2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
82 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83 static char sname##_varname[] = #sname; \
84 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85 struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
99 DEFINE_PER_CPU(struct rcu_data, sname##_data)
101 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
102 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
104 static struct rcu_state *rcu_state_p;
105 LIST_HEAD(rcu_struct_flavors);
107 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
109 module_param(rcu_fanout_leaf, int, 0444);
110 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
111 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
118 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
129 int rcu_scheduler_active __read_mostly;
130 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
144 static int rcu_scheduler_fully_active __read_mostly;
146 #ifdef CONFIG_RCU_BOOST
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
152 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
153 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
155 DEFINE_PER_CPU(char, rcu_cpu_has_work);
157 #endif /* #ifdef CONFIG_RCU_BOOST */
159 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
172 unsigned long rcutorture_testseq;
173 unsigned long rcutorture_vernum;
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
180 static int rcu_gp_in_progress(struct rcu_state *rsp)
182 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
191 void rcu_sched_qs(void)
193 if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) {
194 trace_rcu_grace_period(TPS("rcu_sched"),
195 __this_cpu_read(rcu_sched_data.gpnum),
197 __this_cpu_write(rcu_sched_data.passed_quiesce, 1);
203 if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) {
204 trace_rcu_grace_period(TPS("rcu_bh"),
205 __this_cpu_read(rcu_bh_data.gpnum),
207 __this_cpu_write(rcu_bh_data.passed_quiesce, 1);
211 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
213 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
214 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
215 .dynticks = ATOMIC_INIT(1),
216 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
217 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
218 .dynticks_idle = ATOMIC_INIT(1),
219 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
223 * Let the RCU core know that this CPU has gone through the scheduler,
224 * which is a quiescent state. This is called when the need for a
225 * quiescent state is urgent, so we burn an atomic operation and full
226 * memory barriers to let the RCU core know about it, regardless of what
227 * this CPU might (or might not) do in the near future.
229 * We inform the RCU core by emulating a zero-duration dyntick-idle
230 * period, which we in turn do by incrementing the ->dynticks counter
233 static void rcu_momentary_dyntick_idle(void)
236 struct rcu_data *rdp;
237 struct rcu_dynticks *rdtp;
239 struct rcu_state *rsp;
241 local_irq_save(flags);
244 * Yes, we can lose flag-setting operations. This is OK, because
245 * the flag will be set again after some delay.
247 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
248 raw_cpu_write(rcu_sched_qs_mask, 0);
250 /* Find the flavor that needs a quiescent state. */
251 for_each_rcu_flavor(rsp) {
252 rdp = raw_cpu_ptr(rsp->rda);
253 if (!(resched_mask & rsp->flavor_mask))
255 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
256 if (ACCESS_ONCE(rdp->mynode->completed) !=
257 ACCESS_ONCE(rdp->cond_resched_completed))
261 * Pretend to be momentarily idle for the quiescent state.
262 * This allows the grace-period kthread to record the
263 * quiescent state, with no need for this CPU to do anything
266 rdtp = this_cpu_ptr(&rcu_dynticks);
267 smp_mb__before_atomic(); /* Earlier stuff before QS. */
268 atomic_add(2, &rdtp->dynticks); /* QS. */
269 smp_mb__after_atomic(); /* Later stuff after QS. */
272 local_irq_restore(flags);
276 * Note a context switch. This is a quiescent state for RCU-sched,
277 * and requires special handling for preemptible RCU.
278 * The caller must have disabled preemption.
280 void rcu_note_context_switch(int cpu)
282 trace_rcu_utilization(TPS("Start context switch"));
284 rcu_preempt_note_context_switch(cpu);
285 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
286 rcu_momentary_dyntick_idle();
287 trace_rcu_utilization(TPS("End context switch"));
289 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
291 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
292 static long qhimark = 10000; /* If this many pending, ignore blimit. */
293 static long qlowmark = 100; /* Once only this many pending, use blimit. */
295 module_param(blimit, long, 0444);
296 module_param(qhimark, long, 0444);
297 module_param(qlowmark, long, 0444);
299 static ulong jiffies_till_first_fqs = ULONG_MAX;
300 static ulong jiffies_till_next_fqs = ULONG_MAX;
302 module_param(jiffies_till_first_fqs, ulong, 0644);
303 module_param(jiffies_till_next_fqs, ulong, 0644);
306 * How long the grace period must be before we start recruiting
307 * quiescent-state help from rcu_note_context_switch().
309 static ulong jiffies_till_sched_qs = HZ / 20;
310 module_param(jiffies_till_sched_qs, ulong, 0644);
312 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
313 struct rcu_data *rdp);
314 static void force_qs_rnp(struct rcu_state *rsp,
315 int (*f)(struct rcu_data *rsp, bool *isidle,
316 unsigned long *maxj),
317 bool *isidle, unsigned long *maxj);
318 static void force_quiescent_state(struct rcu_state *rsp);
319 static int rcu_pending(int cpu);
322 * Return the number of RCU-sched batches processed thus far for debug & stats.
324 long rcu_batches_completed_sched(void)
326 return rcu_sched_state.completed;
328 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
331 * Return the number of RCU BH batches processed thus far for debug & stats.
333 long rcu_batches_completed_bh(void)
335 return rcu_bh_state.completed;
337 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
340 * Force a quiescent state.
342 void rcu_force_quiescent_state(void)
344 force_quiescent_state(rcu_state_p);
346 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
349 * Force a quiescent state for RCU BH.
351 void rcu_bh_force_quiescent_state(void)
353 force_quiescent_state(&rcu_bh_state);
355 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
358 * Show the state of the grace-period kthreads.
360 void show_rcu_gp_kthreads(void)
362 struct rcu_state *rsp;
364 for_each_rcu_flavor(rsp) {
365 pr_info("%s: wait state: %d ->state: %#lx\n",
366 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
367 /* sched_show_task(rsp->gp_kthread); */
370 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
373 * Record the number of times rcutorture tests have been initiated and
374 * terminated. This information allows the debugfs tracing stats to be
375 * correlated to the rcutorture messages, even when the rcutorture module
376 * is being repeatedly loaded and unloaded. In other words, we cannot
377 * store this state in rcutorture itself.
379 void rcutorture_record_test_transition(void)
381 rcutorture_testseq++;
382 rcutorture_vernum = 0;
384 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
387 * Send along grace-period-related data for rcutorture diagnostics.
389 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
390 unsigned long *gpnum, unsigned long *completed)
392 struct rcu_state *rsp = NULL;
401 case RCU_SCHED_FLAVOR:
402 rsp = &rcu_sched_state;
408 *flags = ACCESS_ONCE(rsp->gp_flags);
409 *gpnum = ACCESS_ONCE(rsp->gpnum);
410 *completed = ACCESS_ONCE(rsp->completed);
417 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
420 * Record the number of writer passes through the current rcutorture test.
421 * This is also used to correlate debugfs tracing stats with the rcutorture
424 void rcutorture_record_progress(unsigned long vernum)
428 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
431 * Force a quiescent state for RCU-sched.
433 void rcu_sched_force_quiescent_state(void)
435 force_quiescent_state(&rcu_sched_state);
437 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
440 * Does the CPU have callbacks ready to be invoked?
443 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
445 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
446 rdp->nxttail[RCU_DONE_TAIL] != NULL;
450 * Return the root node of the specified rcu_state structure.
452 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
454 return &rsp->node[0];
458 * Is there any need for future grace periods?
459 * Interrupts must be disabled. If the caller does not hold the root
460 * rnp_node structure's ->lock, the results are advisory only.
462 static int rcu_future_needs_gp(struct rcu_state *rsp)
464 struct rcu_node *rnp = rcu_get_root(rsp);
465 int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1;
466 int *fp = &rnp->need_future_gp[idx];
468 return ACCESS_ONCE(*fp);
472 * Does the current CPU require a not-yet-started grace period?
473 * The caller must have disabled interrupts to prevent races with
474 * normal callback registry.
477 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
481 if (rcu_gp_in_progress(rsp))
482 return 0; /* No, a grace period is already in progress. */
483 if (rcu_future_needs_gp(rsp))
484 return 1; /* Yes, a no-CBs CPU needs one. */
485 if (!rdp->nxttail[RCU_NEXT_TAIL])
486 return 0; /* No, this is a no-CBs (or offline) CPU. */
487 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
488 return 1; /* Yes, this CPU has newly registered callbacks. */
489 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
490 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
491 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
492 rdp->nxtcompleted[i]))
493 return 1; /* Yes, CBs for future grace period. */
494 return 0; /* No grace period needed. */
498 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
500 * If the new value of the ->dynticks_nesting counter now is zero,
501 * we really have entered idle, and must do the appropriate accounting.
502 * The caller must have disabled interrupts.
504 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
507 struct rcu_state *rsp;
508 struct rcu_data *rdp;
510 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
511 if (!user && !is_idle_task(current)) {
512 struct task_struct *idle __maybe_unused =
513 idle_task(smp_processor_id());
515 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
516 ftrace_dump(DUMP_ORIG);
517 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
518 current->pid, current->comm,
519 idle->pid, idle->comm); /* must be idle task! */
521 for_each_rcu_flavor(rsp) {
522 rdp = this_cpu_ptr(rsp->rda);
523 do_nocb_deferred_wakeup(rdp);
525 rcu_prepare_for_idle(smp_processor_id());
526 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
527 smp_mb__before_atomic(); /* See above. */
528 atomic_inc(&rdtp->dynticks);
529 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
530 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
531 rcu_dynticks_task_enter();
534 * It is illegal to enter an extended quiescent state while
535 * in an RCU read-side critical section.
537 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
538 "Illegal idle entry in RCU read-side critical section.");
539 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
540 "Illegal idle entry in RCU-bh read-side critical section.");
541 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
542 "Illegal idle entry in RCU-sched read-side critical section.");
546 * Enter an RCU extended quiescent state, which can be either the
547 * idle loop or adaptive-tickless usermode execution.
549 static void rcu_eqs_enter(bool user)
552 struct rcu_dynticks *rdtp;
554 rdtp = this_cpu_ptr(&rcu_dynticks);
555 oldval = rdtp->dynticks_nesting;
556 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
557 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
558 rdtp->dynticks_nesting = 0;
559 rcu_eqs_enter_common(rdtp, oldval, user);
561 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
566 * rcu_idle_enter - inform RCU that current CPU is entering idle
568 * Enter idle mode, in other words, -leave- the mode in which RCU
569 * read-side critical sections can occur. (Though RCU read-side
570 * critical sections can occur in irq handlers in idle, a possibility
571 * handled by irq_enter() and irq_exit().)
573 * We crowbar the ->dynticks_nesting field to zero to allow for
574 * the possibility of usermode upcalls having messed up our count
575 * of interrupt nesting level during the prior busy period.
577 void rcu_idle_enter(void)
581 local_irq_save(flags);
582 rcu_eqs_enter(false);
583 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
584 local_irq_restore(flags);
586 EXPORT_SYMBOL_GPL(rcu_idle_enter);
588 #ifdef CONFIG_RCU_USER_QS
590 * rcu_user_enter - inform RCU that we are resuming userspace.
592 * Enter RCU idle mode right before resuming userspace. No use of RCU
593 * is permitted between this call and rcu_user_exit(). This way the
594 * CPU doesn't need to maintain the tick for RCU maintenance purposes
595 * when the CPU runs in userspace.
597 void rcu_user_enter(void)
601 #endif /* CONFIG_RCU_USER_QS */
604 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
606 * Exit from an interrupt handler, which might possibly result in entering
607 * idle mode, in other words, leaving the mode in which read-side critical
608 * sections can occur.
610 * This code assumes that the idle loop never does anything that might
611 * result in unbalanced calls to irq_enter() and irq_exit(). If your
612 * architecture violates this assumption, RCU will give you what you
613 * deserve, good and hard. But very infrequently and irreproducibly.
615 * Use things like work queues to work around this limitation.
617 * You have been warned.
619 void rcu_irq_exit(void)
623 struct rcu_dynticks *rdtp;
625 local_irq_save(flags);
626 rdtp = this_cpu_ptr(&rcu_dynticks);
627 oldval = rdtp->dynticks_nesting;
628 rdtp->dynticks_nesting--;
629 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
630 if (rdtp->dynticks_nesting)
631 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
633 rcu_eqs_enter_common(rdtp, oldval, true);
634 rcu_sysidle_enter(rdtp, 1);
635 local_irq_restore(flags);
639 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
641 * If the new value of the ->dynticks_nesting counter was previously zero,
642 * we really have exited idle, and must do the appropriate accounting.
643 * The caller must have disabled interrupts.
645 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
648 rcu_dynticks_task_exit();
649 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
650 atomic_inc(&rdtp->dynticks);
651 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
652 smp_mb__after_atomic(); /* See above. */
653 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
654 rcu_cleanup_after_idle(smp_processor_id());
655 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
656 if (!user && !is_idle_task(current)) {
657 struct task_struct *idle __maybe_unused =
658 idle_task(smp_processor_id());
660 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
661 oldval, rdtp->dynticks_nesting);
662 ftrace_dump(DUMP_ORIG);
663 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
664 current->pid, current->comm,
665 idle->pid, idle->comm); /* must be idle task! */
670 * Exit an RCU extended quiescent state, which can be either the
671 * idle loop or adaptive-tickless usermode execution.
673 static void rcu_eqs_exit(bool user)
675 struct rcu_dynticks *rdtp;
678 rdtp = this_cpu_ptr(&rcu_dynticks);
679 oldval = rdtp->dynticks_nesting;
680 WARN_ON_ONCE(oldval < 0);
681 if (oldval & DYNTICK_TASK_NEST_MASK) {
682 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
684 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
685 rcu_eqs_exit_common(rdtp, oldval, user);
690 * rcu_idle_exit - inform RCU that current CPU is leaving idle
692 * Exit idle mode, in other words, -enter- the mode in which RCU
693 * read-side critical sections can occur.
695 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
696 * allow for the possibility of usermode upcalls messing up our count
697 * of interrupt nesting level during the busy period that is just
700 void rcu_idle_exit(void)
704 local_irq_save(flags);
706 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
707 local_irq_restore(flags);
709 EXPORT_SYMBOL_GPL(rcu_idle_exit);
711 #ifdef CONFIG_RCU_USER_QS
713 * rcu_user_exit - inform RCU that we are exiting userspace.
715 * Exit RCU idle mode while entering the kernel because it can
716 * run a RCU read side critical section anytime.
718 void rcu_user_exit(void)
722 #endif /* CONFIG_RCU_USER_QS */
725 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
727 * Enter an interrupt handler, which might possibly result in exiting
728 * idle mode, in other words, entering the mode in which read-side critical
729 * sections can occur.
731 * Note that the Linux kernel is fully capable of entering an interrupt
732 * handler that it never exits, for example when doing upcalls to
733 * user mode! This code assumes that the idle loop never does upcalls to
734 * user mode. If your architecture does do upcalls from the idle loop (or
735 * does anything else that results in unbalanced calls to the irq_enter()
736 * and irq_exit() functions), RCU will give you what you deserve, good
737 * and hard. But very infrequently and irreproducibly.
739 * Use things like work queues to work around this limitation.
741 * You have been warned.
743 void rcu_irq_enter(void)
746 struct rcu_dynticks *rdtp;
749 local_irq_save(flags);
750 rdtp = this_cpu_ptr(&rcu_dynticks);
751 oldval = rdtp->dynticks_nesting;
752 rdtp->dynticks_nesting++;
753 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
755 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
757 rcu_eqs_exit_common(rdtp, oldval, true);
758 rcu_sysidle_exit(rdtp, 1);
759 local_irq_restore(flags);
763 * rcu_nmi_enter - inform RCU of entry to NMI context
765 * If the CPU was idle with dynamic ticks active, and there is no
766 * irq handler running, this updates rdtp->dynticks_nmi to let the
767 * RCU grace-period handling know that the CPU is active.
769 void rcu_nmi_enter(void)
771 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
773 if (rdtp->dynticks_nmi_nesting == 0 &&
774 (atomic_read(&rdtp->dynticks) & 0x1))
776 rdtp->dynticks_nmi_nesting++;
777 smp_mb__before_atomic(); /* Force delay from prior write. */
778 atomic_inc(&rdtp->dynticks);
779 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
780 smp_mb__after_atomic(); /* See above. */
781 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
785 * rcu_nmi_exit - inform RCU of exit from NMI context
787 * If the CPU was idle with dynamic ticks active, and there is no
788 * irq handler running, this updates rdtp->dynticks_nmi to let the
789 * RCU grace-period handling know that the CPU is no longer active.
791 void rcu_nmi_exit(void)
793 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
795 if (rdtp->dynticks_nmi_nesting == 0 ||
796 --rdtp->dynticks_nmi_nesting != 0)
798 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
799 smp_mb__before_atomic(); /* See above. */
800 atomic_inc(&rdtp->dynticks);
801 smp_mb__after_atomic(); /* Force delay to next write. */
802 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
806 * __rcu_is_watching - are RCU read-side critical sections safe?
808 * Return true if RCU is watching the running CPU, which means that
809 * this CPU can safely enter RCU read-side critical sections. Unlike
810 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
811 * least disabled preemption.
813 bool notrace __rcu_is_watching(void)
815 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
819 * rcu_is_watching - see if RCU thinks that the current CPU is idle
821 * If the current CPU is in its idle loop and is neither in an interrupt
822 * or NMI handler, return true.
824 bool notrace rcu_is_watching(void)
829 ret = __rcu_is_watching();
833 EXPORT_SYMBOL_GPL(rcu_is_watching);
835 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
838 * Is the current CPU online? Disable preemption to avoid false positives
839 * that could otherwise happen due to the current CPU number being sampled,
840 * this task being preempted, its old CPU being taken offline, resuming
841 * on some other CPU, then determining that its old CPU is now offline.
842 * It is OK to use RCU on an offline processor during initial boot, hence
843 * the check for rcu_scheduler_fully_active. Note also that it is OK
844 * for a CPU coming online to use RCU for one jiffy prior to marking itself
845 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
846 * offline to continue to use RCU for one jiffy after marking itself
847 * offline in the cpu_online_mask. This leniency is necessary given the
848 * non-atomic nature of the online and offline processing, for example,
849 * the fact that a CPU enters the scheduler after completing the CPU_DYING
852 * This is also why RCU internally marks CPUs online during the
853 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
855 * Disable checking if in an NMI handler because we cannot safely report
856 * errors from NMI handlers anyway.
858 bool rcu_lockdep_current_cpu_online(void)
860 struct rcu_data *rdp;
861 struct rcu_node *rnp;
867 rdp = this_cpu_ptr(&rcu_sched_data);
869 ret = (rdp->grpmask & rnp->qsmaskinit) ||
870 !rcu_scheduler_fully_active;
874 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
876 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
879 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
881 * If the current CPU is idle or running at a first-level (not nested)
882 * interrupt from idle, return true. The caller must have at least
883 * disabled preemption.
885 static int rcu_is_cpu_rrupt_from_idle(void)
887 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
891 * Snapshot the specified CPU's dynticks counter so that we can later
892 * credit them with an implicit quiescent state. Return 1 if this CPU
893 * is in dynticks idle mode, which is an extended quiescent state.
895 static int dyntick_save_progress_counter(struct rcu_data *rdp,
896 bool *isidle, unsigned long *maxj)
898 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
899 rcu_sysidle_check_cpu(rdp, isidle, maxj);
900 if ((rdp->dynticks_snap & 0x1) == 0) {
901 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
909 * This function really isn't for public consumption, but RCU is special in
910 * that context switches can allow the state machine to make progress.
912 extern void resched_cpu(int cpu);
915 * Return true if the specified CPU has passed through a quiescent
916 * state by virtue of being in or having passed through an dynticks
917 * idle state since the last call to dyntick_save_progress_counter()
918 * for this same CPU, or by virtue of having been offline.
920 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
921 bool *isidle, unsigned long *maxj)
927 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
928 snap = (unsigned int)rdp->dynticks_snap;
931 * If the CPU passed through or entered a dynticks idle phase with
932 * no active irq/NMI handlers, then we can safely pretend that the CPU
933 * already acknowledged the request to pass through a quiescent
934 * state. Either way, that CPU cannot possibly be in an RCU
935 * read-side critical section that started before the beginning
936 * of the current RCU grace period.
938 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
939 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
945 * Check for the CPU being offline, but only if the grace period
946 * is old enough. We don't need to worry about the CPU changing
947 * state: If we see it offline even once, it has been through a
950 * The reason for insisting that the grace period be at least
951 * one jiffy old is that CPUs that are not quite online and that
952 * have just gone offline can still execute RCU read-side critical
955 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
956 return 0; /* Grace period is not old enough. */
958 if (cpu_is_offline(rdp->cpu)) {
959 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
965 * A CPU running for an extended time within the kernel can
966 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
967 * even context-switching back and forth between a pair of
968 * in-kernel CPU-bound tasks cannot advance grace periods.
969 * So if the grace period is old enough, make the CPU pay attention.
970 * Note that the unsynchronized assignments to the per-CPU
971 * rcu_sched_qs_mask variable are safe. Yes, setting of
972 * bits can be lost, but they will be set again on the next
973 * force-quiescent-state pass. So lost bit sets do not result
974 * in incorrect behavior, merely in a grace period lasting
975 * a few jiffies longer than it might otherwise. Because
976 * there are at most four threads involved, and because the
977 * updates are only once every few jiffies, the probability of
978 * lossage (and thus of slight grace-period extension) is
981 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
982 * is set too high, we override with half of the RCU CPU stall
985 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
986 if (ULONG_CMP_GE(jiffies,
987 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
988 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
989 if (!(ACCESS_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
990 ACCESS_ONCE(rdp->cond_resched_completed) =
991 ACCESS_ONCE(rdp->mynode->completed);
992 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
993 ACCESS_ONCE(*rcrmp) =
994 ACCESS_ONCE(*rcrmp) + rdp->rsp->flavor_mask;
995 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
996 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
997 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
998 /* Time to beat on that CPU again! */
999 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1000 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1007 static void record_gp_stall_check_time(struct rcu_state *rsp)
1009 unsigned long j = jiffies;
1013 smp_wmb(); /* Record start time before stall time. */
1014 j1 = rcu_jiffies_till_stall_check();
1015 ACCESS_ONCE(rsp->jiffies_stall) = j + j1;
1016 rsp->jiffies_resched = j + j1 / 2;
1020 * Dump stacks of all tasks running on stalled CPUs.
1022 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1025 unsigned long flags;
1026 struct rcu_node *rnp;
1028 rcu_for_each_leaf_node(rsp, rnp) {
1029 raw_spin_lock_irqsave(&rnp->lock, flags);
1030 if (rnp->qsmask != 0) {
1031 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1032 if (rnp->qsmask & (1UL << cpu))
1033 dump_cpu_task(rnp->grplo + cpu);
1035 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1039 static void print_other_cpu_stall(struct rcu_state *rsp)
1043 unsigned long flags;
1045 struct rcu_node *rnp = rcu_get_root(rsp);
1048 /* Only let one CPU complain about others per time interval. */
1050 raw_spin_lock_irqsave(&rnp->lock, flags);
1051 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
1052 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1053 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1056 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1057 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1060 * OK, time to rat on our buddy...
1061 * See Documentation/RCU/stallwarn.txt for info on how to debug
1062 * RCU CPU stall warnings.
1064 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1066 print_cpu_stall_info_begin();
1067 rcu_for_each_leaf_node(rsp, rnp) {
1068 raw_spin_lock_irqsave(&rnp->lock, flags);
1069 ndetected += rcu_print_task_stall(rnp);
1070 if (rnp->qsmask != 0) {
1071 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1072 if (rnp->qsmask & (1UL << cpu)) {
1073 print_cpu_stall_info(rsp,
1078 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1082 * Now rat on any tasks that got kicked up to the root rcu_node
1083 * due to CPU offlining.
1085 rnp = rcu_get_root(rsp);
1086 raw_spin_lock_irqsave(&rnp->lock, flags);
1087 ndetected += rcu_print_task_stall(rnp);
1088 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1090 print_cpu_stall_info_end();
1091 for_each_possible_cpu(cpu)
1092 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1093 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1094 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1095 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1097 pr_err("INFO: Stall ended before state dump start\n");
1099 rcu_dump_cpu_stacks(rsp);
1101 /* Complain about tasks blocking the grace period. */
1103 rcu_print_detail_task_stall(rsp);
1105 force_quiescent_state(rsp); /* Kick them all. */
1108 static void print_cpu_stall(struct rcu_state *rsp)
1111 unsigned long flags;
1112 struct rcu_node *rnp = rcu_get_root(rsp);
1116 * OK, time to rat on ourselves...
1117 * See Documentation/RCU/stallwarn.txt for info on how to debug
1118 * RCU CPU stall warnings.
1120 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1121 print_cpu_stall_info_begin();
1122 print_cpu_stall_info(rsp, smp_processor_id());
1123 print_cpu_stall_info_end();
1124 for_each_possible_cpu(cpu)
1125 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1126 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1127 jiffies - rsp->gp_start,
1128 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1129 rcu_dump_cpu_stacks(rsp);
1131 raw_spin_lock_irqsave(&rnp->lock, flags);
1132 if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall)))
1133 ACCESS_ONCE(rsp->jiffies_stall) = jiffies +
1134 3 * rcu_jiffies_till_stall_check() + 3;
1135 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1138 * Attempt to revive the RCU machinery by forcing a context switch.
1140 * A context switch would normally allow the RCU state machine to make
1141 * progress and it could be we're stuck in kernel space without context
1142 * switches for an entirely unreasonable amount of time.
1144 resched_cpu(smp_processor_id());
1147 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1149 unsigned long completed;
1150 unsigned long gpnum;
1154 struct rcu_node *rnp;
1156 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1161 * Lots of memory barriers to reject false positives.
1163 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1164 * then rsp->gp_start, and finally rsp->completed. These values
1165 * are updated in the opposite order with memory barriers (or
1166 * equivalent) during grace-period initialization and cleanup.
1167 * Now, a false positive can occur if we get an new value of
1168 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1169 * the memory barriers, the only way that this can happen is if one
1170 * grace period ends and another starts between these two fetches.
1171 * Detect this by comparing rsp->completed with the previous fetch
1174 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1175 * and rsp->gp_start suffice to forestall false positives.
1177 gpnum = ACCESS_ONCE(rsp->gpnum);
1178 smp_rmb(); /* Pick up ->gpnum first... */
1179 js = ACCESS_ONCE(rsp->jiffies_stall);
1180 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1181 gps = ACCESS_ONCE(rsp->gp_start);
1182 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1183 completed = ACCESS_ONCE(rsp->completed);
1184 if (ULONG_CMP_GE(completed, gpnum) ||
1185 ULONG_CMP_LT(j, js) ||
1186 ULONG_CMP_GE(gps, js))
1187 return; /* No stall or GP completed since entering function. */
1189 if (rcu_gp_in_progress(rsp) &&
1190 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1192 /* We haven't checked in, so go dump stack. */
1193 print_cpu_stall(rsp);
1195 } else if (rcu_gp_in_progress(rsp) &&
1196 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1198 /* They had a few time units to dump stack, so complain. */
1199 print_other_cpu_stall(rsp);
1204 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1206 * Set the stall-warning timeout way off into the future, thus preventing
1207 * any RCU CPU stall-warning messages from appearing in the current set of
1208 * RCU grace periods.
1210 * The caller must disable hard irqs.
1212 void rcu_cpu_stall_reset(void)
1214 struct rcu_state *rsp;
1216 for_each_rcu_flavor(rsp)
1217 ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2;
1221 * Initialize the specified rcu_data structure's callback list to empty.
1223 static void init_callback_list(struct rcu_data *rdp)
1227 if (init_nocb_callback_list(rdp))
1229 rdp->nxtlist = NULL;
1230 for (i = 0; i < RCU_NEXT_SIZE; i++)
1231 rdp->nxttail[i] = &rdp->nxtlist;
1235 * Determine the value that ->completed will have at the end of the
1236 * next subsequent grace period. This is used to tag callbacks so that
1237 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1238 * been dyntick-idle for an extended period with callbacks under the
1239 * influence of RCU_FAST_NO_HZ.
1241 * The caller must hold rnp->lock with interrupts disabled.
1243 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1244 struct rcu_node *rnp)
1247 * If RCU is idle, we just wait for the next grace period.
1248 * But we can only be sure that RCU is idle if we are looking
1249 * at the root rcu_node structure -- otherwise, a new grace
1250 * period might have started, but just not yet gotten around
1251 * to initializing the current non-root rcu_node structure.
1253 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1254 return rnp->completed + 1;
1257 * Otherwise, wait for a possible partial grace period and
1258 * then the subsequent full grace period.
1260 return rnp->completed + 2;
1264 * Trace-event helper function for rcu_start_future_gp() and
1265 * rcu_nocb_wait_gp().
1267 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1268 unsigned long c, const char *s)
1270 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1271 rnp->completed, c, rnp->level,
1272 rnp->grplo, rnp->grphi, s);
1276 * Start some future grace period, as needed to handle newly arrived
1277 * callbacks. The required future grace periods are recorded in each
1278 * rcu_node structure's ->need_future_gp field. Returns true if there
1279 * is reason to awaken the grace-period kthread.
1281 * The caller must hold the specified rcu_node structure's ->lock.
1283 static bool __maybe_unused
1284 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1285 unsigned long *c_out)
1290 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1293 * Pick up grace-period number for new callbacks. If this
1294 * grace period is already marked as needed, return to the caller.
1296 c = rcu_cbs_completed(rdp->rsp, rnp);
1297 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1298 if (rnp->need_future_gp[c & 0x1]) {
1299 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1304 * If either this rcu_node structure or the root rcu_node structure
1305 * believe that a grace period is in progress, then we must wait
1306 * for the one following, which is in "c". Because our request
1307 * will be noticed at the end of the current grace period, we don't
1308 * need to explicitly start one. We only do the lockless check
1309 * of rnp_root's fields if the current rcu_node structure thinks
1310 * there is no grace period in flight, and because we hold rnp->lock,
1311 * the only possible change is when rnp_root's two fields are
1312 * equal, in which case rnp_root->gpnum might be concurrently
1313 * incremented. But that is OK, as it will just result in our
1314 * doing some extra useless work.
1316 if (rnp->gpnum != rnp->completed ||
1317 ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) {
1318 rnp->need_future_gp[c & 0x1]++;
1319 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1324 * There might be no grace period in progress. If we don't already
1325 * hold it, acquire the root rcu_node structure's lock in order to
1326 * start one (if needed).
1328 if (rnp != rnp_root) {
1329 raw_spin_lock(&rnp_root->lock);
1330 smp_mb__after_unlock_lock();
1334 * Get a new grace-period number. If there really is no grace
1335 * period in progress, it will be smaller than the one we obtained
1336 * earlier. Adjust callbacks as needed. Note that even no-CBs
1337 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1339 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1340 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1341 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1342 rdp->nxtcompleted[i] = c;
1345 * If the needed for the required grace period is already
1346 * recorded, trace and leave.
1348 if (rnp_root->need_future_gp[c & 0x1]) {
1349 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1353 /* Record the need for the future grace period. */
1354 rnp_root->need_future_gp[c & 0x1]++;
1356 /* If a grace period is not already in progress, start one. */
1357 if (rnp_root->gpnum != rnp_root->completed) {
1358 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1360 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1361 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1364 if (rnp != rnp_root)
1365 raw_spin_unlock(&rnp_root->lock);
1373 * Clean up any old requests for the just-ended grace period. Also return
1374 * whether any additional grace periods have been requested. Also invoke
1375 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1376 * waiting for this grace period to complete.
1378 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1380 int c = rnp->completed;
1382 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1384 rcu_nocb_gp_cleanup(rsp, rnp);
1385 rnp->need_future_gp[c & 0x1] = 0;
1386 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1387 trace_rcu_future_gp(rnp, rdp, c,
1388 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1393 * Awaken the grace-period kthread for the specified flavor of RCU.
1394 * Don't do a self-awaken, and don't bother awakening when there is
1395 * nothing for the grace-period kthread to do (as in several CPUs
1396 * raced to awaken, and we lost), and finally don't try to awaken
1397 * a kthread that has not yet been created.
1399 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1401 if (current == rsp->gp_kthread ||
1402 !ACCESS_ONCE(rsp->gp_flags) ||
1405 wake_up(&rsp->gp_wq);
1409 * If there is room, assign a ->completed number to any callbacks on
1410 * this CPU that have not already been assigned. Also accelerate any
1411 * callbacks that were previously assigned a ->completed number that has
1412 * since proven to be too conservative, which can happen if callbacks get
1413 * assigned a ->completed number while RCU is idle, but with reference to
1414 * a non-root rcu_node structure. This function is idempotent, so it does
1415 * not hurt to call it repeatedly. Returns an flag saying that we should
1416 * awaken the RCU grace-period kthread.
1418 * The caller must hold rnp->lock with interrupts disabled.
1420 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1421 struct rcu_data *rdp)
1427 /* If the CPU has no callbacks, nothing to do. */
1428 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1432 * Starting from the sublist containing the callbacks most
1433 * recently assigned a ->completed number and working down, find the
1434 * first sublist that is not assignable to an upcoming grace period.
1435 * Such a sublist has something in it (first two tests) and has
1436 * a ->completed number assigned that will complete sooner than
1437 * the ->completed number for newly arrived callbacks (last test).
1439 * The key point is that any later sublist can be assigned the
1440 * same ->completed number as the newly arrived callbacks, which
1441 * means that the callbacks in any of these later sublist can be
1442 * grouped into a single sublist, whether or not they have already
1443 * been assigned a ->completed number.
1445 c = rcu_cbs_completed(rsp, rnp);
1446 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1447 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1448 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1452 * If there are no sublist for unassigned callbacks, leave.
1453 * At the same time, advance "i" one sublist, so that "i" will
1454 * index into the sublist where all the remaining callbacks should
1457 if (++i >= RCU_NEXT_TAIL)
1461 * Assign all subsequent callbacks' ->completed number to the next
1462 * full grace period and group them all in the sublist initially
1465 for (; i <= RCU_NEXT_TAIL; i++) {
1466 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1467 rdp->nxtcompleted[i] = c;
1469 /* Record any needed additional grace periods. */
1470 ret = rcu_start_future_gp(rnp, rdp, NULL);
1472 /* Trace depending on how much we were able to accelerate. */
1473 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1474 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1476 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1481 * Move any callbacks whose grace period has completed to the
1482 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1483 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1484 * sublist. This function is idempotent, so it does not hurt to
1485 * invoke it repeatedly. As long as it is not invoked -too- often...
1486 * Returns true if the RCU grace-period kthread needs to be awakened.
1488 * The caller must hold rnp->lock with interrupts disabled.
1490 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1491 struct rcu_data *rdp)
1495 /* If the CPU has no callbacks, nothing to do. */
1496 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1500 * Find all callbacks whose ->completed numbers indicate that they
1501 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1503 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1504 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1506 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1508 /* Clean up any sublist tail pointers that were misordered above. */
1509 for (j = RCU_WAIT_TAIL; j < i; j++)
1510 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1512 /* Copy down callbacks to fill in empty sublists. */
1513 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1514 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1516 rdp->nxttail[j] = rdp->nxttail[i];
1517 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1520 /* Classify any remaining callbacks. */
1521 return rcu_accelerate_cbs(rsp, rnp, rdp);
1525 * Update CPU-local rcu_data state to record the beginnings and ends of
1526 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1527 * structure corresponding to the current CPU, and must have irqs disabled.
1528 * Returns true if the grace-period kthread needs to be awakened.
1530 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1531 struct rcu_data *rdp)
1535 /* Handle the ends of any preceding grace periods first. */
1536 if (rdp->completed == rnp->completed) {
1538 /* No grace period end, so just accelerate recent callbacks. */
1539 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1543 /* Advance callbacks. */
1544 ret = rcu_advance_cbs(rsp, rnp, rdp);
1546 /* Remember that we saw this grace-period completion. */
1547 rdp->completed = rnp->completed;
1548 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1551 if (rdp->gpnum != rnp->gpnum) {
1553 * If the current grace period is waiting for this CPU,
1554 * set up to detect a quiescent state, otherwise don't
1555 * go looking for one.
1557 rdp->gpnum = rnp->gpnum;
1558 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1559 rdp->passed_quiesce = 0;
1560 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1561 zero_cpu_stall_ticks(rdp);
1566 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1568 unsigned long flags;
1570 struct rcu_node *rnp;
1572 local_irq_save(flags);
1574 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1575 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1576 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1577 local_irq_restore(flags);
1580 smp_mb__after_unlock_lock();
1581 needwake = __note_gp_changes(rsp, rnp, rdp);
1582 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1584 rcu_gp_kthread_wake(rsp);
1588 * Initialize a new grace period. Return 0 if no grace period required.
1590 static int rcu_gp_init(struct rcu_state *rsp)
1592 struct rcu_data *rdp;
1593 struct rcu_node *rnp = rcu_get_root(rsp);
1595 rcu_bind_gp_kthread();
1596 raw_spin_lock_irq(&rnp->lock);
1597 smp_mb__after_unlock_lock();
1598 if (!ACCESS_ONCE(rsp->gp_flags)) {
1599 /* Spurious wakeup, tell caller to go back to sleep. */
1600 raw_spin_unlock_irq(&rnp->lock);
1603 ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */
1605 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1607 * Grace period already in progress, don't start another.
1608 * Not supposed to be able to happen.
1610 raw_spin_unlock_irq(&rnp->lock);
1614 /* Advance to a new grace period and initialize state. */
1615 record_gp_stall_check_time(rsp);
1616 /* Record GP times before starting GP, hence smp_store_release(). */
1617 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1618 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1619 raw_spin_unlock_irq(&rnp->lock);
1621 /* Exclude any concurrent CPU-hotplug operations. */
1622 mutex_lock(&rsp->onoff_mutex);
1623 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1626 * Set the quiescent-state-needed bits in all the rcu_node
1627 * structures for all currently online CPUs in breadth-first order,
1628 * starting from the root rcu_node structure, relying on the layout
1629 * of the tree within the rsp->node[] array. Note that other CPUs
1630 * will access only the leaves of the hierarchy, thus seeing that no
1631 * grace period is in progress, at least until the corresponding
1632 * leaf node has been initialized. In addition, we have excluded
1633 * CPU-hotplug operations.
1635 * The grace period cannot complete until the initialization
1636 * process finishes, because this kthread handles both.
1638 rcu_for_each_node_breadth_first(rsp, rnp) {
1639 raw_spin_lock_irq(&rnp->lock);
1640 smp_mb__after_unlock_lock();
1641 rdp = this_cpu_ptr(rsp->rda);
1642 rcu_preempt_check_blocked_tasks(rnp);
1643 rnp->qsmask = rnp->qsmaskinit;
1644 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1645 WARN_ON_ONCE(rnp->completed != rsp->completed);
1646 ACCESS_ONCE(rnp->completed) = rsp->completed;
1647 if (rnp == rdp->mynode)
1648 (void)__note_gp_changes(rsp, rnp, rdp);
1649 rcu_preempt_boost_start_gp(rnp);
1650 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1651 rnp->level, rnp->grplo,
1652 rnp->grphi, rnp->qsmask);
1653 raw_spin_unlock_irq(&rnp->lock);
1654 cond_resched_rcu_qs();
1657 mutex_unlock(&rsp->onoff_mutex);
1662 * Do one round of quiescent-state forcing.
1664 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1666 int fqs_state = fqs_state_in;
1667 bool isidle = false;
1669 struct rcu_node *rnp = rcu_get_root(rsp);
1672 if (fqs_state == RCU_SAVE_DYNTICK) {
1673 /* Collect dyntick-idle snapshots. */
1674 if (is_sysidle_rcu_state(rsp)) {
1676 maxj = jiffies - ULONG_MAX / 4;
1678 force_qs_rnp(rsp, dyntick_save_progress_counter,
1680 rcu_sysidle_report_gp(rsp, isidle, maxj);
1681 fqs_state = RCU_FORCE_QS;
1683 /* Handle dyntick-idle and offline CPUs. */
1685 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1687 /* Clear flag to prevent immediate re-entry. */
1688 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1689 raw_spin_lock_irq(&rnp->lock);
1690 smp_mb__after_unlock_lock();
1691 ACCESS_ONCE(rsp->gp_flags) &= ~RCU_GP_FLAG_FQS;
1692 raw_spin_unlock_irq(&rnp->lock);
1698 * Clean up after the old grace period.
1700 static void rcu_gp_cleanup(struct rcu_state *rsp)
1702 unsigned long gp_duration;
1703 bool needgp = false;
1705 struct rcu_data *rdp;
1706 struct rcu_node *rnp = rcu_get_root(rsp);
1708 raw_spin_lock_irq(&rnp->lock);
1709 smp_mb__after_unlock_lock();
1710 gp_duration = jiffies - rsp->gp_start;
1711 if (gp_duration > rsp->gp_max)
1712 rsp->gp_max = gp_duration;
1715 * We know the grace period is complete, but to everyone else
1716 * it appears to still be ongoing. But it is also the case
1717 * that to everyone else it looks like there is nothing that
1718 * they can do to advance the grace period. It is therefore
1719 * safe for us to drop the lock in order to mark the grace
1720 * period as completed in all of the rcu_node structures.
1722 raw_spin_unlock_irq(&rnp->lock);
1725 * Propagate new ->completed value to rcu_node structures so
1726 * that other CPUs don't have to wait until the start of the next
1727 * grace period to process their callbacks. This also avoids
1728 * some nasty RCU grace-period initialization races by forcing
1729 * the end of the current grace period to be completely recorded in
1730 * all of the rcu_node structures before the beginning of the next
1731 * grace period is recorded in any of the rcu_node structures.
1733 rcu_for_each_node_breadth_first(rsp, rnp) {
1734 raw_spin_lock_irq(&rnp->lock);
1735 smp_mb__after_unlock_lock();
1736 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1737 rdp = this_cpu_ptr(rsp->rda);
1738 if (rnp == rdp->mynode)
1739 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
1740 /* smp_mb() provided by prior unlock-lock pair. */
1741 nocb += rcu_future_gp_cleanup(rsp, rnp);
1742 raw_spin_unlock_irq(&rnp->lock);
1743 cond_resched_rcu_qs();
1745 rnp = rcu_get_root(rsp);
1746 raw_spin_lock_irq(&rnp->lock);
1747 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1748 rcu_nocb_gp_set(rnp, nocb);
1750 /* Declare grace period done. */
1751 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1752 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1753 rsp->fqs_state = RCU_GP_IDLE;
1754 rdp = this_cpu_ptr(rsp->rda);
1755 /* Advance CBs to reduce false positives below. */
1756 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
1757 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
1758 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1759 trace_rcu_grace_period(rsp->name,
1760 ACCESS_ONCE(rsp->gpnum),
1763 raw_spin_unlock_irq(&rnp->lock);
1767 * Body of kthread that handles grace periods.
1769 static int __noreturn rcu_gp_kthread(void *arg)
1775 struct rcu_state *rsp = arg;
1776 struct rcu_node *rnp = rcu_get_root(rsp);
1780 /* Handle grace-period start. */
1782 trace_rcu_grace_period(rsp->name,
1783 ACCESS_ONCE(rsp->gpnum),
1785 rsp->gp_state = RCU_GP_WAIT_GPS;
1786 wait_event_interruptible(rsp->gp_wq,
1787 ACCESS_ONCE(rsp->gp_flags) &
1789 /* Locking provides needed memory barrier. */
1790 if (rcu_gp_init(rsp))
1792 cond_resched_rcu_qs();
1793 flush_signals(current);
1794 trace_rcu_grace_period(rsp->name,
1795 ACCESS_ONCE(rsp->gpnum),
1799 /* Handle quiescent-state forcing. */
1800 fqs_state = RCU_SAVE_DYNTICK;
1801 j = jiffies_till_first_fqs;
1804 jiffies_till_first_fqs = HZ;
1809 rsp->jiffies_force_qs = jiffies + j;
1810 trace_rcu_grace_period(rsp->name,
1811 ACCESS_ONCE(rsp->gpnum),
1813 rsp->gp_state = RCU_GP_WAIT_FQS;
1814 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1815 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1817 (!ACCESS_ONCE(rnp->qsmask) &&
1818 !rcu_preempt_blocked_readers_cgp(rnp)),
1820 /* Locking provides needed memory barriers. */
1821 /* If grace period done, leave loop. */
1822 if (!ACCESS_ONCE(rnp->qsmask) &&
1823 !rcu_preempt_blocked_readers_cgp(rnp))
1825 /* If time for quiescent-state forcing, do it. */
1826 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1827 (gf & RCU_GP_FLAG_FQS)) {
1828 trace_rcu_grace_period(rsp->name,
1829 ACCESS_ONCE(rsp->gpnum),
1831 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1832 trace_rcu_grace_period(rsp->name,
1833 ACCESS_ONCE(rsp->gpnum),
1835 cond_resched_rcu_qs();
1837 /* Deal with stray signal. */
1838 cond_resched_rcu_qs();
1839 flush_signals(current);
1840 trace_rcu_grace_period(rsp->name,
1841 ACCESS_ONCE(rsp->gpnum),
1844 j = jiffies_till_next_fqs;
1847 jiffies_till_next_fqs = HZ;
1850 jiffies_till_next_fqs = 1;
1854 /* Handle grace-period end. */
1855 rcu_gp_cleanup(rsp);
1860 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1861 * in preparation for detecting the next grace period. The caller must hold
1862 * the root node's ->lock and hard irqs must be disabled.
1864 * Note that it is legal for a dying CPU (which is marked as offline) to
1865 * invoke this function. This can happen when the dying CPU reports its
1868 * Returns true if the grace-period kthread must be awakened.
1871 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1872 struct rcu_data *rdp)
1874 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1876 * Either we have not yet spawned the grace-period
1877 * task, this CPU does not need another grace period,
1878 * or a grace period is already in progress.
1879 * Either way, don't start a new grace period.
1883 ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT;
1884 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1888 * We can't do wakeups while holding the rnp->lock, as that
1889 * could cause possible deadlocks with the rq->lock. Defer
1890 * the wakeup to our caller.
1896 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1897 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1898 * is invoked indirectly from rcu_advance_cbs(), which would result in
1899 * endless recursion -- or would do so if it wasn't for the self-deadlock
1900 * that is encountered beforehand.
1902 * Returns true if the grace-period kthread needs to be awakened.
1904 static bool rcu_start_gp(struct rcu_state *rsp)
1906 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1907 struct rcu_node *rnp = rcu_get_root(rsp);
1911 * If there is no grace period in progress right now, any
1912 * callbacks we have up to this point will be satisfied by the
1913 * next grace period. Also, advancing the callbacks reduces the
1914 * probability of false positives from cpu_needs_another_gp()
1915 * resulting in pointless grace periods. So, advance callbacks
1916 * then start the grace period!
1918 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
1919 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
1924 * Report a full set of quiescent states to the specified rcu_state
1925 * data structure. This involves cleaning up after the prior grace
1926 * period and letting rcu_start_gp() start up the next grace period
1927 * if one is needed. Note that the caller must hold rnp->lock, which
1928 * is released before return.
1930 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1931 __releases(rcu_get_root(rsp)->lock)
1933 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1934 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1935 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1939 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1940 * Allows quiescent states for a group of CPUs to be reported at one go
1941 * to the specified rcu_node structure, though all the CPUs in the group
1942 * must be represented by the same rcu_node structure (which need not be
1943 * a leaf rcu_node structure, though it often will be). That structure's
1944 * lock must be held upon entry, and it is released before return.
1947 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1948 struct rcu_node *rnp, unsigned long flags)
1949 __releases(rnp->lock)
1951 struct rcu_node *rnp_c;
1953 /* Walk up the rcu_node hierarchy. */
1955 if (!(rnp->qsmask & mask)) {
1957 /* Our bit has already been cleared, so done. */
1958 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1961 rnp->qsmask &= ~mask;
1962 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1963 mask, rnp->qsmask, rnp->level,
1964 rnp->grplo, rnp->grphi,
1966 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1968 /* Other bits still set at this level, so done. */
1969 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1972 mask = rnp->grpmask;
1973 if (rnp->parent == NULL) {
1975 /* No more levels. Exit loop holding root lock. */
1979 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1982 raw_spin_lock_irqsave(&rnp->lock, flags);
1983 smp_mb__after_unlock_lock();
1984 WARN_ON_ONCE(rnp_c->qsmask);
1988 * Get here if we are the last CPU to pass through a quiescent
1989 * state for this grace period. Invoke rcu_report_qs_rsp()
1990 * to clean up and start the next grace period if one is needed.
1992 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1996 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1997 * structure. This must be either called from the specified CPU, or
1998 * called when the specified CPU is known to be offline (and when it is
1999 * also known that no other CPU is concurrently trying to help the offline
2000 * CPU). The lastcomp argument is used to make sure we are still in the
2001 * grace period of interest. We don't want to end the current grace period
2002 * based on quiescent states detected in an earlier grace period!
2005 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2007 unsigned long flags;
2010 struct rcu_node *rnp;
2013 raw_spin_lock_irqsave(&rnp->lock, flags);
2014 smp_mb__after_unlock_lock();
2015 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
2016 rnp->completed == rnp->gpnum) {
2019 * The grace period in which this quiescent state was
2020 * recorded has ended, so don't report it upwards.
2021 * We will instead need a new quiescent state that lies
2022 * within the current grace period.
2024 rdp->passed_quiesce = 0; /* need qs for new gp. */
2025 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2028 mask = rdp->grpmask;
2029 if ((rnp->qsmask & mask) == 0) {
2030 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2032 rdp->qs_pending = 0;
2035 * This GP can't end until cpu checks in, so all of our
2036 * callbacks can be processed during the next GP.
2038 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2040 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
2042 rcu_gp_kthread_wake(rsp);
2047 * Check to see if there is a new grace period of which this CPU
2048 * is not yet aware, and if so, set up local rcu_data state for it.
2049 * Otherwise, see if this CPU has just passed through its first
2050 * quiescent state for this grace period, and record that fact if so.
2053 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2055 /* Check for grace-period ends and beginnings. */
2056 note_gp_changes(rsp, rdp);
2059 * Does this CPU still need to do its part for current grace period?
2060 * If no, return and let the other CPUs do their part as well.
2062 if (!rdp->qs_pending)
2066 * Was there a quiescent state since the beginning of the grace
2067 * period? If no, then exit and wait for the next call.
2069 if (!rdp->passed_quiesce)
2073 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2076 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2079 #ifdef CONFIG_HOTPLUG_CPU
2082 * Send the specified CPU's RCU callbacks to the orphanage. The
2083 * specified CPU must be offline, and the caller must hold the
2087 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2088 struct rcu_node *rnp, struct rcu_data *rdp)
2090 /* No-CBs CPUs do not have orphanable callbacks. */
2091 if (rcu_is_nocb_cpu(rdp->cpu))
2095 * Orphan the callbacks. First adjust the counts. This is safe
2096 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2097 * cannot be running now. Thus no memory barrier is required.
2099 if (rdp->nxtlist != NULL) {
2100 rsp->qlen_lazy += rdp->qlen_lazy;
2101 rsp->qlen += rdp->qlen;
2102 rdp->n_cbs_orphaned += rdp->qlen;
2104 ACCESS_ONCE(rdp->qlen) = 0;
2108 * Next, move those callbacks still needing a grace period to
2109 * the orphanage, where some other CPU will pick them up.
2110 * Some of the callbacks might have gone partway through a grace
2111 * period, but that is too bad. They get to start over because we
2112 * cannot assume that grace periods are synchronized across CPUs.
2113 * We don't bother updating the ->nxttail[] array yet, instead
2114 * we just reset the whole thing later on.
2116 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2117 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2118 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2119 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2123 * Then move the ready-to-invoke callbacks to the orphanage,
2124 * where some other CPU will pick them up. These will not be
2125 * required to pass though another grace period: They are done.
2127 if (rdp->nxtlist != NULL) {
2128 *rsp->orphan_donetail = rdp->nxtlist;
2129 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2132 /* Finally, initialize the rcu_data structure's list to empty. */
2133 init_callback_list(rdp);
2137 * Adopt the RCU callbacks from the specified rcu_state structure's
2138 * orphanage. The caller must hold the ->orphan_lock.
2140 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2143 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2145 /* No-CBs CPUs are handled specially. */
2146 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2149 /* Do the accounting first. */
2150 rdp->qlen_lazy += rsp->qlen_lazy;
2151 rdp->qlen += rsp->qlen;
2152 rdp->n_cbs_adopted += rsp->qlen;
2153 if (rsp->qlen_lazy != rsp->qlen)
2154 rcu_idle_count_callbacks_posted();
2159 * We do not need a memory barrier here because the only way we
2160 * can get here if there is an rcu_barrier() in flight is if
2161 * we are the task doing the rcu_barrier().
2164 /* First adopt the ready-to-invoke callbacks. */
2165 if (rsp->orphan_donelist != NULL) {
2166 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2167 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2168 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2169 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2170 rdp->nxttail[i] = rsp->orphan_donetail;
2171 rsp->orphan_donelist = NULL;
2172 rsp->orphan_donetail = &rsp->orphan_donelist;
2175 /* And then adopt the callbacks that still need a grace period. */
2176 if (rsp->orphan_nxtlist != NULL) {
2177 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2178 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2179 rsp->orphan_nxtlist = NULL;
2180 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2185 * Trace the fact that this CPU is going offline.
2187 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2189 RCU_TRACE(unsigned long mask);
2190 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2191 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2193 RCU_TRACE(mask = rdp->grpmask);
2194 trace_rcu_grace_period(rsp->name,
2195 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2200 * The CPU has been completely removed, and some other CPU is reporting
2201 * this fact from process context. Do the remainder of the cleanup,
2202 * including orphaning the outgoing CPU's RCU callbacks, and also
2203 * adopting them. There can only be one CPU hotplug operation at a time,
2204 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2206 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2208 unsigned long flags;
2210 int need_report = 0;
2211 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2212 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2214 /* Adjust any no-longer-needed kthreads. */
2215 rcu_boost_kthread_setaffinity(rnp, -1);
2217 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2219 /* Exclude any attempts to start a new grace period. */
2220 mutex_lock(&rsp->onoff_mutex);
2221 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2223 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2224 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2225 rcu_adopt_orphan_cbs(rsp, flags);
2227 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2228 mask = rdp->grpmask; /* rnp->grplo is constant. */
2230 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2231 smp_mb__after_unlock_lock();
2232 rnp->qsmaskinit &= ~mask;
2233 if (rnp->qsmaskinit != 0) {
2234 if (rnp != rdp->mynode)
2235 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2238 if (rnp == rdp->mynode)
2239 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2241 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2242 mask = rnp->grpmask;
2244 } while (rnp != NULL);
2247 * We still hold the leaf rcu_node structure lock here, and
2248 * irqs are still disabled. The reason for this subterfuge is
2249 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2250 * held leads to deadlock.
2252 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2254 if (need_report & RCU_OFL_TASKS_NORM_GP)
2255 rcu_report_unblock_qs_rnp(rnp, flags);
2257 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2258 if (need_report & RCU_OFL_TASKS_EXP_GP)
2259 rcu_report_exp_rnp(rsp, rnp, true);
2260 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2261 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2262 cpu, rdp->qlen, rdp->nxtlist);
2263 init_callback_list(rdp);
2264 /* Disallow further callbacks on this CPU. */
2265 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2266 mutex_unlock(&rsp->onoff_mutex);
2269 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2271 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2275 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2279 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2282 * Invoke any RCU callbacks that have made it to the end of their grace
2283 * period. Thottle as specified by rdp->blimit.
2285 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2287 unsigned long flags;
2288 struct rcu_head *next, *list, **tail;
2289 long bl, count, count_lazy;
2292 /* If no callbacks are ready, just return. */
2293 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2294 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2295 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2296 need_resched(), is_idle_task(current),
2297 rcu_is_callbacks_kthread());
2302 * Extract the list of ready callbacks, disabling to prevent
2303 * races with call_rcu() from interrupt handlers.
2305 local_irq_save(flags);
2306 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2308 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2309 list = rdp->nxtlist;
2310 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2311 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2312 tail = rdp->nxttail[RCU_DONE_TAIL];
2313 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2314 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2315 rdp->nxttail[i] = &rdp->nxtlist;
2316 local_irq_restore(flags);
2318 /* Invoke callbacks. */
2319 count = count_lazy = 0;
2323 debug_rcu_head_unqueue(list);
2324 if (__rcu_reclaim(rsp->name, list))
2327 /* Stop only if limit reached and CPU has something to do. */
2328 if (++count >= bl &&
2330 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2334 local_irq_save(flags);
2335 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2336 is_idle_task(current),
2337 rcu_is_callbacks_kthread());
2339 /* Update count, and requeue any remaining callbacks. */
2341 *tail = rdp->nxtlist;
2342 rdp->nxtlist = list;
2343 for (i = 0; i < RCU_NEXT_SIZE; i++)
2344 if (&rdp->nxtlist == rdp->nxttail[i])
2345 rdp->nxttail[i] = tail;
2349 smp_mb(); /* List handling before counting for rcu_barrier(). */
2350 rdp->qlen_lazy -= count_lazy;
2351 ACCESS_ONCE(rdp->qlen) = rdp->qlen - count;
2352 rdp->n_cbs_invoked += count;
2354 /* Reinstate batch limit if we have worked down the excess. */
2355 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2356 rdp->blimit = blimit;
2358 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2359 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2360 rdp->qlen_last_fqs_check = 0;
2361 rdp->n_force_qs_snap = rsp->n_force_qs;
2362 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2363 rdp->qlen_last_fqs_check = rdp->qlen;
2364 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2366 local_irq_restore(flags);
2368 /* Re-invoke RCU core processing if there are callbacks remaining. */
2369 if (cpu_has_callbacks_ready_to_invoke(rdp))
2374 * Check to see if this CPU is in a non-context-switch quiescent state
2375 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2376 * Also schedule RCU core processing.
2378 * This function must be called from hardirq context. It is normally
2379 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2380 * false, there is no point in invoking rcu_check_callbacks().
2382 void rcu_check_callbacks(int cpu, int user)
2384 trace_rcu_utilization(TPS("Start scheduler-tick"));
2385 increment_cpu_stall_ticks();
2386 if (user || rcu_is_cpu_rrupt_from_idle()) {
2389 * Get here if this CPU took its interrupt from user
2390 * mode or from the idle loop, and if this is not a
2391 * nested interrupt. In this case, the CPU is in
2392 * a quiescent state, so note it.
2394 * No memory barrier is required here because both
2395 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2396 * variables that other CPUs neither access nor modify,
2397 * at least not while the corresponding CPU is online.
2403 } else if (!in_softirq()) {
2406 * Get here if this CPU did not take its interrupt from
2407 * softirq, in other words, if it is not interrupting
2408 * a rcu_bh read-side critical section. This is an _bh
2409 * critical section, so note it.
2414 rcu_preempt_check_callbacks(cpu);
2415 if (rcu_pending(cpu))
2418 rcu_note_voluntary_context_switch(current);
2419 trace_rcu_utilization(TPS("End scheduler-tick"));
2423 * Scan the leaf rcu_node structures, processing dyntick state for any that
2424 * have not yet encountered a quiescent state, using the function specified.
2425 * Also initiate boosting for any threads blocked on the root rcu_node.
2427 * The caller must have suppressed start of new grace periods.
2429 static void force_qs_rnp(struct rcu_state *rsp,
2430 int (*f)(struct rcu_data *rsp, bool *isidle,
2431 unsigned long *maxj),
2432 bool *isidle, unsigned long *maxj)
2436 unsigned long flags;
2438 struct rcu_node *rnp;
2440 rcu_for_each_leaf_node(rsp, rnp) {
2441 cond_resched_rcu_qs();
2443 raw_spin_lock_irqsave(&rnp->lock, flags);
2444 smp_mb__after_unlock_lock();
2445 if (!rcu_gp_in_progress(rsp)) {
2446 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2449 if (rnp->qsmask == 0) {
2450 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2455 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2456 if ((rnp->qsmask & bit) != 0) {
2457 if ((rnp->qsmaskinit & bit) != 0)
2459 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2465 /* rcu_report_qs_rnp() releases rnp->lock. */
2466 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2469 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2471 rnp = rcu_get_root(rsp);
2472 if (rnp->qsmask == 0) {
2473 raw_spin_lock_irqsave(&rnp->lock, flags);
2474 smp_mb__after_unlock_lock();
2475 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2480 * Force quiescent states on reluctant CPUs, and also detect which
2481 * CPUs are in dyntick-idle mode.
2483 static void force_quiescent_state(struct rcu_state *rsp)
2485 unsigned long flags;
2487 struct rcu_node *rnp;
2488 struct rcu_node *rnp_old = NULL;
2490 /* Funnel through hierarchy to reduce memory contention. */
2491 rnp = __this_cpu_read(rsp->rda->mynode);
2492 for (; rnp != NULL; rnp = rnp->parent) {
2493 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2494 !raw_spin_trylock(&rnp->fqslock);
2495 if (rnp_old != NULL)
2496 raw_spin_unlock(&rnp_old->fqslock);
2498 rsp->n_force_qs_lh++;
2503 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2505 /* Reached the root of the rcu_node tree, acquire lock. */
2506 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2507 smp_mb__after_unlock_lock();
2508 raw_spin_unlock(&rnp_old->fqslock);
2509 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2510 rsp->n_force_qs_lh++;
2511 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2512 return; /* Someone beat us to it. */
2514 ACCESS_ONCE(rsp->gp_flags) |= RCU_GP_FLAG_FQS;
2515 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2516 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2520 * This does the RCU core processing work for the specified rcu_state
2521 * and rcu_data structures. This may be called only from the CPU to
2522 * whom the rdp belongs.
2525 __rcu_process_callbacks(struct rcu_state *rsp)
2527 unsigned long flags;
2529 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2531 WARN_ON_ONCE(rdp->beenonline == 0);
2533 /* Update RCU state based on any recent quiescent states. */
2534 rcu_check_quiescent_state(rsp, rdp);
2536 /* Does this CPU require a not-yet-started grace period? */
2537 local_irq_save(flags);
2538 if (cpu_needs_another_gp(rsp, rdp)) {
2539 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2540 needwake = rcu_start_gp(rsp);
2541 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2543 rcu_gp_kthread_wake(rsp);
2545 local_irq_restore(flags);
2548 /* If there are callbacks ready, invoke them. */
2549 if (cpu_has_callbacks_ready_to_invoke(rdp))
2550 invoke_rcu_callbacks(rsp, rdp);
2552 /* Do any needed deferred wakeups of rcuo kthreads. */
2553 do_nocb_deferred_wakeup(rdp);
2557 * Do RCU core processing for the current CPU.
2559 static void rcu_process_callbacks(struct softirq_action *unused)
2561 struct rcu_state *rsp;
2563 if (cpu_is_offline(smp_processor_id()))
2565 trace_rcu_utilization(TPS("Start RCU core"));
2566 for_each_rcu_flavor(rsp)
2567 __rcu_process_callbacks(rsp);
2568 trace_rcu_utilization(TPS("End RCU core"));
2572 * Schedule RCU callback invocation. If the specified type of RCU
2573 * does not support RCU priority boosting, just do a direct call,
2574 * otherwise wake up the per-CPU kernel kthread. Note that because we
2575 * are running on the current CPU with interrupts disabled, the
2576 * rcu_cpu_kthread_task cannot disappear out from under us.
2578 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2580 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2582 if (likely(!rsp->boost)) {
2583 rcu_do_batch(rsp, rdp);
2586 invoke_rcu_callbacks_kthread();
2589 static void invoke_rcu_core(void)
2591 if (cpu_online(smp_processor_id()))
2592 raise_softirq(RCU_SOFTIRQ);
2596 * Handle any core-RCU processing required by a call_rcu() invocation.
2598 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2599 struct rcu_head *head, unsigned long flags)
2604 * If called from an extended quiescent state, invoke the RCU
2605 * core in order to force a re-evaluation of RCU's idleness.
2607 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2610 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2611 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2615 * Force the grace period if too many callbacks or too long waiting.
2616 * Enforce hysteresis, and don't invoke force_quiescent_state()
2617 * if some other CPU has recently done so. Also, don't bother
2618 * invoking force_quiescent_state() if the newly enqueued callback
2619 * is the only one waiting for a grace period to complete.
2621 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2623 /* Are we ignoring a completed grace period? */
2624 note_gp_changes(rsp, rdp);
2626 /* Start a new grace period if one not already started. */
2627 if (!rcu_gp_in_progress(rsp)) {
2628 struct rcu_node *rnp_root = rcu_get_root(rsp);
2630 raw_spin_lock(&rnp_root->lock);
2631 smp_mb__after_unlock_lock();
2632 needwake = rcu_start_gp(rsp);
2633 raw_spin_unlock(&rnp_root->lock);
2635 rcu_gp_kthread_wake(rsp);
2637 /* Give the grace period a kick. */
2638 rdp->blimit = LONG_MAX;
2639 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2640 *rdp->nxttail[RCU_DONE_TAIL] != head)
2641 force_quiescent_state(rsp);
2642 rdp->n_force_qs_snap = rsp->n_force_qs;
2643 rdp->qlen_last_fqs_check = rdp->qlen;
2649 * RCU callback function to leak a callback.
2651 static void rcu_leak_callback(struct rcu_head *rhp)
2656 * Helper function for call_rcu() and friends. The cpu argument will
2657 * normally be -1, indicating "currently running CPU". It may specify
2658 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2659 * is expected to specify a CPU.
2662 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2663 struct rcu_state *rsp, int cpu, bool lazy)
2665 unsigned long flags;
2666 struct rcu_data *rdp;
2668 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
2669 if (debug_rcu_head_queue(head)) {
2670 /* Probable double call_rcu(), so leak the callback. */
2671 ACCESS_ONCE(head->func) = rcu_leak_callback;
2672 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2679 * Opportunistically note grace-period endings and beginnings.
2680 * Note that we might see a beginning right after we see an
2681 * end, but never vice versa, since this CPU has to pass through
2682 * a quiescent state betweentimes.
2684 local_irq_save(flags);
2685 rdp = this_cpu_ptr(rsp->rda);
2687 /* Add the callback to our list. */
2688 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2692 rdp = per_cpu_ptr(rsp->rda, cpu);
2693 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2694 WARN_ON_ONCE(offline);
2695 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2696 local_irq_restore(flags);
2699 ACCESS_ONCE(rdp->qlen) = rdp->qlen + 1;
2703 rcu_idle_count_callbacks_posted();
2704 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2705 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2706 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2708 if (__is_kfree_rcu_offset((unsigned long)func))
2709 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2710 rdp->qlen_lazy, rdp->qlen);
2712 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2714 /* Go handle any RCU core processing required. */
2715 __call_rcu_core(rsp, rdp, head, flags);
2716 local_irq_restore(flags);
2720 * Queue an RCU-sched callback for invocation after a grace period.
2722 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2724 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2726 EXPORT_SYMBOL_GPL(call_rcu_sched);
2729 * Queue an RCU callback for invocation after a quicker grace period.
2731 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2733 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2735 EXPORT_SYMBOL_GPL(call_rcu_bh);
2738 * Queue an RCU callback for lazy invocation after a grace period.
2739 * This will likely be later named something like "call_rcu_lazy()",
2740 * but this change will require some way of tagging the lazy RCU
2741 * callbacks in the list of pending callbacks. Until then, this
2742 * function may only be called from __kfree_rcu().
2744 void kfree_call_rcu(struct rcu_head *head,
2745 void (*func)(struct rcu_head *rcu))
2747 __call_rcu(head, func, rcu_state_p, -1, 1);
2749 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2752 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2753 * any blocking grace-period wait automatically implies a grace period
2754 * if there is only one CPU online at any point time during execution
2755 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2756 * occasionally incorrectly indicate that there are multiple CPUs online
2757 * when there was in fact only one the whole time, as this just adds
2758 * some overhead: RCU still operates correctly.
2760 static inline int rcu_blocking_is_gp(void)
2764 might_sleep(); /* Check for RCU read-side critical section. */
2766 ret = num_online_cpus() <= 1;
2772 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2774 * Control will return to the caller some time after a full rcu-sched
2775 * grace period has elapsed, in other words after all currently executing
2776 * rcu-sched read-side critical sections have completed. These read-side
2777 * critical sections are delimited by rcu_read_lock_sched() and
2778 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2779 * local_irq_disable(), and so on may be used in place of
2780 * rcu_read_lock_sched().
2782 * This means that all preempt_disable code sequences, including NMI and
2783 * non-threaded hardware-interrupt handlers, in progress on entry will
2784 * have completed before this primitive returns. However, this does not
2785 * guarantee that softirq handlers will have completed, since in some
2786 * kernels, these handlers can run in process context, and can block.
2788 * Note that this guarantee implies further memory-ordering guarantees.
2789 * On systems with more than one CPU, when synchronize_sched() returns,
2790 * each CPU is guaranteed to have executed a full memory barrier since the
2791 * end of its last RCU-sched read-side critical section whose beginning
2792 * preceded the call to synchronize_sched(). In addition, each CPU having
2793 * an RCU read-side critical section that extends beyond the return from
2794 * synchronize_sched() is guaranteed to have executed a full memory barrier
2795 * after the beginning of synchronize_sched() and before the beginning of
2796 * that RCU read-side critical section. Note that these guarantees include
2797 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2798 * that are executing in the kernel.
2800 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2801 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2802 * to have executed a full memory barrier during the execution of
2803 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2804 * again only if the system has more than one CPU).
2806 * This primitive provides the guarantees made by the (now removed)
2807 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2808 * guarantees that rcu_read_lock() sections will have completed.
2809 * In "classic RCU", these two guarantees happen to be one and
2810 * the same, but can differ in realtime RCU implementations.
2812 void synchronize_sched(void)
2814 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2815 !lock_is_held(&rcu_lock_map) &&
2816 !lock_is_held(&rcu_sched_lock_map),
2817 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2818 if (rcu_blocking_is_gp())
2821 synchronize_sched_expedited();
2823 wait_rcu_gp(call_rcu_sched);
2825 EXPORT_SYMBOL_GPL(synchronize_sched);
2828 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2830 * Control will return to the caller some time after a full rcu_bh grace
2831 * period has elapsed, in other words after all currently executing rcu_bh
2832 * read-side critical sections have completed. RCU read-side critical
2833 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2834 * and may be nested.
2836 * See the description of synchronize_sched() for more detailed information
2837 * on memory ordering guarantees.
2839 void synchronize_rcu_bh(void)
2841 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2842 !lock_is_held(&rcu_lock_map) &&
2843 !lock_is_held(&rcu_sched_lock_map),
2844 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2845 if (rcu_blocking_is_gp())
2848 synchronize_rcu_bh_expedited();
2850 wait_rcu_gp(call_rcu_bh);
2852 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2855 * get_state_synchronize_rcu - Snapshot current RCU state
2857 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2858 * to determine whether or not a full grace period has elapsed in the
2861 unsigned long get_state_synchronize_rcu(void)
2864 * Any prior manipulation of RCU-protected data must happen
2865 * before the load from ->gpnum.
2870 * Make sure this load happens before the purportedly
2871 * time-consuming work between get_state_synchronize_rcu()
2872 * and cond_synchronize_rcu().
2874 return smp_load_acquire(&rcu_state_p->gpnum);
2876 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2879 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2881 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2883 * If a full RCU grace period has elapsed since the earlier call to
2884 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2885 * synchronize_rcu() to wait for a full grace period.
2887 * Yes, this function does not take counter wrap into account. But
2888 * counter wrap is harmless. If the counter wraps, we have waited for
2889 * more than 2 billion grace periods (and way more on a 64-bit system!),
2890 * so waiting for one additional grace period should be just fine.
2892 void cond_synchronize_rcu(unsigned long oldstate)
2894 unsigned long newstate;
2897 * Ensure that this load happens before any RCU-destructive
2898 * actions the caller might carry out after we return.
2900 newstate = smp_load_acquire(&rcu_state_p->completed);
2901 if (ULONG_CMP_GE(oldstate, newstate))
2904 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2906 static int synchronize_sched_expedited_cpu_stop(void *data)
2909 * There must be a full memory barrier on each affected CPU
2910 * between the time that try_stop_cpus() is called and the
2911 * time that it returns.
2913 * In the current initial implementation of cpu_stop, the
2914 * above condition is already met when the control reaches
2915 * this point and the following smp_mb() is not strictly
2916 * necessary. Do smp_mb() anyway for documentation and
2917 * robustness against future implementation changes.
2919 smp_mb(); /* See above comment block. */
2924 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2926 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2927 * approach to force the grace period to end quickly. This consumes
2928 * significant time on all CPUs and is unfriendly to real-time workloads,
2929 * so is thus not recommended for any sort of common-case code. In fact,
2930 * if you are using synchronize_sched_expedited() in a loop, please
2931 * restructure your code to batch your updates, and then use a single
2932 * synchronize_sched() instead.
2934 * Note that it is illegal to call this function while holding any lock
2935 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2936 * to call this function from a CPU-hotplug notifier. Failing to observe
2937 * these restriction will result in deadlock.
2939 * This implementation can be thought of as an application of ticket
2940 * locking to RCU, with sync_sched_expedited_started and
2941 * sync_sched_expedited_done taking on the roles of the halves
2942 * of the ticket-lock word. Each task atomically increments
2943 * sync_sched_expedited_started upon entry, snapshotting the old value,
2944 * then attempts to stop all the CPUs. If this succeeds, then each
2945 * CPU will have executed a context switch, resulting in an RCU-sched
2946 * grace period. We are then done, so we use atomic_cmpxchg() to
2947 * update sync_sched_expedited_done to match our snapshot -- but
2948 * only if someone else has not already advanced past our snapshot.
2950 * On the other hand, if try_stop_cpus() fails, we check the value
2951 * of sync_sched_expedited_done. If it has advanced past our
2952 * initial snapshot, then someone else must have forced a grace period
2953 * some time after we took our snapshot. In this case, our work is
2954 * done for us, and we can simply return. Otherwise, we try again,
2955 * but keep our initial snapshot for purposes of checking for someone
2956 * doing our work for us.
2958 * If we fail too many times in a row, we fall back to synchronize_sched().
2960 void synchronize_sched_expedited(void)
2962 long firstsnap, s, snap;
2964 struct rcu_state *rsp = &rcu_sched_state;
2967 * If we are in danger of counter wrap, just do synchronize_sched().
2968 * By allowing sync_sched_expedited_started to advance no more than
2969 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2970 * that more than 3.5 billion CPUs would be required to force a
2971 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2972 * course be required on a 64-bit system.
2974 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2975 (ulong)atomic_long_read(&rsp->expedited_done) +
2977 synchronize_sched();
2978 atomic_long_inc(&rsp->expedited_wrap);
2983 * Take a ticket. Note that atomic_inc_return() implies a
2984 * full memory barrier.
2986 snap = atomic_long_inc_return(&rsp->expedited_start);
2989 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2992 * Each pass through the following loop attempts to force a
2993 * context switch on each CPU.
2995 while (try_stop_cpus(cpu_online_mask,
2996 synchronize_sched_expedited_cpu_stop,
2999 atomic_long_inc(&rsp->expedited_tryfail);
3001 /* Check to see if someone else did our work for us. */
3002 s = atomic_long_read(&rsp->expedited_done);
3003 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3004 /* ensure test happens before caller kfree */
3005 smp_mb__before_atomic(); /* ^^^ */
3006 atomic_long_inc(&rsp->expedited_workdone1);
3010 /* No joy, try again later. Or just synchronize_sched(). */
3011 if (trycount++ < 10) {
3012 udelay(trycount * num_online_cpus());
3014 wait_rcu_gp(call_rcu_sched);
3015 atomic_long_inc(&rsp->expedited_normal);
3019 /* Recheck to see if someone else did our work for us. */
3020 s = atomic_long_read(&rsp->expedited_done);
3021 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3022 /* ensure test happens before caller kfree */
3023 smp_mb__before_atomic(); /* ^^^ */
3024 atomic_long_inc(&rsp->expedited_workdone2);
3029 * Refetching sync_sched_expedited_started allows later
3030 * callers to piggyback on our grace period. We retry
3031 * after they started, so our grace period works for them,
3032 * and they started after our first try, so their grace
3033 * period works for us.
3036 snap = atomic_long_read(&rsp->expedited_start);
3037 smp_mb(); /* ensure read is before try_stop_cpus(). */
3039 atomic_long_inc(&rsp->expedited_stoppedcpus);
3042 * Everyone up to our most recent fetch is covered by our grace
3043 * period. Update the counter, but only if our work is still
3044 * relevant -- which it won't be if someone who started later
3045 * than we did already did their update.
3048 atomic_long_inc(&rsp->expedited_done_tries);
3049 s = atomic_long_read(&rsp->expedited_done);
3050 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
3051 /* ensure test happens before caller kfree */
3052 smp_mb__before_atomic(); /* ^^^ */
3053 atomic_long_inc(&rsp->expedited_done_lost);
3056 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
3057 atomic_long_inc(&rsp->expedited_done_exit);
3061 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3064 * Check to see if there is any immediate RCU-related work to be done
3065 * by the current CPU, for the specified type of RCU, returning 1 if so.
3066 * The checks are in order of increasing expense: checks that can be
3067 * carried out against CPU-local state are performed first. However,
3068 * we must check for CPU stalls first, else we might not get a chance.
3070 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3072 struct rcu_node *rnp = rdp->mynode;
3074 rdp->n_rcu_pending++;
3076 /* Check for CPU stalls, if enabled. */
3077 check_cpu_stall(rsp, rdp);
3079 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3080 if (rcu_nohz_full_cpu(rsp))
3083 /* Is the RCU core waiting for a quiescent state from this CPU? */
3084 if (rcu_scheduler_fully_active &&
3085 rdp->qs_pending && !rdp->passed_quiesce) {
3086 rdp->n_rp_qs_pending++;
3087 } else if (rdp->qs_pending && rdp->passed_quiesce) {
3088 rdp->n_rp_report_qs++;
3092 /* Does this CPU have callbacks ready to invoke? */
3093 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3094 rdp->n_rp_cb_ready++;
3098 /* Has RCU gone idle with this CPU needing another grace period? */
3099 if (cpu_needs_another_gp(rsp, rdp)) {
3100 rdp->n_rp_cpu_needs_gp++;
3104 /* Has another RCU grace period completed? */
3105 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3106 rdp->n_rp_gp_completed++;
3110 /* Has a new RCU grace period started? */
3111 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
3112 rdp->n_rp_gp_started++;
3116 /* Does this CPU need a deferred NOCB wakeup? */
3117 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3118 rdp->n_rp_nocb_defer_wakeup++;
3123 rdp->n_rp_need_nothing++;
3128 * Check to see if there is any immediate RCU-related work to be done
3129 * by the current CPU, returning 1 if so. This function is part of the
3130 * RCU implementation; it is -not- an exported member of the RCU API.
3132 static int rcu_pending(int cpu)
3134 struct rcu_state *rsp;
3136 for_each_rcu_flavor(rsp)
3137 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
3143 * Return true if the specified CPU has any callback. If all_lazy is
3144 * non-NULL, store an indication of whether all callbacks are lazy.
3145 * (If there are no callbacks, all of them are deemed to be lazy.)
3147 static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
3151 struct rcu_data *rdp;
3152 struct rcu_state *rsp;
3154 for_each_rcu_flavor(rsp) {
3155 rdp = per_cpu_ptr(rsp->rda, cpu);
3159 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3170 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3171 * the compiler is expected to optimize this away.
3173 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3174 int cpu, unsigned long done)
3176 trace_rcu_barrier(rsp->name, s, cpu,
3177 atomic_read(&rsp->barrier_cpu_count), done);
3181 * RCU callback function for _rcu_barrier(). If we are last, wake
3182 * up the task executing _rcu_barrier().
3184 static void rcu_barrier_callback(struct rcu_head *rhp)
3186 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3187 struct rcu_state *rsp = rdp->rsp;
3189 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3190 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
3191 complete(&rsp->barrier_completion);
3193 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
3198 * Called with preemption disabled, and from cross-cpu IRQ context.
3200 static void rcu_barrier_func(void *type)
3202 struct rcu_state *rsp = type;
3203 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3205 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
3206 atomic_inc(&rsp->barrier_cpu_count);
3207 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3211 * Orchestrate the specified type of RCU barrier, waiting for all
3212 * RCU callbacks of the specified type to complete.
3214 static void _rcu_barrier(struct rcu_state *rsp)
3217 struct rcu_data *rdp;
3218 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3219 unsigned long snap_done;
3221 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3223 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3224 mutex_lock(&rsp->barrier_mutex);
3227 * Ensure that all prior references, including to ->n_barrier_done,
3228 * are ordered before the _rcu_barrier() machinery.
3230 smp_mb(); /* See above block comment. */
3233 * Recheck ->n_barrier_done to see if others did our work for us.
3234 * This means checking ->n_barrier_done for an even-to-odd-to-even
3235 * transition. The "if" expression below therefore rounds the old
3236 * value up to the next even number and adds two before comparing.
3238 snap_done = rsp->n_barrier_done;
3239 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3242 * If the value in snap is odd, we needed to wait for the current
3243 * rcu_barrier() to complete, then wait for the next one, in other
3244 * words, we need the value of snap_done to be three larger than
3245 * the value of snap. On the other hand, if the value in snap is
3246 * even, we only had to wait for the next rcu_barrier() to complete,
3247 * in other words, we need the value of snap_done to be only two
3248 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3249 * this for us (thank you, Linus!).
3251 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3252 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3253 smp_mb(); /* caller's subsequent code after above check. */
3254 mutex_unlock(&rsp->barrier_mutex);
3259 * Increment ->n_barrier_done to avoid duplicate work. Use
3260 * ACCESS_ONCE() to prevent the compiler from speculating
3261 * the increment to precede the early-exit check.
3263 ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1;
3264 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3265 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3266 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3269 * Initialize the count to one rather than to zero in order to
3270 * avoid a too-soon return to zero in case of a short grace period
3271 * (or preemption of this task). Exclude CPU-hotplug operations
3272 * to ensure that no offline CPU has callbacks queued.
3274 init_completion(&rsp->barrier_completion);
3275 atomic_set(&rsp->barrier_cpu_count, 1);
3279 * Force each CPU with callbacks to register a new callback.
3280 * When that callback is invoked, we will know that all of the
3281 * corresponding CPU's preceding callbacks have been invoked.
3283 for_each_possible_cpu(cpu) {
3284 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3286 rdp = per_cpu_ptr(rsp->rda, cpu);
3287 if (rcu_is_nocb_cpu(cpu)) {
3288 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3289 rsp->n_barrier_done);
3290 atomic_inc(&rsp->barrier_cpu_count);
3291 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3293 } else if (ACCESS_ONCE(rdp->qlen)) {
3294 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3295 rsp->n_barrier_done);
3296 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3298 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3299 rsp->n_barrier_done);
3305 * Now that we have an rcu_barrier_callback() callback on each
3306 * CPU, and thus each counted, remove the initial count.
3308 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3309 complete(&rsp->barrier_completion);
3311 /* Increment ->n_barrier_done to prevent duplicate work. */
3312 smp_mb(); /* Keep increment after above mechanism. */
3313 ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1;
3314 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3315 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3316 smp_mb(); /* Keep increment before caller's subsequent code. */
3318 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3319 wait_for_completion(&rsp->barrier_completion);
3321 /* Other rcu_barrier() invocations can now safely proceed. */
3322 mutex_unlock(&rsp->barrier_mutex);
3326 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3328 void rcu_barrier_bh(void)
3330 _rcu_barrier(&rcu_bh_state);
3332 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3335 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3337 void rcu_barrier_sched(void)
3339 _rcu_barrier(&rcu_sched_state);
3341 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3344 * Do boot-time initialization of a CPU's per-CPU RCU data.
3347 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3349 unsigned long flags;
3350 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3351 struct rcu_node *rnp = rcu_get_root(rsp);
3353 /* Set up local state, ensuring consistent view of global state. */
3354 raw_spin_lock_irqsave(&rnp->lock, flags);
3355 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3356 init_callback_list(rdp);
3358 ACCESS_ONCE(rdp->qlen) = 0;
3359 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3360 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3361 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3364 rcu_boot_init_nocb_percpu_data(rdp);
3365 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3369 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3370 * offline event can be happening at a given time. Note also that we
3371 * can accept some slop in the rsp->completed access due to the fact
3372 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3375 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3377 unsigned long flags;
3379 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3380 struct rcu_node *rnp = rcu_get_root(rsp);
3382 /* Exclude new grace periods. */
3383 mutex_lock(&rsp->onoff_mutex);
3385 /* Set up local state, ensuring consistent view of global state. */
3386 raw_spin_lock_irqsave(&rnp->lock, flags);
3387 rdp->beenonline = 1; /* We have now been online. */
3388 rdp->qlen_last_fqs_check = 0;
3389 rdp->n_force_qs_snap = rsp->n_force_qs;
3390 rdp->blimit = blimit;
3391 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3392 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3393 rcu_sysidle_init_percpu_data(rdp->dynticks);
3394 atomic_set(&rdp->dynticks->dynticks,
3395 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3396 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3398 /* Add CPU to rcu_node bitmasks. */
3400 mask = rdp->grpmask;
3402 /* Exclude any attempts to start a new GP on small systems. */
3403 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3404 rnp->qsmaskinit |= mask;
3405 mask = rnp->grpmask;
3406 if (rnp == rdp->mynode) {
3408 * If there is a grace period in progress, we will
3409 * set up to wait for it next time we run the
3412 rdp->gpnum = rnp->completed;
3413 rdp->completed = rnp->completed;
3414 rdp->passed_quiesce = 0;
3415 rdp->qs_pending = 0;
3416 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3418 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3420 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3421 local_irq_restore(flags);
3423 mutex_unlock(&rsp->onoff_mutex);
3426 static void rcu_prepare_cpu(int cpu)
3428 struct rcu_state *rsp;
3430 for_each_rcu_flavor(rsp)
3431 rcu_init_percpu_data(cpu, rsp);
3435 * Handle CPU online/offline notification events.
3437 static int rcu_cpu_notify(struct notifier_block *self,
3438 unsigned long action, void *hcpu)
3440 long cpu = (long)hcpu;
3441 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3442 struct rcu_node *rnp = rdp->mynode;
3443 struct rcu_state *rsp;
3445 trace_rcu_utilization(TPS("Start CPU hotplug"));
3447 case CPU_UP_PREPARE:
3448 case CPU_UP_PREPARE_FROZEN:
3449 rcu_prepare_cpu(cpu);
3450 rcu_prepare_kthreads(cpu);
3453 case CPU_DOWN_FAILED:
3454 rcu_boost_kthread_setaffinity(rnp, -1);
3456 case CPU_DOWN_PREPARE:
3457 rcu_boost_kthread_setaffinity(rnp, cpu);
3460 case CPU_DYING_FROZEN:
3461 for_each_rcu_flavor(rsp)
3462 rcu_cleanup_dying_cpu(rsp);
3465 case CPU_DEAD_FROZEN:
3466 case CPU_UP_CANCELED:
3467 case CPU_UP_CANCELED_FROZEN:
3468 for_each_rcu_flavor(rsp)
3469 rcu_cleanup_dead_cpu(cpu, rsp);
3474 trace_rcu_utilization(TPS("End CPU hotplug"));
3478 static int rcu_pm_notify(struct notifier_block *self,
3479 unsigned long action, void *hcpu)
3482 case PM_HIBERNATION_PREPARE:
3483 case PM_SUSPEND_PREPARE:
3484 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3487 case PM_POST_HIBERNATION:
3488 case PM_POST_SUSPEND:
3498 * Spawn the kthread that handles this RCU flavor's grace periods.
3500 static int __init rcu_spawn_gp_kthread(void)
3502 unsigned long flags;
3503 struct rcu_node *rnp;
3504 struct rcu_state *rsp;
3505 struct task_struct *t;
3507 for_each_rcu_flavor(rsp) {
3508 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3510 rnp = rcu_get_root(rsp);
3511 raw_spin_lock_irqsave(&rnp->lock, flags);
3512 rsp->gp_kthread = t;
3513 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3514 rcu_spawn_nocb_kthreads(rsp);
3518 early_initcall(rcu_spawn_gp_kthread);
3521 * This function is invoked towards the end of the scheduler's initialization
3522 * process. Before this is called, the idle task might contain
3523 * RCU read-side critical sections (during which time, this idle
3524 * task is booting the system). After this function is called, the
3525 * idle tasks are prohibited from containing RCU read-side critical
3526 * sections. This function also enables RCU lockdep checking.
3528 void rcu_scheduler_starting(void)
3530 WARN_ON(num_online_cpus() != 1);
3531 WARN_ON(nr_context_switches() > 0);
3532 rcu_scheduler_active = 1;
3536 * Compute the per-level fanout, either using the exact fanout specified
3537 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3539 #ifdef CONFIG_RCU_FANOUT_EXACT
3540 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3544 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3545 for (i = rcu_num_lvls - 2; i >= 0; i--)
3546 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3548 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3549 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3556 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3557 ccur = rsp->levelcnt[i];
3558 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3562 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3565 * Helper function for rcu_init() that initializes one rcu_state structure.
3567 static void __init rcu_init_one(struct rcu_state *rsp,
3568 struct rcu_data __percpu *rda)
3570 static const char * const buf[] = {
3574 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3575 static const char * const fqs[] = {
3579 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3580 static u8 fl_mask = 0x1;
3584 struct rcu_node *rnp;
3586 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3588 /* Silence gcc 4.8 warning about array index out of range. */
3589 if (rcu_num_lvls > RCU_NUM_LVLS)
3590 panic("rcu_init_one: rcu_num_lvls overflow");
3592 /* Initialize the level-tracking arrays. */
3594 for (i = 0; i < rcu_num_lvls; i++)
3595 rsp->levelcnt[i] = num_rcu_lvl[i];
3596 for (i = 1; i < rcu_num_lvls; i++)
3597 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3598 rcu_init_levelspread(rsp);
3599 rsp->flavor_mask = fl_mask;
3602 /* Initialize the elements themselves, starting from the leaves. */
3604 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3605 cpustride *= rsp->levelspread[i];
3606 rnp = rsp->level[i];
3607 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3608 raw_spin_lock_init(&rnp->lock);
3609 lockdep_set_class_and_name(&rnp->lock,
3610 &rcu_node_class[i], buf[i]);
3611 raw_spin_lock_init(&rnp->fqslock);
3612 lockdep_set_class_and_name(&rnp->fqslock,
3613 &rcu_fqs_class[i], fqs[i]);
3614 rnp->gpnum = rsp->gpnum;
3615 rnp->completed = rsp->completed;
3617 rnp->qsmaskinit = 0;
3618 rnp->grplo = j * cpustride;
3619 rnp->grphi = (j + 1) * cpustride - 1;
3620 if (rnp->grphi >= nr_cpu_ids)
3621 rnp->grphi = nr_cpu_ids - 1;
3627 rnp->grpnum = j % rsp->levelspread[i - 1];
3628 rnp->grpmask = 1UL << rnp->grpnum;
3629 rnp->parent = rsp->level[i - 1] +
3630 j / rsp->levelspread[i - 1];
3633 INIT_LIST_HEAD(&rnp->blkd_tasks);
3634 rcu_init_one_nocb(rnp);
3639 init_waitqueue_head(&rsp->gp_wq);
3640 rnp = rsp->level[rcu_num_lvls - 1];
3641 for_each_possible_cpu(i) {
3642 while (i > rnp->grphi)
3644 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3645 rcu_boot_init_percpu_data(i, rsp);
3647 list_add(&rsp->flavors, &rcu_struct_flavors);
3651 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3652 * replace the definitions in tree.h because those are needed to size
3653 * the ->node array in the rcu_state structure.
3655 static void __init rcu_init_geometry(void)
3661 int rcu_capacity[MAX_RCU_LVLS + 1];
3664 * Initialize any unspecified boot parameters.
3665 * The default values of jiffies_till_first_fqs and
3666 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3667 * value, which is a function of HZ, then adding one for each
3668 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3670 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3671 if (jiffies_till_first_fqs == ULONG_MAX)
3672 jiffies_till_first_fqs = d;
3673 if (jiffies_till_next_fqs == ULONG_MAX)
3674 jiffies_till_next_fqs = d;
3676 /* If the compile-time values are accurate, just leave. */
3677 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3678 nr_cpu_ids == NR_CPUS)
3680 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3681 rcu_fanout_leaf, nr_cpu_ids);
3684 * Compute number of nodes that can be handled an rcu_node tree
3685 * with the given number of levels. Setting rcu_capacity[0] makes
3686 * some of the arithmetic easier.
3688 rcu_capacity[0] = 1;
3689 rcu_capacity[1] = rcu_fanout_leaf;
3690 for (i = 2; i <= MAX_RCU_LVLS; i++)
3691 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3694 * The boot-time rcu_fanout_leaf parameter is only permitted
3695 * to increase the leaf-level fanout, not decrease it. Of course,
3696 * the leaf-level fanout cannot exceed the number of bits in
3697 * the rcu_node masks. Finally, the tree must be able to accommodate
3698 * the configured number of CPUs. Complain and fall back to the
3699 * compile-time values if these limits are exceeded.
3701 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3702 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3703 n > rcu_capacity[MAX_RCU_LVLS]) {
3708 /* Calculate the number of rcu_nodes at each level of the tree. */
3709 for (i = 1; i <= MAX_RCU_LVLS; i++)
3710 if (n <= rcu_capacity[i]) {
3711 for (j = 0; j <= i; j++)
3713 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3715 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3720 /* Calculate the total number of rcu_node structures. */
3722 for (i = 0; i <= MAX_RCU_LVLS; i++)
3723 rcu_num_nodes += num_rcu_lvl[i];
3727 void __init rcu_init(void)
3731 rcu_bootup_announce();
3732 rcu_init_geometry();
3733 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3734 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3735 __rcu_init_preempt();
3736 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3739 * We don't need protection against CPU-hotplug here because
3740 * this is called early in boot, before either interrupts
3741 * or the scheduler are operational.
3743 cpu_notifier(rcu_cpu_notify, 0);
3744 pm_notifier(rcu_pm_notify, 0);
3745 for_each_online_cpu(cpu)
3746 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3749 #include "tree_plugin.h"