2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
72 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
82 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83 static char sname##_varname[] = #sname; \
84 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85 struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
99 DEFINE_PER_CPU(struct rcu_data, sname##_data)
101 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
102 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
104 static struct rcu_state
*rcu_state_p
;
105 LIST_HEAD(rcu_struct_flavors
);
107 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
109 module_param(rcu_fanout_leaf
, int, 0444);
110 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
111 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
118 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
129 int rcu_scheduler_active __read_mostly
;
130 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
144 static int rcu_scheduler_fully_active __read_mostly
;
146 #ifdef CONFIG_RCU_BOOST
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
152 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
153 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
154 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
155 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
157 #endif /* #ifdef CONFIG_RCU_BOOST */
159 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
172 unsigned long rcutorture_testseq
;
173 unsigned long rcutorture_vernum
;
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
180 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
182 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
191 void rcu_sched_qs(int cpu
)
193 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
195 if (rdp
->passed_quiesce
== 0)
196 trace_rcu_grace_period(TPS("rcu_sched"), rdp
->gpnum
, TPS("cpuqs"));
197 rdp
->passed_quiesce
= 1;
200 void rcu_bh_qs(int cpu
)
202 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
204 if (rdp
->passed_quiesce
== 0)
205 trace_rcu_grace_period(TPS("rcu_bh"), rdp
->gpnum
, TPS("cpuqs"));
206 rdp
->passed_quiesce
= 1;
209 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
211 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
212 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
213 .dynticks
= ATOMIC_INIT(1),
214 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
215 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
216 .dynticks_idle
= ATOMIC_INIT(1),
217 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
221 * Let the RCU core know that this CPU has gone through the scheduler,
222 * which is a quiescent state. This is called when the need for a
223 * quiescent state is urgent, so we burn an atomic operation and full
224 * memory barriers to let the RCU core know about it, regardless of what
225 * this CPU might (or might not) do in the near future.
227 * We inform the RCU core by emulating a zero-duration dyntick-idle
228 * period, which we in turn do by incrementing the ->dynticks counter
231 static void rcu_momentary_dyntick_idle(void)
234 struct rcu_data
*rdp
;
235 struct rcu_dynticks
*rdtp
;
237 struct rcu_state
*rsp
;
239 local_irq_save(flags
);
242 * Yes, we can lose flag-setting operations. This is OK, because
243 * the flag will be set again after some delay.
245 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
246 raw_cpu_write(rcu_sched_qs_mask
, 0);
248 /* Find the flavor that needs a quiescent state. */
249 for_each_rcu_flavor(rsp
) {
250 rdp
= raw_cpu_ptr(rsp
->rda
);
251 if (!(resched_mask
& rsp
->flavor_mask
))
253 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
254 if (ACCESS_ONCE(rdp
->mynode
->completed
) !=
255 ACCESS_ONCE(rdp
->cond_resched_completed
))
259 * Pretend to be momentarily idle for the quiescent state.
260 * This allows the grace-period kthread to record the
261 * quiescent state, with no need for this CPU to do anything
264 rdtp
= this_cpu_ptr(&rcu_dynticks
);
265 smp_mb__before_atomic(); /* Earlier stuff before QS. */
266 atomic_add(2, &rdtp
->dynticks
); /* QS. */
267 smp_mb__after_atomic(); /* Later stuff after QS. */
270 local_irq_restore(flags
);
274 * Note a context switch. This is a quiescent state for RCU-sched,
275 * and requires special handling for preemptible RCU.
276 * The caller must have disabled preemption.
278 void rcu_note_context_switch(int cpu
)
280 trace_rcu_utilization(TPS("Start context switch"));
282 rcu_preempt_note_context_switch(cpu
);
283 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
284 rcu_momentary_dyntick_idle();
285 trace_rcu_utilization(TPS("End context switch"));
287 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
289 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
290 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
291 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
293 module_param(blimit
, long, 0444);
294 module_param(qhimark
, long, 0444);
295 module_param(qlowmark
, long, 0444);
297 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
298 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
300 module_param(jiffies_till_first_fqs
, ulong
, 0644);
301 module_param(jiffies_till_next_fqs
, ulong
, 0644);
304 * How long the grace period must be before we start recruiting
305 * quiescent-state help from rcu_note_context_switch().
307 static ulong jiffies_till_sched_qs
= HZ
/ 20;
308 module_param(jiffies_till_sched_qs
, ulong
, 0644);
310 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
311 struct rcu_data
*rdp
);
312 static void force_qs_rnp(struct rcu_state
*rsp
,
313 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
314 unsigned long *maxj
),
315 bool *isidle
, unsigned long *maxj
);
316 static void force_quiescent_state(struct rcu_state
*rsp
);
317 static int rcu_pending(int cpu
);
320 * Return the number of RCU-sched batches processed thus far for debug & stats.
322 long rcu_batches_completed_sched(void)
324 return rcu_sched_state
.completed
;
326 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
329 * Return the number of RCU BH batches processed thus far for debug & stats.
331 long rcu_batches_completed_bh(void)
333 return rcu_bh_state
.completed
;
335 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
338 * Force a quiescent state.
340 void rcu_force_quiescent_state(void)
342 force_quiescent_state(rcu_state_p
);
344 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
347 * Force a quiescent state for RCU BH.
349 void rcu_bh_force_quiescent_state(void)
351 force_quiescent_state(&rcu_bh_state
);
353 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
356 * Show the state of the grace-period kthreads.
358 void show_rcu_gp_kthreads(void)
360 struct rcu_state
*rsp
;
362 for_each_rcu_flavor(rsp
) {
363 pr_info("%s: wait state: %d ->state: %#lx\n",
364 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
365 /* sched_show_task(rsp->gp_kthread); */
368 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
371 * Record the number of times rcutorture tests have been initiated and
372 * terminated. This information allows the debugfs tracing stats to be
373 * correlated to the rcutorture messages, even when the rcutorture module
374 * is being repeatedly loaded and unloaded. In other words, we cannot
375 * store this state in rcutorture itself.
377 void rcutorture_record_test_transition(void)
379 rcutorture_testseq
++;
380 rcutorture_vernum
= 0;
382 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
385 * Send along grace-period-related data for rcutorture diagnostics.
387 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
388 unsigned long *gpnum
, unsigned long *completed
)
390 struct rcu_state
*rsp
= NULL
;
399 case RCU_SCHED_FLAVOR
:
400 rsp
= &rcu_sched_state
;
406 *flags
= ACCESS_ONCE(rsp
->gp_flags
);
407 *gpnum
= ACCESS_ONCE(rsp
->gpnum
);
408 *completed
= ACCESS_ONCE(rsp
->completed
);
415 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
418 * Record the number of writer passes through the current rcutorture test.
419 * This is also used to correlate debugfs tracing stats with the rcutorture
422 void rcutorture_record_progress(unsigned long vernum
)
426 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
429 * Force a quiescent state for RCU-sched.
431 void rcu_sched_force_quiescent_state(void)
433 force_quiescent_state(&rcu_sched_state
);
435 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
438 * Does the CPU have callbacks ready to be invoked?
441 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
443 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
444 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
448 * Return the root node of the specified rcu_state structure.
450 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
452 return &rsp
->node
[0];
456 * Is there any need for future grace periods?
457 * Interrupts must be disabled. If the caller does not hold the root
458 * rnp_node structure's ->lock, the results are advisory only.
460 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
462 struct rcu_node
*rnp
= rcu_get_root(rsp
);
463 int idx
= (ACCESS_ONCE(rnp
->completed
) + 1) & 0x1;
464 int *fp
= &rnp
->need_future_gp
[idx
];
466 return ACCESS_ONCE(*fp
);
470 * Does the current CPU require a not-yet-started grace period?
471 * The caller must have disabled interrupts to prevent races with
472 * normal callback registry.
475 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
479 if (rcu_gp_in_progress(rsp
))
480 return 0; /* No, a grace period is already in progress. */
481 if (rcu_future_needs_gp(rsp
))
482 return 1; /* Yes, a no-CBs CPU needs one. */
483 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
484 return 0; /* No, this is a no-CBs (or offline) CPU. */
485 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
486 return 1; /* Yes, this CPU has newly registered callbacks. */
487 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
488 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
489 ULONG_CMP_LT(ACCESS_ONCE(rsp
->completed
),
490 rdp
->nxtcompleted
[i
]))
491 return 1; /* Yes, CBs for future grace period. */
492 return 0; /* No grace period needed. */
496 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
498 * If the new value of the ->dynticks_nesting counter now is zero,
499 * we really have entered idle, and must do the appropriate accounting.
500 * The caller must have disabled interrupts.
502 static void rcu_eqs_enter_common(struct rcu_dynticks
*rdtp
, long long oldval
,
505 struct rcu_state
*rsp
;
506 struct rcu_data
*rdp
;
508 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
509 if (!user
&& !is_idle_task(current
)) {
510 struct task_struct
*idle __maybe_unused
=
511 idle_task(smp_processor_id());
513 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
514 ftrace_dump(DUMP_ORIG
);
515 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
516 current
->pid
, current
->comm
,
517 idle
->pid
, idle
->comm
); /* must be idle task! */
519 for_each_rcu_flavor(rsp
) {
520 rdp
= this_cpu_ptr(rsp
->rda
);
521 do_nocb_deferred_wakeup(rdp
);
523 rcu_prepare_for_idle(smp_processor_id());
524 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
525 smp_mb__before_atomic(); /* See above. */
526 atomic_inc(&rdtp
->dynticks
);
527 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
528 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
531 * It is illegal to enter an extended quiescent state while
532 * in an RCU read-side critical section.
534 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
535 "Illegal idle entry in RCU read-side critical section.");
536 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
537 "Illegal idle entry in RCU-bh read-side critical section.");
538 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
539 "Illegal idle entry in RCU-sched read-side critical section.");
543 * Enter an RCU extended quiescent state, which can be either the
544 * idle loop or adaptive-tickless usermode execution.
546 static void rcu_eqs_enter(bool user
)
549 struct rcu_dynticks
*rdtp
;
551 rdtp
= this_cpu_ptr(&rcu_dynticks
);
552 oldval
= rdtp
->dynticks_nesting
;
553 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
554 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
555 rdtp
->dynticks_nesting
= 0;
556 rcu_eqs_enter_common(rdtp
, oldval
, user
);
558 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
563 * rcu_idle_enter - inform RCU that current CPU is entering idle
565 * Enter idle mode, in other words, -leave- the mode in which RCU
566 * read-side critical sections can occur. (Though RCU read-side
567 * critical sections can occur in irq handlers in idle, a possibility
568 * handled by irq_enter() and irq_exit().)
570 * We crowbar the ->dynticks_nesting field to zero to allow for
571 * the possibility of usermode upcalls having messed up our count
572 * of interrupt nesting level during the prior busy period.
574 void rcu_idle_enter(void)
578 local_irq_save(flags
);
579 rcu_eqs_enter(false);
580 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks
), 0);
581 local_irq_restore(flags
);
583 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
585 #ifdef CONFIG_RCU_USER_QS
587 * rcu_user_enter - inform RCU that we are resuming userspace.
589 * Enter RCU idle mode right before resuming userspace. No use of RCU
590 * is permitted between this call and rcu_user_exit(). This way the
591 * CPU doesn't need to maintain the tick for RCU maintenance purposes
592 * when the CPU runs in userspace.
594 void rcu_user_enter(void)
598 #endif /* CONFIG_RCU_USER_QS */
601 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
603 * Exit from an interrupt handler, which might possibly result in entering
604 * idle mode, in other words, leaving the mode in which read-side critical
605 * sections can occur.
607 * This code assumes that the idle loop never does anything that might
608 * result in unbalanced calls to irq_enter() and irq_exit(). If your
609 * architecture violates this assumption, RCU will give you what you
610 * deserve, good and hard. But very infrequently and irreproducibly.
612 * Use things like work queues to work around this limitation.
614 * You have been warned.
616 void rcu_irq_exit(void)
620 struct rcu_dynticks
*rdtp
;
622 local_irq_save(flags
);
623 rdtp
= this_cpu_ptr(&rcu_dynticks
);
624 oldval
= rdtp
->dynticks_nesting
;
625 rdtp
->dynticks_nesting
--;
626 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
627 if (rdtp
->dynticks_nesting
)
628 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
630 rcu_eqs_enter_common(rdtp
, oldval
, true);
631 rcu_sysidle_enter(rdtp
, 1);
632 local_irq_restore(flags
);
636 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
638 * If the new value of the ->dynticks_nesting counter was previously zero,
639 * we really have exited idle, and must do the appropriate accounting.
640 * The caller must have disabled interrupts.
642 static void rcu_eqs_exit_common(struct rcu_dynticks
*rdtp
, long long oldval
,
645 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
646 atomic_inc(&rdtp
->dynticks
);
647 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
648 smp_mb__after_atomic(); /* See above. */
649 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
650 rcu_cleanup_after_idle(smp_processor_id());
651 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
652 if (!user
&& !is_idle_task(current
)) {
653 struct task_struct
*idle __maybe_unused
=
654 idle_task(smp_processor_id());
656 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
657 oldval
, rdtp
->dynticks_nesting
);
658 ftrace_dump(DUMP_ORIG
);
659 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
660 current
->pid
, current
->comm
,
661 idle
->pid
, idle
->comm
); /* must be idle task! */
666 * Exit an RCU extended quiescent state, which can be either the
667 * idle loop or adaptive-tickless usermode execution.
669 static void rcu_eqs_exit(bool user
)
671 struct rcu_dynticks
*rdtp
;
674 rdtp
= this_cpu_ptr(&rcu_dynticks
);
675 oldval
= rdtp
->dynticks_nesting
;
676 WARN_ON_ONCE(oldval
< 0);
677 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
678 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
680 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
681 rcu_eqs_exit_common(rdtp
, oldval
, user
);
686 * rcu_idle_exit - inform RCU that current CPU is leaving idle
688 * Exit idle mode, in other words, -enter- the mode in which RCU
689 * read-side critical sections can occur.
691 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
692 * allow for the possibility of usermode upcalls messing up our count
693 * of interrupt nesting level during the busy period that is just
696 void rcu_idle_exit(void)
700 local_irq_save(flags
);
702 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks
), 0);
703 local_irq_restore(flags
);
705 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
707 #ifdef CONFIG_RCU_USER_QS
709 * rcu_user_exit - inform RCU that we are exiting userspace.
711 * Exit RCU idle mode while entering the kernel because it can
712 * run a RCU read side critical section anytime.
714 void rcu_user_exit(void)
718 #endif /* CONFIG_RCU_USER_QS */
721 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
723 * Enter an interrupt handler, which might possibly result in exiting
724 * idle mode, in other words, entering the mode in which read-side critical
725 * sections can occur.
727 * Note that the Linux kernel is fully capable of entering an interrupt
728 * handler that it never exits, for example when doing upcalls to
729 * user mode! This code assumes that the idle loop never does upcalls to
730 * user mode. If your architecture does do upcalls from the idle loop (or
731 * does anything else that results in unbalanced calls to the irq_enter()
732 * and irq_exit() functions), RCU will give you what you deserve, good
733 * and hard. But very infrequently and irreproducibly.
735 * Use things like work queues to work around this limitation.
737 * You have been warned.
739 void rcu_irq_enter(void)
742 struct rcu_dynticks
*rdtp
;
745 local_irq_save(flags
);
746 rdtp
= this_cpu_ptr(&rcu_dynticks
);
747 oldval
= rdtp
->dynticks_nesting
;
748 rdtp
->dynticks_nesting
++;
749 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
751 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
753 rcu_eqs_exit_common(rdtp
, oldval
, true);
754 rcu_sysidle_exit(rdtp
, 1);
755 local_irq_restore(flags
);
759 * rcu_nmi_enter - inform RCU of entry to NMI context
761 * If the CPU was idle with dynamic ticks active, and there is no
762 * irq handler running, this updates rdtp->dynticks_nmi to let the
763 * RCU grace-period handling know that the CPU is active.
765 void rcu_nmi_enter(void)
767 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
769 if (rdtp
->dynticks_nmi_nesting
== 0 &&
770 (atomic_read(&rdtp
->dynticks
) & 0x1))
772 rdtp
->dynticks_nmi_nesting
++;
773 smp_mb__before_atomic(); /* Force delay from prior write. */
774 atomic_inc(&rdtp
->dynticks
);
775 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
776 smp_mb__after_atomic(); /* See above. */
777 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
781 * rcu_nmi_exit - inform RCU of exit from NMI context
783 * If the CPU was idle with dynamic ticks active, and there is no
784 * irq handler running, this updates rdtp->dynticks_nmi to let the
785 * RCU grace-period handling know that the CPU is no longer active.
787 void rcu_nmi_exit(void)
789 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
791 if (rdtp
->dynticks_nmi_nesting
== 0 ||
792 --rdtp
->dynticks_nmi_nesting
!= 0)
794 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
795 smp_mb__before_atomic(); /* See above. */
796 atomic_inc(&rdtp
->dynticks
);
797 smp_mb__after_atomic(); /* Force delay to next write. */
798 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
802 * __rcu_is_watching - are RCU read-side critical sections safe?
804 * Return true if RCU is watching the running CPU, which means that
805 * this CPU can safely enter RCU read-side critical sections. Unlike
806 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
807 * least disabled preemption.
809 bool notrace
__rcu_is_watching(void)
811 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
815 * rcu_is_watching - see if RCU thinks that the current CPU is idle
817 * If the current CPU is in its idle loop and is neither in an interrupt
818 * or NMI handler, return true.
820 bool notrace
rcu_is_watching(void)
825 ret
= __rcu_is_watching();
829 EXPORT_SYMBOL_GPL(rcu_is_watching
);
831 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
834 * Is the current CPU online? Disable preemption to avoid false positives
835 * that could otherwise happen due to the current CPU number being sampled,
836 * this task being preempted, its old CPU being taken offline, resuming
837 * on some other CPU, then determining that its old CPU is now offline.
838 * It is OK to use RCU on an offline processor during initial boot, hence
839 * the check for rcu_scheduler_fully_active. Note also that it is OK
840 * for a CPU coming online to use RCU for one jiffy prior to marking itself
841 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
842 * offline to continue to use RCU for one jiffy after marking itself
843 * offline in the cpu_online_mask. This leniency is necessary given the
844 * non-atomic nature of the online and offline processing, for example,
845 * the fact that a CPU enters the scheduler after completing the CPU_DYING
848 * This is also why RCU internally marks CPUs online during the
849 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
851 * Disable checking if in an NMI handler because we cannot safely report
852 * errors from NMI handlers anyway.
854 bool rcu_lockdep_current_cpu_online(void)
856 struct rcu_data
*rdp
;
857 struct rcu_node
*rnp
;
863 rdp
= this_cpu_ptr(&rcu_sched_data
);
865 ret
= (rdp
->grpmask
& rnp
->qsmaskinit
) ||
866 !rcu_scheduler_fully_active
;
870 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
872 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
875 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
877 * If the current CPU is idle or running at a first-level (not nested)
878 * interrupt from idle, return true. The caller must have at least
879 * disabled preemption.
881 static int rcu_is_cpu_rrupt_from_idle(void)
883 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
887 * Snapshot the specified CPU's dynticks counter so that we can later
888 * credit them with an implicit quiescent state. Return 1 if this CPU
889 * is in dynticks idle mode, which is an extended quiescent state.
891 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
892 bool *isidle
, unsigned long *maxj
)
894 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
895 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
896 if ((rdp
->dynticks_snap
& 0x1) == 0) {
897 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
905 * This function really isn't for public consumption, but RCU is special in
906 * that context switches can allow the state machine to make progress.
908 extern void resched_cpu(int cpu
);
911 * Return true if the specified CPU has passed through a quiescent
912 * state by virtue of being in or having passed through an dynticks
913 * idle state since the last call to dyntick_save_progress_counter()
914 * for this same CPU, or by virtue of having been offline.
916 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
917 bool *isidle
, unsigned long *maxj
)
923 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
924 snap
= (unsigned int)rdp
->dynticks_snap
;
927 * If the CPU passed through or entered a dynticks idle phase with
928 * no active irq/NMI handlers, then we can safely pretend that the CPU
929 * already acknowledged the request to pass through a quiescent
930 * state. Either way, that CPU cannot possibly be in an RCU
931 * read-side critical section that started before the beginning
932 * of the current RCU grace period.
934 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
935 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
941 * Check for the CPU being offline, but only if the grace period
942 * is old enough. We don't need to worry about the CPU changing
943 * state: If we see it offline even once, it has been through a
946 * The reason for insisting that the grace period be at least
947 * one jiffy old is that CPUs that are not quite online and that
948 * have just gone offline can still execute RCU read-side critical
951 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
952 return 0; /* Grace period is not old enough. */
954 if (cpu_is_offline(rdp
->cpu
)) {
955 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
961 * A CPU running for an extended time within the kernel can
962 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
963 * even context-switching back and forth between a pair of
964 * in-kernel CPU-bound tasks cannot advance grace periods.
965 * So if the grace period is old enough, make the CPU pay attention.
966 * Note that the unsynchronized assignments to the per-CPU
967 * rcu_sched_qs_mask variable are safe. Yes, setting of
968 * bits can be lost, but they will be set again on the next
969 * force-quiescent-state pass. So lost bit sets do not result
970 * in incorrect behavior, merely in a grace period lasting
971 * a few jiffies longer than it might otherwise. Because
972 * there are at most four threads involved, and because the
973 * updates are only once every few jiffies, the probability of
974 * lossage (and thus of slight grace-period extension) is
977 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
978 * is set too high, we override with half of the RCU CPU stall
981 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
982 if (ULONG_CMP_GE(jiffies
,
983 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
984 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
985 if (!(ACCESS_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
986 ACCESS_ONCE(rdp
->cond_resched_completed
) =
987 ACCESS_ONCE(rdp
->mynode
->completed
);
988 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
989 ACCESS_ONCE(*rcrmp
) =
990 ACCESS_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
;
991 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
992 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
993 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
994 /* Time to beat on that CPU again! */
995 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
996 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1003 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1005 unsigned long j
= jiffies
;
1009 smp_wmb(); /* Record start time before stall time. */
1010 j1
= rcu_jiffies_till_stall_check();
1011 ACCESS_ONCE(rsp
->jiffies_stall
) = j
+ j1
;
1012 rsp
->jiffies_resched
= j
+ j1
/ 2;
1016 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
1017 * for architectures that do not implement trigger_all_cpu_backtrace().
1018 * The NMI-triggered stack traces are more accurate because they are
1019 * printed by the target CPU.
1021 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1024 unsigned long flags
;
1025 struct rcu_node
*rnp
;
1027 rcu_for_each_leaf_node(rsp
, rnp
) {
1028 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1029 if (rnp
->qsmask
!= 0) {
1030 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1031 if (rnp
->qsmask
& (1UL << cpu
))
1032 dump_cpu_task(rnp
->grplo
+ cpu
);
1034 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1038 static void print_other_cpu_stall(struct rcu_state
*rsp
)
1042 unsigned long flags
;
1044 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1047 /* Only let one CPU complain about others per time interval. */
1049 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1050 delta
= jiffies
- ACCESS_ONCE(rsp
->jiffies_stall
);
1051 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1052 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1055 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3;
1056 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1059 * OK, time to rat on our buddy...
1060 * See Documentation/RCU/stallwarn.txt for info on how to debug
1061 * RCU CPU stall warnings.
1063 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1065 print_cpu_stall_info_begin();
1066 rcu_for_each_leaf_node(rsp
, rnp
) {
1067 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1068 ndetected
+= rcu_print_task_stall(rnp
);
1069 if (rnp
->qsmask
!= 0) {
1070 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1071 if (rnp
->qsmask
& (1UL << cpu
)) {
1072 print_cpu_stall_info(rsp
,
1077 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1081 * Now rat on any tasks that got kicked up to the root rcu_node
1082 * due to CPU offlining.
1084 rnp
= rcu_get_root(rsp
);
1085 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1086 ndetected
+= rcu_print_task_stall(rnp
);
1087 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1089 print_cpu_stall_info_end();
1090 for_each_possible_cpu(cpu
)
1091 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1092 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1093 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1094 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1096 pr_err("INFO: Stall ended before state dump start\n");
1097 else if (!trigger_all_cpu_backtrace())
1098 rcu_dump_cpu_stacks(rsp
);
1100 /* Complain about tasks blocking the grace period. */
1102 rcu_print_detail_task_stall(rsp
);
1104 force_quiescent_state(rsp
); /* Kick them all. */
1107 static void print_cpu_stall(struct rcu_state
*rsp
)
1110 unsigned long flags
;
1111 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1115 * OK, time to rat on ourselves...
1116 * See Documentation/RCU/stallwarn.txt for info on how to debug
1117 * RCU CPU stall warnings.
1119 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1120 print_cpu_stall_info_begin();
1121 print_cpu_stall_info(rsp
, smp_processor_id());
1122 print_cpu_stall_info_end();
1123 for_each_possible_cpu(cpu
)
1124 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1125 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1126 jiffies
- rsp
->gp_start
,
1127 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1128 if (!trigger_all_cpu_backtrace())
1131 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1132 if (ULONG_CMP_GE(jiffies
, ACCESS_ONCE(rsp
->jiffies_stall
)))
1133 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+
1134 3 * rcu_jiffies_till_stall_check() + 3;
1135 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1138 * Attempt to revive the RCU machinery by forcing a context switch.
1140 * A context switch would normally allow the RCU state machine to make
1141 * progress and it could be we're stuck in kernel space without context
1142 * switches for an entirely unreasonable amount of time.
1144 resched_cpu(smp_processor_id());
1147 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1149 unsigned long completed
;
1150 unsigned long gpnum
;
1154 struct rcu_node
*rnp
;
1156 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1161 * Lots of memory barriers to reject false positives.
1163 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1164 * then rsp->gp_start, and finally rsp->completed. These values
1165 * are updated in the opposite order with memory barriers (or
1166 * equivalent) during grace-period initialization and cleanup.
1167 * Now, a false positive can occur if we get an new value of
1168 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1169 * the memory barriers, the only way that this can happen is if one
1170 * grace period ends and another starts between these two fetches.
1171 * Detect this by comparing rsp->completed with the previous fetch
1174 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1175 * and rsp->gp_start suffice to forestall false positives.
1177 gpnum
= ACCESS_ONCE(rsp
->gpnum
);
1178 smp_rmb(); /* Pick up ->gpnum first... */
1179 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
1180 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1181 gps
= ACCESS_ONCE(rsp
->gp_start
);
1182 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1183 completed
= ACCESS_ONCE(rsp
->completed
);
1184 if (ULONG_CMP_GE(completed
, gpnum
) ||
1185 ULONG_CMP_LT(j
, js
) ||
1186 ULONG_CMP_GE(gps
, js
))
1187 return; /* No stall or GP completed since entering function. */
1189 if (rcu_gp_in_progress(rsp
) &&
1190 (ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1192 /* We haven't checked in, so go dump stack. */
1193 print_cpu_stall(rsp
);
1195 } else if (rcu_gp_in_progress(rsp
) &&
1196 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1198 /* They had a few time units to dump stack, so complain. */
1199 print_other_cpu_stall(rsp
);
1204 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1206 * Set the stall-warning timeout way off into the future, thus preventing
1207 * any RCU CPU stall-warning messages from appearing in the current set of
1208 * RCU grace periods.
1210 * The caller must disable hard irqs.
1212 void rcu_cpu_stall_reset(void)
1214 struct rcu_state
*rsp
;
1216 for_each_rcu_flavor(rsp
)
1217 ACCESS_ONCE(rsp
->jiffies_stall
) = jiffies
+ ULONG_MAX
/ 2;
1221 * Initialize the specified rcu_data structure's callback list to empty.
1223 static void init_callback_list(struct rcu_data
*rdp
)
1227 if (init_nocb_callback_list(rdp
))
1229 rdp
->nxtlist
= NULL
;
1230 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1231 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1235 * Determine the value that ->completed will have at the end of the
1236 * next subsequent grace period. This is used to tag callbacks so that
1237 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1238 * been dyntick-idle for an extended period with callbacks under the
1239 * influence of RCU_FAST_NO_HZ.
1241 * The caller must hold rnp->lock with interrupts disabled.
1243 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1244 struct rcu_node
*rnp
)
1247 * If RCU is idle, we just wait for the next grace period.
1248 * But we can only be sure that RCU is idle if we are looking
1249 * at the root rcu_node structure -- otherwise, a new grace
1250 * period might have started, but just not yet gotten around
1251 * to initializing the current non-root rcu_node structure.
1253 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1254 return rnp
->completed
+ 1;
1257 * Otherwise, wait for a possible partial grace period and
1258 * then the subsequent full grace period.
1260 return rnp
->completed
+ 2;
1264 * Trace-event helper function for rcu_start_future_gp() and
1265 * rcu_nocb_wait_gp().
1267 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1268 unsigned long c
, const char *s
)
1270 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1271 rnp
->completed
, c
, rnp
->level
,
1272 rnp
->grplo
, rnp
->grphi
, s
);
1276 * Start some future grace period, as needed to handle newly arrived
1277 * callbacks. The required future grace periods are recorded in each
1278 * rcu_node structure's ->need_future_gp field. Returns true if there
1279 * is reason to awaken the grace-period kthread.
1281 * The caller must hold the specified rcu_node structure's ->lock.
1283 static bool __maybe_unused
1284 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1285 unsigned long *c_out
)
1290 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1293 * Pick up grace-period number for new callbacks. If this
1294 * grace period is already marked as needed, return to the caller.
1296 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1297 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1298 if (rnp
->need_future_gp
[c
& 0x1]) {
1299 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1304 * If either this rcu_node structure or the root rcu_node structure
1305 * believe that a grace period is in progress, then we must wait
1306 * for the one following, which is in "c". Because our request
1307 * will be noticed at the end of the current grace period, we don't
1308 * need to explicitly start one.
1310 if (rnp
->gpnum
!= rnp
->completed
||
1311 ACCESS_ONCE(rnp
->gpnum
) != ACCESS_ONCE(rnp
->completed
)) {
1312 rnp
->need_future_gp
[c
& 0x1]++;
1313 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1318 * There might be no grace period in progress. If we don't already
1319 * hold it, acquire the root rcu_node structure's lock in order to
1320 * start one (if needed).
1322 if (rnp
!= rnp_root
) {
1323 raw_spin_lock(&rnp_root
->lock
);
1324 smp_mb__after_unlock_lock();
1328 * Get a new grace-period number. If there really is no grace
1329 * period in progress, it will be smaller than the one we obtained
1330 * earlier. Adjust callbacks as needed. Note that even no-CBs
1331 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1333 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1334 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1335 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1336 rdp
->nxtcompleted
[i
] = c
;
1339 * If the needed for the required grace period is already
1340 * recorded, trace and leave.
1342 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1343 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1347 /* Record the need for the future grace period. */
1348 rnp_root
->need_future_gp
[c
& 0x1]++;
1350 /* If a grace period is not already in progress, start one. */
1351 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1352 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1354 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1355 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1358 if (rnp
!= rnp_root
)
1359 raw_spin_unlock(&rnp_root
->lock
);
1367 * Clean up any old requests for the just-ended grace period. Also return
1368 * whether any additional grace periods have been requested. Also invoke
1369 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1370 * waiting for this grace period to complete.
1372 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1374 int c
= rnp
->completed
;
1376 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1378 rcu_nocb_gp_cleanup(rsp
, rnp
);
1379 rnp
->need_future_gp
[c
& 0x1] = 0;
1380 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1381 trace_rcu_future_gp(rnp
, rdp
, c
,
1382 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1387 * Awaken the grace-period kthread for the specified flavor of RCU.
1388 * Don't do a self-awaken, and don't bother awakening when there is
1389 * nothing for the grace-period kthread to do (as in several CPUs
1390 * raced to awaken, and we lost), and finally don't try to awaken
1391 * a kthread that has not yet been created.
1393 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1395 if (current
== rsp
->gp_kthread
||
1396 !ACCESS_ONCE(rsp
->gp_flags
) ||
1399 wake_up(&rsp
->gp_wq
);
1403 * If there is room, assign a ->completed number to any callbacks on
1404 * this CPU that have not already been assigned. Also accelerate any
1405 * callbacks that were previously assigned a ->completed number that has
1406 * since proven to be too conservative, which can happen if callbacks get
1407 * assigned a ->completed number while RCU is idle, but with reference to
1408 * a non-root rcu_node structure. This function is idempotent, so it does
1409 * not hurt to call it repeatedly. Returns an flag saying that we should
1410 * awaken the RCU grace-period kthread.
1412 * The caller must hold rnp->lock with interrupts disabled.
1414 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1415 struct rcu_data
*rdp
)
1421 /* If the CPU has no callbacks, nothing to do. */
1422 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1426 * Starting from the sublist containing the callbacks most
1427 * recently assigned a ->completed number and working down, find the
1428 * first sublist that is not assignable to an upcoming grace period.
1429 * Such a sublist has something in it (first two tests) and has
1430 * a ->completed number assigned that will complete sooner than
1431 * the ->completed number for newly arrived callbacks (last test).
1433 * The key point is that any later sublist can be assigned the
1434 * same ->completed number as the newly arrived callbacks, which
1435 * means that the callbacks in any of these later sublist can be
1436 * grouped into a single sublist, whether or not they have already
1437 * been assigned a ->completed number.
1439 c
= rcu_cbs_completed(rsp
, rnp
);
1440 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1441 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1442 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1446 * If there are no sublist for unassigned callbacks, leave.
1447 * At the same time, advance "i" one sublist, so that "i" will
1448 * index into the sublist where all the remaining callbacks should
1451 if (++i
>= RCU_NEXT_TAIL
)
1455 * Assign all subsequent callbacks' ->completed number to the next
1456 * full grace period and group them all in the sublist initially
1459 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1460 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1461 rdp
->nxtcompleted
[i
] = c
;
1463 /* Record any needed additional grace periods. */
1464 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1466 /* Trace depending on how much we were able to accelerate. */
1467 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1468 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1470 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1475 * Move any callbacks whose grace period has completed to the
1476 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1477 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1478 * sublist. This function is idempotent, so it does not hurt to
1479 * invoke it repeatedly. As long as it is not invoked -too- often...
1480 * Returns true if the RCU grace-period kthread needs to be awakened.
1482 * The caller must hold rnp->lock with interrupts disabled.
1484 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1485 struct rcu_data
*rdp
)
1489 /* If the CPU has no callbacks, nothing to do. */
1490 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1494 * Find all callbacks whose ->completed numbers indicate that they
1495 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1497 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1498 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1500 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1502 /* Clean up any sublist tail pointers that were misordered above. */
1503 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1504 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1506 /* Copy down callbacks to fill in empty sublists. */
1507 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1508 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1510 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1511 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1514 /* Classify any remaining callbacks. */
1515 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1519 * Update CPU-local rcu_data state to record the beginnings and ends of
1520 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1521 * structure corresponding to the current CPU, and must have irqs disabled.
1522 * Returns true if the grace-period kthread needs to be awakened.
1524 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1525 struct rcu_data
*rdp
)
1529 /* Handle the ends of any preceding grace periods first. */
1530 if (rdp
->completed
== rnp
->completed
) {
1532 /* No grace period end, so just accelerate recent callbacks. */
1533 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1537 /* Advance callbacks. */
1538 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1540 /* Remember that we saw this grace-period completion. */
1541 rdp
->completed
= rnp
->completed
;
1542 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1545 if (rdp
->gpnum
!= rnp
->gpnum
) {
1547 * If the current grace period is waiting for this CPU,
1548 * set up to detect a quiescent state, otherwise don't
1549 * go looking for one.
1551 rdp
->gpnum
= rnp
->gpnum
;
1552 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1553 rdp
->passed_quiesce
= 0;
1554 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1555 zero_cpu_stall_ticks(rdp
);
1560 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1562 unsigned long flags
;
1564 struct rcu_node
*rnp
;
1566 local_irq_save(flags
);
1568 if ((rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) &&
1569 rdp
->completed
== ACCESS_ONCE(rnp
->completed
)) || /* w/out lock. */
1570 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1571 local_irq_restore(flags
);
1574 smp_mb__after_unlock_lock();
1575 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1576 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1578 rcu_gp_kthread_wake(rsp
);
1582 * Initialize a new grace period. Return 0 if no grace period required.
1584 static int rcu_gp_init(struct rcu_state
*rsp
)
1586 struct rcu_data
*rdp
;
1587 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1589 rcu_bind_gp_kthread();
1590 raw_spin_lock_irq(&rnp
->lock
);
1591 smp_mb__after_unlock_lock();
1592 if (!ACCESS_ONCE(rsp
->gp_flags
)) {
1593 /* Spurious wakeup, tell caller to go back to sleep. */
1594 raw_spin_unlock_irq(&rnp
->lock
);
1597 ACCESS_ONCE(rsp
->gp_flags
) = 0; /* Clear all flags: New grace period. */
1599 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1601 * Grace period already in progress, don't start another.
1602 * Not supposed to be able to happen.
1604 raw_spin_unlock_irq(&rnp
->lock
);
1608 /* Advance to a new grace period and initialize state. */
1609 record_gp_stall_check_time(rsp
);
1610 /* Record GP times before starting GP, hence smp_store_release(). */
1611 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1612 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1613 raw_spin_unlock_irq(&rnp
->lock
);
1615 /* Exclude any concurrent CPU-hotplug operations. */
1616 mutex_lock(&rsp
->onoff_mutex
);
1617 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1620 * Set the quiescent-state-needed bits in all the rcu_node
1621 * structures for all currently online CPUs in breadth-first order,
1622 * starting from the root rcu_node structure, relying on the layout
1623 * of the tree within the rsp->node[] array. Note that other CPUs
1624 * will access only the leaves of the hierarchy, thus seeing that no
1625 * grace period is in progress, at least until the corresponding
1626 * leaf node has been initialized. In addition, we have excluded
1627 * CPU-hotplug operations.
1629 * The grace period cannot complete until the initialization
1630 * process finishes, because this kthread handles both.
1632 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1633 raw_spin_lock_irq(&rnp
->lock
);
1634 smp_mb__after_unlock_lock();
1635 rdp
= this_cpu_ptr(rsp
->rda
);
1636 rcu_preempt_check_blocked_tasks(rnp
);
1637 rnp
->qsmask
= rnp
->qsmaskinit
;
1638 ACCESS_ONCE(rnp
->gpnum
) = rsp
->gpnum
;
1639 WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
);
1640 ACCESS_ONCE(rnp
->completed
) = rsp
->completed
;
1641 if (rnp
== rdp
->mynode
)
1642 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1643 rcu_preempt_boost_start_gp(rnp
);
1644 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1645 rnp
->level
, rnp
->grplo
,
1646 rnp
->grphi
, rnp
->qsmask
);
1647 raw_spin_unlock_irq(&rnp
->lock
);
1648 #ifdef CONFIG_PROVE_RCU_DELAY
1649 if ((prandom_u32() % (rcu_num_nodes
+ 1)) == 0 &&
1650 system_state
== SYSTEM_RUNNING
)
1652 #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1656 mutex_unlock(&rsp
->onoff_mutex
);
1661 * Do one round of quiescent-state forcing.
1663 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1665 int fqs_state
= fqs_state_in
;
1666 bool isidle
= false;
1668 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1671 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1672 /* Collect dyntick-idle snapshots. */
1673 if (is_sysidle_rcu_state(rsp
)) {
1675 maxj
= jiffies
- ULONG_MAX
/ 4;
1677 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1679 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1680 fqs_state
= RCU_FORCE_QS
;
1682 /* Handle dyntick-idle and offline CPUs. */
1684 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1686 /* Clear flag to prevent immediate re-entry. */
1687 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1688 raw_spin_lock_irq(&rnp
->lock
);
1689 smp_mb__after_unlock_lock();
1690 ACCESS_ONCE(rsp
->gp_flags
) &= ~RCU_GP_FLAG_FQS
;
1691 raw_spin_unlock_irq(&rnp
->lock
);
1697 * Clean up after the old grace period.
1699 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1701 unsigned long gp_duration
;
1702 bool needgp
= false;
1704 struct rcu_data
*rdp
;
1705 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1707 raw_spin_lock_irq(&rnp
->lock
);
1708 smp_mb__after_unlock_lock();
1709 gp_duration
= jiffies
- rsp
->gp_start
;
1710 if (gp_duration
> rsp
->gp_max
)
1711 rsp
->gp_max
= gp_duration
;
1714 * We know the grace period is complete, but to everyone else
1715 * it appears to still be ongoing. But it is also the case
1716 * that to everyone else it looks like there is nothing that
1717 * they can do to advance the grace period. It is therefore
1718 * safe for us to drop the lock in order to mark the grace
1719 * period as completed in all of the rcu_node structures.
1721 raw_spin_unlock_irq(&rnp
->lock
);
1724 * Propagate new ->completed value to rcu_node structures so
1725 * that other CPUs don't have to wait until the start of the next
1726 * grace period to process their callbacks. This also avoids
1727 * some nasty RCU grace-period initialization races by forcing
1728 * the end of the current grace period to be completely recorded in
1729 * all of the rcu_node structures before the beginning of the next
1730 * grace period is recorded in any of the rcu_node structures.
1732 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1733 raw_spin_lock_irq(&rnp
->lock
);
1734 smp_mb__after_unlock_lock();
1735 ACCESS_ONCE(rnp
->completed
) = rsp
->gpnum
;
1736 rdp
= this_cpu_ptr(rsp
->rda
);
1737 if (rnp
== rdp
->mynode
)
1738 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1739 /* smp_mb() provided by prior unlock-lock pair. */
1740 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1741 raw_spin_unlock_irq(&rnp
->lock
);
1744 rnp
= rcu_get_root(rsp
);
1745 raw_spin_lock_irq(&rnp
->lock
);
1746 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1747 rcu_nocb_gp_set(rnp
, nocb
);
1749 /* Declare grace period done. */
1750 ACCESS_ONCE(rsp
->completed
) = rsp
->gpnum
;
1751 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1752 rsp
->fqs_state
= RCU_GP_IDLE
;
1753 rdp
= this_cpu_ptr(rsp
->rda
);
1754 /* Advance CBs to reduce false positives below. */
1755 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1756 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1757 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1758 trace_rcu_grace_period(rsp
->name
,
1759 ACCESS_ONCE(rsp
->gpnum
),
1762 raw_spin_unlock_irq(&rnp
->lock
);
1766 * Body of kthread that handles grace periods.
1768 static int __noreturn
rcu_gp_kthread(void *arg
)
1774 struct rcu_state
*rsp
= arg
;
1775 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1779 /* Handle grace-period start. */
1781 trace_rcu_grace_period(rsp
->name
,
1782 ACCESS_ONCE(rsp
->gpnum
),
1784 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
1785 wait_event_interruptible(rsp
->gp_wq
,
1786 ACCESS_ONCE(rsp
->gp_flags
) &
1788 /* Locking provides needed memory barrier. */
1789 if (rcu_gp_init(rsp
))
1792 flush_signals(current
);
1793 trace_rcu_grace_period(rsp
->name
,
1794 ACCESS_ONCE(rsp
->gpnum
),
1798 /* Handle quiescent-state forcing. */
1799 fqs_state
= RCU_SAVE_DYNTICK
;
1800 j
= jiffies_till_first_fqs
;
1803 jiffies_till_first_fqs
= HZ
;
1808 rsp
->jiffies_force_qs
= jiffies
+ j
;
1809 trace_rcu_grace_period(rsp
->name
,
1810 ACCESS_ONCE(rsp
->gpnum
),
1812 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
1813 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
1814 ((gf
= ACCESS_ONCE(rsp
->gp_flags
)) &
1816 (!ACCESS_ONCE(rnp
->qsmask
) &&
1817 !rcu_preempt_blocked_readers_cgp(rnp
)),
1819 /* Locking provides needed memory barriers. */
1820 /* If grace period done, leave loop. */
1821 if (!ACCESS_ONCE(rnp
->qsmask
) &&
1822 !rcu_preempt_blocked_readers_cgp(rnp
))
1824 /* If time for quiescent-state forcing, do it. */
1825 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
1826 (gf
& RCU_GP_FLAG_FQS
)) {
1827 trace_rcu_grace_period(rsp
->name
,
1828 ACCESS_ONCE(rsp
->gpnum
),
1830 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
1831 trace_rcu_grace_period(rsp
->name
,
1832 ACCESS_ONCE(rsp
->gpnum
),
1836 /* Deal with stray signal. */
1838 flush_signals(current
);
1839 trace_rcu_grace_period(rsp
->name
,
1840 ACCESS_ONCE(rsp
->gpnum
),
1843 j
= jiffies_till_next_fqs
;
1846 jiffies_till_next_fqs
= HZ
;
1849 jiffies_till_next_fqs
= 1;
1853 /* Handle grace-period end. */
1854 rcu_gp_cleanup(rsp
);
1859 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1860 * in preparation for detecting the next grace period. The caller must hold
1861 * the root node's ->lock and hard irqs must be disabled.
1863 * Note that it is legal for a dying CPU (which is marked as offline) to
1864 * invoke this function. This can happen when the dying CPU reports its
1867 * Returns true if the grace-period kthread must be awakened.
1870 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1871 struct rcu_data
*rdp
)
1873 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
1875 * Either we have not yet spawned the grace-period
1876 * task, this CPU does not need another grace period,
1877 * or a grace period is already in progress.
1878 * Either way, don't start a new grace period.
1882 ACCESS_ONCE(rsp
->gp_flags
) = RCU_GP_FLAG_INIT
;
1883 trace_rcu_grace_period(rsp
->name
, ACCESS_ONCE(rsp
->gpnum
),
1887 * We can't do wakeups while holding the rnp->lock, as that
1888 * could cause possible deadlocks with the rq->lock. Defer
1889 * the wakeup to our caller.
1895 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1896 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1897 * is invoked indirectly from rcu_advance_cbs(), which would result in
1898 * endless recursion -- or would do so if it wasn't for the self-deadlock
1899 * that is encountered beforehand.
1901 * Returns true if the grace-period kthread needs to be awakened.
1903 static bool rcu_start_gp(struct rcu_state
*rsp
)
1905 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1906 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1910 * If there is no grace period in progress right now, any
1911 * callbacks we have up to this point will be satisfied by the
1912 * next grace period. Also, advancing the callbacks reduces the
1913 * probability of false positives from cpu_needs_another_gp()
1914 * resulting in pointless grace periods. So, advance callbacks
1915 * then start the grace period!
1917 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
1918 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
1923 * Report a full set of quiescent states to the specified rcu_state
1924 * data structure. This involves cleaning up after the prior grace
1925 * period and letting rcu_start_gp() start up the next grace period
1926 * if one is needed. Note that the caller must hold rnp->lock, which
1927 * is released before return.
1929 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
1930 __releases(rcu_get_root(rsp
)->lock
)
1932 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
1933 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
1934 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
1938 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1939 * Allows quiescent states for a group of CPUs to be reported at one go
1940 * to the specified rcu_node structure, though all the CPUs in the group
1941 * must be represented by the same rcu_node structure (which need not be
1942 * a leaf rcu_node structure, though it often will be). That structure's
1943 * lock must be held upon entry, and it is released before return.
1946 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
1947 struct rcu_node
*rnp
, unsigned long flags
)
1948 __releases(rnp
->lock
)
1950 struct rcu_node
*rnp_c
;
1952 /* Walk up the rcu_node hierarchy. */
1954 if (!(rnp
->qsmask
& mask
)) {
1956 /* Our bit has already been cleared, so done. */
1957 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1960 rnp
->qsmask
&= ~mask
;
1961 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
1962 mask
, rnp
->qsmask
, rnp
->level
,
1963 rnp
->grplo
, rnp
->grphi
,
1965 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1967 /* Other bits still set at this level, so done. */
1968 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1971 mask
= rnp
->grpmask
;
1972 if (rnp
->parent
== NULL
) {
1974 /* No more levels. Exit loop holding root lock. */
1978 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1981 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1982 smp_mb__after_unlock_lock();
1983 WARN_ON_ONCE(rnp_c
->qsmask
);
1987 * Get here if we are the last CPU to pass through a quiescent
1988 * state for this grace period. Invoke rcu_report_qs_rsp()
1989 * to clean up and start the next grace period if one is needed.
1991 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
1995 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1996 * structure. This must be either called from the specified CPU, or
1997 * called when the specified CPU is known to be offline (and when it is
1998 * also known that no other CPU is concurrently trying to help the offline
1999 * CPU). The lastcomp argument is used to make sure we are still in the
2000 * grace period of interest. We don't want to end the current grace period
2001 * based on quiescent states detected in an earlier grace period!
2004 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2006 unsigned long flags
;
2009 struct rcu_node
*rnp
;
2012 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2013 smp_mb__after_unlock_lock();
2014 if (rdp
->passed_quiesce
== 0 || rdp
->gpnum
!= rnp
->gpnum
||
2015 rnp
->completed
== rnp
->gpnum
) {
2018 * The grace period in which this quiescent state was
2019 * recorded has ended, so don't report it upwards.
2020 * We will instead need a new quiescent state that lies
2021 * within the current grace period.
2023 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2024 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2027 mask
= rdp
->grpmask
;
2028 if ((rnp
->qsmask
& mask
) == 0) {
2029 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2031 rdp
->qs_pending
= 0;
2034 * This GP can't end until cpu checks in, so all of our
2035 * callbacks can be processed during the next GP.
2037 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2039 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
2041 rcu_gp_kthread_wake(rsp
);
2046 * Check to see if there is a new grace period of which this CPU
2047 * is not yet aware, and if so, set up local rcu_data state for it.
2048 * Otherwise, see if this CPU has just passed through its first
2049 * quiescent state for this grace period, and record that fact if so.
2052 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2054 /* Check for grace-period ends and beginnings. */
2055 note_gp_changes(rsp
, rdp
);
2058 * Does this CPU still need to do its part for current grace period?
2059 * If no, return and let the other CPUs do their part as well.
2061 if (!rdp
->qs_pending
)
2065 * Was there a quiescent state since the beginning of the grace
2066 * period? If no, then exit and wait for the next call.
2068 if (!rdp
->passed_quiesce
)
2072 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2075 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2078 #ifdef CONFIG_HOTPLUG_CPU
2081 * Send the specified CPU's RCU callbacks to the orphanage. The
2082 * specified CPU must be offline, and the caller must hold the
2086 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2087 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2089 /* No-CBs CPUs do not have orphanable callbacks. */
2090 if (rcu_is_nocb_cpu(rdp
->cpu
))
2094 * Orphan the callbacks. First adjust the counts. This is safe
2095 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2096 * cannot be running now. Thus no memory barrier is required.
2098 if (rdp
->nxtlist
!= NULL
) {
2099 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2100 rsp
->qlen
+= rdp
->qlen
;
2101 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2103 ACCESS_ONCE(rdp
->qlen
) = 0;
2107 * Next, move those callbacks still needing a grace period to
2108 * the orphanage, where some other CPU will pick them up.
2109 * Some of the callbacks might have gone partway through a grace
2110 * period, but that is too bad. They get to start over because we
2111 * cannot assume that grace periods are synchronized across CPUs.
2112 * We don't bother updating the ->nxttail[] array yet, instead
2113 * we just reset the whole thing later on.
2115 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2116 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2117 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2118 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2122 * Then move the ready-to-invoke callbacks to the orphanage,
2123 * where some other CPU will pick them up. These will not be
2124 * required to pass though another grace period: They are done.
2126 if (rdp
->nxtlist
!= NULL
) {
2127 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2128 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2131 /* Finally, initialize the rcu_data structure's list to empty. */
2132 init_callback_list(rdp
);
2136 * Adopt the RCU callbacks from the specified rcu_state structure's
2137 * orphanage. The caller must hold the ->orphan_lock.
2139 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2142 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2144 /* No-CBs CPUs are handled specially. */
2145 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2148 /* Do the accounting first. */
2149 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2150 rdp
->qlen
+= rsp
->qlen
;
2151 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2152 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2153 rcu_idle_count_callbacks_posted();
2158 * We do not need a memory barrier here because the only way we
2159 * can get here if there is an rcu_barrier() in flight is if
2160 * we are the task doing the rcu_barrier().
2163 /* First adopt the ready-to-invoke callbacks. */
2164 if (rsp
->orphan_donelist
!= NULL
) {
2165 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2166 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2167 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2168 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2169 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2170 rsp
->orphan_donelist
= NULL
;
2171 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2174 /* And then adopt the callbacks that still need a grace period. */
2175 if (rsp
->orphan_nxtlist
!= NULL
) {
2176 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2177 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2178 rsp
->orphan_nxtlist
= NULL
;
2179 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2184 * Trace the fact that this CPU is going offline.
2186 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2188 RCU_TRACE(unsigned long mask
);
2189 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2190 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2192 RCU_TRACE(mask
= rdp
->grpmask
);
2193 trace_rcu_grace_period(rsp
->name
,
2194 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2199 * The CPU has been completely removed, and some other CPU is reporting
2200 * this fact from process context. Do the remainder of the cleanup,
2201 * including orphaning the outgoing CPU's RCU callbacks, and also
2202 * adopting them. There can only be one CPU hotplug operation at a time,
2203 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2205 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2207 unsigned long flags
;
2209 int need_report
= 0;
2210 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2211 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2213 /* Adjust any no-longer-needed kthreads. */
2214 rcu_boost_kthread_setaffinity(rnp
, -1);
2216 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2218 /* Exclude any attempts to start a new grace period. */
2219 mutex_lock(&rsp
->onoff_mutex
);
2220 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2222 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2223 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2224 rcu_adopt_orphan_cbs(rsp
, flags
);
2226 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2227 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
2229 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2230 smp_mb__after_unlock_lock();
2231 rnp
->qsmaskinit
&= ~mask
;
2232 if (rnp
->qsmaskinit
!= 0) {
2233 if (rnp
!= rdp
->mynode
)
2234 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2237 if (rnp
== rdp
->mynode
)
2238 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
2240 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2241 mask
= rnp
->grpmask
;
2243 } while (rnp
!= NULL
);
2246 * We still hold the leaf rcu_node structure lock here, and
2247 * irqs are still disabled. The reason for this subterfuge is
2248 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2249 * held leads to deadlock.
2251 raw_spin_unlock(&rsp
->orphan_lock
); /* irqs remain disabled. */
2253 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
2254 rcu_report_unblock_qs_rnp(rnp
, flags
);
2256 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2257 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
2258 rcu_report_exp_rnp(rsp
, rnp
, true);
2259 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2260 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2261 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2262 init_callback_list(rdp
);
2263 /* Disallow further callbacks on this CPU. */
2264 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2265 mutex_unlock(&rsp
->onoff_mutex
);
2268 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2270 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2274 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2278 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2281 * Invoke any RCU callbacks that have made it to the end of their grace
2282 * period. Thottle as specified by rdp->blimit.
2284 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2286 unsigned long flags
;
2287 struct rcu_head
*next
, *list
, **tail
;
2288 long bl
, count
, count_lazy
;
2291 /* If no callbacks are ready, just return. */
2292 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2293 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2294 trace_rcu_batch_end(rsp
->name
, 0, !!ACCESS_ONCE(rdp
->nxtlist
),
2295 need_resched(), is_idle_task(current
),
2296 rcu_is_callbacks_kthread());
2301 * Extract the list of ready callbacks, disabling to prevent
2302 * races with call_rcu() from interrupt handlers.
2304 local_irq_save(flags
);
2305 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2307 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2308 list
= rdp
->nxtlist
;
2309 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2310 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2311 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2312 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2313 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2314 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2315 local_irq_restore(flags
);
2317 /* Invoke callbacks. */
2318 count
= count_lazy
= 0;
2322 debug_rcu_head_unqueue(list
);
2323 if (__rcu_reclaim(rsp
->name
, list
))
2326 /* Stop only if limit reached and CPU has something to do. */
2327 if (++count
>= bl
&&
2329 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2333 local_irq_save(flags
);
2334 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2335 is_idle_task(current
),
2336 rcu_is_callbacks_kthread());
2338 /* Update count, and requeue any remaining callbacks. */
2340 *tail
= rdp
->nxtlist
;
2341 rdp
->nxtlist
= list
;
2342 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2343 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2344 rdp
->nxttail
[i
] = tail
;
2348 smp_mb(); /* List handling before counting for rcu_barrier(). */
2349 rdp
->qlen_lazy
-= count_lazy
;
2350 ACCESS_ONCE(rdp
->qlen
) -= count
;
2351 rdp
->n_cbs_invoked
+= count
;
2353 /* Reinstate batch limit if we have worked down the excess. */
2354 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2355 rdp
->blimit
= blimit
;
2357 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2358 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2359 rdp
->qlen_last_fqs_check
= 0;
2360 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2361 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2362 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2363 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2365 local_irq_restore(flags
);
2367 /* Re-invoke RCU core processing if there are callbacks remaining. */
2368 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2373 * Check to see if this CPU is in a non-context-switch quiescent state
2374 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2375 * Also schedule RCU core processing.
2377 * This function must be called from hardirq context. It is normally
2378 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2379 * false, there is no point in invoking rcu_check_callbacks().
2381 void rcu_check_callbacks(int cpu
, int user
)
2383 trace_rcu_utilization(TPS("Start scheduler-tick"));
2384 increment_cpu_stall_ticks();
2385 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2388 * Get here if this CPU took its interrupt from user
2389 * mode or from the idle loop, and if this is not a
2390 * nested interrupt. In this case, the CPU is in
2391 * a quiescent state, so note it.
2393 * No memory barrier is required here because both
2394 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2395 * variables that other CPUs neither access nor modify,
2396 * at least not while the corresponding CPU is online.
2402 } else if (!in_softirq()) {
2405 * Get here if this CPU did not take its interrupt from
2406 * softirq, in other words, if it is not interrupting
2407 * a rcu_bh read-side critical section. This is an _bh
2408 * critical section, so note it.
2413 rcu_preempt_check_callbacks(cpu
);
2414 if (rcu_pending(cpu
))
2416 trace_rcu_utilization(TPS("End scheduler-tick"));
2420 * Scan the leaf rcu_node structures, processing dyntick state for any that
2421 * have not yet encountered a quiescent state, using the function specified.
2422 * Also initiate boosting for any threads blocked on the root rcu_node.
2424 * The caller must have suppressed start of new grace periods.
2426 static void force_qs_rnp(struct rcu_state
*rsp
,
2427 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2428 unsigned long *maxj
),
2429 bool *isidle
, unsigned long *maxj
)
2433 unsigned long flags
;
2435 struct rcu_node
*rnp
;
2437 rcu_for_each_leaf_node(rsp
, rnp
) {
2440 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2441 smp_mb__after_unlock_lock();
2442 if (!rcu_gp_in_progress(rsp
)) {
2443 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2446 if (rnp
->qsmask
== 0) {
2447 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
2452 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2453 if ((rnp
->qsmask
& bit
) != 0) {
2454 if ((rnp
->qsmaskinit
& bit
) != 0)
2456 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2462 /* rcu_report_qs_rnp() releases rnp->lock. */
2463 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
2466 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2468 rnp
= rcu_get_root(rsp
);
2469 if (rnp
->qsmask
== 0) {
2470 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2471 smp_mb__after_unlock_lock();
2472 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
2477 * Force quiescent states on reluctant CPUs, and also detect which
2478 * CPUs are in dyntick-idle mode.
2480 static void force_quiescent_state(struct rcu_state
*rsp
)
2482 unsigned long flags
;
2484 struct rcu_node
*rnp
;
2485 struct rcu_node
*rnp_old
= NULL
;
2487 /* Funnel through hierarchy to reduce memory contention. */
2488 rnp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id())->mynode
;
2489 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2490 ret
= (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2491 !raw_spin_trylock(&rnp
->fqslock
);
2492 if (rnp_old
!= NULL
)
2493 raw_spin_unlock(&rnp_old
->fqslock
);
2495 ACCESS_ONCE(rsp
->n_force_qs_lh
)++;
2500 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2502 /* Reached the root of the rcu_node tree, acquire lock. */
2503 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2504 smp_mb__after_unlock_lock();
2505 raw_spin_unlock(&rnp_old
->fqslock
);
2506 if (ACCESS_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2507 ACCESS_ONCE(rsp
->n_force_qs_lh
)++;
2508 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2509 return; /* Someone beat us to it. */
2511 ACCESS_ONCE(rsp
->gp_flags
) |= RCU_GP_FLAG_FQS
;
2512 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2513 wake_up(&rsp
->gp_wq
); /* Memory barrier implied by wake_up() path. */
2517 * This does the RCU core processing work for the specified rcu_state
2518 * and rcu_data structures. This may be called only from the CPU to
2519 * whom the rdp belongs.
2522 __rcu_process_callbacks(struct rcu_state
*rsp
)
2524 unsigned long flags
;
2526 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2528 WARN_ON_ONCE(rdp
->beenonline
== 0);
2530 /* Update RCU state based on any recent quiescent states. */
2531 rcu_check_quiescent_state(rsp
, rdp
);
2533 /* Does this CPU require a not-yet-started grace period? */
2534 local_irq_save(flags
);
2535 if (cpu_needs_another_gp(rsp
, rdp
)) {
2536 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2537 needwake
= rcu_start_gp(rsp
);
2538 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2540 rcu_gp_kthread_wake(rsp
);
2542 local_irq_restore(flags
);
2545 /* If there are callbacks ready, invoke them. */
2546 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2547 invoke_rcu_callbacks(rsp
, rdp
);
2549 /* Do any needed deferred wakeups of rcuo kthreads. */
2550 do_nocb_deferred_wakeup(rdp
);
2554 * Do RCU core processing for the current CPU.
2556 static void rcu_process_callbacks(struct softirq_action
*unused
)
2558 struct rcu_state
*rsp
;
2560 if (cpu_is_offline(smp_processor_id()))
2562 trace_rcu_utilization(TPS("Start RCU core"));
2563 for_each_rcu_flavor(rsp
)
2564 __rcu_process_callbacks(rsp
);
2565 trace_rcu_utilization(TPS("End RCU core"));
2569 * Schedule RCU callback invocation. If the specified type of RCU
2570 * does not support RCU priority boosting, just do a direct call,
2571 * otherwise wake up the per-CPU kernel kthread. Note that because we
2572 * are running on the current CPU with interrupts disabled, the
2573 * rcu_cpu_kthread_task cannot disappear out from under us.
2575 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2577 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
2579 if (likely(!rsp
->boost
)) {
2580 rcu_do_batch(rsp
, rdp
);
2583 invoke_rcu_callbacks_kthread();
2586 static void invoke_rcu_core(void)
2588 if (cpu_online(smp_processor_id()))
2589 raise_softirq(RCU_SOFTIRQ
);
2593 * Handle any core-RCU processing required by a call_rcu() invocation.
2595 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2596 struct rcu_head
*head
, unsigned long flags
)
2601 * If called from an extended quiescent state, invoke the RCU
2602 * core in order to force a re-evaluation of RCU's idleness.
2604 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2607 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2608 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2612 * Force the grace period if too many callbacks or too long waiting.
2613 * Enforce hysteresis, and don't invoke force_quiescent_state()
2614 * if some other CPU has recently done so. Also, don't bother
2615 * invoking force_quiescent_state() if the newly enqueued callback
2616 * is the only one waiting for a grace period to complete.
2618 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2620 /* Are we ignoring a completed grace period? */
2621 note_gp_changes(rsp
, rdp
);
2623 /* Start a new grace period if one not already started. */
2624 if (!rcu_gp_in_progress(rsp
)) {
2625 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2627 raw_spin_lock(&rnp_root
->lock
);
2628 smp_mb__after_unlock_lock();
2629 needwake
= rcu_start_gp(rsp
);
2630 raw_spin_unlock(&rnp_root
->lock
);
2632 rcu_gp_kthread_wake(rsp
);
2634 /* Give the grace period a kick. */
2635 rdp
->blimit
= LONG_MAX
;
2636 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2637 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2638 force_quiescent_state(rsp
);
2639 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2640 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2646 * RCU callback function to leak a callback.
2648 static void rcu_leak_callback(struct rcu_head
*rhp
)
2653 * Helper function for call_rcu() and friends. The cpu argument will
2654 * normally be -1, indicating "currently running CPU". It may specify
2655 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2656 * is expected to specify a CPU.
2659 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2660 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2662 unsigned long flags
;
2663 struct rcu_data
*rdp
;
2665 WARN_ON_ONCE((unsigned long)head
& 0x3); /* Misaligned rcu_head! */
2666 if (debug_rcu_head_queue(head
)) {
2667 /* Probable double call_rcu(), so leak the callback. */
2668 ACCESS_ONCE(head
->func
) = rcu_leak_callback
;
2669 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2676 * Opportunistically note grace-period endings and beginnings.
2677 * Note that we might see a beginning right after we see an
2678 * end, but never vice versa, since this CPU has to pass through
2679 * a quiescent state betweentimes.
2681 local_irq_save(flags
);
2682 rdp
= this_cpu_ptr(rsp
->rda
);
2684 /* Add the callback to our list. */
2685 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2689 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2690 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
2691 WARN_ON_ONCE(offline
);
2692 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2693 local_irq_restore(flags
);
2696 ACCESS_ONCE(rdp
->qlen
)++;
2700 rcu_idle_count_callbacks_posted();
2701 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2702 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
2703 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
2705 if (__is_kfree_rcu_offset((unsigned long)func
))
2706 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
2707 rdp
->qlen_lazy
, rdp
->qlen
);
2709 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
2711 /* Go handle any RCU core processing required. */
2712 __call_rcu_core(rsp
, rdp
, head
, flags
);
2713 local_irq_restore(flags
);
2717 * Queue an RCU-sched callback for invocation after a grace period.
2719 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2721 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
2723 EXPORT_SYMBOL_GPL(call_rcu_sched
);
2726 * Queue an RCU callback for invocation after a quicker grace period.
2728 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
2730 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
2732 EXPORT_SYMBOL_GPL(call_rcu_bh
);
2735 * Queue an RCU callback for lazy invocation after a grace period.
2736 * This will likely be later named something like "call_rcu_lazy()",
2737 * but this change will require some way of tagging the lazy RCU
2738 * callbacks in the list of pending callbacks. Until then, this
2739 * function may only be called from __kfree_rcu().
2741 void kfree_call_rcu(struct rcu_head
*head
,
2742 void (*func
)(struct rcu_head
*rcu
))
2744 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
2746 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
2749 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2750 * any blocking grace-period wait automatically implies a grace period
2751 * if there is only one CPU online at any point time during execution
2752 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2753 * occasionally incorrectly indicate that there are multiple CPUs online
2754 * when there was in fact only one the whole time, as this just adds
2755 * some overhead: RCU still operates correctly.
2757 static inline int rcu_blocking_is_gp(void)
2761 might_sleep(); /* Check for RCU read-side critical section. */
2763 ret
= num_online_cpus() <= 1;
2769 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2771 * Control will return to the caller some time after a full rcu-sched
2772 * grace period has elapsed, in other words after all currently executing
2773 * rcu-sched read-side critical sections have completed. These read-side
2774 * critical sections are delimited by rcu_read_lock_sched() and
2775 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2776 * local_irq_disable(), and so on may be used in place of
2777 * rcu_read_lock_sched().
2779 * This means that all preempt_disable code sequences, including NMI and
2780 * non-threaded hardware-interrupt handlers, in progress on entry will
2781 * have completed before this primitive returns. However, this does not
2782 * guarantee that softirq handlers will have completed, since in some
2783 * kernels, these handlers can run in process context, and can block.
2785 * Note that this guarantee implies further memory-ordering guarantees.
2786 * On systems with more than one CPU, when synchronize_sched() returns,
2787 * each CPU is guaranteed to have executed a full memory barrier since the
2788 * end of its last RCU-sched read-side critical section whose beginning
2789 * preceded the call to synchronize_sched(). In addition, each CPU having
2790 * an RCU read-side critical section that extends beyond the return from
2791 * synchronize_sched() is guaranteed to have executed a full memory barrier
2792 * after the beginning of synchronize_sched() and before the beginning of
2793 * that RCU read-side critical section. Note that these guarantees include
2794 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2795 * that are executing in the kernel.
2797 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2798 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2799 * to have executed a full memory barrier during the execution of
2800 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2801 * again only if the system has more than one CPU).
2803 * This primitive provides the guarantees made by the (now removed)
2804 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2805 * guarantees that rcu_read_lock() sections will have completed.
2806 * In "classic RCU", these two guarantees happen to be one and
2807 * the same, but can differ in realtime RCU implementations.
2809 void synchronize_sched(void)
2811 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2812 !lock_is_held(&rcu_lock_map
) &&
2813 !lock_is_held(&rcu_sched_lock_map
),
2814 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2815 if (rcu_blocking_is_gp())
2818 synchronize_sched_expedited();
2820 wait_rcu_gp(call_rcu_sched
);
2822 EXPORT_SYMBOL_GPL(synchronize_sched
);
2825 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2827 * Control will return to the caller some time after a full rcu_bh grace
2828 * period has elapsed, in other words after all currently executing rcu_bh
2829 * read-side critical sections have completed. RCU read-side critical
2830 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2831 * and may be nested.
2833 * See the description of synchronize_sched() for more detailed information
2834 * on memory ordering guarantees.
2836 void synchronize_rcu_bh(void)
2838 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
2839 !lock_is_held(&rcu_lock_map
) &&
2840 !lock_is_held(&rcu_sched_lock_map
),
2841 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2842 if (rcu_blocking_is_gp())
2845 synchronize_rcu_bh_expedited();
2847 wait_rcu_gp(call_rcu_bh
);
2849 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
2852 * get_state_synchronize_rcu - Snapshot current RCU state
2854 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2855 * to determine whether or not a full grace period has elapsed in the
2858 unsigned long get_state_synchronize_rcu(void)
2861 * Any prior manipulation of RCU-protected data must happen
2862 * before the load from ->gpnum.
2867 * Make sure this load happens before the purportedly
2868 * time-consuming work between get_state_synchronize_rcu()
2869 * and cond_synchronize_rcu().
2871 return smp_load_acquire(&rcu_state_p
->gpnum
);
2873 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
2876 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2878 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2880 * If a full RCU grace period has elapsed since the earlier call to
2881 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2882 * synchronize_rcu() to wait for a full grace period.
2884 * Yes, this function does not take counter wrap into account. But
2885 * counter wrap is harmless. If the counter wraps, we have waited for
2886 * more than 2 billion grace periods (and way more on a 64-bit system!),
2887 * so waiting for one additional grace period should be just fine.
2889 void cond_synchronize_rcu(unsigned long oldstate
)
2891 unsigned long newstate
;
2894 * Ensure that this load happens before any RCU-destructive
2895 * actions the caller might carry out after we return.
2897 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
2898 if (ULONG_CMP_GE(oldstate
, newstate
))
2901 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
2903 static int synchronize_sched_expedited_cpu_stop(void *data
)
2906 * There must be a full memory barrier on each affected CPU
2907 * between the time that try_stop_cpus() is called and the
2908 * time that it returns.
2910 * In the current initial implementation of cpu_stop, the
2911 * above condition is already met when the control reaches
2912 * this point and the following smp_mb() is not strictly
2913 * necessary. Do smp_mb() anyway for documentation and
2914 * robustness against future implementation changes.
2916 smp_mb(); /* See above comment block. */
2921 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2923 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2924 * approach to force the grace period to end quickly. This consumes
2925 * significant time on all CPUs and is unfriendly to real-time workloads,
2926 * so is thus not recommended for any sort of common-case code. In fact,
2927 * if you are using synchronize_sched_expedited() in a loop, please
2928 * restructure your code to batch your updates, and then use a single
2929 * synchronize_sched() instead.
2931 * Note that it is illegal to call this function while holding any lock
2932 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2933 * to call this function from a CPU-hotplug notifier. Failing to observe
2934 * these restriction will result in deadlock.
2936 * This implementation can be thought of as an application of ticket
2937 * locking to RCU, with sync_sched_expedited_started and
2938 * sync_sched_expedited_done taking on the roles of the halves
2939 * of the ticket-lock word. Each task atomically increments
2940 * sync_sched_expedited_started upon entry, snapshotting the old value,
2941 * then attempts to stop all the CPUs. If this succeeds, then each
2942 * CPU will have executed a context switch, resulting in an RCU-sched
2943 * grace period. We are then done, so we use atomic_cmpxchg() to
2944 * update sync_sched_expedited_done to match our snapshot -- but
2945 * only if someone else has not already advanced past our snapshot.
2947 * On the other hand, if try_stop_cpus() fails, we check the value
2948 * of sync_sched_expedited_done. If it has advanced past our
2949 * initial snapshot, then someone else must have forced a grace period
2950 * some time after we took our snapshot. In this case, our work is
2951 * done for us, and we can simply return. Otherwise, we try again,
2952 * but keep our initial snapshot for purposes of checking for someone
2953 * doing our work for us.
2955 * If we fail too many times in a row, we fall back to synchronize_sched().
2957 void synchronize_sched_expedited(void)
2959 long firstsnap
, s
, snap
;
2961 struct rcu_state
*rsp
= &rcu_sched_state
;
2964 * If we are in danger of counter wrap, just do synchronize_sched().
2965 * By allowing sync_sched_expedited_started to advance no more than
2966 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2967 * that more than 3.5 billion CPUs would be required to force a
2968 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2969 * course be required on a 64-bit system.
2971 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
2972 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
2974 synchronize_sched();
2975 atomic_long_inc(&rsp
->expedited_wrap
);
2980 * Take a ticket. Note that atomic_inc_return() implies a
2981 * full memory barrier.
2983 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
2986 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2989 * Each pass through the following loop attempts to force a
2990 * context switch on each CPU.
2992 while (try_stop_cpus(cpu_online_mask
,
2993 synchronize_sched_expedited_cpu_stop
,
2996 atomic_long_inc(&rsp
->expedited_tryfail
);
2998 /* Check to see if someone else did our work for us. */
2999 s
= atomic_long_read(&rsp
->expedited_done
);
3000 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3001 /* ensure test happens before caller kfree */
3002 smp_mb__before_atomic(); /* ^^^ */
3003 atomic_long_inc(&rsp
->expedited_workdone1
);
3007 /* No joy, try again later. Or just synchronize_sched(). */
3008 if (trycount
++ < 10) {
3009 udelay(trycount
* num_online_cpus());
3011 wait_rcu_gp(call_rcu_sched
);
3012 atomic_long_inc(&rsp
->expedited_normal
);
3016 /* Recheck to see if someone else did our work for us. */
3017 s
= atomic_long_read(&rsp
->expedited_done
);
3018 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3019 /* ensure test happens before caller kfree */
3020 smp_mb__before_atomic(); /* ^^^ */
3021 atomic_long_inc(&rsp
->expedited_workdone2
);
3026 * Refetching sync_sched_expedited_started allows later
3027 * callers to piggyback on our grace period. We retry
3028 * after they started, so our grace period works for them,
3029 * and they started after our first try, so their grace
3030 * period works for us.
3033 snap
= atomic_long_read(&rsp
->expedited_start
);
3034 smp_mb(); /* ensure read is before try_stop_cpus(). */
3036 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3039 * Everyone up to our most recent fetch is covered by our grace
3040 * period. Update the counter, but only if our work is still
3041 * relevant -- which it won't be if someone who started later
3042 * than we did already did their update.
3045 atomic_long_inc(&rsp
->expedited_done_tries
);
3046 s
= atomic_long_read(&rsp
->expedited_done
);
3047 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3048 /* ensure test happens before caller kfree */
3049 smp_mb__before_atomic(); /* ^^^ */
3050 atomic_long_inc(&rsp
->expedited_done_lost
);
3053 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3054 atomic_long_inc(&rsp
->expedited_done_exit
);
3058 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3061 * Check to see if there is any immediate RCU-related work to be done
3062 * by the current CPU, for the specified type of RCU, returning 1 if so.
3063 * The checks are in order of increasing expense: checks that can be
3064 * carried out against CPU-local state are performed first. However,
3065 * we must check for CPU stalls first, else we might not get a chance.
3067 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3069 struct rcu_node
*rnp
= rdp
->mynode
;
3071 rdp
->n_rcu_pending
++;
3073 /* Check for CPU stalls, if enabled. */
3074 check_cpu_stall(rsp
, rdp
);
3076 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3077 if (rcu_nohz_full_cpu(rsp
))
3080 /* Is the RCU core waiting for a quiescent state from this CPU? */
3081 if (rcu_scheduler_fully_active
&&
3082 rdp
->qs_pending
&& !rdp
->passed_quiesce
) {
3083 rdp
->n_rp_qs_pending
++;
3084 } else if (rdp
->qs_pending
&& rdp
->passed_quiesce
) {
3085 rdp
->n_rp_report_qs
++;
3089 /* Does this CPU have callbacks ready to invoke? */
3090 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3091 rdp
->n_rp_cb_ready
++;
3095 /* Has RCU gone idle with this CPU needing another grace period? */
3096 if (cpu_needs_another_gp(rsp
, rdp
)) {
3097 rdp
->n_rp_cpu_needs_gp
++;
3101 /* Has another RCU grace period completed? */
3102 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3103 rdp
->n_rp_gp_completed
++;
3107 /* Has a new RCU grace period started? */
3108 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
3109 rdp
->n_rp_gp_started
++;
3113 /* Does this CPU need a deferred NOCB wakeup? */
3114 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3115 rdp
->n_rp_nocb_defer_wakeup
++;
3120 rdp
->n_rp_need_nothing
++;
3125 * Check to see if there is any immediate RCU-related work to be done
3126 * by the current CPU, returning 1 if so. This function is part of the
3127 * RCU implementation; it is -not- an exported member of the RCU API.
3129 static int rcu_pending(int cpu
)
3131 struct rcu_state
*rsp
;
3133 for_each_rcu_flavor(rsp
)
3134 if (__rcu_pending(rsp
, per_cpu_ptr(rsp
->rda
, cpu
)))
3140 * Return true if the specified CPU has any callback. If all_lazy is
3141 * non-NULL, store an indication of whether all callbacks are lazy.
3142 * (If there are no callbacks, all of them are deemed to be lazy.)
3144 static int __maybe_unused
rcu_cpu_has_callbacks(int cpu
, bool *all_lazy
)
3148 struct rcu_data
*rdp
;
3149 struct rcu_state
*rsp
;
3151 for_each_rcu_flavor(rsp
) {
3152 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3156 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3167 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3168 * the compiler is expected to optimize this away.
3170 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3171 int cpu
, unsigned long done
)
3173 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3174 atomic_read(&rsp
->barrier_cpu_count
), done
);
3178 * RCU callback function for _rcu_barrier(). If we are last, wake
3179 * up the task executing _rcu_barrier().
3181 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3183 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3184 struct rcu_state
*rsp
= rdp
->rsp
;
3186 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3187 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3188 complete(&rsp
->barrier_completion
);
3190 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3195 * Called with preemption disabled, and from cross-cpu IRQ context.
3197 static void rcu_barrier_func(void *type
)
3199 struct rcu_state
*rsp
= type
;
3200 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3202 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3203 atomic_inc(&rsp
->barrier_cpu_count
);
3204 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3208 * Orchestrate the specified type of RCU barrier, waiting for all
3209 * RCU callbacks of the specified type to complete.
3211 static void _rcu_barrier(struct rcu_state
*rsp
)
3214 struct rcu_data
*rdp
;
3215 unsigned long snap
= ACCESS_ONCE(rsp
->n_barrier_done
);
3216 unsigned long snap_done
;
3218 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3220 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3221 mutex_lock(&rsp
->barrier_mutex
);
3224 * Ensure that all prior references, including to ->n_barrier_done,
3225 * are ordered before the _rcu_barrier() machinery.
3227 smp_mb(); /* See above block comment. */
3230 * Recheck ->n_barrier_done to see if others did our work for us.
3231 * This means checking ->n_barrier_done for an even-to-odd-to-even
3232 * transition. The "if" expression below therefore rounds the old
3233 * value up to the next even number and adds two before comparing.
3235 snap_done
= rsp
->n_barrier_done
;
3236 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3239 * If the value in snap is odd, we needed to wait for the current
3240 * rcu_barrier() to complete, then wait for the next one, in other
3241 * words, we need the value of snap_done to be three larger than
3242 * the value of snap. On the other hand, if the value in snap is
3243 * even, we only had to wait for the next rcu_barrier() to complete,
3244 * in other words, we need the value of snap_done to be only two
3245 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3246 * this for us (thank you, Linus!).
3248 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3249 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3250 smp_mb(); /* caller's subsequent code after above check. */
3251 mutex_unlock(&rsp
->barrier_mutex
);
3256 * Increment ->n_barrier_done to avoid duplicate work. Use
3257 * ACCESS_ONCE() to prevent the compiler from speculating
3258 * the increment to precede the early-exit check.
3260 ACCESS_ONCE(rsp
->n_barrier_done
)++;
3261 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3262 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3263 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3266 * Initialize the count to one rather than to zero in order to
3267 * avoid a too-soon return to zero in case of a short grace period
3268 * (or preemption of this task). Exclude CPU-hotplug operations
3269 * to ensure that no offline CPU has callbacks queued.
3271 init_completion(&rsp
->barrier_completion
);
3272 atomic_set(&rsp
->barrier_cpu_count
, 1);
3276 * Force each CPU with callbacks to register a new callback.
3277 * When that callback is invoked, we will know that all of the
3278 * corresponding CPU's preceding callbacks have been invoked.
3280 for_each_possible_cpu(cpu
) {
3281 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3283 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3284 if (rcu_is_nocb_cpu(cpu
)) {
3285 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3286 rsp
->n_barrier_done
);
3287 atomic_inc(&rsp
->barrier_cpu_count
);
3288 __call_rcu(&rdp
->barrier_head
, rcu_barrier_callback
,
3290 } else if (ACCESS_ONCE(rdp
->qlen
)) {
3291 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3292 rsp
->n_barrier_done
);
3293 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3295 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3296 rsp
->n_barrier_done
);
3302 * Now that we have an rcu_barrier_callback() callback on each
3303 * CPU, and thus each counted, remove the initial count.
3305 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3306 complete(&rsp
->barrier_completion
);
3308 /* Increment ->n_barrier_done to prevent duplicate work. */
3309 smp_mb(); /* Keep increment after above mechanism. */
3310 ACCESS_ONCE(rsp
->n_barrier_done
)++;
3311 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3312 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3313 smp_mb(); /* Keep increment before caller's subsequent code. */
3315 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3316 wait_for_completion(&rsp
->barrier_completion
);
3318 /* Other rcu_barrier() invocations can now safely proceed. */
3319 mutex_unlock(&rsp
->barrier_mutex
);
3323 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3325 void rcu_barrier_bh(void)
3327 _rcu_barrier(&rcu_bh_state
);
3329 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3332 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3334 void rcu_barrier_sched(void)
3336 _rcu_barrier(&rcu_sched_state
);
3338 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3341 * Do boot-time initialization of a CPU's per-CPU RCU data.
3344 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3346 unsigned long flags
;
3347 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3348 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3350 /* Set up local state, ensuring consistent view of global state. */
3351 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3352 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3353 init_callback_list(rdp
);
3355 ACCESS_ONCE(rdp
->qlen
) = 0;
3356 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3357 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3358 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3361 rcu_boot_init_nocb_percpu_data(rdp
);
3362 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3366 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3367 * offline event can be happening at a given time. Note also that we
3368 * can accept some slop in the rsp->completed access due to the fact
3369 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3372 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3374 unsigned long flags
;
3376 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3377 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3379 /* Exclude new grace periods. */
3380 mutex_lock(&rsp
->onoff_mutex
);
3382 /* Set up local state, ensuring consistent view of global state. */
3383 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3384 rdp
->beenonline
= 1; /* We have now been online. */
3385 rdp
->qlen_last_fqs_check
= 0;
3386 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3387 rdp
->blimit
= blimit
;
3388 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3389 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3390 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3391 atomic_set(&rdp
->dynticks
->dynticks
,
3392 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3393 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3395 /* Add CPU to rcu_node bitmasks. */
3397 mask
= rdp
->grpmask
;
3399 /* Exclude any attempts to start a new GP on small systems. */
3400 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3401 rnp
->qsmaskinit
|= mask
;
3402 mask
= rnp
->grpmask
;
3403 if (rnp
== rdp
->mynode
) {
3405 * If there is a grace period in progress, we will
3406 * set up to wait for it next time we run the
3409 rdp
->gpnum
= rnp
->completed
;
3410 rdp
->completed
= rnp
->completed
;
3411 rdp
->passed_quiesce
= 0;
3412 rdp
->qs_pending
= 0;
3413 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3415 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
3417 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
3418 local_irq_restore(flags
);
3420 mutex_unlock(&rsp
->onoff_mutex
);
3423 static void rcu_prepare_cpu(int cpu
)
3425 struct rcu_state
*rsp
;
3427 for_each_rcu_flavor(rsp
)
3428 rcu_init_percpu_data(cpu
, rsp
);
3432 * Handle CPU online/offline notification events.
3434 static int rcu_cpu_notify(struct notifier_block
*self
,
3435 unsigned long action
, void *hcpu
)
3437 long cpu
= (long)hcpu
;
3438 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3439 struct rcu_node
*rnp
= rdp
->mynode
;
3440 struct rcu_state
*rsp
;
3442 trace_rcu_utilization(TPS("Start CPU hotplug"));
3444 case CPU_UP_PREPARE
:
3445 case CPU_UP_PREPARE_FROZEN
:
3446 rcu_prepare_cpu(cpu
);
3447 rcu_prepare_kthreads(cpu
);
3450 case CPU_DOWN_FAILED
:
3451 rcu_boost_kthread_setaffinity(rnp
, -1);
3453 case CPU_DOWN_PREPARE
:
3454 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3457 case CPU_DYING_FROZEN
:
3458 for_each_rcu_flavor(rsp
)
3459 rcu_cleanup_dying_cpu(rsp
);
3462 case CPU_DEAD_FROZEN
:
3463 case CPU_UP_CANCELED
:
3464 case CPU_UP_CANCELED_FROZEN
:
3465 for_each_rcu_flavor(rsp
)
3466 rcu_cleanup_dead_cpu(cpu
, rsp
);
3471 trace_rcu_utilization(TPS("End CPU hotplug"));
3475 static int rcu_pm_notify(struct notifier_block
*self
,
3476 unsigned long action
, void *hcpu
)
3479 case PM_HIBERNATION_PREPARE
:
3480 case PM_SUSPEND_PREPARE
:
3481 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3484 case PM_POST_HIBERNATION
:
3485 case PM_POST_SUSPEND
:
3495 * Spawn the kthread that handles this RCU flavor's grace periods.
3497 static int __init
rcu_spawn_gp_kthread(void)
3499 unsigned long flags
;
3500 struct rcu_node
*rnp
;
3501 struct rcu_state
*rsp
;
3502 struct task_struct
*t
;
3504 for_each_rcu_flavor(rsp
) {
3505 t
= kthread_run(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3507 rnp
= rcu_get_root(rsp
);
3508 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3509 rsp
->gp_kthread
= t
;
3510 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3511 rcu_spawn_nocb_kthreads(rsp
);
3515 early_initcall(rcu_spawn_gp_kthread
);
3518 * This function is invoked towards the end of the scheduler's initialization
3519 * process. Before this is called, the idle task might contain
3520 * RCU read-side critical sections (during which time, this idle
3521 * task is booting the system). After this function is called, the
3522 * idle tasks are prohibited from containing RCU read-side critical
3523 * sections. This function also enables RCU lockdep checking.
3525 void rcu_scheduler_starting(void)
3527 WARN_ON(num_online_cpus() != 1);
3528 WARN_ON(nr_context_switches() > 0);
3529 rcu_scheduler_active
= 1;
3533 * Compute the per-level fanout, either using the exact fanout specified
3534 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3536 #ifdef CONFIG_RCU_FANOUT_EXACT
3537 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3541 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3542 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3543 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3545 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3546 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3553 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3554 ccur
= rsp
->levelcnt
[i
];
3555 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3559 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3562 * Helper function for rcu_init() that initializes one rcu_state structure.
3564 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3565 struct rcu_data __percpu
*rda
)
3567 static char *buf
[] = { "rcu_node_0",
3570 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3571 static char *fqs
[] = { "rcu_node_fqs_0",
3574 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3575 static u8 fl_mask
= 0x1;
3579 struct rcu_node
*rnp
;
3581 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3583 /* Silence gcc 4.8 warning about array index out of range. */
3584 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3585 panic("rcu_init_one: rcu_num_lvls overflow");
3587 /* Initialize the level-tracking arrays. */
3589 for (i
= 0; i
< rcu_num_lvls
; i
++)
3590 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3591 for (i
= 1; i
< rcu_num_lvls
; i
++)
3592 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3593 rcu_init_levelspread(rsp
);
3594 rsp
->flavor_mask
= fl_mask
;
3597 /* Initialize the elements themselves, starting from the leaves. */
3599 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3600 cpustride
*= rsp
->levelspread
[i
];
3601 rnp
= rsp
->level
[i
];
3602 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3603 raw_spin_lock_init(&rnp
->lock
);
3604 lockdep_set_class_and_name(&rnp
->lock
,
3605 &rcu_node_class
[i
], buf
[i
]);
3606 raw_spin_lock_init(&rnp
->fqslock
);
3607 lockdep_set_class_and_name(&rnp
->fqslock
,
3608 &rcu_fqs_class
[i
], fqs
[i
]);
3609 rnp
->gpnum
= rsp
->gpnum
;
3610 rnp
->completed
= rsp
->completed
;
3612 rnp
->qsmaskinit
= 0;
3613 rnp
->grplo
= j
* cpustride
;
3614 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
3615 if (rnp
->grphi
>= nr_cpu_ids
)
3616 rnp
->grphi
= nr_cpu_ids
- 1;
3622 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
3623 rnp
->grpmask
= 1UL << rnp
->grpnum
;
3624 rnp
->parent
= rsp
->level
[i
- 1] +
3625 j
/ rsp
->levelspread
[i
- 1];
3628 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
3629 rcu_init_one_nocb(rnp
);
3634 init_waitqueue_head(&rsp
->gp_wq
);
3635 rnp
= rsp
->level
[rcu_num_lvls
- 1];
3636 for_each_possible_cpu(i
) {
3637 while (i
> rnp
->grphi
)
3639 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
3640 rcu_boot_init_percpu_data(i
, rsp
);
3642 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
3646 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3647 * replace the definitions in tree.h because those are needed to size
3648 * the ->node array in the rcu_state structure.
3650 static void __init
rcu_init_geometry(void)
3656 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
3659 * Initialize any unspecified boot parameters.
3660 * The default values of jiffies_till_first_fqs and
3661 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3662 * value, which is a function of HZ, then adding one for each
3663 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3665 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
3666 if (jiffies_till_first_fqs
== ULONG_MAX
)
3667 jiffies_till_first_fqs
= d
;
3668 if (jiffies_till_next_fqs
== ULONG_MAX
)
3669 jiffies_till_next_fqs
= d
;
3671 /* If the compile-time values are accurate, just leave. */
3672 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
3673 nr_cpu_ids
== NR_CPUS
)
3675 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3676 rcu_fanout_leaf
, nr_cpu_ids
);
3679 * Compute number of nodes that can be handled an rcu_node tree
3680 * with the given number of levels. Setting rcu_capacity[0] makes
3681 * some of the arithmetic easier.
3683 rcu_capacity
[0] = 1;
3684 rcu_capacity
[1] = rcu_fanout_leaf
;
3685 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
3686 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
3689 * The boot-time rcu_fanout_leaf parameter is only permitted
3690 * to increase the leaf-level fanout, not decrease it. Of course,
3691 * the leaf-level fanout cannot exceed the number of bits in
3692 * the rcu_node masks. Finally, the tree must be able to accommodate
3693 * the configured number of CPUs. Complain and fall back to the
3694 * compile-time values if these limits are exceeded.
3696 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
3697 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
3698 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
3703 /* Calculate the number of rcu_nodes at each level of the tree. */
3704 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
3705 if (n
<= rcu_capacity
[i
]) {
3706 for (j
= 0; j
<= i
; j
++)
3708 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
3710 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
3715 /* Calculate the total number of rcu_node structures. */
3717 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
3718 rcu_num_nodes
+= num_rcu_lvl
[i
];
3722 void __init
rcu_init(void)
3726 rcu_bootup_announce();
3727 rcu_init_geometry();
3728 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
3729 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
3730 __rcu_init_preempt();
3731 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
3734 * We don't need protection against CPU-hotplug here because
3735 * this is called early in boot, before either interrupts
3736 * or the scheduler are operational.
3738 cpu_notifier(rcu_cpu_notify
, 0);
3739 pm_notifier(rcu_pm_notify
, 0);
3740 for_each_online_cpu(cpu
)
3741 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
3744 #include "tree_plugin.h"