cannot disappear out from under us.
2228 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2230 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2232 if (likely(!rsp->boost)) {
2233 rcu_do_batch(rsp, rdp);
2236 invoke_rcu_callbacks_kthread();
2239 static void invoke_rcu_core(void)
2241 if (cpu_online(smp_processor_id()))
2242 raise_softirq(RCU_SOFTIRQ);
2246 * Handle any core-RCU processing required by a call_rcu() invocation.
2248 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2249 struct rcu_head *head, unsigned long flags)
2252 * If called from an extended quiescent state, invoke the RCU
2253 * core in order to force a re-evaluation of RCU's idleness.
2255 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2258 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2259 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2263 * Force the grace period if too many callbacks or too long waiting.
2264 * Enforce hysteresis, and don't invoke force_quiescent_state()
2265 * if some other CPU has recently done so. Also, don't bother
2266 * invoking force_quiescent_state() if the newly enqueued callback
2267 * is the only one waiting for a grace period to complete.
2269 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2271 /* Are we ignoring a completed grace period? */
2272 note_gp_changes(rsp, rdp);
2274 /* Start a new grace period if one not already started. */
2275 if (!rcu_gp_in_progress(rsp)) {
2276 struct rcu_node *rnp_root = rcu_get_root(rsp);
2278 raw_spin_lock(&rnp_root->lock);
2280 raw_spin_unlock(&rnp_root->lock);
2282 /* Give the grace period a kick. */
2283 rdp->blimit = LONG_MAX;
2284 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2285 *rdp->nxttail[RCU_DONE_TAIL] != head)
2286 force_quiescent_state(rsp);
2287 rdp->n_force_qs_snap = rsp->n_force_qs;
2288 rdp->qlen_last_fqs_check = rdp->qlen;
2294 * Helper function for call_rcu() and friends. The cpu argument will
2295 * normally be -1, indicating "currently running CPU". It may specify
2296 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2297 * is expected to specify a CPU.
2300 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2301 struct rcu_state *rsp, int cpu, bool lazy)
2303 unsigned long flags;
2304 struct rcu_data *rdp;
2306 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2307 debug_rcu_head_queue(head);
2312 * Opportunistically note grace-period endings and beginnings.
2313 * Note that we might see a beginning right after we see an
2314 * end, but never vice versa, since this CPU has to pass through
2315 * a quiescent state betweentimes.
2317 local_irq_save(flags);
2318 rdp = this_cpu_ptr(rsp->rda);
2320 /* Add the callback to our list. */
2321 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2325 rdp = per_cpu_ptr(rsp->rda, cpu);
2326 offline = !__call_rcu_nocb(rdp, head, lazy);
2327 WARN_ON_ONCE(offline);
2328 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2329 local_irq_restore(flags);
2332 ACCESS_ONCE(rdp->qlen)++;
2336 rcu_idle_count_callbacks_posted();
2337 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2338 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2339 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2341 if (__is_kfree_rcu_offset((unsigned long)func))
2342 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2343 rdp->qlen_lazy, rdp->qlen);
2345 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2347 /* Go handle any RCU core processing required. */
2348 __call_rcu_core(rsp, rdp, head, flags);
2349 local_irq_restore(flags);
2353 * Queue an RCU-sched callback for invocation after a grace period.
2355 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2357 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2359 EXPORT_SYMBOL_GPL(call_rcu_sched);
2362 * Queue an RCU callback for invocation after a quicker grace period.
2364 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2366 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2368 EXPORT_SYMBOL_GPL(call_rcu_bh);
2371 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2372 * any blocking grace-period wait automatically implies a grace period
2373 * if there is only one CPU online at any point time during execution
2374 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2375 * occasionally incorrectly indicate that there are multiple CPUs online
2376 * when there was in fact only one the whole time, as this just adds
2377 * some overhead: RCU still operates correctly.
2379 static inline int rcu_blocking_is_gp(void)
2383 might_sleep(); /* Check for RCU read-side critical section. */
2385 ret = num_online_cpus() <= 1;
2391 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2393 * Control will return to the caller some time after a full rcu-sched
2394 * grace period has elapsed, in other words after all currently executing
2395 * rcu-sched read-side critical sections have completed. These read-side
2396 * critical sections are delimited by rcu_read_lock_sched() and
2397 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2398 * local_irq_disable(), and so on may be used in place of
2399 * rcu_read_lock_sched().
2401 * This means that all preempt_disable code sequences, including NMI and
2402 * non-threaded hardware-interrupt handlers, in progress on entry will
2403 * have completed before this primitive returns. However, this does not
2404 * guarantee that softirq handlers will have completed, since in some
2405 * kernels, these handlers can run in process context, and can block.
2407 * Note that this guarantee implies further memory-ordering guarantees.
2408 * On systems with more than one CPU, when synchronize_sched() returns,
2409 * each CPU is guaranteed to have executed a full memory barrier since the
2410 * end of its last RCU-sched read-side critical section whose beginning
2411 * preceded the call to synchronize_sched(). In addition, each CPU having
2412 * an RCU read-side critical section that extends beyond the return from
2413 * synchronize_sched() is guaranteed to have executed a full memory barrier
2414 * after the beginning of synchronize_sched() and before the beginning of
2415 * that RCU read-side critical section. Note that these guarantees include
2416 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2417 * that are executing in the kernel.
2419 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2420 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2421 * to have executed a full memory barrier during the execution of
2422 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2423 * again only if the system has more than one CPU).
2425 * This primitive provides the guarantees made by the (now removed)
2426 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2427 * guarantees that rcu_read_lock() sections will have completed.
2428 * In "classic RCU", these two guarantees happen to be one and
2429 * the same, but can differ in realtime RCU implementations.
2431 void synchronize_sched(void)
2433 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2434 !lock_is_held(&rcu_lock_map) &&
2435 !lock_is_held(&rcu_sched_lock_map),
2436 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2437 if (rcu_blocking_is_gp())
2440 synchronize_sched_expedited();
2442 wait_rcu_gp(call_rcu_sched);
2444 EXPORT_SYMBOL_GPL(synchronize_sched);
2447 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2449 * Control will return to the caller some time after a full rcu_bh grace
2450 * period has elapsed, in other words after all currently executing rcu_bh
2451 * read-side critical sections have completed. RCU read-side critical
2452 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2453 * and may be nested.
2455 * See the description of synchronize_sched() for more detailed information
2456 * on memory ordering guarantees.
2458 void synchronize_rcu_bh(void)
2460 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2461 !lock_is_held(&rcu_lock_map) &&
2462 !lock_is_held(&rcu_sched_lock_map),
2463 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2464 if (rcu_blocking_is_gp())
2467 synchronize_rcu_bh_expedited();
2469 wait_rcu_gp(call_rcu_bh);
2471 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2473 static int synchronize_sched_expedited_cpu_stop(void *data)
2476 * There must be a full memory barrier on each affected CPU
2477 * between the time that try_stop_cpus() is called and the
2478 * time that it returns.
2480 * In the current initial implementation of cpu_stop, the
2481 * above condition is already met when the control reaches
2482 * this point and the following smp_mb() is not strictly
2483 * necessary. Do smp_mb() anyway for documentation and
2484 * robustness against future implementation changes.
2486 smp_mb(); /* See above comment block. */
2491 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2493 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2494 * approach to force the grace period to end quickly. This consumes
2495 * significant time on all CPUs and is unfriendly to real-time workloads,
2496 * so is thus not recommended for any sort of common-case code. In fact,
2497 * if you are using synchronize_sched_expedited() in a loop, please
2498 * restructure your code to batch your updates, and then use a single
2499 * synchronize_sched() instead.
2501 * Note that it is illegal to call this function while holding any lock
2502 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2503 * to call this function from a CPU-hotplug notifier. Failing to observe
2504 * these restriction will result in deadlock.
2506 * This implementation can be thought of as an application of ticket
2507 * locking to RCU, with sync_sched_expedited_started and
2508 * sync_sched_expedited_done taking on the roles of the halves
2509 * of the ticket-lock word. Each task atomically increments
2510 * sync_sched_expedited_started upon entry, snapshotting the old value,
2511 * then attempts to stop all the CPUs. If this succeeds, then each
2512 * CPU will have executed a context switch, resulting in an RCU-sched
2513 * grace period. We are then done, so we use atomic_cmpxchg() to
2514 * update sync_sched_expedited_done to match our snapshot -- but
2515 * only if someone else has not already advanced past our snapshot.
2517 * On the other hand, if try_stop_cpus() fails, we check the value
2518 * of sync_sched_expedited_done. If it has advanced past our
2519 * initial snapshot, then someone else must have forced a grace period
2520 * some time after we took our snapshot. In this case, our work is
2521 * done for us, and we can simply return. Otherwise, we try again,
2522 * but keep our initial snapshot for purposes of checking for someone
2523 * doing our work for us.
2525 * If we fail too many times in a row, we fall back to synchronize_sched().
2527 void synchronize_sched_expedited(void)
2529 long firstsnap, s, snap;
2531 struct rcu_state *rsp = &rcu_sched_state;
2534 * If we are in danger of counter wrap, just do synchronize_sched().
2535 * By allowing sync_sched_expedited_started to advance no more than
2536 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2537 * that more than 3.5 billion CPUs would be required to force a
2538 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2539 * course be required on a 64-bit system.
2541 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2542 (ulong)atomic_long_read(&rsp->expedited_done) +
2544 synchronize_sched();
2545 atomic_long_inc(&rsp->expedited_wrap);
2550 * Take a ticket. Note that atomic_inc_return() implies a
2551 * full memory barrier.
2553 snap = atomic_long_inc_return(&rsp->expedited_start);
2556 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2559 * Each pass through the following loop attempts to force a
2560 * context switch on each CPU.
2562 while (try_stop_cpus(cpu_online_mask,
2563 synchronize_sched_expedited_cpu_stop,
2566 atomic_long_inc(&rsp->expedited_tryfail);
2568 /* Check to see if someone else did our work for us. */
2569 s = atomic_long_read(&rsp->expedited_done);
2570 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2571 /* ensure test happens before caller kfree */
2572 smp_mb__before_atomic_inc(); /* ^^^ */
2573 atomic_long_inc(&rsp->expedited_workdone1);
2577 /* No joy, try again later. Or just synchronize_sched(). */
2578 if (trycount++ < 10) {
2579 udelay(trycount * num_online_cpus());
2581 wait_rcu_gp(call_rcu_sched);
2582 atomic_long_inc(&rsp->expedited_normal);
2586 /* Recheck to see if someone else did our work for us. */
2587 s = atomic_long_read(&rsp->expedited_done);
2588 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2589 /* ensure test happens before caller kfree */
2590 smp_mb__before_atomic_inc(); /* ^^^ */
2591 atomic_long_inc(&rsp->expedited_workdone2);
2596 * Refetching sync_sched_expedited_started allows later
2597 * callers to piggyback on our grace period. We retry
2598 * after they started, so our grace period works for them,
2599 * and they started after our first try, so their grace
2600 * period works for us.
2603 snap = atomic_long_read(&rsp->expedited_start);
2604 smp_mb(); /* ensure read is before try_stop_cpus(). */
2606 atomic_long_inc(&rsp->expedited_stoppedcpus);
2609 * Everyone up to our most recent fetch is covered by our grace
2610 * period. Update the counter, but only if our work is still
2611 * relevant -- which it won't be if someone who started later
2612 * than we did already did their update.
2615 atomic_long_inc(&rsp->expedited_done_tries);
2616 s = atomic_long_read(&rsp->expedited_done);
2617 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2618 /* ensure test happens before caller kfree */
2619 smp_mb__before_atomic_inc(); /* ^^^ */
2620 atomic_long_inc(&rsp->expedited_done_lost);
2623 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2624 atomic_long_inc(&rsp->expedited_done_exit);
2628 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2631 * Check to see if there is any immediate RCU-related work to be done
2632 * by the current CPU, for the specified type of RCU, returning 1 if so.
2633 * The checks are in order of increasing expense: checks that can be
2634 * carried out against CPU-local state are performed first. However,
2635 * we must check for CPU stalls first, else we might not get a chance.
2637 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2639 struct rcu_node *rnp = rdp->mynode;
2641 rdp->n_rcu_pending++;
2643 /* Check for CPU stalls, if enabled. */
2644 check_cpu_stall(rsp, rdp);
2646 /* Is the RCU core waiting for a quiescent state from this CPU? */
2647 if (rcu_scheduler_fully_active &&
2648 rdp->qs_pending && !rdp->passed_quiesce) {
2649 rdp->n_rp_qs_pending++;
2650 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2651 rdp->n_rp_report_qs++;
2655 /* Does this CPU have callbacks ready to invoke? */
2656 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2657 rdp->n_rp_cb_ready++;
2661 /* Has RCU gone idle with this CPU needing another grace period? */
2662 if (cpu_needs_another_gp(rsp, rdp)) {
2663 rdp->n_rp_cpu_needs_gp++;
2667 /* Has another RCU grace period completed? */
2668 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2669 rdp->n_rp_gp_completed++;
2673 /* Has a new RCU grace period started? */
2674 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2675 rdp->n_rp_gp_started++;
2680 rdp->n_rp_need_nothing++;
2685 * Check to see if there is any immediate RCU-related work to be done
2686 * by the current CPU, returning 1 if so. This function is part of the
2687 * RCU implementation; it is -not- an exported member of the RCU API.
2689 static int rcu_pending(int cpu)
2691 struct rcu_state *rsp;
2693 for_each_rcu_flavor(rsp)
2694 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2700 * Return true if the specified CPU has any callback. If all_lazy is
2701 * non-NULL, store an indication of whether all callbacks are lazy.
2702 * (If there are no callbacks, all of them are deemed to be lazy.)
2704 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2708 struct rcu_data *rdp;
2709 struct rcu_state *rsp;
2711 for_each_rcu_flavor(rsp) {
2712 rdp = per_cpu_ptr(rsp->rda, cpu);
2713 if (rdp->qlen != rdp->qlen_lazy)
2724 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2725 * the compiler is expected to optimize this away.
2727 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2728 int cpu, unsigned long done)
2730 trace_rcu_barrier(rsp->name, s, cpu,
2731 atomic_read(&rsp->barrier_cpu_count), done);
2735 * RCU callback function for _rcu_barrier(). If we are last, wake
2736 * up the task executing _rcu_barrier().
2738 static void rcu_barrier_callback(struct rcu_head *rhp)
2740 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2741 struct rcu_state *rsp = rdp->rsp;
2743 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2744 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2745 complete(&rsp->barrier_completion);
2747 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2752 * Called with preemption disabled, and from cross-cpu IRQ context.
2754 static void rcu_barrier_func(void *type)
2756 struct rcu_state *rsp = type;
2757 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2759 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2760 atomic_inc(&rsp->barrier_cpu_count);
2761 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2765 * Orchestrate the specified type of RCU barrier, waiting for all
2766 * RCU callbacks of the specified type to complete.
2768 static void _rcu_barrier(struct rcu_state *rsp)
2771 struct rcu_data *rdp;
2772 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2773 unsigned long snap_done;
2775 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2777 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2778 mutex_lock(&rsp->barrier_mutex);
2781 * Ensure that all prior references, including to ->n_barrier_done,
2782 * are ordered before the _rcu_barrier() machinery.
2784 smp_mb(); /* See above block comment. */
2787 * Recheck ->n_barrier_done to see if others did our work for us.
2788 * This means checking ->n_barrier_done for an even-to-odd-to-even
2789 * transition. The "if" expression below therefore rounds the old
2790 * value up to the next even number and adds two before comparing.
2792 snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2793 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2794 if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2795 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2796 smp_mb(); /* caller's subsequent code after above check. */
2797 mutex_unlock(&rsp->barrier_mutex);
2802 * Increment ->n_barrier_done to avoid duplicate work. Use
2803 * ACCESS_ONCE() to prevent the compiler from speculating
2804 * the increment to precede the early-exit check.
2806 ACCESS_ONCE(rsp->n_barrier_done)++;
2807 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2808 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2809 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2812 * Initialize the count to one rather than to zero in order to
2813 * avoid a too-soon return to zero in case of a short grace period
2814 * (or preemption of this task). Exclude CPU-hotplug operations
2815 * to ensure that no offline CPU has callbacks queued.
2817 init_completion(&rsp->barrier_completion);
2818 atomic_set(&rsp->barrier_cpu_count, 1);
2822 * Force each CPU with callbacks to register a new callback.
2823 * When that callback is invoked, we will know that all of the
2824 * corresponding CPU's preceding callbacks have been invoked.
2826 for_each_possible_cpu(cpu) {
2827 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2829 rdp = per_cpu_ptr(rsp->rda, cpu);
2830 if (rcu_is_nocb_cpu(cpu)) {
2831 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2832 rsp->n_barrier_done);
2833 atomic_inc(&rsp->barrier_cpu_count);
2834 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2836 } else if (ACCESS_ONCE(rdp->qlen)) {
2837 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2838 rsp->n_barrier_done);
2839 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2841 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2842 rsp->n_barrier_done);
2848 * Now that we have an rcu_barrier_callback() callback on each
2849 * CPU, and thus each counted, remove the initial count.
2851 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2852 complete(&rsp->barrier_completion);
2854 /* Increment ->n_barrier_done to prevent duplicate work. */
2855 smp_mb(); /* Keep increment after above mechanism. */
2856 ACCESS_ONCE(rsp->n_barrier_done)++;
2857 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2858 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2859 smp_mb(); /* Keep increment before caller's subsequent code. */
2861 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2862 wait_for_completion(&rsp->barrier_completion);
2864 /* Other rcu_barrier() invocations can now safely proceed. */
2865 mutex_unlock(&rsp->barrier_mutex);
2869 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2871 void rcu_barrier_bh(void)
2873 _rcu_barrier(&rcu_bh_state);
2875 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2878 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2880 void rcu_barrier_sched(void)
2882 _rcu_barrier(&rcu_sched_state);
2884 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2887 * Do boot-time initialization of a CPU's per-CPU RCU data.
2890 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2892 unsigned long flags;
2893 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2894 struct rcu_node *rnp = rcu_get_root(rsp);
2896 /* Set up local state, ensuring consistent view of global state. */
2897 raw_spin_lock_irqsave(&rnp->lock, flags);
2898 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2899 init_callback_list(rdp);
2901 ACCESS_ONCE(rdp->qlen) = 0;
2902 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2903 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2904 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2907 rcu_boot_init_nocb_percpu_data(rdp);
2908 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2912 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2913 * offline event can be happening at a given time. Note also that we
2914 * can accept some slop in the rsp->completed access due to the fact
2915 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2918 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2920 unsigned long flags;
2922 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2923 struct rcu_node *rnp = rcu_get_root(rsp);
2925 /* Exclude new grace periods. */
2926 mutex_lock(&rsp->onoff_mutex);
2928 /* Set up local state, ensuring consistent view of global state. */
2929 raw_spin_lock_irqsave(&rnp->lock, flags);
2930 rdp->beenonline = 1; /* We have now been online. */
2931 rdp->preemptible = preemptible;
2932 rdp->qlen_last_fqs_check = 0;
2933 rdp->n_force_qs_snap = rsp->n_force_qs;
2934 rdp->blimit = blimit;
2935 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
2936 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2937 rcu_sysidle_init_percpu_data(rdp->dynticks);
2938 atomic_set(&rdp->dynticks->dynticks,
2939 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2940 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2942 /* Add CPU to rcu_node bitmasks. */
2944 mask = rdp->grpmask;
2946 /* Exclude any attempts to start a new GP on small systems. */
2947 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2948 rnp->qsmaskinit |= mask;
2949 mask = rnp->grpmask;
2950 if (rnp == rdp->mynode) {
2952 * If there is a grace period in progress, we will
2953 * set up to wait for it next time we run the
2956 rdp->gpnum = rnp->completed;
2957 rdp->completed = rnp->completed;
2958 rdp->passed_quiesce = 0;
2959 rdp->qs_pending = 0;
2960 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
2962 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2964 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2965 local_irq_restore(flags);
2967 mutex_unlock(&rsp->onoff_mutex);
2970 static void rcu_prepare_cpu(int cpu)
2972 struct rcu_state *rsp;
2974 for_each_rcu_flavor(rsp)
2975 rcu_init_percpu_data(cpu, rsp,
2976 strcmp(rsp->name, "rcu_preempt") == 0);
2980 * Handle CPU online/offline notification events.
2982 static int rcu_cpu_notify(struct notifier_block *self,
2983 unsigned long action, void *hcpu)
2985 long cpu = (long)hcpu;
2986 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2987 struct rcu_node *rnp = rdp->mynode;
2988 struct rcu_state *rsp;
2990 trace_rcu_utilization(TPS("Start CPU hotplug"));
2992 case CPU_UP_PREPARE:
2993 case CPU_UP_PREPARE_FROZEN:
2994 rcu_prepare_cpu(cpu);
2995 rcu_prepare_kthreads(cpu);
2998 case CPU_DOWN_FAILED:
2999 rcu_boost_kthread_setaffinity(rnp, -1);
3001 case CPU_DOWN_PREPARE:
3002 rcu_boost_kthread_setaffinity(rnp, cpu);
3005 case CPU_DYING_FROZEN:
3006 for_each_rcu_flavor(rsp)
3007 rcu_cleanup_dying_cpu(rsp);
3010 case CPU_DEAD_FROZEN:
3011 case CPU_UP_CANCELED:
3012 case CPU_UP_CANCELED_FROZEN:
3013 for_each_rcu_flavor(rsp)
3014 rcu_cleanup_dead_cpu(cpu, rsp);
3019 trace_rcu_utilization(TPS("End CPU hotplug"));
3024 * Spawn the kthread that handles this RCU flavor's grace periods.
3026 static int __init rcu_spawn_gp_kthread(void)
3028 unsigned long flags;
3029 struct rcu_node *rnp;
3030 struct rcu_state *rsp;
3031 struct task_struct *t;
3033 for_each_rcu_flavor(rsp) {
3034 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3036 rnp = rcu_get_root(rsp);
3037 raw_spin_lock_irqsave(&rnp->lock, flags);
3038 rsp->gp_kthread = t;
3039 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3040 rcu_spawn_nocb_kthreads(rsp);
3044 early_initcall(rcu_spawn_gp_kthread);
3047 * This function is invoked towards the end of the scheduler's initialization
3048 * process. Before this is called, the idle task might contain
3049 * RCU read-side critical sections (during which time, this idle
3050 * task is booting the system). After this function is called, the
3051 * idle tasks are prohibited from containing RCU read-side critical
3052 * sections. This function also enables RCU lockdep checking.
3054 void rcu_scheduler_starting(void)
3056 WARN_ON(num_online_cpus() != 1);
3057 WARN_ON(nr_context_switches() > 0);
3058 rcu_scheduler_active = 1;
3062 * Compute the per-level fanout, either using the exact fanout specified
3063 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3065 #ifdef CONFIG_RCU_FANOUT_EXACT
3066 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3070 for (i = rcu_num_lvls - 1; i > 0; i--)
3071 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3072 rsp->levelspread[0] = rcu_fanout_leaf;
3074 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3075 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3082 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3083 ccur = rsp->levelcnt[i];
3084 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3088 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3091 * Helper function for rcu_init() that initializes one rcu_state structure.
3093 static void __init rcu_init_one(struct rcu_state *rsp,
3094 struct rcu_data __percpu *rda)
3096 static char *buf[] = { "rcu_node_0",
3099 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3100 static char *fqs[] = { "rcu_node_fqs_0",
3103 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3107 struct rcu_node *rnp;
3109 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3111 /* Silence gcc 4.8 warning about array index out of range. */
3112 if (rcu_num_lvls > RCU_NUM_LVLS)
3113 panic("rcu_init_one: rcu_num_lvls overflow");
3115 /* Initialize the level-tracking arrays. */
3117 for (i = 0; i < rcu_num_lvls; i++)
3118 rsp->levelcnt[i] = num_rcu_lvl[i];
3119 for (i = 1; i < rcu_num_lvls; i++)
3120 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3121 rcu_init_levelspread(rsp);
3123 /* Initialize the elements themselves, starting from the leaves. */
3125 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3126 cpustride *= rsp->levelspread[i];
3127 rnp = rsp->level[i];
3128 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3129 raw_spin_lock_init(&rnp->lock);
3130 lockdep_set_class_and_name(&rnp->lock,
3131 &rcu_node_class[i], buf[i]);
3132 raw_spin_lock_init(&rnp->fqslock);
3133 lockdep_set_class_and_name(&rnp->fqslock,
3134 &rcu_fqs_class[i], fqs[i]);
3135 rnp->gpnum = rsp->gpnum;
3136 rnp->completed = rsp->completed;
3138 rnp->qsmaskinit = 0;
3139 rnp->grplo = j * cpustride;
3140 rnp->grphi = (j + 1) * cpustride - 1;
3141 if (rnp->grphi >= NR_CPUS)
3142 rnp->grphi = NR_CPUS - 1;
3148 rnp->grpnum = j % rsp->levelspread[i - 1];
3149 rnp->grpmask = 1UL << rnp->grpnum;
3150 rnp->parent = rsp->level[i - 1] +
3151 j / rsp->levelspread[i - 1];
3154 INIT_LIST_HEAD(&rnp->blkd_tasks);
3155 rcu_init_one_nocb(rnp);
3160 init_waitqueue_head(&rsp->gp_wq);
3161 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3162 rnp = rsp->level[rcu_num_lvls - 1];
3163 for_each_possible_cpu(i) {
3164 while (i > rnp->grphi)
3166 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3167 rcu_boot_init_percpu_data(i, rsp);
3169 list_add(&rsp->flavors, &rcu_struct_flavors);
3173 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3174 * replace the definitions in rcutree.h because those are needed to size
3175 * the ->node array in the rcu_state structure.
3177 static void __init rcu_init_geometry(void)
3183 int rcu_capacity[MAX_RCU_LVLS + 1];
3186 * Initialize any unspecified boot parameters.
3187 * The default values of jiffies_till_first_fqs and
3188 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3189 * value, which is a function of HZ, then adding one for each
3190 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3192 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3193 if (jiffies_till_first_fqs == ULONG_MAX)
3194 jiffies_till_first_fqs = d;
3195 if (jiffies_till_next_fqs == ULONG_MAX)
3196 jiffies_till_next_fqs = d;
3198 /* If the compile-time values are accurate, just leave. */
3199 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3200 nr_cpu_ids == NR_CPUS)
3204 * Compute number of nodes that can be handled an rcu_node tree
3205 * with the given number of levels. Setting rcu_capacity[0] makes
3206 * some of the arithmetic easier.
3208 rcu_capacity[0] = 1;
3209 rcu_capacity[1] = rcu_fanout_leaf;
3210 for (i = 2; i <= MAX_RCU_LVLS; i++)
3211 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3214 * The boot-time rcu_fanout_leaf parameter is only permitted
3215 * to increase the leaf-level fanout, not decrease it. Of course,
3216 * the leaf-level fanout cannot exceed the number of bits in
3217 * the rcu_node masks. Finally, the tree must be able to accommodate
3218 * the configured number of CPUs. Complain and fall back to the
3219 * compile-time values if these limits are exceeded.
3221 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3222 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3223 n > rcu_capacity[MAX_RCU_LVLS]) {
3228 /* Calculate the number of rcu_nodes at each level of the tree. */
3229 for (i = 1; i <= MAX_RCU_LVLS; i++)
3230 if (n <= rcu_capacity[i]) {
3231 for (j = 0; j <= i; j++)
3233 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3235 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3240 /* Calculate the total number of rcu_node structures. */
3242 for (i = 0; i <= MAX_RCU_LVLS; i++)
3243 rcu_num_nodes += num_rcu_lvl[i];
3247 void __init rcu_init(void)
3251 rcu_bootup_announce();
3252 rcu_init_geometry();
3253 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3254 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3255 __rcu_init_preempt();
3256 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3259 * We don't need protection against CPU-hotplug here because
3260 * this is called early in boot, before either interrupts
3261 * or the scheduler are operational.
3263 cpu_notifier(rcu_cpu_notify, 0);
3264 for_each_online_cpu(cpu)
3265 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3268 #include "rcutree_plugin.h"