2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/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>
61 #include <trace/events/rcu.h>
65 MODULE_ALIAS("rcutree");
66 #ifdef MODULE_PARAM_PREFIX
67 #undef MODULE_PARAM_PREFIX
69 #define MODULE_PARAM_PREFIX "rcutree."
71 /* Data structures. */
73 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
74 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
77 * In order to export the rcu_state name to the tracing tools, it
78 * needs to be added in the __tracepoint_string section.
79 * This requires defining a separate variable tp_<sname>_varname
80 * that points to the string being used, and this will allow
81 * the tracing userspace tools to be able to decipher the string
82 * address to the matching string.
84 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
87 struct rcu_state sname##_state = { \
88 .level = { &sname##_state.node[0] }, \
90 .fqs_state = RCU_GP_IDLE, \
91 .gpnum = 0UL - 300UL, \
92 .completed = 0UL - 300UL, \
93 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
94 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
95 .orphan_donetail = &sname##_state.orphan_donelist, \
96 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
97 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
98 .name = sname##_varname, \
101 DEFINE_PER_CPU(struct rcu_data, sname##_data)
103 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
104 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
106 static struct rcu_state *rcu_state;
107 LIST_HEAD(rcu_struct_flavors);
109 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
110 static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
111 module_param(rcu_fanout_leaf, int, 0444);
112 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
113 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
120 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
123 * The rcu_scheduler_active variable transitions from zero to one just
124 * before the first task is spawned. So when this variable is zero, RCU
125 * can assume that there is but one task, allowing RCU to (for example)
126 * optimize synchronize_sched() to a simple barrier(). When this variable
127 * is one, RCU must actually do all the hard work required to detect real
128 * grace periods. This variable is also used to suppress boot-time false
129 * positives from lockdep-RCU error checking.
131 int rcu_scheduler_active __read_mostly;
132 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
135 * The rcu_scheduler_fully_active variable transitions from zero to one
136 * during the early_initcall() processing, which is after the scheduler
137 * is capable of creating new tasks. So RCU processing (for example,
138 * creating tasks for RCU priority boosting) must be delayed until after
139 * rcu_scheduler_fully_active transitions from zero to one. We also
140 * currently delay invocation of any RCU callbacks until after this point.
142 * It might later prove better for people registering RCU callbacks during
143 * early boot to take responsibility for these callbacks, but one step at
146 static int rcu_scheduler_fully_active __read_mostly;
148 #ifdef CONFIG_RCU_BOOST
151 * Control variables for per-CPU and per-rcu_node kthreads. These
152 * handle all flavors of RCU.
154 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
155 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
156 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
157 DEFINE_PER_CPU(char, rcu_cpu_has_work);
159 #endif /* #ifdef CONFIG_RCU_BOOST */
161 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
162 static void invoke_rcu_core(void);
163 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
166 * Track the rcutorture test sequence number and the update version
167 * number within a given test. The rcutorture_testseq is incremented
168 * on every rcutorture module load and unload, so has an odd value
169 * when a test is running. The rcutorture_vernum is set to zero
170 * when rcutorture starts and is incremented on each rcutorture update.
171 * These variables enable correlating rcutorture output with the
172 * RCU tracing information.
174 unsigned long rcutorture_testseq;
175 unsigned long rcutorture_vernum;
178 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
179 * permit this function to be invoked without holding the root rcu_node
180 * structure's ->lock, but of course results can be subject to change.
182 static int rcu_gp_in_progress(struct rcu_state *rsp)
184 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
188 * Note a quiescent state. Because we do not need to know
189 * how many quiescent states passed, just if there was at least
190 * one since the start of the grace period, this just sets a flag.
191 * The caller must have disabled preemption.
193 void rcu_sched_qs(int cpu)
195 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
197 if (rdp->passed_quiesce == 0)
198 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
199 rdp->passed_quiesce = 1;
202 void rcu_bh_qs(int cpu)
204 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
206 if (rdp->passed_quiesce == 0)
207 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
208 rdp->passed_quiesce = 1;
212 * Note a context switch. This is a quiescent state for RCU-sched,
213 * and requires special handling for preemptible RCU.
214 * The caller must have disabled preemption.
216 void rcu_note_context_switch(int cpu)
218 trace_rcu_utilization(TPS("Start context switch"));
220 rcu_preempt_note_context_switch(cpu);
221 trace_rcu_utilization(TPS("End context switch"));
223 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
225 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
226 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
227 .dynticks = ATOMIC_INIT(1),
228 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
229 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
230 .dynticks_idle = ATOMIC_INIT(1),
231 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
234 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
235 static long qhimark = 10000; /* If this many pending, ignore blimit. */
236 static long qlowmark = 100; /* Once only this many pending, use blimit. */
238 module_param(blimit, long, 0444);
239 module_param(qhimark, long, 0444);
240 module_param(qlowmark, long, 0444);
242 static ulong jiffies_till_first_fqs = ULONG_MAX;
243 static ulong jiffies_till_next_fqs = ULONG_MAX;
245 module_param(jiffies_till_first_fqs, ulong, 0644);
246 module_param(jiffies_till_next_fqs, ulong, 0644);
248 static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
249 struct rcu_data *rdp);
250 static void force_qs_rnp(struct rcu_state *rsp,
251 int (*f)(struct rcu_data *rsp, bool *isidle,
252 unsigned long *maxj),
253 bool *isidle, unsigned long *maxj);
254 static void force_quiescent_state(struct rcu_state *rsp);
255 static int rcu_pending(int cpu);
258 * Return the number of RCU-sched batches processed thus far for debug & stats.
260 long rcu_batches_completed_sched(void)
262 return rcu_sched_state.completed;
264 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
267 * Return the number of RCU BH batches processed thus far for debug & stats.
269 long rcu_batches_completed_bh(void)
271 return rcu_bh_state.completed;
273 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
276 * Force a quiescent state for RCU BH.
278 void rcu_bh_force_quiescent_state(void)
280 force_quiescent_state(&rcu_bh_state);
282 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
285 * Record the number of times rcutorture tests have been initiated and
286 * terminated. This information allows the debugfs tracing stats to be
287 * correlated to the rcutorture messages, even when the rcutorture module
288 * is being repeatedly loaded and unloaded. In other words, we cannot
289 * store this state in rcutorture itself.
291 void rcutorture_record_test_transition(void)
293 rcutorture_testseq++;
294 rcutorture_vernum = 0;
296 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
299 * Record the number of writer passes through the current rcutorture test.
300 * This is also used to correlate debugfs tracing stats with the rcutorture
303 void rcutorture_record_progress(unsigned long vernum)
307 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
310 * Force a quiescent state for RCU-sched.
312 void rcu_sched_force_quiescent_state(void)
314 force_quiescent_state(&rcu_sched_state);
316 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
319 * Does the CPU have callbacks ready to be invoked?
322 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
324 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
325 rdp->nxttail[RCU_DONE_TAIL] != NULL;
329 * Does the current CPU require a not-yet-started grace period?
330 * The caller must have disabled interrupts to prevent races with
331 * normal callback registry.
334 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
338 if (rcu_gp_in_progress(rsp))
339 return 0; /* No, a grace period is already in progress. */
340 if (rcu_nocb_needs_gp(rsp))
341 return 1; /* Yes, a no-CBs CPU needs one. */
342 if (!rdp->nxttail[RCU_NEXT_TAIL])
343 return 0; /* No, this is a no-CBs (or offline) CPU. */
344 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
345 return 1; /* Yes, this CPU has newly registered callbacks. */
346 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
347 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
348 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
349 rdp->nxtcompleted[i]))
350 return 1; /* Yes, CBs for future grace period. */
351 return 0; /* No grace period needed. */
355 * Return the root node of the specified rcu_state structure.
357 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
359 return &rsp->node[0];
363 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
365 * If the new value of the ->dynticks_nesting counter now is zero,
366 * we really have entered idle, and must do the appropriate accounting.
367 * The caller must have disabled interrupts.
369 static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
372 struct rcu_state *rsp;
373 struct rcu_data *rdp;
375 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
376 if (!user && !is_idle_task(current)) {
377 struct task_struct *idle __maybe_unused =
378 idle_task(smp_processor_id());
380 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
381 ftrace_dump(DUMP_ORIG);
382 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
383 current->pid, current->comm,
384 idle->pid, idle->comm); /* must be idle task! */
386 for_each_rcu_flavor(rsp) {
387 rdp = this_cpu_ptr(rsp->rda);
388 do_nocb_deferred_wakeup(rdp);
390 rcu_prepare_for_idle(smp_processor_id());
391 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
392 smp_mb__before_atomic_inc(); /* See above. */
393 atomic_inc(&rdtp->dynticks);
394 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
395 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
398 * It is illegal to enter an extended quiescent state while
399 * in an RCU read-side critical section.
401 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
402 "Illegal idle entry in RCU read-side critical section.");
403 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
404 "Illegal idle entry in RCU-bh read-side critical section.");
405 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
406 "Illegal idle entry in RCU-sched read-side critical section.");
410 * Enter an RCU extended quiescent state, which can be either the
411 * idle loop or adaptive-tickless usermode execution.
413 static void rcu_eqs_enter(bool user)
416 struct rcu_dynticks *rdtp;
418 rdtp = this_cpu_ptr(&rcu_dynticks);
419 oldval = rdtp->dynticks_nesting;
420 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
421 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
422 rdtp->dynticks_nesting = 0;
423 rcu_eqs_enter_common(rdtp, oldval, user);
425 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
430 * rcu_idle_enter - inform RCU that current CPU is entering idle
432 * Enter idle mode, in other words, -leave- the mode in which RCU
433 * read-side critical sections can occur. (Though RCU read-side
434 * critical sections can occur in irq handlers in idle, a possibility
435 * handled by irq_enter() and irq_exit().)
437 * We crowbar the ->dynticks_nesting field to zero to allow for
438 * the possibility of usermode upcalls having messed up our count
439 * of interrupt nesting level during the prior busy period.
441 void rcu_idle_enter(void)
445 local_irq_save(flags);
446 rcu_eqs_enter(false);
447 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
448 local_irq_restore(flags);
450 EXPORT_SYMBOL_GPL(rcu_idle_enter);
452 #ifdef CONFIG_RCU_USER_QS
454 * rcu_user_enter - inform RCU that we are resuming userspace.
456 * Enter RCU idle mode right before resuming userspace. No use of RCU
457 * is permitted between this call and rcu_user_exit(). This way the
458 * CPU doesn't need to maintain the tick for RCU maintenance purposes
459 * when the CPU runs in userspace.
461 void rcu_user_enter(void)
465 #endif /* CONFIG_RCU_USER_QS */
468 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
470 * Exit from an interrupt handler, which might possibly result in entering
471 * idle mode, in other words, leaving the mode in which read-side critical
472 * sections can occur.
474 * This code assumes that the idle loop never does anything that might
475 * result in unbalanced calls to irq_enter() and irq_exit(). If your
476 * architecture violates this assumption, RCU will give you what you
477 * deserve, good and hard. But very infrequently and irreproducibly.
479 * Use things like work queues to work around this limitation.
481 * You have been warned.
483 void rcu_irq_exit(void)
487 struct rcu_dynticks *rdtp;
489 local_irq_save(flags);
490 rdtp = this_cpu_ptr(&rcu_dynticks);
491 oldval = rdtp->dynticks_nesting;
492 rdtp->dynticks_nesting--;
493 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
494 if (rdtp->dynticks_nesting)
495 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
497 rcu_eqs_enter_common(rdtp, oldval, true);
498 rcu_sysidle_enter(rdtp, 1);
499 local_irq_restore(flags);
503 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
505 * If the new value of the ->dynticks_nesting counter was previously zero,
506 * we really have exited idle, and must do the appropriate accounting.
507 * The caller must have disabled interrupts.
509 static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
512 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
513 atomic_inc(&rdtp->dynticks);
514 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
515 smp_mb__after_atomic_inc(); /* See above. */
516 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
517 rcu_cleanup_after_idle(smp_processor_id());
518 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
519 if (!user && !is_idle_task(current)) {
520 struct task_struct *idle __maybe_unused =
521 idle_task(smp_processor_id());
523 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
524 oldval, rdtp->dynticks_nesting);
525 ftrace_dump(DUMP_ORIG);
526 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
527 current->pid, current->comm,
528 idle->pid, idle->comm); /* must be idle task! */
533 * Exit an RCU extended quiescent state, which can be either the
534 * idle loop or adaptive-tickless usermode execution.
536 static void rcu_eqs_exit(bool user)
538 struct rcu_dynticks *rdtp;
541 rdtp = this_cpu_ptr(&rcu_dynticks);
542 oldval = rdtp->dynticks_nesting;
543 WARN_ON_ONCE(oldval < 0);
544 if (oldval & DYNTICK_TASK_NEST_MASK) {
545 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
547 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
548 rcu_eqs_exit_common(rdtp, oldval, user);
553 * rcu_idle_exit - inform RCU that current CPU is leaving idle
555 * Exit idle mode, in other words, -enter- the mode in which RCU
556 * read-side critical sections can occur.
558 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
559 * allow for the possibility of usermode upcalls messing up our count
560 * of interrupt nesting level during the busy period that is just
563 void rcu_idle_exit(void)
567 local_irq_save(flags);
569 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
570 local_irq_restore(flags);
572 EXPORT_SYMBOL_GPL(rcu_idle_exit);
574 #ifdef CONFIG_RCU_USER_QS
576 * rcu_user_exit - inform RCU that we are exiting userspace.
578 * Exit RCU idle mode while entering the kernel because it can
579 * run a RCU read side critical section anytime.
581 void rcu_user_exit(void)
585 #endif /* CONFIG_RCU_USER_QS */
588 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
590 * Enter an interrupt handler, which might possibly result in exiting
591 * idle mode, in other words, entering the mode in which read-side critical
592 * sections can occur.
594 * Note that the Linux kernel is fully capable of entering an interrupt
595 * handler that it never exits, for example when doing upcalls to
596 * user mode! This code assumes that the idle loop never does upcalls to
597 * user mode. If your architecture does do upcalls from the idle loop (or
598 * does anything else that results in unbalanced calls to the irq_enter()
599 * and irq_exit() functions), RCU will give you what you deserve, good
600 * and hard. But very infrequently and irreproducibly.
602 * Use things like work queues to work around this limitation.
604 * You have been warned.
606 void rcu_irq_enter(void)
609 struct rcu_dynticks *rdtp;
612 local_irq_save(flags);
613 rdtp = this_cpu_ptr(&rcu_dynticks);
614 oldval = rdtp->dynticks_nesting;
615 rdtp->dynticks_nesting++;
616 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
618 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
620 rcu_eqs_exit_common(rdtp, oldval, true);
621 rcu_sysidle_exit(rdtp, 1);
622 local_irq_restore(flags);
626 * rcu_nmi_enter - inform RCU of entry to NMI context
628 * If the CPU was idle with dynamic ticks active, and there is no
629 * irq handler running, this updates rdtp->dynticks_nmi to let the
630 * RCU grace-period handling know that the CPU is active.
632 void rcu_nmi_enter(void)
634 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
636 if (rdtp->dynticks_nmi_nesting == 0 &&
637 (atomic_read(&rdtp->dynticks) & 0x1))
639 rdtp->dynticks_nmi_nesting++;
640 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
641 atomic_inc(&rdtp->dynticks);
642 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
643 smp_mb__after_atomic_inc(); /* See above. */
644 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
648 * rcu_nmi_exit - inform RCU of exit from NMI context
650 * If the CPU was idle with dynamic ticks active, and there is no
651 * irq handler running, this updates rdtp->dynticks_nmi to let the
652 * RCU grace-period handling know that the CPU is no longer active.
654 void rcu_nmi_exit(void)
656 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
658 if (rdtp->dynticks_nmi_nesting == 0 ||
659 --rdtp->dynticks_nmi_nesting != 0)
661 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
662 smp_mb__before_atomic_inc(); /* See above. */
663 atomic_inc(&rdtp->dynticks);
664 smp_mb__after_atomic_inc(); /* Force delay to next write. */
665 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
669 * __rcu_is_watching - are RCU read-side critical sections safe?
671 * Return true if RCU is watching the running CPU, which means that
672 * this CPU can safely enter RCU read-side critical sections. Unlike
673 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
674 * least disabled preemption.
676 bool notrace __rcu_is_watching(void)
678 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
682 * rcu_is_watching - see if RCU thinks that the current CPU is idle
684 * If the current CPU is in its idle loop and is neither in an interrupt
685 * or NMI handler, return true.
687 bool notrace rcu_is_watching(void)
692 ret = __rcu_is_watching();
696 EXPORT_SYMBOL_GPL(rcu_is_watching);
698 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
701 * Is the current CPU online? Disable preemption to avoid false positives
702 * that could otherwise happen due to the current CPU number being sampled,
703 * this task being preempted, its old CPU being taken offline, resuming
704 * on some other CPU, then determining that its old CPU is now offline.
705 * It is OK to use RCU on an offline processor during initial boot, hence
706 * the check for rcu_scheduler_fully_active. Note also that it is OK
707 * for a CPU coming online to use RCU for one jiffy prior to marking itself
708 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
709 * offline to continue to use RCU for one jiffy after marking itself
710 * offline in the cpu_online_mask. This leniency is necessary given the
711 * non-atomic nature of the online and offline processing, for example,
712 * the fact that a CPU enters the scheduler after completing the CPU_DYING
715 * This is also why RCU internally marks CPUs online during the
716 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
718 * Disable checking if in an NMI handler because we cannot safely report
719 * errors from NMI handlers anyway.
721 bool rcu_lockdep_current_cpu_online(void)
723 struct rcu_data *rdp;
724 struct rcu_node *rnp;
730 rdp = this_cpu_ptr(&rcu_sched_data);
732 ret = (rdp->grpmask & rnp->qsmaskinit) ||
733 !rcu_scheduler_fully_active;
737 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
739 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
742 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
744 * If the current CPU is idle or running at a first-level (not nested)
745 * interrupt from idle, return true. The caller must have at least
746 * disabled preemption.
748 static int rcu_is_cpu_rrupt_from_idle(void)
750 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
754 * Snapshot the specified CPU's dynticks counter so that we can later
755 * credit them with an implicit quiescent state. Return 1 if this CPU
756 * is in dynticks idle mode, which is an extended quiescent state.
758 static int dyntick_save_progress_counter(struct rcu_data *rdp,
759 bool *isidle, unsigned long *maxj)
761 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
762 rcu_sysidle_check_cpu(rdp, isidle, maxj);
763 return (rdp->dynticks_snap & 0x1) == 0;
767 * This function really isn't for public consumption, but RCU is special in
768 * that context switches can allow the state machine to make progress.
770 extern void resched_cpu(int cpu);
773 * Return true if the specified CPU has passed through a quiescent
774 * state by virtue of being in or having passed through an dynticks
775 * idle state since the last call to dyntick_save_progress_counter()
776 * for this same CPU, or by virtue of having been offline.
778 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
779 bool *isidle, unsigned long *maxj)
784 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
785 snap = (unsigned int)rdp->dynticks_snap;
788 * If the CPU passed through or entered a dynticks idle phase with
789 * no active irq/NMI handlers, then we can safely pretend that the CPU
790 * already acknowledged the request to pass through a quiescent
791 * state. Either way, that CPU cannot possibly be in an RCU
792 * read-side critical section that started before the beginning
793 * of the current RCU grace period.
795 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
796 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
802 * Check for the CPU being offline, but only if the grace period
803 * is old enough. We don't need to worry about the CPU changing
804 * state: If we see it offline even once, it has been through a
807 * The reason for insisting that the grace period be at least
808 * one jiffy old is that CPUs that are not quite online and that
809 * have just gone offline can still execute RCU read-side critical
812 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
813 return 0; /* Grace period is not old enough. */
815 if (cpu_is_offline(rdp->cpu)) {
816 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
822 * There is a possibility that a CPU in adaptive-ticks state
823 * might run in the kernel with the scheduling-clock tick disabled
824 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
825 * force the CPU to restart the scheduling-clock tick in this
826 * CPU is in this state.
828 rcu_kick_nohz_cpu(rdp->cpu);
831 * Alternatively, the CPU might be running in the kernel
832 * for an extended period of time without a quiescent state.
833 * Attempt to force the CPU through the scheduler to gain the
834 * needed quiescent state, but only if the grace period has gone
835 * on for an uncommonly long time. If there are many stuck CPUs,
836 * we will beat on the first one until it gets unstuck, then move
837 * to the next. Only do this for the primary flavor of RCU.
839 if (rdp->rsp == rcu_state &&
840 ULONG_CMP_GE(ACCESS_ONCE(jiffies), rdp->rsp->jiffies_resched)) {
841 rdp->rsp->jiffies_resched += 5;
842 resched_cpu(rdp->cpu);
848 static void record_gp_stall_check_time(struct rcu_state *rsp)
850 unsigned long j = ACCESS_ONCE(jiffies);
854 smp_wmb(); /* Record start time before stall time. */
855 j1 = rcu_jiffies_till_stall_check();
856 rsp->jiffies_stall = j + j1;
857 rsp->jiffies_resched = j + j1 / 2;
861 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
862 * for architectures that do not implement trigger_all_cpu_backtrace().
863 * The NMI-triggered stack traces are more accurate because they are
864 * printed by the target CPU.
866 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
870 struct rcu_node *rnp;
872 rcu_for_each_leaf_node(rsp, rnp) {
873 raw_spin_lock_irqsave(&rnp->lock, flags);
874 if (rnp->qsmask != 0) {
875 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
876 if (rnp->qsmask & (1UL << cpu))
877 dump_cpu_task(rnp->grplo + cpu);
879 raw_spin_unlock_irqrestore(&rnp->lock, flags);
883 static void print_other_cpu_stall(struct rcu_state *rsp)
889 struct rcu_node *rnp = rcu_get_root(rsp);
892 /* Only let one CPU complain about others per time interval. */
894 raw_spin_lock_irqsave(&rnp->lock, flags);
895 delta = jiffies - rsp->jiffies_stall;
896 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
897 raw_spin_unlock_irqrestore(&rnp->lock, flags);
900 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
901 raw_spin_unlock_irqrestore(&rnp->lock, flags);
904 * OK, time to rat on our buddy...
905 * See Documentation/RCU/stallwarn.txt for info on how to debug
906 * RCU CPU stall warnings.
908 pr_err("INFO: %s detected stalls on CPUs/tasks:",
910 print_cpu_stall_info_begin();
911 rcu_for_each_leaf_node(rsp, rnp) {
912 raw_spin_lock_irqsave(&rnp->lock, flags);
913 ndetected += rcu_print_task_stall(rnp);
914 if (rnp->qsmask != 0) {
915 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
916 if (rnp->qsmask & (1UL << cpu)) {
917 print_cpu_stall_info(rsp,
922 raw_spin_unlock_irqrestore(&rnp->lock, flags);
926 * Now rat on any tasks that got kicked up to the root rcu_node
927 * due to CPU offlining.
929 rnp = rcu_get_root(rsp);
930 raw_spin_lock_irqsave(&rnp->lock, flags);
931 ndetected += rcu_print_task_stall(rnp);
932 raw_spin_unlock_irqrestore(&rnp->lock, flags);
934 print_cpu_stall_info_end();
935 for_each_possible_cpu(cpu)
936 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
937 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
938 smp_processor_id(), (long)(jiffies - rsp->gp_start),
939 rsp->gpnum, rsp->completed, totqlen);
941 pr_err("INFO: Stall ended before state dump start\n");
942 else if (!trigger_all_cpu_backtrace())
943 rcu_dump_cpu_stacks(rsp);
945 /* Complain about tasks blocking the grace period. */
947 rcu_print_detail_task_stall(rsp);
949 force_quiescent_state(rsp); /* Kick them all. */
953 * This function really isn't for public consumption, but RCU is special in
954 * that context switches can allow the state machine to make progress.
956 extern void resched_cpu(int cpu);
958 static void print_cpu_stall(struct rcu_state *rsp)
962 struct rcu_node *rnp = rcu_get_root(rsp);
966 * OK, time to rat on ourselves...
967 * See Documentation/RCU/stallwarn.txt for info on how to debug
968 * RCU CPU stall warnings.
970 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
971 print_cpu_stall_info_begin();
972 print_cpu_stall_info(rsp, smp_processor_id());
973 print_cpu_stall_info_end();
974 for_each_possible_cpu(cpu)
975 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
976 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
977 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
978 if (!trigger_all_cpu_backtrace())
981 raw_spin_lock_irqsave(&rnp->lock, flags);
982 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
983 rsp->jiffies_stall = jiffies +
984 3 * rcu_jiffies_till_stall_check() + 3;
985 raw_spin_unlock_irqrestore(&rnp->lock, flags);
988 * Attempt to revive the RCU machinery by forcing a context switch.
990 * A context switch would normally allow the RCU state machine to make
991 * progress and it could be we're stuck in kernel space without context
992 * switches for an entirely unreasonable amount of time.
994 resched_cpu(smp_processor_id());
997 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
999 unsigned long completed;
1000 unsigned long gpnum;
1004 struct rcu_node *rnp;
1006 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1008 j = ACCESS_ONCE(jiffies);
1011 * Lots of memory barriers to reject false positives.
1013 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1014 * then rsp->gp_start, and finally rsp->completed. These values
1015 * are updated in the opposite order with memory barriers (or
1016 * equivalent) during grace-period initialization and cleanup.
1017 * Now, a false positive can occur if we get an new value of
1018 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1019 * the memory barriers, the only way that this can happen is if one
1020 * grace period ends and another starts between these two fetches.
1021 * Detect this by comparing rsp->completed with the previous fetch
1024 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1025 * and rsp->gp_start suffice to forestall false positives.
1027 gpnum = ACCESS_ONCE(rsp->gpnum);
1028 smp_rmb(); /* Pick up ->gpnum first... */
1029 js = ACCESS_ONCE(rsp->jiffies_stall);
1030 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1031 gps = ACCESS_ONCE(rsp->gp_start);
1032 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1033 completed = ACCESS_ONCE(rsp->completed);
1034 if (ULONG_CMP_GE(completed, gpnum) ||
1035 ULONG_CMP_LT(j, js) ||
1036 ULONG_CMP_GE(gps, js))
1037 return; /* No stall or GP completed since entering function. */
1039 if (rcu_gp_in_progress(rsp) &&
1040 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1042 /* We haven't checked in, so go dump stack. */
1043 print_cpu_stall(rsp);
1045 } else if (rcu_gp_in_progress(rsp) &&
1046 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1048 /* They had a few time units to dump stack, so complain. */
1049 print_other_cpu_stall(rsp);
1054 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1056 * Set the stall-warning timeout way off into the future, thus preventing
1057 * any RCU CPU stall-warning messages from appearing in the current set of
1058 * RCU grace periods.
1060 * The caller must disable hard irqs.
1062 void rcu_cpu_stall_reset(void)
1064 struct rcu_state *rsp;
1066 for_each_rcu_flavor(rsp)
1067 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1071 * Initialize the specified rcu_data structure's callback list to empty.
1073 static void init_callback_list(struct rcu_data *rdp)
1077 if (init_nocb_callback_list(rdp))
1079 rdp->nxtlist = NULL;
1080 for (i = 0; i < RCU_NEXT_SIZE; i++)
1081 rdp->nxttail[i] = &rdp->nxtlist;
1085 * Determine the value that ->completed will have at the end of the
1086 * next subsequent grace period. This is used to tag callbacks so that
1087 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1088 * been dyntick-idle for an extended period with callbacks under the
1089 * influence of RCU_FAST_NO_HZ.
1091 * The caller must hold rnp->lock with interrupts disabled.
1093 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1094 struct rcu_node *rnp)
1097 * If RCU is idle, we just wait for the next grace period.
1098 * But we can only be sure that RCU is idle if we are looking
1099 * at the root rcu_node structure -- otherwise, a new grace
1100 * period might have started, but just not yet gotten around
1101 * to initializing the current non-root rcu_node structure.
1103 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1104 return rnp->completed + 1;
1107 * Otherwise, wait for a possible partial grace period and
1108 * then the subsequent full grace period.
1110 return rnp->completed + 2;
1114 * Trace-event helper function for rcu_start_future_gp() and
1115 * rcu_nocb_wait_gp().
1117 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1118 unsigned long c, const char *s)
1120 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1121 rnp->completed, c, rnp->level,
1122 rnp->grplo, rnp->grphi, s);
1126 * Start some future grace period, as needed to handle newly arrived
1127 * callbacks. The required future grace periods are recorded in each
1128 * rcu_node structure's ->need_future_gp field.
1130 * The caller must hold the specified rcu_node structure's ->lock.
1132 static unsigned long __maybe_unused
1133 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1137 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1140 * Pick up grace-period number for new callbacks. If this
1141 * grace period is already marked as needed, return to the caller.
1143 c = rcu_cbs_completed(rdp->rsp, rnp);
1144 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1145 if (rnp->need_future_gp[c & 0x1]) {
1146 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1151 * If either this rcu_node structure or the root rcu_node structure
1152 * believe that a grace period is in progress, then we must wait
1153 * for the one following, which is in "c". Because our request
1154 * will be noticed at the end of the current grace period, we don't
1155 * need to explicitly start one.
1157 if (rnp->gpnum != rnp->completed ||
1158 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1159 rnp->need_future_gp[c & 0x1]++;
1160 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1165 * There might be no grace period in progress. If we don't already
1166 * hold it, acquire the root rcu_node structure's lock in order to
1167 * start one (if needed).
1169 if (rnp != rnp_root) {
1170 raw_spin_lock(&rnp_root->lock);
1171 smp_mb__after_unlock_lock();
1175 * Get a new grace-period number. If there really is no grace
1176 * period in progress, it will be smaller than the one we obtained
1177 * earlier. Adjust callbacks as needed. Note that even no-CBs
1178 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1180 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1181 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1182 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1183 rdp->nxtcompleted[i] = c;
1186 * If the needed for the required grace period is already
1187 * recorded, trace and leave.
1189 if (rnp_root->need_future_gp[c & 0x1]) {
1190 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1194 /* Record the need for the future grace period. */
1195 rnp_root->need_future_gp[c & 0x1]++;
1197 /* If a grace period is not already in progress, start one. */
1198 if (rnp_root->gpnum != rnp_root->completed) {
1199 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1201 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1202 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1205 if (rnp != rnp_root)
1206 raw_spin_unlock(&rnp_root->lock);
1211 * Clean up any old requests for the just-ended grace period. Also return
1212 * whether any additional grace periods have been requested. Also invoke
1213 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1214 * waiting for this grace period to complete.
1216 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1218 int c = rnp->completed;
1220 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1222 rcu_nocb_gp_cleanup(rsp, rnp);
1223 rnp->need_future_gp[c & 0x1] = 0;
1224 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1225 trace_rcu_future_gp(rnp, rdp, c,
1226 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1231 * Awaken the grace-period kthread for the specified flavor of RCU.
1232 * Don't do a self-awaken, and don't bother awakening when there is
1233 * nothing for the grace-period kthread to do (as in several CPUs
1234 * raced to awaken, and we lost), and finally don't try to awaken
1235 * a kthread that has not yet been created.
1237 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1239 if (current == rsp->gp_kthread ||
1240 !ACCESS_ONCE(rsp->gp_flags) ||
1243 wake_up(&rsp->gp_wq);
1247 * If there is room, assign a ->completed number to any callbacks on
1248 * this CPU that have not already been assigned. Also accelerate any
1249 * callbacks that were previously assigned a ->completed number that has
1250 * since proven to be too conservative, which can happen if callbacks get
1251 * assigned a ->completed number while RCU is idle, but with reference to
1252 * a non-root rcu_node structure. This function is idempotent, so it does
1253 * not hurt to call it repeatedly.
1255 * The caller must hold rnp->lock with interrupts disabled.
1257 static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1258 struct rcu_data *rdp)
1263 /* If the CPU has no callbacks, nothing to do. */
1264 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1268 * Starting from the sublist containing the callbacks most
1269 * recently assigned a ->completed number and working down, find the
1270 * first sublist that is not assignable to an upcoming grace period.
1271 * Such a sublist has something in it (first two tests) and has
1272 * a ->completed number assigned that will complete sooner than
1273 * the ->completed number for newly arrived callbacks (last test).
1275 * The key point is that any later sublist can be assigned the
1276 * same ->completed number as the newly arrived callbacks, which
1277 * means that the callbacks in any of these later sublist can be
1278 * grouped into a single sublist, whether or not they have already
1279 * been assigned a ->completed number.
1281 c = rcu_cbs_completed(rsp, rnp);
1282 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1283 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1284 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1288 * If there are no sublist for unassigned callbacks, leave.
1289 * At the same time, advance "i" one sublist, so that "i" will
1290 * index into the sublist where all the remaining callbacks should
1293 if (++i >= RCU_NEXT_TAIL)
1297 * Assign all subsequent callbacks' ->completed number to the next
1298 * full grace period and group them all in the sublist initially
1301 for (; i <= RCU_NEXT_TAIL; i++) {
1302 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1303 rdp->nxtcompleted[i] = c;
1305 /* Record any needed additional grace periods. */
1306 rcu_start_future_gp(rnp, rdp);
1308 /* Trace depending on how much we were able to accelerate. */
1309 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1310 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1312 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1316 * Move any callbacks whose grace period has completed to the
1317 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1318 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1319 * sublist. This function is idempotent, so it does not hurt to
1320 * invoke it repeatedly. As long as it is not invoked -too- often...
1322 * The caller must hold rnp->lock with interrupts disabled.
1324 static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1325 struct rcu_data *rdp)
1329 /* If the CPU has no callbacks, nothing to do. */
1330 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1334 * Find all callbacks whose ->completed numbers indicate that they
1335 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1337 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1338 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1340 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1342 /* Clean up any sublist tail pointers that were misordered above. */
1343 for (j = RCU_WAIT_TAIL; j < i; j++)
1344 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1346 /* Copy down callbacks to fill in empty sublists. */
1347 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1348 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1350 rdp->nxttail[j] = rdp->nxttail[i];
1351 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1354 /* Classify any remaining callbacks. */
1355 rcu_accelerate_cbs(rsp, rnp, rdp);
1359 * Update CPU-local rcu_data state to record the beginnings and ends of
1360 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1361 * structure corresponding to the current CPU, and must have irqs disabled.
1363 static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1365 /* Handle the ends of any preceding grace periods first. */
1366 if (rdp->completed == rnp->completed) {
1368 /* No grace period end, so just accelerate recent callbacks. */
1369 rcu_accelerate_cbs(rsp, rnp, rdp);
1373 /* Advance callbacks. */
1374 rcu_advance_cbs(rsp, rnp, rdp);
1376 /* Remember that we saw this grace-period completion. */
1377 rdp->completed = rnp->completed;
1378 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1381 if (rdp->gpnum != rnp->gpnum) {
1383 * If the current grace period is waiting for this CPU,
1384 * set up to detect a quiescent state, otherwise don't
1385 * go looking for one.
1387 rdp->gpnum = rnp->gpnum;
1388 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1389 rdp->passed_quiesce = 0;
1390 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1391 zero_cpu_stall_ticks(rdp);
1395 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1397 unsigned long flags;
1398 struct rcu_node *rnp;
1400 local_irq_save(flags);
1402 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1403 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1404 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1405 local_irq_restore(flags);
1408 smp_mb__after_unlock_lock();
1409 __note_gp_changes(rsp, rnp, rdp);
1410 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1414 * Initialize a new grace period. Return 0 if no grace period required.
1416 static int rcu_gp_init(struct rcu_state *rsp)
1418 struct rcu_data *rdp;
1419 struct rcu_node *rnp = rcu_get_root(rsp);
1421 rcu_bind_gp_kthread();
1422 raw_spin_lock_irq(&rnp->lock);
1423 smp_mb__after_unlock_lock();
1424 if (rsp->gp_flags == 0) {
1425 /* Spurious wakeup, tell caller to go back to sleep. */
1426 raw_spin_unlock_irq(&rnp->lock);
1429 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1431 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1433 * Grace period already in progress, don't start another.
1434 * Not supposed to be able to happen.
1436 raw_spin_unlock_irq(&rnp->lock);
1440 /* Advance to a new grace period and initialize state. */
1441 record_gp_stall_check_time(rsp);
1442 smp_wmb(); /* Record GP times before starting GP. */
1444 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1445 raw_spin_unlock_irq(&rnp->lock);
1447 /* Exclude any concurrent CPU-hotplug operations. */
1448 mutex_lock(&rsp->onoff_mutex);
1451 * Set the quiescent-state-needed bits in all the rcu_node
1452 * structures for all currently online CPUs in breadth-first order,
1453 * starting from the root rcu_node structure, relying on the layout
1454 * of the tree within the rsp->node[] array. Note that other CPUs
1455 * will access only the leaves of the hierarchy, thus seeing that no
1456 * grace period is in progress, at least until the corresponding
1457 * leaf node has been initialized. In addition, we have excluded
1458 * CPU-hotplug operations.
1460 * The grace period cannot complete until the initialization
1461 * process finishes, because this kthread handles both.
1463 rcu_for_each_node_breadth_first(rsp, rnp) {
1464 raw_spin_lock_irq(&rnp->lock);
1465 smp_mb__after_unlock_lock();
1466 rdp = this_cpu_ptr(rsp->rda);
1467 rcu_preempt_check_blocked_tasks(rnp);
1468 rnp->qsmask = rnp->qsmaskinit;
1469 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1470 WARN_ON_ONCE(rnp->completed != rsp->completed);
1471 ACCESS_ONCE(rnp->completed) = rsp->completed;
1472 if (rnp == rdp->mynode)
1473 __note_gp_changes(rsp, rnp, rdp);
1474 rcu_preempt_boost_start_gp(rnp);
1475 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1476 rnp->level, rnp->grplo,
1477 rnp->grphi, rnp->qsmask);
1478 raw_spin_unlock_irq(&rnp->lock);
1479 #ifdef CONFIG_PROVE_RCU_DELAY
1480 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1481 system_state == SYSTEM_RUNNING)
1483 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1487 mutex_unlock(&rsp->onoff_mutex);
1492 * Do one round of quiescent-state forcing.
1494 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1496 int fqs_state = fqs_state_in;
1497 bool isidle = false;
1499 struct rcu_node *rnp = rcu_get_root(rsp);
1502 if (fqs_state == RCU_SAVE_DYNTICK) {
1503 /* Collect dyntick-idle snapshots. */
1504 if (is_sysidle_rcu_state(rsp)) {
1506 maxj = jiffies - ULONG_MAX / 4;
1508 force_qs_rnp(rsp, dyntick_save_progress_counter,
1510 rcu_sysidle_report_gp(rsp, isidle, maxj);
1511 fqs_state = RCU_FORCE_QS;
1513 /* Handle dyntick-idle and offline CPUs. */
1515 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1517 /* Clear flag to prevent immediate re-entry. */
1518 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1519 raw_spin_lock_irq(&rnp->lock);
1520 smp_mb__after_unlock_lock();
1521 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1522 raw_spin_unlock_irq(&rnp->lock);
1528 * Clean up after the old grace period.
1530 static void rcu_gp_cleanup(struct rcu_state *rsp)
1532 unsigned long gp_duration;
1534 struct rcu_data *rdp;
1535 struct rcu_node *rnp = rcu_get_root(rsp);
1537 raw_spin_lock_irq(&rnp->lock);
1538 smp_mb__after_unlock_lock();
1539 gp_duration = jiffies - rsp->gp_start;
1540 if (gp_duration > rsp->gp_max)
1541 rsp->gp_max = gp_duration;
1544 * We know the grace period is complete, but to everyone else
1545 * it appears to still be ongoing. But it is also the case
1546 * that to everyone else it looks like there is nothing that
1547 * they can do to advance the grace period. It is therefore
1548 * safe for us to drop the lock in order to mark the grace
1549 * period as completed in all of the rcu_node structures.
1551 raw_spin_unlock_irq(&rnp->lock);
1554 * Propagate new ->completed value to rcu_node structures so
1555 * that other CPUs don't have to wait until the start of the next
1556 * grace period to process their callbacks. This also avoids
1557 * some nasty RCU grace-period initialization races by forcing
1558 * the end of the current grace period to be completely recorded in
1559 * all of the rcu_node structures before the beginning of the next
1560 * grace period is recorded in any of the rcu_node structures.
1562 rcu_for_each_node_breadth_first(rsp, rnp) {
1563 raw_spin_lock_irq(&rnp->lock);
1564 smp_mb__after_unlock_lock();
1565 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1566 rdp = this_cpu_ptr(rsp->rda);
1567 if (rnp == rdp->mynode)
1568 __note_gp_changes(rsp, rnp, rdp);
1569 /* smp_mb() provided by prior unlock-lock pair. */
1570 nocb += rcu_future_gp_cleanup(rsp, rnp);
1571 raw_spin_unlock_irq(&rnp->lock);
1574 rnp = rcu_get_root(rsp);
1575 raw_spin_lock_irq(&rnp->lock);
1576 smp_mb__after_unlock_lock();
1577 rcu_nocb_gp_set(rnp, nocb);
1579 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1580 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1581 rsp->fqs_state = RCU_GP_IDLE;
1582 rdp = this_cpu_ptr(rsp->rda);
1583 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1584 if (cpu_needs_another_gp(rsp, rdp)) {
1585 rsp->gp_flags = RCU_GP_FLAG_INIT;
1586 trace_rcu_grace_period(rsp->name,
1587 ACCESS_ONCE(rsp->gpnum),
1590 raw_spin_unlock_irq(&rnp->lock);
1594 * Body of kthread that handles grace periods.
1596 static int __noreturn rcu_gp_kthread(void *arg)
1602 struct rcu_state *rsp = arg;
1603 struct rcu_node *rnp = rcu_get_root(rsp);
1607 /* Handle grace-period start. */
1609 trace_rcu_grace_period(rsp->name,
1610 ACCESS_ONCE(rsp->gpnum),
1612 wait_event_interruptible(rsp->gp_wq,
1613 ACCESS_ONCE(rsp->gp_flags) &
1615 /* Locking provides needed memory barrier. */
1616 if (rcu_gp_init(rsp))
1619 flush_signals(current);
1620 trace_rcu_grace_period(rsp->name,
1621 ACCESS_ONCE(rsp->gpnum),
1625 /* Handle quiescent-state forcing. */
1626 fqs_state = RCU_SAVE_DYNTICK;
1627 j = jiffies_till_first_fqs;
1630 jiffies_till_first_fqs = HZ;
1635 rsp->jiffies_force_qs = jiffies + j;
1636 trace_rcu_grace_period(rsp->name,
1637 ACCESS_ONCE(rsp->gpnum),
1639 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1640 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1642 (!ACCESS_ONCE(rnp->qsmask) &&
1643 !rcu_preempt_blocked_readers_cgp(rnp)),
1645 /* Locking provides needed memory barriers. */
1646 /* If grace period done, leave loop. */
1647 if (!ACCESS_ONCE(rnp->qsmask) &&
1648 !rcu_preempt_blocked_readers_cgp(rnp))
1650 /* If time for quiescent-state forcing, do it. */
1651 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1652 (gf & RCU_GP_FLAG_FQS)) {
1653 trace_rcu_grace_period(rsp->name,
1654 ACCESS_ONCE(rsp->gpnum),
1656 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1657 trace_rcu_grace_period(rsp->name,
1658 ACCESS_ONCE(rsp->gpnum),
1662 /* Deal with stray signal. */
1664 flush_signals(current);
1665 trace_rcu_grace_period(rsp->name,
1666 ACCESS_ONCE(rsp->gpnum),
1669 j = jiffies_till_next_fqs;
1672 jiffies_till_next_fqs = HZ;
1675 jiffies_till_next_fqs = 1;
1679 /* Handle grace-period end. */
1680 rcu_gp_cleanup(rsp);
1684 static void rsp_wakeup(struct irq_work *work)
1686 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1688 /* Wake up rcu_gp_kthread() to start the grace period. */
1689 rcu_gp_kthread_wake(rsp);
1693 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1694 * in preparation for detecting the next grace period. The caller must hold
1695 * the root node's ->lock and hard irqs must be disabled.
1697 * Note that it is legal for a dying CPU (which is marked as offline) to
1698 * invoke this function. This can happen when the dying CPU reports its
1702 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1703 struct rcu_data *rdp)
1705 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1707 * Either we have not yet spawned the grace-period
1708 * task, this CPU does not need another grace period,
1709 * or a grace period is already in progress.
1710 * Either way, don't start a new grace period.
1714 rsp->gp_flags = RCU_GP_FLAG_INIT;
1715 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1719 * We can't do wakeups while holding the rnp->lock, as that
1720 * could cause possible deadlocks with the rq->lock. Defer
1721 * the wakeup to interrupt context. And don't bother waking
1722 * up the running kthread.
1724 if (current != rsp->gp_kthread)
1725 irq_work_queue(&rsp->wakeup_work);
1729 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1730 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1731 * is invoked indirectly from rcu_advance_cbs(), which would result in
1732 * endless recursion -- or would do so if it wasn't for the self-deadlock
1733 * that is encountered beforehand.
1736 rcu_start_gp(struct rcu_state *rsp)
1738 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1739 struct rcu_node *rnp = rcu_get_root(rsp);
1742 * If there is no grace period in progress right now, any
1743 * callbacks we have up to this point will be satisfied by the
1744 * next grace period. Also, advancing the callbacks reduces the
1745 * probability of false positives from cpu_needs_another_gp()
1746 * resulting in pointless grace periods. So, advance callbacks
1747 * then start the grace period!
1749 rcu_advance_cbs(rsp, rnp, rdp);
1750 rcu_start_gp_advanced(rsp, rnp, rdp);
1754 * Report a full set of quiescent states to the specified rcu_state
1755 * data structure. This involves cleaning up after the prior grace
1756 * period and letting rcu_start_gp() start up the next grace period
1757 * if one is needed. Note that the caller must hold rnp->lock, which
1758 * is released before return.
1760 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1761 __releases(rcu_get_root(rsp)->lock)
1763 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1764 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1765 rcu_gp_kthread_wake(rsp);
1769 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1770 * Allows quiescent states for a group of CPUs to be reported at one go
1771 * to the specified rcu_node structure, though all the CPUs in the group
1772 * must be represented by the same rcu_node structure (which need not be
1773 * a leaf rcu_node structure, though it often will be). That structure's
1774 * lock must be held upon entry, and it is released before return.
1777 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1778 struct rcu_node *rnp, unsigned long flags)
1779 __releases(rnp->lock)
1781 struct rcu_node *rnp_c;
1783 /* Walk up the rcu_node hierarchy. */
1785 if (!(rnp->qsmask & mask)) {
1787 /* Our bit has already been cleared, so done. */
1788 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1791 rnp->qsmask &= ~mask;
1792 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1793 mask, rnp->qsmask, rnp->level,
1794 rnp->grplo, rnp->grphi,
1796 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1798 /* Other bits still set at this level, so done. */
1799 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1802 mask = rnp->grpmask;
1803 if (rnp->parent == NULL) {
1805 /* No more levels. Exit loop holding root lock. */
1809 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1812 raw_spin_lock_irqsave(&rnp->lock, flags);
1813 smp_mb__after_unlock_lock();
1814 WARN_ON_ONCE(rnp_c->qsmask);
1818 * Get here if we are the last CPU to pass through a quiescent
1819 * state for this grace period. Invoke rcu_report_qs_rsp()
1820 * to clean up and start the next grace period if one is needed.
1822 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1826 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1827 * structure. This must be either called from the specified CPU, or
1828 * called when the specified CPU is known to be offline (and when it is
1829 * also known that no other CPU is concurrently trying to help the offline
1830 * CPU). The lastcomp argument is used to make sure we are still in the
1831 * grace period of interest. We don't want to end the current grace period
1832 * based on quiescent states detected in an earlier grace period!
1835 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1837 unsigned long flags;
1839 struct rcu_node *rnp;
1842 raw_spin_lock_irqsave(&rnp->lock, flags);
1843 smp_mb__after_unlock_lock();
1844 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1845 rnp->completed == rnp->gpnum) {
1848 * The grace period in which this quiescent state was
1849 * recorded has ended, so don't report it upwards.
1850 * We will instead need a new quiescent state that lies
1851 * within the current grace period.
1853 rdp->passed_quiesce = 0; /* need qs for new gp. */
1854 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1857 mask = rdp->grpmask;
1858 if ((rnp->qsmask & mask) == 0) {
1859 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1861 rdp->qs_pending = 0;
1864 * This GP can't end until cpu checks in, so all of our
1865 * callbacks can be processed during the next GP.
1867 rcu_accelerate_cbs(rsp, rnp, rdp);
1869 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1874 * Check to see if there is a new grace period of which this CPU
1875 * is not yet aware, and if so, set up local rcu_data state for it.
1876 * Otherwise, see if this CPU has just passed through its first
1877 * quiescent state for this grace period, and record that fact if so.
1880 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1882 /* Check for grace-period ends and beginnings. */
1883 note_gp_changes(rsp, rdp);
1886 * Does this CPU still need to do its part for current grace period?
1887 * If no, return and let the other CPUs do their part as well.
1889 if (!rdp->qs_pending)
1893 * Was there a quiescent state since the beginning of the grace
1894 * period? If no, then exit and wait for the next call.
1896 if (!rdp->passed_quiesce)
1900 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1903 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1906 #ifdef CONFIG_HOTPLUG_CPU
1909 * Send the specified CPU's RCU callbacks to the orphanage. The
1910 * specified CPU must be offline, and the caller must hold the
1914 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1915 struct rcu_node *rnp, struct rcu_data *rdp)
1917 /* No-CBs CPUs do not have orphanable callbacks. */
1918 if (rcu_is_nocb_cpu(rdp->cpu))
1922 * Orphan the callbacks. First adjust the counts. This is safe
1923 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1924 * cannot be running now. Thus no memory barrier is required.
1926 if (rdp->nxtlist != NULL) {
1927 rsp->qlen_lazy += rdp->qlen_lazy;
1928 rsp->qlen += rdp->qlen;
1929 rdp->n_cbs_orphaned += rdp->qlen;
1931 ACCESS_ONCE(rdp->qlen) = 0;
1935 * Next, move those callbacks still needing a grace period to
1936 * the orphanage, where some other CPU will pick them up.
1937 * Some of the callbacks might have gone partway through a grace
1938 * period, but that is too bad. They get to start over because we
1939 * cannot assume that grace periods are synchronized across CPUs.
1940 * We don't bother updating the ->nxttail[] array yet, instead
1941 * we just reset the whole thing later on.
1943 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1944 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1945 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1946 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1950 * Then move the ready-to-invoke callbacks to the orphanage,
1951 * where some other CPU will pick them up. These will not be
1952 * required to pass though another grace period: They are done.
1954 if (rdp->nxtlist != NULL) {
1955 *rsp->orphan_donetail = rdp->nxtlist;
1956 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1959 /* Finally, initialize the rcu_data structure's list to empty. */
1960 init_callback_list(rdp);
1964 * Adopt the RCU callbacks from the specified rcu_state structure's
1965 * orphanage. The caller must hold the ->orphan_lock.
1967 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
1970 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1972 /* No-CBs CPUs are handled specially. */
1973 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
1976 /* Do the accounting first. */
1977 rdp->qlen_lazy += rsp->qlen_lazy;
1978 rdp->qlen += rsp->qlen;
1979 rdp->n_cbs_adopted += rsp->qlen;
1980 if (rsp->qlen_lazy != rsp->qlen)
1981 rcu_idle_count_callbacks_posted();
1986 * We do not need a memory barrier here because the only way we
1987 * can get here if there is an rcu_barrier() in flight is if
1988 * we are the task doing the rcu_barrier().
1991 /* First adopt the ready-to-invoke callbacks. */
1992 if (rsp->orphan_donelist != NULL) {
1993 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1994 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1995 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1996 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1997 rdp->nxttail[i] = rsp->orphan_donetail;
1998 rsp->orphan_donelist = NULL;
1999 rsp->orphan_donetail = &rsp->orphan_donelist;
2002 /* And then adopt the callbacks that still need a grace period. */
2003 if (rsp->orphan_nxtlist != NULL) {
2004 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2005 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2006 rsp->orphan_nxtlist = NULL;
2007 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2012 * Trace the fact that this CPU is going offline.
2014 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2016 RCU_TRACE(unsigned long mask);
2017 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2018 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2020 RCU_TRACE(mask = rdp->grpmask);
2021 trace_rcu_grace_period(rsp->name,
2022 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2027 * The CPU has been completely removed, and some other CPU is reporting
2028 * this fact from process context. Do the remainder of the cleanup,
2029 * including orphaning the outgoing CPU's RCU callbacks, and also
2030 * adopting them. There can only be one CPU hotplug operation at a time,
2031 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2033 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2035 unsigned long flags;
2037 int need_report = 0;
2038 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2039 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2041 /* Adjust any no-longer-needed kthreads. */
2042 rcu_boost_kthread_setaffinity(rnp, -1);
2044 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2046 /* Exclude any attempts to start a new grace period. */
2047 mutex_lock(&rsp->onoff_mutex);
2048 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2050 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2051 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2052 rcu_adopt_orphan_cbs(rsp, flags);
2054 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2055 mask = rdp->grpmask; /* rnp->grplo is constant. */
2057 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2058 smp_mb__after_unlock_lock();
2059 rnp->qsmaskinit &= ~mask;
2060 if (rnp->qsmaskinit != 0) {
2061 if (rnp != rdp->mynode)
2062 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2065 if (rnp == rdp->mynode)
2066 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2068 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2069 mask = rnp->grpmask;
2071 } while (rnp != NULL);
2074 * We still hold the leaf rcu_node structure lock here, and
2075 * irqs are still disabled. The reason for this subterfuge is
2076 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2077 * held leads to deadlock.
2079 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2081 if (need_report & RCU_OFL_TASKS_NORM_GP)
2082 rcu_report_unblock_qs_rnp(rnp, flags);
2084 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2085 if (need_report & RCU_OFL_TASKS_EXP_GP)
2086 rcu_report_exp_rnp(rsp, rnp, true);
2087 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2088 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2089 cpu, rdp->qlen, rdp->nxtlist);
2090 init_callback_list(rdp);
2091 /* Disallow further callbacks on this CPU. */
2092 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2093 mutex_unlock(&rsp->onoff_mutex);
2096 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2098 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2102 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2106 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2109 * Invoke any RCU callbacks that have made it to the end of their grace
2110 * period. Thottle as specified by rdp->blimit.
2112 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2114 unsigned long flags;
2115 struct rcu_head *next, *list, **tail;
2116 long bl, count, count_lazy;
2119 /* If no callbacks are ready, just return. */
2120 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2121 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2122 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2123 need_resched(), is_idle_task(current),
2124 rcu_is_callbacks_kthread());
2129 * Extract the list of ready callbacks, disabling to prevent
2130 * races with call_rcu() from interrupt handlers.
2132 local_irq_save(flags);
2133 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2135 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2136 list = rdp->nxtlist;
2137 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2138 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2139 tail = rdp->nxttail[RCU_DONE_TAIL];
2140 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2141 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2142 rdp->nxttail[i] = &rdp->nxtlist;
2143 local_irq_restore(flags);
2145 /* Invoke callbacks. */
2146 count = count_lazy = 0;
2150 debug_rcu_head_unqueue(list);
2151 if (__rcu_reclaim(rsp->name, list))
2154 /* Stop only if limit reached and CPU has something to do. */
2155 if (++count >= bl &&
2157 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2161 local_irq_save(flags);
2162 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2163 is_idle_task(current),
2164 rcu_is_callbacks_kthread());
2166 /* Update count, and requeue any remaining callbacks. */
2168 *tail = rdp->nxtlist;
2169 rdp->nxtlist = list;
2170 for (i = 0; i < RCU_NEXT_SIZE; i++)
2171 if (&rdp->nxtlist == rdp->nxttail[i])
2172 rdp->nxttail[i] = tail;
2176 smp_mb(); /* List handling before counting for rcu_barrier(). */
2177 rdp->qlen_lazy -= count_lazy;
2178 ACCESS_ONCE(rdp->qlen) -= count;
2179 rdp->n_cbs_invoked += count;
2181 /* Reinstate batch limit if we have worked down the excess. */
2182 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2183 rdp->blimit = blimit;
2185 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2186 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2187 rdp->qlen_last_fqs_check = 0;
2188 rdp->n_force_qs_snap = rsp->n_force_qs;
2189 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2190 rdp->qlen_last_fqs_check = rdp->qlen;
2191 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2193 local_irq_restore(flags);
2195 /* Re-invoke RCU core processing if there are callbacks remaining. */
2196 if (cpu_has_callbacks_ready_to_invoke(rdp))
2201 * Check to see if this CPU is in a non-context-switch quiescent state
2202 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2203 * Also schedule RCU core processing.
2205 * This function must be called from hardirq context. It is normally
2206 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2207 * false, there is no point in invoking rcu_check_callbacks().
2209 void rcu_check_callbacks(int cpu, int user)
2211 trace_rcu_utilization(TPS("Start scheduler-tick"));
2212 increment_cpu_stall_ticks();
2213 if (user || rcu_is_cpu_rrupt_from_idle()) {
2216 * Get here if this CPU took its interrupt from user
2217 * mode or from the idle loop, and if this is not a
2218 * nested interrupt. In this case, the CPU is in
2219 * a quiescent state, so note it.
2221 * No memory barrier is required here because both
2222 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2223 * variables that other CPUs neither access nor modify,
2224 * at least not while the corresponding CPU is online.
2230 } else if (!in_softirq()) {
2233 * Get here if this CPU did not take its interrupt from
2234 * softirq, in other words, if it is not interrupting
2235 * a rcu_bh read-side critical section. This is an _bh
2236 * critical section, so note it.
2241 rcu_preempt_check_callbacks(cpu);
2242 if (rcu_pending(cpu))
2244 trace_rcu_utilization(TPS("End scheduler-tick"));
2248 * Scan the leaf rcu_node structures, processing dyntick state for any that
2249 * have not yet encountered a quiescent state, using the function specified.
2250 * Also initiate boosting for any threads blocked on the root rcu_node.
2252 * The caller must have suppressed start of new grace periods.
2254 static void force_qs_rnp(struct rcu_state *rsp,
2255 int (*f)(struct rcu_data *rsp, bool *isidle,
2256 unsigned long *maxj),
2257 bool *isidle, unsigned long *maxj)
2261 unsigned long flags;
2263 struct rcu_node *rnp;
2265 rcu_for_each_leaf_node(rsp, rnp) {
2268 raw_spin_lock_irqsave(&rnp->lock, flags);
2269 smp_mb__after_unlock_lock();
2270 if (!rcu_gp_in_progress(rsp)) {
2271 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2274 if (rnp->qsmask == 0) {
2275 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2280 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2281 if ((rnp->qsmask & bit) != 0) {
2282 if ((rnp->qsmaskinit & bit) != 0)
2284 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2290 /* rcu_report_qs_rnp() releases rnp->lock. */
2291 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2294 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2296 rnp = rcu_get_root(rsp);
2297 if (rnp->qsmask == 0) {
2298 raw_spin_lock_irqsave(&rnp->lock, flags);
2299 smp_mb__after_unlock_lock();
2300 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2305 * Force quiescent states on reluctant CPUs, and also detect which
2306 * CPUs are in dyntick-idle mode.
2308 static void force_quiescent_state(struct rcu_state *rsp)
2310 unsigned long flags;
2312 struct rcu_node *rnp;
2313 struct rcu_node *rnp_old = NULL;
2315 /* Funnel through hierarchy to reduce memory contention. */
2316 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2317 for (; rnp != NULL; rnp = rnp->parent) {
2318 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2319 !raw_spin_trylock(&rnp->fqslock);
2320 if (rnp_old != NULL)
2321 raw_spin_unlock(&rnp_old->fqslock);
2323 rsp->n_force_qs_lh++;
2328 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2330 /* Reached the root of the rcu_node tree, acquire lock. */
2331 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2332 smp_mb__after_unlock_lock();
2333 raw_spin_unlock(&rnp_old->fqslock);
2334 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2335 rsp->n_force_qs_lh++;
2336 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2337 return; /* Someone beat us to it. */
2339 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2340 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2341 rcu_gp_kthread_wake(rsp);
2345 * This does the RCU core processing work for the specified rcu_state
2346 * and rcu_data structures. This may be called only from the CPU to
2347 * whom the rdp belongs.
2350 __rcu_process_callbacks(struct rcu_state *rsp)
2352 unsigned long flags;
2353 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2355 WARN_ON_ONCE(rdp->beenonline == 0);
2357 /* Update RCU state based on any recent quiescent states. */
2358 rcu_check_quiescent_state(rsp, rdp);
2360 /* Does this CPU require a not-yet-started grace period? */
2361 local_irq_save(flags);
2362 if (cpu_needs_another_gp(rsp, rdp)) {
2363 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2365 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2367 local_irq_restore(flags);
2370 /* If there are callbacks ready, invoke them. */
2371 if (cpu_has_callbacks_ready_to_invoke(rdp))
2372 invoke_rcu_callbacks(rsp, rdp);
2374 /* Do any needed deferred wakeups of rcuo kthreads. */
2375 do_nocb_deferred_wakeup(rdp);
2379 * Do RCU core processing for the current CPU.
2381 static void rcu_process_callbacks(struct softirq_action *unused)
2383 struct rcu_state *rsp;
2385 if (cpu_is_offline(smp_processor_id()))
2387 trace_rcu_utilization(TPS("Start RCU core"));
2388 for_each_rcu_flavor(rsp)
2389 __rcu_process_callbacks(rsp);
2390 trace_rcu_utilization(TPS("End RCU core"));
2394 * Schedule RCU callback invocation. If the specified type of RCU
2395 * does not support RCU priority boosting, just do a direct call,
2396 * otherwise wake up the per-CPU kernel kthread. Note that because we
2397 * are running on the current CPU with interrupts disabled, the
2398 * rcu_cpu_kthread_task cannot disappear out from under us.
2400 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2402 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2404 if (likely(!rsp->boost)) {
2405 rcu_do_batch(rsp, rdp);
2408 invoke_rcu_callbacks_kthread();
2411 static void invoke_rcu_core(void)
2413 if (cpu_online(smp_processor_id()))
2414 raise_softirq(RCU_SOFTIRQ);
2418 * Handle any core-RCU processing required by a call_rcu() invocation.
2420 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2421 struct rcu_head *head, unsigned long flags)
2424 * If called from an extended quiescent state, invoke the RCU
2425 * core in order to force a re-evaluation of RCU's idleness.
2427 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2430 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2431 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2435 * Force the grace period if too many callbacks or too long waiting.
2436 * Enforce hysteresis, and don't invoke force_quiescent_state()
2437 * if some other CPU has recently done so. Also, don't bother
2438 * invoking force_quiescent_state() if the newly enqueued callback
2439 * is the only one waiting for a grace period to complete.
2441 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2443 /* Are we ignoring a completed grace period? */
2444 note_gp_changes(rsp, rdp);
2446 /* Start a new grace period if one not already started. */
2447 if (!rcu_gp_in_progress(rsp)) {
2448 struct rcu_node *rnp_root = rcu_get_root(rsp);
2450 raw_spin_lock(&rnp_root->lock);
2451 smp_mb__after_unlock_lock();
2453 raw_spin_unlock(&rnp_root->lock);
2455 /* Give the grace period a kick. */
2456 rdp->blimit = LONG_MAX;
2457 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2458 *rdp->nxttail[RCU_DONE_TAIL] != head)
2459 force_quiescent_state(rsp);
2460 rdp->n_force_qs_snap = rsp->n_force_qs;
2461 rdp->qlen_last_fqs_check = rdp->qlen;
2467 * RCU callback function to leak a callback.
2469 static void rcu_leak_callback(struct rcu_head *rhp)
2474 * Helper function for call_rcu() and friends. The cpu argument will
2475 * normally be -1, indicating "currently running CPU". It may specify
2476 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2477 * is expected to specify a CPU.
2480 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2481 struct rcu_state *rsp, int cpu, bool lazy)
2483 unsigned long flags;
2484 struct rcu_data *rdp;
2486 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2487 if (debug_rcu_head_queue(head)) {
2488 /* Probable double call_rcu(), so leak the callback. */
2489 ACCESS_ONCE(head->func) = rcu_leak_callback;
2490 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2497 * Opportunistically note grace-period endings and beginnings.
2498 * Note that we might see a beginning right after we see an
2499 * end, but never vice versa, since this CPU has to pass through
2500 * a quiescent state betweentimes.
2502 local_irq_save(flags);
2503 rdp = this_cpu_ptr(rsp->rda);
2505 /* Add the callback to our list. */
2506 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2510 rdp = per_cpu_ptr(rsp->rda, cpu);
2511 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2512 WARN_ON_ONCE(offline);
2513 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2514 local_irq_restore(flags);
2517 ACCESS_ONCE(rdp->qlen)++;
2521 rcu_idle_count_callbacks_posted();
2522 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2523 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2524 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2526 if (__is_kfree_rcu_offset((unsigned long)func))
2527 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2528 rdp->qlen_lazy, rdp->qlen);
2530 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2532 /* Go handle any RCU core processing required. */
2533 __call_rcu_core(rsp, rdp, head, flags);
2534 local_irq_restore(flags);
2538 * Queue an RCU-sched callback for invocation after a grace period.
2540 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2542 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2544 EXPORT_SYMBOL_GPL(call_rcu_sched);
2547 * Queue an RCU callback for invocation after a quicker grace period.
2549 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2551 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2553 EXPORT_SYMBOL_GPL(call_rcu_bh);
2556 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2557 * any blocking grace-period wait automatically implies a grace period
2558 * if there is only one CPU online at any point time during execution
2559 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2560 * occasionally incorrectly indicate that there are multiple CPUs online
2561 * when there was in fact only one the whole time, as this just adds
2562 * some overhead: RCU still operates correctly.
2564 static inline int rcu_blocking_is_gp(void)
2568 might_sleep(); /* Check for RCU read-side critical section. */
2570 ret = num_online_cpus() <= 1;
2576 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2578 * Control will return to the caller some time after a full rcu-sched
2579 * grace period has elapsed, in other words after all currently executing
2580 * rcu-sched read-side critical sections have completed. These read-side
2581 * critical sections are delimited by rcu_read_lock_sched() and
2582 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2583 * local_irq_disable(), and so on may be used in place of
2584 * rcu_read_lock_sched().
2586 * This means that all preempt_disable code sequences, including NMI and
2587 * non-threaded hardware-interrupt handlers, in progress on entry will
2588 * have completed before this primitive returns. However, this does not
2589 * guarantee that softirq handlers will have completed, since in some
2590 * kernels, these handlers can run in process context, and can block.
2592 * Note that this guarantee implies further memory-ordering guarantees.
2593 * On systems with more than one CPU, when synchronize_sched() returns,
2594 * each CPU is guaranteed to have executed a full memory barrier since the
2595 * end of its last RCU-sched read-side critical section whose beginning
2596 * preceded the call to synchronize_sched(). In addition, each CPU having
2597 * an RCU read-side critical section that extends beyond the return from
2598 * synchronize_sched() is guaranteed to have executed a full memory barrier
2599 * after the beginning of synchronize_sched() and before the beginning of
2600 * that RCU read-side critical section. Note that these guarantees include
2601 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2602 * that are executing in the kernel.
2604 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2605 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2606 * to have executed a full memory barrier during the execution of
2607 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2608 * again only if the system has more than one CPU).
2610 * This primitive provides the guarantees made by the (now removed)
2611 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2612 * guarantees that rcu_read_lock() sections will have completed.
2613 * In "classic RCU", these two guarantees happen to be one and
2614 * the same, but can differ in realtime RCU implementations.
2616 void synchronize_sched(void)
2618 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2619 !lock_is_held(&rcu_lock_map) &&
2620 !lock_is_held(&rcu_sched_lock_map),
2621 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2622 if (rcu_blocking_is_gp())
2625 synchronize_sched_expedited();
2627 wait_rcu_gp(call_rcu_sched);
2629 EXPORT_SYMBOL_GPL(synchronize_sched);
2632 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2634 * Control will return to the caller some time after a full rcu_bh grace
2635 * period has elapsed, in other words after all currently executing rcu_bh
2636 * read-side critical sections have completed. RCU read-side critical
2637 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2638 * and may be nested.
2640 * See the description of synchronize_sched() for more detailed information
2641 * on memory ordering guarantees.
2643 void synchronize_rcu_bh(void)
2645 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2646 !lock_is_held(&rcu_lock_map) &&
2647 !lock_is_held(&rcu_sched_lock_map),
2648 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2649 if (rcu_blocking_is_gp())
2652 synchronize_rcu_bh_expedited();
2654 wait_rcu_gp(call_rcu_bh);
2656 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2658 static int synchronize_sched_expedited_cpu_stop(void *data)
2661 * There must be a full memory barrier on each affected CPU
2662 * between the time that try_stop_cpus() is called and the
2663 * time that it returns.
2665 * In the current initial implementation of cpu_stop, the
2666 * above condition is already met when the control reaches
2667 * this point and the following smp_mb() is not strictly
2668 * necessary. Do smp_mb() anyway for documentation and
2669 * robustness against future implementation changes.
2671 smp_mb(); /* See above comment block. */
2676 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2678 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2679 * approach to force the grace period to end quickly. This consumes
2680 * significant time on all CPUs and is unfriendly to real-time workloads,
2681 * so is thus not recommended for any sort of common-case code. In fact,
2682 * if you are using synchronize_sched_expedited() in a loop, please
2683 * restructure your code to batch your updates, and then use a single
2684 * synchronize_sched() instead.
2686 * Note that it is illegal to call this function while holding any lock
2687 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2688 * to call this function from a CPU-hotplug notifier. Failing to observe
2689 * these restriction will result in deadlock.
2691 * This implementation can be thought of as an application of ticket
2692 * locking to RCU, with sync_sched_expedited_started and
2693 * sync_sched_expedited_done taking on the roles of the halves
2694 * of the ticket-lock word. Each task atomically increments
2695 * sync_sched_expedited_started upon entry, snapshotting the old value,
2696 * then attempts to stop all the CPUs. If this succeeds, then each
2697 * CPU will have executed a context switch, resulting in an RCU-sched
2698 * grace period. We are then done, so we use atomic_cmpxchg() to
2699 * update sync_sched_expedited_done to match our snapshot -- but
2700 * only if someone else has not already advanced past our snapshot.
2702 * On the other hand, if try_stop_cpus() fails, we check the value
2703 * of sync_sched_expedited_done. If it has advanced past our
2704 * initial snapshot, then someone else must have forced a grace period
2705 * some time after we took our snapshot. In this case, our work is
2706 * done for us, and we can simply return. Otherwise, we try again,
2707 * but keep our initial snapshot for purposes of checking for someone
2708 * doing our work for us.
2710 * If we fail too many times in a row, we fall back to synchronize_sched().
2712 void synchronize_sched_expedited(void)
2714 long firstsnap, s, snap;
2716 struct rcu_state *rsp = &rcu_sched_state;
2719 * If we are in danger of counter wrap, just do synchronize_sched().
2720 * By allowing sync_sched_expedited_started to advance no more than
2721 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2722 * that more than 3.5 billion CPUs would be required to force a
2723 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2724 * course be required on a 64-bit system.
2726 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2727 (ulong)atomic_long_read(&rsp->expedited_done) +
2729 synchronize_sched();
2730 atomic_long_inc(&rsp->expedited_wrap);
2735 * Take a ticket. Note that atomic_inc_return() implies a
2736 * full memory barrier.
2738 snap = atomic_long_inc_return(&rsp->expedited_start);
2741 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2744 * Each pass through the following loop attempts to force a
2745 * context switch on each CPU.
2747 while (try_stop_cpus(cpu_online_mask,
2748 synchronize_sched_expedited_cpu_stop,
2751 atomic_long_inc(&rsp->expedited_tryfail);
2753 /* Check to see if someone else did our work for us. */
2754 s = atomic_long_read(&rsp->expedited_done);
2755 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2756 /* ensure test happens before caller kfree */
2757 smp_mb__before_atomic_inc(); /* ^^^ */
2758 atomic_long_inc(&rsp->expedited_workdone1);
2762 /* No joy, try again later. Or just synchronize_sched(). */
2763 if (trycount++ < 10) {
2764 udelay(trycount * num_online_cpus());
2766 wait_rcu_gp(call_rcu_sched);
2767 atomic_long_inc(&rsp->expedited_normal);
2771 /* Recheck to see if someone else did our work for us. */
2772 s = atomic_long_read(&rsp->expedited_done);
2773 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2774 /* ensure test happens before caller kfree */
2775 smp_mb__before_atomic_inc(); /* ^^^ */
2776 atomic_long_inc(&rsp->expedited_workdone2);
2781 * Refetching sync_sched_expedited_started allows later
2782 * callers to piggyback on our grace period. We retry
2783 * after they started, so our grace period works for them,
2784 * and they started after our first try, so their grace
2785 * period works for us.
2788 snap = atomic_long_read(&rsp->expedited_start);
2789 smp_mb(); /* ensure read is before try_stop_cpus(). */
2791 atomic_long_inc(&rsp->expedited_stoppedcpus);
2794 * Everyone up to our most recent fetch is covered by our grace
2795 * period. Update the counter, but only if our work is still
2796 * relevant -- which it won't be if someone who started later
2797 * than we did already did their update.
2800 atomic_long_inc(&rsp->expedited_done_tries);
2801 s = atomic_long_read(&rsp->expedited_done);
2802 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2803 /* ensure test happens before caller kfree */
2804 smp_mb__before_atomic_inc(); /* ^^^ */
2805 atomic_long_inc(&rsp->expedited_done_lost);
2808 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2809 atomic_long_inc(&rsp->expedited_done_exit);
2813 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2816 * Check to see if there is any immediate RCU-related work to be done
2817 * by the current CPU, for the specified type of RCU, returning 1 if so.
2818 * The checks are in order of increasing expense: checks that can be
2819 * carried out against CPU-local state are performed first. However,
2820 * we must check for CPU stalls first, else we might not get a chance.
2822 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2824 struct rcu_node *rnp = rdp->mynode;
2826 rdp->n_rcu_pending++;
2828 /* Check for CPU stalls, if enabled. */
2829 check_cpu_stall(rsp, rdp);
2831 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2832 if (rcu_nohz_full_cpu(rsp))
2835 /* Is the RCU core waiting for a quiescent state from this CPU? */
2836 if (rcu_scheduler_fully_active &&
2837 rdp->qs_pending && !rdp->passed_quiesce) {
2838 rdp->n_rp_qs_pending++;
2839 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2840 rdp->n_rp_report_qs++;
2844 /* Does this CPU have callbacks ready to invoke? */
2845 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2846 rdp->n_rp_cb_ready++;
2850 /* Has RCU gone idle with this CPU needing another grace period? */
2851 if (cpu_needs_another_gp(rsp, rdp)) {
2852 rdp->n_rp_cpu_needs_gp++;
2856 /* Has another RCU grace period completed? */
2857 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2858 rdp->n_rp_gp_completed++;
2862 /* Has a new RCU grace period started? */
2863 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2864 rdp->n_rp_gp_started++;
2868 /* Does this CPU need a deferred NOCB wakeup? */
2869 if (rcu_nocb_need_deferred_wakeup(rdp)) {
2870 rdp->n_rp_nocb_defer_wakeup++;
2875 rdp->n_rp_need_nothing++;
2880 * Check to see if there is any immediate RCU-related work to be done
2881 * by the current CPU, returning 1 if so. This function is part of the
2882 * RCU implementation; it is -not- an exported member of the RCU API.
2884 static int rcu_pending(int cpu)
2886 struct rcu_state *rsp;
2888 for_each_rcu_flavor(rsp)
2889 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2895 * Return true if the specified CPU has any callback. If all_lazy is
2896 * non-NULL, store an indication of whether all callbacks are lazy.
2897 * (If there are no callbacks, all of them are deemed to be lazy.)
2899 static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2903 struct rcu_data *rdp;
2904 struct rcu_state *rsp;
2906 for_each_rcu_flavor(rsp) {
2907 rdp = per_cpu_ptr(rsp->rda, cpu);
2911 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2922 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2923 * the compiler is expected to optimize this away.
2925 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2926 int cpu, unsigned long done)
2928 trace_rcu_barrier(rsp->name, s, cpu,
2929 atomic_read(&rsp->barrier_cpu_count), done);
2933 * RCU callback function for _rcu_barrier(). If we are last, wake
2934 * up the task executing _rcu_barrier().
2936 static void rcu_barrier_callback(struct rcu_head *rhp)
2938 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2939 struct rcu_state *rsp = rdp->rsp;
2941 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2942 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2943 complete(&rsp->barrier_completion);
2945 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2950 * Called with preemption disabled, and from cross-cpu IRQ context.
2952 static void rcu_barrier_func(void *type)
2954 struct rcu_state *rsp = type;
2955 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2957 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2958 atomic_inc(&rsp->barrier_cpu_count);
2959 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2963 * Orchestrate the specified type of RCU barrier, waiting for all
2964 * RCU callbacks of the specified type to complete.
2966 static void _rcu_barrier(struct rcu_state *rsp)
2969 struct rcu_data *rdp;
2970 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2971 unsigned long snap_done;
2973 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2975 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2976 mutex_lock(&rsp->barrier_mutex);
2979 * Ensure that all prior references, including to ->n_barrier_done,
2980 * are ordered before the _rcu_barrier() machinery.
2982 smp_mb(); /* See above block comment. */
2985 * Recheck ->n_barrier_done to see if others did our work for us.
2986 * This means checking ->n_barrier_done for an even-to-odd-to-even
2987 * transition. The "if" expression below therefore rounds the old
2988 * value up to the next even number and adds two before comparing.
2990 snap_done = rsp->n_barrier_done;
2991 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2994 * If the value in snap is odd, we needed to wait for the current
2995 * rcu_barrier() to complete, then wait for the next one, in other
2996 * words, we need the value of snap_done to be three larger than
2997 * the value of snap. On the other hand, if the value in snap is
2998 * even, we only had to wait for the next rcu_barrier() to complete,
2999 * in other words, we need the value of snap_done to be only two
3000 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3001 * this for us (thank you, Linus!).
3003 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3004 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3005 smp_mb(); /* caller's subsequent code after above check. */
3006 mutex_unlock(&rsp->barrier_mutex);
3011 * Increment ->n_barrier_done to avoid duplicate work. Use
3012 * ACCESS_ONCE() to prevent the compiler from speculating
3013 * the increment to precede the early-exit check.
3015 ACCESS_ONCE(rsp->n_barrier_done)++;
3016 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3017 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3018 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3021 * Initialize the count to one rather than to zero in order to
3022 * avoid a too-soon return to zero in case of a short grace period
3023 * (or preemption of this task). Exclude CPU-hotplug operations
3024 * to ensure that no offline CPU has callbacks queued.
3026 init_completion(&rsp->barrier_completion);
3027 atomic_set(&rsp->barrier_cpu_count, 1);
3031 * Force each CPU with callbacks to register a new callback.
3032 * When that callback is invoked, we will know that all of the
3033 * corresponding CPU's preceding callbacks have been invoked.
3035 for_each_possible_cpu(cpu) {
3036 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3038 rdp = per_cpu_ptr(rsp->rda, cpu);
3039 if (rcu_is_nocb_cpu(cpu)) {
3040 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3041 rsp->n_barrier_done);
3042 atomic_inc(&rsp->barrier_cpu_count);
3043 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3045 } else if (ACCESS_ONCE(rdp->qlen)) {
3046 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3047 rsp->n_barrier_done);
3048 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3050 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3051 rsp->n_barrier_done);
3057 * Now that we have an rcu_barrier_callback() callback on each
3058 * CPU, and thus each counted, remove the initial count.
3060 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3061 complete(&rsp->barrier_completion);
3063 /* Increment ->n_barrier_done to prevent duplicate work. */
3064 smp_mb(); /* Keep increment after above mechanism. */
3065 ACCESS_ONCE(rsp->n_barrier_done)++;
3066 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3067 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3068 smp_mb(); /* Keep increment before caller's subsequent code. */
3070 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3071 wait_for_completion(&rsp->barrier_completion);
3073 /* Other rcu_barrier() invocations can now safely proceed. */
3074 mutex_unlock(&rsp->barrier_mutex);
3078 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3080 void rcu_barrier_bh(void)
3082 _rcu_barrier(&rcu_bh_state);
3084 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3087 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3089 void rcu_barrier_sched(void)
3091 _rcu_barrier(&rcu_sched_state);
3093 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3096 * Do boot-time initialization of a CPU's per-CPU RCU data.
3099 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3101 unsigned long flags;
3102 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3103 struct rcu_node *rnp = rcu_get_root(rsp);
3105 /* Set up local state, ensuring consistent view of global state. */
3106 raw_spin_lock_irqsave(&rnp->lock, flags);
3107 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3108 init_callback_list(rdp);
3110 ACCESS_ONCE(rdp->qlen) = 0;
3111 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3112 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3113 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3116 rcu_boot_init_nocb_percpu_data(rdp);
3117 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3121 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3122 * offline event can be happening at a given time. Note also that we
3123 * can accept some slop in the rsp->completed access due to the fact
3124 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3127 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3129 unsigned long flags;
3131 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3132 struct rcu_node *rnp = rcu_get_root(rsp);
3134 /* Exclude new grace periods. */
3135 mutex_lock(&rsp->onoff_mutex);
3137 /* Set up local state, ensuring consistent view of global state. */
3138 raw_spin_lock_irqsave(&rnp->lock, flags);
3139 rdp->beenonline = 1; /* We have now been online. */
3140 rdp->preemptible = preemptible;
3141 rdp->qlen_last_fqs_check = 0;
3142 rdp->n_force_qs_snap = rsp->n_force_qs;
3143 rdp->blimit = blimit;
3144 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3145 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3146 rcu_sysidle_init_percpu_data(rdp->dynticks);
3147 atomic_set(&rdp->dynticks->dynticks,
3148 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3149 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3151 /* Add CPU to rcu_node bitmasks. */
3153 mask = rdp->grpmask;
3155 /* Exclude any attempts to start a new GP on small systems. */
3156 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3157 rnp->qsmaskinit |= mask;
3158 mask = rnp->grpmask;
3159 if (rnp == rdp->mynode) {
3161 * If there is a grace period in progress, we will
3162 * set up to wait for it next time we run the
3165 rdp->gpnum = rnp->completed;
3166 rdp->completed = rnp->completed;
3167 rdp->passed_quiesce = 0;
3168 rdp->qs_pending = 0;
3169 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3171 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3173 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3174 local_irq_restore(flags);
3176 mutex_unlock(&rsp->onoff_mutex);
3179 static void rcu_prepare_cpu(int cpu)
3181 struct rcu_state *rsp;
3183 for_each_rcu_flavor(rsp)
3184 rcu_init_percpu_data(cpu, rsp,
3185 strcmp(rsp->name, "rcu_preempt") == 0);
3189 * Handle CPU online/offline notification events.
3191 static int rcu_cpu_notify(struct notifier_block *self,
3192 unsigned long action, void *hcpu)
3194 long cpu = (long)hcpu;
3195 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3196 struct rcu_node *rnp = rdp->mynode;
3197 struct rcu_state *rsp;
3199 trace_rcu_utilization(TPS("Start CPU hotplug"));
3201 case CPU_UP_PREPARE:
3202 case CPU_UP_PREPARE_FROZEN:
3203 rcu_prepare_cpu(cpu);
3204 rcu_prepare_kthreads(cpu);
3207 case CPU_DOWN_FAILED:
3208 rcu_boost_kthread_setaffinity(rnp, -1);
3210 case CPU_DOWN_PREPARE:
3211 rcu_boost_kthread_setaffinity(rnp, cpu);
3214 case CPU_DYING_FROZEN:
3215 for_each_rcu_flavor(rsp)
3216 rcu_cleanup_dying_cpu(rsp);
3219 case CPU_DEAD_FROZEN:
3220 case CPU_UP_CANCELED:
3221 case CPU_UP_CANCELED_FROZEN:
3222 for_each_rcu_flavor(rsp)
3223 rcu_cleanup_dead_cpu(cpu, rsp);
3228 trace_rcu_utilization(TPS("End CPU hotplug"));
3232 static int rcu_pm_notify(struct notifier_block *self,
3233 unsigned long action, void *hcpu)
3236 case PM_HIBERNATION_PREPARE:
3237 case PM_SUSPEND_PREPARE:
3238 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3241 case PM_POST_HIBERNATION:
3242 case PM_POST_SUSPEND:
3252 * Spawn the kthread that handles this RCU flavor's grace periods.
3254 static int __init rcu_spawn_gp_kthread(void)
3256 unsigned long flags;
3257 struct rcu_node *rnp;
3258 struct rcu_state *rsp;
3259 struct task_struct *t;
3261 for_each_rcu_flavor(rsp) {
3262 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3264 rnp = rcu_get_root(rsp);
3265 raw_spin_lock_irqsave(&rnp->lock, flags);
3266 rsp->gp_kthread = t;
3267 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3268 rcu_spawn_nocb_kthreads(rsp);
3272 early_initcall(rcu_spawn_gp_kthread);
3275 * This function is invoked towards the end of the scheduler's initialization
3276 * process. Before this is called, the idle task might contain
3277 * RCU read-side critical sections (during which time, this idle
3278 * task is booting the system). After this function is called, the
3279 * idle tasks are prohibited from containing RCU read-side critical
3280 * sections. This function also enables RCU lockdep checking.
3282 void rcu_scheduler_starting(void)
3284 WARN_ON(num_online_cpus() != 1);
3285 WARN_ON(nr_context_switches() > 0);
3286 rcu_scheduler_active = 1;
3290 * Compute the per-level fanout, either using the exact fanout specified
3291 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3293 #ifdef CONFIG_RCU_FANOUT_EXACT
3294 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3298 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3299 for (i = rcu_num_lvls - 2; i >= 0; i--)
3300 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3302 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3303 static void __init rcu_init_levelspread(struct rcu_state *rsp)
3310 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3311 ccur = rsp->levelcnt[i];
3312 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3316 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3319 * Helper function for rcu_init() that initializes one rcu_state structure.
3321 static void __init rcu_init_one(struct rcu_state *rsp,
3322 struct rcu_data __percpu *rda)
3324 static char *buf[] = { "rcu_node_0",
3327 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3328 static char *fqs[] = { "rcu_node_fqs_0",
3331 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3335 struct rcu_node *rnp;
3337 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3339 /* Silence gcc 4.8 warning about array index out of range. */
3340 if (rcu_num_lvls > RCU_NUM_LVLS)
3341 panic("rcu_init_one: rcu_num_lvls overflow");
3343 /* Initialize the level-tracking arrays. */
3345 for (i = 0; i < rcu_num_lvls; i++)
3346 rsp->levelcnt[i] = num_rcu_lvl[i];
3347 for (i = 1; i < rcu_num_lvls; i++)
3348 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3349 rcu_init_levelspread(rsp);
3351 /* Initialize the elements themselves, starting from the leaves. */
3353 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3354 cpustride *= rsp->levelspread[i];
3355 rnp = rsp->level[i];
3356 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3357 raw_spin_lock_init(&rnp->lock);
3358 lockdep_set_class_and_name(&rnp->lock,
3359 &rcu_node_class[i], buf[i]);
3360 raw_spin_lock_init(&rnp->fqslock);
3361 lockdep_set_class_and_name(&rnp->fqslock,
3362 &rcu_fqs_class[i], fqs[i]);
3363 rnp->gpnum = rsp->gpnum;
3364 rnp->completed = rsp->completed;
3366 rnp->qsmaskinit = 0;
3367 rnp->grplo = j * cpustride;
3368 rnp->grphi = (j + 1) * cpustride - 1;
3369 if (rnp->grphi >= NR_CPUS)
3370 rnp->grphi = NR_CPUS - 1;
3376 rnp->grpnum = j % rsp->levelspread[i - 1];
3377 rnp->grpmask = 1UL << rnp->grpnum;
3378 rnp->parent = rsp->level[i - 1] +
3379 j / rsp->levelspread[i - 1];
3382 INIT_LIST_HEAD(&rnp->blkd_tasks);
3383 rcu_init_one_nocb(rnp);
3388 init_waitqueue_head(&rsp->gp_wq);
3389 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3390 rnp = rsp->level[rcu_num_lvls - 1];
3391 for_each_possible_cpu(i) {
3392 while (i > rnp->grphi)
3394 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3395 rcu_boot_init_percpu_data(i, rsp);
3397 list_add(&rsp->flavors, &rcu_struct_flavors);
3401 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3402 * replace the definitions in tree.h because those are needed to size
3403 * the ->node array in the rcu_state structure.
3405 static void __init rcu_init_geometry(void)
3411 int rcu_capacity[MAX_RCU_LVLS + 1];
3414 * Initialize any unspecified boot parameters.
3415 * The default values of jiffies_till_first_fqs and
3416 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3417 * value, which is a function of HZ, then adding one for each
3418 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3420 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3421 if (jiffies_till_first_fqs == ULONG_MAX)
3422 jiffies_till_first_fqs = d;
3423 if (jiffies_till_next_fqs == ULONG_MAX)
3424 jiffies_till_next_fqs = d;
3426 /* If the compile-time values are accurate, just leave. */
3427 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3428 nr_cpu_ids == NR_CPUS)
3430 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3431 rcu_fanout_leaf, nr_cpu_ids);
3434 * Compute number of nodes that can be handled an rcu_node tree
3435 * with the given number of levels. Setting rcu_capacity[0] makes
3436 * some of the arithmetic easier.
3438 rcu_capacity[0] = 1;
3439 rcu_capacity[1] = rcu_fanout_leaf;
3440 for (i = 2; i <= MAX_RCU_LVLS; i++)
3441 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3444 * The boot-time rcu_fanout_leaf parameter is only permitted
3445 * to increase the leaf-level fanout, not decrease it. Of course,
3446 * the leaf-level fanout cannot exceed the number of bits in
3447 * the rcu_node masks. Finally, the tree must be able to accommodate
3448 * the configured number of CPUs. Complain and fall back to the
3449 * compile-time values if these limits are exceeded.
3451 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3452 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3453 n > rcu_capacity[MAX_RCU_LVLS]) {
3458 /* Calculate the number of rcu_nodes at each level of the tree. */
3459 for (i = 1; i <= MAX_RCU_LVLS; i++)
3460 if (n <= rcu_capacity[i]) {
3461 for (j = 0; j <= i; j++)
3463 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3465 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3470 /* Calculate the total number of rcu_node structures. */
3472 for (i = 0; i <= MAX_RCU_LVLS; i++)
3473 rcu_num_nodes += num_rcu_lvl[i];
3477 void __init rcu_init(void)
3481 rcu_bootup_announce();
3482 rcu_init_geometry();
3483 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3484 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3485 __rcu_init_preempt();
3486 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3489 * We don't need protection against CPU-hotplug here because
3490 * this is called early in boot, before either interrupts
3491 * or the scheduler are operational.
3493 cpu_notifier(rcu_cpu_notify, 0);
3494 pm_notifier(rcu_pm_notify, 0);
3495 for_each_online_cpu(cpu)
3496 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3499 #include "tree_plugin.h"