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.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
95 struct rcu_state sname##_state = { \
96 .level = { &sname##_state.node[0] }, \
97 .rda = &sname##_data, \
99 .fqs_state = RCU_GP_IDLE, \
100 .gpnum = 0UL - 300UL, \
101 .completed = 0UL - 300UL, \
102 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
103 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
104 .orphan_donetail = &sname##_state.orphan_donelist, \
105 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
106 .name = RCU_STATE_NAME(sname), \
110 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
111 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
113 static struct rcu_state
*const rcu_state_p
;
114 static struct rcu_data __percpu
*const rcu_data_p
;
115 LIST_HEAD(rcu_struct_flavors
);
117 /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
118 static int rcu_fanout_leaf
= CONFIG_RCU_FANOUT_LEAF
;
119 module_param(rcu_fanout_leaf
, int, 0444);
120 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
121 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
128 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
131 * The rcu_scheduler_active variable transitions from zero to one just
132 * before the first task is spawned. So when this variable is zero, RCU
133 * can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_sched() to a simple barrier(). When this variable
135 * is one, RCU must actually do all the hard work required to detect real
136 * grace periods. This variable is also used to suppress boot-time false
137 * positives from lockdep-RCU error checking.
139 int rcu_scheduler_active __read_mostly
;
140 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
143 * The rcu_scheduler_fully_active variable transitions from zero to one
144 * during the early_initcall() processing, which is after the scheduler
145 * is capable of creating new tasks. So RCU processing (for example,
146 * creating tasks for RCU priority boosting) must be delayed until after
147 * rcu_scheduler_fully_active transitions from zero to one. We also
148 * currently delay invocation of any RCU callbacks until after this point.
150 * It might later prove better for people registering RCU callbacks during
151 * early boot to take responsibility for these callbacks, but one step at
154 static int rcu_scheduler_fully_active __read_mostly
;
156 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
157 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
158 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
159 static void invoke_rcu_core(void);
160 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
162 /* rcuc/rcub kthread realtime priority */
163 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
164 module_param(kthread_prio
, int, 0644);
166 /* Delay in jiffies for grace-period initialization delays, debug only. */
167 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
168 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
169 module_param(gp_init_delay
, int, 0644);
170 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
171 static const int gp_init_delay
;
172 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
173 #define PER_RCU_NODE_PERIOD 10 /* Number of grace periods between delays. */
176 * Track the rcutorture test sequence number and the update version
177 * number within a given test. The rcutorture_testseq is incremented
178 * on every rcutorture module load and unload, so has an odd value
179 * when a test is running. The rcutorture_vernum is set to zero
180 * when rcutorture starts and is incremented on each rcutorture update.
181 * These variables enable correlating rcutorture output with the
182 * RCU tracing information.
184 unsigned long rcutorture_testseq
;
185 unsigned long rcutorture_vernum
;
188 * Compute the mask of online CPUs for the specified rcu_node structure.
189 * This will not be stable unless the rcu_node structure's ->lock is
190 * held, but the bit corresponding to the current CPU will be stable
193 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
195 return READ_ONCE(rnp
->qsmaskinitnext
);
199 * Return true if an RCU grace period is in progress. The READ_ONCE()s
200 * permit this function to be invoked without holding the root rcu_node
201 * structure's ->lock, but of course results can be subject to change.
203 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
205 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
209 * Note a quiescent state. Because we do not need to know
210 * how many quiescent states passed, just if there was at least
211 * one since the start of the grace period, this just sets a flag.
212 * The caller must have disabled preemption.
214 void rcu_sched_qs(void)
216 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
217 trace_rcu_grace_period(TPS("rcu_sched"),
218 __this_cpu_read(rcu_sched_data
.gpnum
),
220 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
226 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
227 trace_rcu_grace_period(TPS("rcu_bh"),
228 __this_cpu_read(rcu_bh_data
.gpnum
),
230 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
234 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
236 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
237 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
238 .dynticks
= ATOMIC_INIT(1),
239 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
240 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
241 .dynticks_idle
= ATOMIC_INIT(1),
242 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
245 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
246 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
249 * Let the RCU core know that this CPU has gone through the scheduler,
250 * which is a quiescent state. This is called when the need for a
251 * quiescent state is urgent, so we burn an atomic operation and full
252 * memory barriers to let the RCU core know about it, regardless of what
253 * this CPU might (or might not) do in the near future.
255 * We inform the RCU core by emulating a zero-duration dyntick-idle
256 * period, which we in turn do by incrementing the ->dynticks counter
259 static void rcu_momentary_dyntick_idle(void)
262 struct rcu_data
*rdp
;
263 struct rcu_dynticks
*rdtp
;
265 struct rcu_state
*rsp
;
267 local_irq_save(flags
);
270 * Yes, we can lose flag-setting operations. This is OK, because
271 * the flag will be set again after some delay.
273 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
274 raw_cpu_write(rcu_sched_qs_mask
, 0);
276 /* Find the flavor that needs a quiescent state. */
277 for_each_rcu_flavor(rsp
) {
278 rdp
= raw_cpu_ptr(rsp
->rda
);
279 if (!(resched_mask
& rsp
->flavor_mask
))
281 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
282 if (READ_ONCE(rdp
->mynode
->completed
) !=
283 READ_ONCE(rdp
->cond_resched_completed
))
287 * Pretend to be momentarily idle for the quiescent state.
288 * This allows the grace-period kthread to record the
289 * quiescent state, with no need for this CPU to do anything
292 rdtp
= this_cpu_ptr(&rcu_dynticks
);
293 smp_mb__before_atomic(); /* Earlier stuff before QS. */
294 atomic_add(2, &rdtp
->dynticks
); /* QS. */
295 smp_mb__after_atomic(); /* Later stuff after QS. */
298 local_irq_restore(flags
);
302 * Note a context switch. This is a quiescent state for RCU-sched,
303 * and requires special handling for preemptible RCU.
304 * The caller must have disabled preemption.
306 void rcu_note_context_switch(void)
308 trace_rcu_utilization(TPS("Start context switch"));
310 rcu_preempt_note_context_switch();
311 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
312 rcu_momentary_dyntick_idle();
313 trace_rcu_utilization(TPS("End context switch"));
315 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
318 * Register a quiescent state for all RCU flavors. If there is an
319 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
320 * dyntick-idle quiescent state visible to other CPUs (but only for those
321 * RCU flavors in desperate need of a quiescent state, which will normally
322 * be none of them). Either way, do a lightweight quiescent state for
325 void rcu_all_qs(void)
327 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
328 rcu_momentary_dyntick_idle();
329 this_cpu_inc(rcu_qs_ctr
);
331 EXPORT_SYMBOL_GPL(rcu_all_qs
);
333 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
334 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
335 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
337 module_param(blimit
, long, 0444);
338 module_param(qhimark
, long, 0444);
339 module_param(qlowmark
, long, 0444);
341 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
342 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
344 module_param(jiffies_till_first_fqs
, ulong
, 0644);
345 module_param(jiffies_till_next_fqs
, ulong
, 0644);
348 * How long the grace period must be before we start recruiting
349 * quiescent-state help from rcu_note_context_switch().
351 static ulong jiffies_till_sched_qs
= HZ
/ 20;
352 module_param(jiffies_till_sched_qs
, ulong
, 0644);
354 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
355 struct rcu_data
*rdp
);
356 static void force_qs_rnp(struct rcu_state
*rsp
,
357 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
358 unsigned long *maxj
),
359 bool *isidle
, unsigned long *maxj
);
360 static void force_quiescent_state(struct rcu_state
*rsp
);
361 static int rcu_pending(void);
364 * Return the number of RCU batches started thus far for debug & stats.
366 unsigned long rcu_batches_started(void)
368 return rcu_state_p
->gpnum
;
370 EXPORT_SYMBOL_GPL(rcu_batches_started
);
373 * Return the number of RCU-sched batches started thus far for debug & stats.
375 unsigned long rcu_batches_started_sched(void)
377 return rcu_sched_state
.gpnum
;
379 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
382 * Return the number of RCU BH batches started thus far for debug & stats.
384 unsigned long rcu_batches_started_bh(void)
386 return rcu_bh_state
.gpnum
;
388 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
391 * Return the number of RCU batches completed thus far for debug & stats.
393 unsigned long rcu_batches_completed(void)
395 return rcu_state_p
->completed
;
397 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
400 * Return the number of RCU-sched batches completed thus far for debug & stats.
402 unsigned long rcu_batches_completed_sched(void)
404 return rcu_sched_state
.completed
;
406 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
409 * Return the number of RCU BH batches completed thus far for debug & stats.
411 unsigned long rcu_batches_completed_bh(void)
413 return rcu_bh_state
.completed
;
415 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
418 * Force a quiescent state.
420 void rcu_force_quiescent_state(void)
422 force_quiescent_state(rcu_state_p
);
424 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
427 * Force a quiescent state for RCU BH.
429 void rcu_bh_force_quiescent_state(void)
431 force_quiescent_state(&rcu_bh_state
);
433 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
436 * Force a quiescent state for RCU-sched.
438 void rcu_sched_force_quiescent_state(void)
440 force_quiescent_state(&rcu_sched_state
);
442 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
445 * Show the state of the grace-period kthreads.
447 void show_rcu_gp_kthreads(void)
449 struct rcu_state
*rsp
;
451 for_each_rcu_flavor(rsp
) {
452 pr_info("%s: wait state: %d ->state: %#lx\n",
453 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
454 /* sched_show_task(rsp->gp_kthread); */
457 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
460 * Record the number of times rcutorture tests have been initiated and
461 * terminated. This information allows the debugfs tracing stats to be
462 * correlated to the rcutorture messages, even when the rcutorture module
463 * is being repeatedly loaded and unloaded. In other words, we cannot
464 * store this state in rcutorture itself.
466 void rcutorture_record_test_transition(void)
468 rcutorture_testseq
++;
469 rcutorture_vernum
= 0;
471 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
474 * Send along grace-period-related data for rcutorture diagnostics.
476 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
477 unsigned long *gpnum
, unsigned long *completed
)
479 struct rcu_state
*rsp
= NULL
;
488 case RCU_SCHED_FLAVOR
:
489 rsp
= &rcu_sched_state
;
495 *flags
= READ_ONCE(rsp
->gp_flags
);
496 *gpnum
= READ_ONCE(rsp
->gpnum
);
497 *completed
= READ_ONCE(rsp
->completed
);
504 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
507 * Record the number of writer passes through the current rcutorture test.
508 * This is also used to correlate debugfs tracing stats with the rcutorture
511 void rcutorture_record_progress(unsigned long vernum
)
515 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
518 * Does the CPU have callbacks ready to be invoked?
521 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
523 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
524 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
528 * Return the root node of the specified rcu_state structure.
530 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
532 return &rsp
->node
[0];
536 * Is there any need for future grace periods?
537 * Interrupts must be disabled. If the caller does not hold the root
538 * rnp_node structure's ->lock, the results are advisory only.
540 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
542 struct rcu_node
*rnp
= rcu_get_root(rsp
);
543 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
544 int *fp
= &rnp
->need_future_gp
[idx
];
546 return READ_ONCE(*fp
);
550 * Does the current CPU require a not-yet-started grace period?
551 * The caller must have disabled interrupts to prevent races with
552 * normal callback registry.
555 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
559 if (rcu_gp_in_progress(rsp
))
560 return 0; /* No, a grace period is already in progress. */
561 if (rcu_future_needs_gp(rsp
))
562 return 1; /* Yes, a no-CBs CPU needs one. */
563 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
564 return 0; /* No, this is a no-CBs (or offline) CPU. */
565 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
566 return 1; /* Yes, this CPU has newly registered callbacks. */
567 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
568 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
569 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
570 rdp
->nxtcompleted
[i
]))
571 return 1; /* Yes, CBs for future grace period. */
572 return 0; /* No grace period needed. */
576 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
578 * If the new value of the ->dynticks_nesting counter now is zero,
579 * we really have entered idle, and must do the appropriate accounting.
580 * The caller must have disabled interrupts.
582 static void rcu_eqs_enter_common(long long oldval
, bool user
)
584 struct rcu_state
*rsp
;
585 struct rcu_data
*rdp
;
586 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
588 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
589 if (!user
&& !is_idle_task(current
)) {
590 struct task_struct
*idle __maybe_unused
=
591 idle_task(smp_processor_id());
593 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
594 ftrace_dump(DUMP_ORIG
);
595 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
596 current
->pid
, current
->comm
,
597 idle
->pid
, idle
->comm
); /* must be idle task! */
599 for_each_rcu_flavor(rsp
) {
600 rdp
= this_cpu_ptr(rsp
->rda
);
601 do_nocb_deferred_wakeup(rdp
);
603 rcu_prepare_for_idle();
604 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
605 smp_mb__before_atomic(); /* See above. */
606 atomic_inc(&rdtp
->dynticks
);
607 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
608 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
609 rcu_dynticks_task_enter();
612 * It is illegal to enter an extended quiescent state while
613 * in an RCU read-side critical section.
615 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
616 "Illegal idle entry in RCU read-side critical section.");
617 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
618 "Illegal idle entry in RCU-bh read-side critical section.");
619 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
620 "Illegal idle entry in RCU-sched read-side critical section.");
624 * Enter an RCU extended quiescent state, which can be either the
625 * idle loop or adaptive-tickless usermode execution.
627 static void rcu_eqs_enter(bool user
)
630 struct rcu_dynticks
*rdtp
;
632 rdtp
= this_cpu_ptr(&rcu_dynticks
);
633 oldval
= rdtp
->dynticks_nesting
;
634 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
635 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
636 rdtp
->dynticks_nesting
= 0;
637 rcu_eqs_enter_common(oldval
, user
);
639 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
644 * rcu_idle_enter - inform RCU that current CPU is entering idle
646 * Enter idle mode, in other words, -leave- the mode in which RCU
647 * read-side critical sections can occur. (Though RCU read-side
648 * critical sections can occur in irq handlers in idle, a possibility
649 * handled by irq_enter() and irq_exit().)
651 * We crowbar the ->dynticks_nesting field to zero to allow for
652 * the possibility of usermode upcalls having messed up our count
653 * of interrupt nesting level during the prior busy period.
655 void rcu_idle_enter(void)
659 local_irq_save(flags
);
660 rcu_eqs_enter(false);
661 rcu_sysidle_enter(0);
662 local_irq_restore(flags
);
664 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
666 #ifdef CONFIG_RCU_USER_QS
668 * rcu_user_enter - inform RCU that we are resuming userspace.
670 * Enter RCU idle mode right before resuming userspace. No use of RCU
671 * is permitted between this call and rcu_user_exit(). This way the
672 * CPU doesn't need to maintain the tick for RCU maintenance purposes
673 * when the CPU runs in userspace.
675 void rcu_user_enter(void)
679 #endif /* CONFIG_RCU_USER_QS */
682 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
684 * Exit from an interrupt handler, which might possibly result in entering
685 * idle mode, in other words, leaving the mode in which read-side critical
686 * sections can occur.
688 * This code assumes that the idle loop never does anything that might
689 * result in unbalanced calls to irq_enter() and irq_exit(). If your
690 * architecture violates this assumption, RCU will give you what you
691 * deserve, good and hard. But very infrequently and irreproducibly.
693 * Use things like work queues to work around this limitation.
695 * You have been warned.
697 void rcu_irq_exit(void)
701 struct rcu_dynticks
*rdtp
;
703 local_irq_save(flags
);
704 rdtp
= this_cpu_ptr(&rcu_dynticks
);
705 oldval
= rdtp
->dynticks_nesting
;
706 rdtp
->dynticks_nesting
--;
707 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
708 if (rdtp
->dynticks_nesting
)
709 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
711 rcu_eqs_enter_common(oldval
, true);
712 rcu_sysidle_enter(1);
713 local_irq_restore(flags
);
717 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
719 * If the new value of the ->dynticks_nesting counter was previously zero,
720 * we really have exited idle, and must do the appropriate accounting.
721 * The caller must have disabled interrupts.
723 static void rcu_eqs_exit_common(long long oldval
, int user
)
725 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
727 rcu_dynticks_task_exit();
728 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
729 atomic_inc(&rdtp
->dynticks
);
730 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
731 smp_mb__after_atomic(); /* See above. */
732 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
733 rcu_cleanup_after_idle();
734 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
735 if (!user
&& !is_idle_task(current
)) {
736 struct task_struct
*idle __maybe_unused
=
737 idle_task(smp_processor_id());
739 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
740 oldval
, rdtp
->dynticks_nesting
);
741 ftrace_dump(DUMP_ORIG
);
742 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
743 current
->pid
, current
->comm
,
744 idle
->pid
, idle
->comm
); /* must be idle task! */
749 * Exit an RCU extended quiescent state, which can be either the
750 * idle loop or adaptive-tickless usermode execution.
752 static void rcu_eqs_exit(bool user
)
754 struct rcu_dynticks
*rdtp
;
757 rdtp
= this_cpu_ptr(&rcu_dynticks
);
758 oldval
= rdtp
->dynticks_nesting
;
759 WARN_ON_ONCE(oldval
< 0);
760 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
761 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
763 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
764 rcu_eqs_exit_common(oldval
, user
);
769 * rcu_idle_exit - inform RCU that current CPU is leaving idle
771 * Exit idle mode, in other words, -enter- the mode in which RCU
772 * read-side critical sections can occur.
774 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
775 * allow for the possibility of usermode upcalls messing up our count
776 * of interrupt nesting level during the busy period that is just
779 void rcu_idle_exit(void)
783 local_irq_save(flags
);
786 local_irq_restore(flags
);
788 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
790 #ifdef CONFIG_RCU_USER_QS
792 * rcu_user_exit - inform RCU that we are exiting userspace.
794 * Exit RCU idle mode while entering the kernel because it can
795 * run a RCU read side critical section anytime.
797 void rcu_user_exit(void)
801 #endif /* CONFIG_RCU_USER_QS */
804 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
806 * Enter an interrupt handler, which might possibly result in exiting
807 * idle mode, in other words, entering the mode in which read-side critical
808 * sections can occur.
810 * Note that the Linux kernel is fully capable of entering an interrupt
811 * handler that it never exits, for example when doing upcalls to
812 * user mode! This code assumes that the idle loop never does upcalls to
813 * user mode. If your architecture does do upcalls from the idle loop (or
814 * does anything else that results in unbalanced calls to the irq_enter()
815 * and irq_exit() functions), RCU will give you what you deserve, good
816 * and hard. But very infrequently and irreproducibly.
818 * Use things like work queues to work around this limitation.
820 * You have been warned.
822 void rcu_irq_enter(void)
825 struct rcu_dynticks
*rdtp
;
828 local_irq_save(flags
);
829 rdtp
= this_cpu_ptr(&rcu_dynticks
);
830 oldval
= rdtp
->dynticks_nesting
;
831 rdtp
->dynticks_nesting
++;
832 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
834 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
836 rcu_eqs_exit_common(oldval
, true);
838 local_irq_restore(flags
);
842 * rcu_nmi_enter - inform RCU of entry to NMI context
844 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
845 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
846 * that the CPU is active. This implementation permits nested NMIs, as
847 * long as the nesting level does not overflow an int. (You will probably
848 * run out of stack space first.)
850 void rcu_nmi_enter(void)
852 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
855 /* Complain about underflow. */
856 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
859 * If idle from RCU viewpoint, atomically increment ->dynticks
860 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
861 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
862 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
863 * to be in the outermost NMI handler that interrupted an RCU-idle
864 * period (observation due to Andy Lutomirski).
866 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
867 smp_mb__before_atomic(); /* Force delay from prior write. */
868 atomic_inc(&rdtp
->dynticks
);
869 /* atomic_inc() before later RCU read-side crit sects */
870 smp_mb__after_atomic(); /* See above. */
871 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
874 rdtp
->dynticks_nmi_nesting
+= incby
;
879 * rcu_nmi_exit - inform RCU of exit from NMI context
881 * If we are returning from the outermost NMI handler that interrupted an
882 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
883 * to let the RCU grace-period handling know that the CPU is back to
886 void rcu_nmi_exit(void)
888 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
891 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
892 * (We are exiting an NMI handler, so RCU better be paying attention
895 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
896 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
899 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
900 * leave it in non-RCU-idle state.
902 if (rdtp
->dynticks_nmi_nesting
!= 1) {
903 rdtp
->dynticks_nmi_nesting
-= 2;
907 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
908 rdtp
->dynticks_nmi_nesting
= 0;
909 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
910 smp_mb__before_atomic(); /* See above. */
911 atomic_inc(&rdtp
->dynticks
);
912 smp_mb__after_atomic(); /* Force delay to next write. */
913 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
917 * __rcu_is_watching - are RCU read-side critical sections safe?
919 * Return true if RCU is watching the running CPU, which means that
920 * this CPU can safely enter RCU read-side critical sections. Unlike
921 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
922 * least disabled preemption.
924 bool notrace
__rcu_is_watching(void)
926 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
930 * rcu_is_watching - see if RCU thinks that the current CPU is idle
932 * If the current CPU is in its idle loop and is neither in an interrupt
933 * or NMI handler, return true.
935 bool notrace
rcu_is_watching(void)
940 ret
= __rcu_is_watching();
944 EXPORT_SYMBOL_GPL(rcu_is_watching
);
946 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
949 * Is the current CPU online? Disable preemption to avoid false positives
950 * that could otherwise happen due to the current CPU number being sampled,
951 * this task being preempted, its old CPU being taken offline, resuming
952 * on some other CPU, then determining that its old CPU is now offline.
953 * It is OK to use RCU on an offline processor during initial boot, hence
954 * the check for rcu_scheduler_fully_active. Note also that it is OK
955 * for a CPU coming online to use RCU for one jiffy prior to marking itself
956 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
957 * offline to continue to use RCU for one jiffy after marking itself
958 * offline in the cpu_online_mask. This leniency is necessary given the
959 * non-atomic nature of the online and offline processing, for example,
960 * the fact that a CPU enters the scheduler after completing the CPU_DYING
963 * This is also why RCU internally marks CPUs online during the
964 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
966 * Disable checking if in an NMI handler because we cannot safely report
967 * errors from NMI handlers anyway.
969 bool rcu_lockdep_current_cpu_online(void)
971 struct rcu_data
*rdp
;
972 struct rcu_node
*rnp
;
978 rdp
= this_cpu_ptr(&rcu_sched_data
);
980 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
981 !rcu_scheduler_fully_active
;
985 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
987 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
990 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
992 * If the current CPU is idle or running at a first-level (not nested)
993 * interrupt from idle, return true. The caller must have at least
994 * disabled preemption.
996 static int rcu_is_cpu_rrupt_from_idle(void)
998 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1002 * Snapshot the specified CPU's dynticks counter so that we can later
1003 * credit them with an implicit quiescent state. Return 1 if this CPU
1004 * is in dynticks idle mode, which is an extended quiescent state.
1006 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1007 bool *isidle
, unsigned long *maxj
)
1009 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1010 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1011 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1012 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1015 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1016 rdp
->mynode
->gpnum
))
1017 WRITE_ONCE(rdp
->gpwrap
, true);
1023 * Return true if the specified CPU has passed through a quiescent
1024 * state by virtue of being in or having passed through an dynticks
1025 * idle state since the last call to dyntick_save_progress_counter()
1026 * for this same CPU, or by virtue of having been offline.
1028 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1029 bool *isidle
, unsigned long *maxj
)
1035 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1036 snap
= (unsigned int)rdp
->dynticks_snap
;
1039 * If the CPU passed through or entered a dynticks idle phase with
1040 * no active irq/NMI handlers, then we can safely pretend that the CPU
1041 * already acknowledged the request to pass through a quiescent
1042 * state. Either way, that CPU cannot possibly be in an RCU
1043 * read-side critical section that started before the beginning
1044 * of the current RCU grace period.
1046 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1047 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1048 rdp
->dynticks_fqs
++;
1053 * Check for the CPU being offline, but only if the grace period
1054 * is old enough. We don't need to worry about the CPU changing
1055 * state: If we see it offline even once, it has been through a
1058 * The reason for insisting that the grace period be at least
1059 * one jiffy old is that CPUs that are not quite online and that
1060 * have just gone offline can still execute RCU read-side critical
1063 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1064 return 0; /* Grace period is not old enough. */
1066 if (cpu_is_offline(rdp
->cpu
)) {
1067 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1073 * A CPU running for an extended time within the kernel can
1074 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1075 * even context-switching back and forth between a pair of
1076 * in-kernel CPU-bound tasks cannot advance grace periods.
1077 * So if the grace period is old enough, make the CPU pay attention.
1078 * Note that the unsynchronized assignments to the per-CPU
1079 * rcu_sched_qs_mask variable are safe. Yes, setting of
1080 * bits can be lost, but they will be set again on the next
1081 * force-quiescent-state pass. So lost bit sets do not result
1082 * in incorrect behavior, merely in a grace period lasting
1083 * a few jiffies longer than it might otherwise. Because
1084 * there are at most four threads involved, and because the
1085 * updates are only once every few jiffies, the probability of
1086 * lossage (and thus of slight grace-period extension) is
1089 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1090 * is set too high, we override with half of the RCU CPU stall
1093 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1094 if (ULONG_CMP_GE(jiffies
,
1095 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1096 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1097 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1098 WRITE_ONCE(rdp
->cond_resched_completed
,
1099 READ_ONCE(rdp
->mynode
->completed
));
1100 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1102 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1103 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1104 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1105 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1106 /* Time to beat on that CPU again! */
1107 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1108 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1115 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1117 unsigned long j
= jiffies
;
1121 smp_wmb(); /* Record start time before stall time. */
1122 j1
= rcu_jiffies_till_stall_check();
1123 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1124 rsp
->jiffies_resched
= j
+ j1
/ 2;
1125 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1129 * Complain about starvation of grace-period kthread.
1131 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1137 gpa
= READ_ONCE(rsp
->gp_activity
);
1138 if (j
- gpa
> 2 * HZ
)
1139 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x\n",
1141 rsp
->gpnum
, rsp
->completed
, rsp
->gp_flags
);
1145 * Dump stacks of all tasks running on stalled CPUs.
1147 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1150 unsigned long flags
;
1151 struct rcu_node
*rnp
;
1153 rcu_for_each_leaf_node(rsp
, rnp
) {
1154 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1155 if (rnp
->qsmask
!= 0) {
1156 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1157 if (rnp
->qsmask
& (1UL << cpu
))
1158 dump_cpu_task(rnp
->grplo
+ cpu
);
1160 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1164 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1168 unsigned long flags
;
1172 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1175 /* Only let one CPU complain about others per time interval. */
1177 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1178 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1179 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1180 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1183 WRITE_ONCE(rsp
->jiffies_stall
,
1184 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1185 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1188 * OK, time to rat on our buddy...
1189 * See Documentation/RCU/stallwarn.txt for info on how to debug
1190 * RCU CPU stall warnings.
1192 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1194 print_cpu_stall_info_begin();
1195 rcu_for_each_leaf_node(rsp
, rnp
) {
1196 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1197 ndetected
+= rcu_print_task_stall(rnp
);
1198 if (rnp
->qsmask
!= 0) {
1199 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1200 if (rnp
->qsmask
& (1UL << cpu
)) {
1201 print_cpu_stall_info(rsp
,
1206 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1209 print_cpu_stall_info_end();
1210 for_each_possible_cpu(cpu
)
1211 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1212 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1213 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1214 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1216 rcu_dump_cpu_stacks(rsp
);
1218 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1219 READ_ONCE(rsp
->completed
) == gpnum
) {
1220 pr_err("INFO: Stall ended before state dump start\n");
1223 gpa
= READ_ONCE(rsp
->gp_activity
);
1224 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1225 rsp
->name
, j
- gpa
, j
, gpa
,
1226 jiffies_till_next_fqs
,
1227 rcu_get_root(rsp
)->qsmask
);
1228 /* In this case, the current CPU might be at fault. */
1229 sched_show_task(current
);
1233 /* Complain about tasks blocking the grace period. */
1234 rcu_print_detail_task_stall(rsp
);
1236 rcu_check_gp_kthread_starvation(rsp
);
1238 force_quiescent_state(rsp
); /* Kick them all. */
1241 static void print_cpu_stall(struct rcu_state
*rsp
)
1244 unsigned long flags
;
1245 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1249 * OK, time to rat on ourselves...
1250 * See Documentation/RCU/stallwarn.txt for info on how to debug
1251 * RCU CPU stall warnings.
1253 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1254 print_cpu_stall_info_begin();
1255 print_cpu_stall_info(rsp
, smp_processor_id());
1256 print_cpu_stall_info_end();
1257 for_each_possible_cpu(cpu
)
1258 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1259 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1260 jiffies
- rsp
->gp_start
,
1261 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1263 rcu_check_gp_kthread_starvation(rsp
);
1265 rcu_dump_cpu_stacks(rsp
);
1267 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1268 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1269 WRITE_ONCE(rsp
->jiffies_stall
,
1270 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1271 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1274 * Attempt to revive the RCU machinery by forcing a context switch.
1276 * A context switch would normally allow the RCU state machine to make
1277 * progress and it could be we're stuck in kernel space without context
1278 * switches for an entirely unreasonable amount of time.
1280 resched_cpu(smp_processor_id());
1283 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1285 unsigned long completed
;
1286 unsigned long gpnum
;
1290 struct rcu_node
*rnp
;
1292 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1297 * Lots of memory barriers to reject false positives.
1299 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1300 * then rsp->gp_start, and finally rsp->completed. These values
1301 * are updated in the opposite order with memory barriers (or
1302 * equivalent) during grace-period initialization and cleanup.
1303 * Now, a false positive can occur if we get an new value of
1304 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1305 * the memory barriers, the only way that this can happen is if one
1306 * grace period ends and another starts between these two fetches.
1307 * Detect this by comparing rsp->completed with the previous fetch
1310 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1311 * and rsp->gp_start suffice to forestall false positives.
1313 gpnum
= READ_ONCE(rsp
->gpnum
);
1314 smp_rmb(); /* Pick up ->gpnum first... */
1315 js
= READ_ONCE(rsp
->jiffies_stall
);
1316 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1317 gps
= READ_ONCE(rsp
->gp_start
);
1318 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1319 completed
= READ_ONCE(rsp
->completed
);
1320 if (ULONG_CMP_GE(completed
, gpnum
) ||
1321 ULONG_CMP_LT(j
, js
) ||
1322 ULONG_CMP_GE(gps
, js
))
1323 return; /* No stall or GP completed since entering function. */
1325 if (rcu_gp_in_progress(rsp
) &&
1326 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1328 /* We haven't checked in, so go dump stack. */
1329 print_cpu_stall(rsp
);
1331 } else if (rcu_gp_in_progress(rsp
) &&
1332 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1334 /* They had a few time units to dump stack, so complain. */
1335 print_other_cpu_stall(rsp
, gpnum
);
1340 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1342 * Set the stall-warning timeout way off into the future, thus preventing
1343 * any RCU CPU stall-warning messages from appearing in the current set of
1344 * RCU grace periods.
1346 * The caller must disable hard irqs.
1348 void rcu_cpu_stall_reset(void)
1350 struct rcu_state
*rsp
;
1352 for_each_rcu_flavor(rsp
)
1353 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1357 * Initialize the specified rcu_data structure's default callback list
1358 * to empty. The default callback list is the one that is not used by
1359 * no-callbacks CPUs.
1361 static void init_default_callback_list(struct rcu_data
*rdp
)
1365 rdp
->nxtlist
= NULL
;
1366 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1367 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1371 * Initialize the specified rcu_data structure's callback list to empty.
1373 static void init_callback_list(struct rcu_data
*rdp
)
1375 if (init_nocb_callback_list(rdp
))
1377 init_default_callback_list(rdp
);
1381 * Determine the value that ->completed will have at the end of the
1382 * next subsequent grace period. This is used to tag callbacks so that
1383 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1384 * been dyntick-idle for an extended period with callbacks under the
1385 * influence of RCU_FAST_NO_HZ.
1387 * The caller must hold rnp->lock with interrupts disabled.
1389 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1390 struct rcu_node
*rnp
)
1393 * If RCU is idle, we just wait for the next grace period.
1394 * But we can only be sure that RCU is idle if we are looking
1395 * at the root rcu_node structure -- otherwise, a new grace
1396 * period might have started, but just not yet gotten around
1397 * to initializing the current non-root rcu_node structure.
1399 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1400 return rnp
->completed
+ 1;
1403 * Otherwise, wait for a possible partial grace period and
1404 * then the subsequent full grace period.
1406 return rnp
->completed
+ 2;
1410 * Trace-event helper function for rcu_start_future_gp() and
1411 * rcu_nocb_wait_gp().
1413 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1414 unsigned long c
, const char *s
)
1416 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1417 rnp
->completed
, c
, rnp
->level
,
1418 rnp
->grplo
, rnp
->grphi
, s
);
1422 * Start some future grace period, as needed to handle newly arrived
1423 * callbacks. The required future grace periods are recorded in each
1424 * rcu_node structure's ->need_future_gp field. Returns true if there
1425 * is reason to awaken the grace-period kthread.
1427 * The caller must hold the specified rcu_node structure's ->lock.
1429 static bool __maybe_unused
1430 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1431 unsigned long *c_out
)
1436 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1439 * Pick up grace-period number for new callbacks. If this
1440 * grace period is already marked as needed, return to the caller.
1442 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1443 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1444 if (rnp
->need_future_gp
[c
& 0x1]) {
1445 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1450 * If either this rcu_node structure or the root rcu_node structure
1451 * believe that a grace period is in progress, then we must wait
1452 * for the one following, which is in "c". Because our request
1453 * will be noticed at the end of the current grace period, we don't
1454 * need to explicitly start one. We only do the lockless check
1455 * of rnp_root's fields if the current rcu_node structure thinks
1456 * there is no grace period in flight, and because we hold rnp->lock,
1457 * the only possible change is when rnp_root's two fields are
1458 * equal, in which case rnp_root->gpnum might be concurrently
1459 * incremented. But that is OK, as it will just result in our
1460 * doing some extra useless work.
1462 if (rnp
->gpnum
!= rnp
->completed
||
1463 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1464 rnp
->need_future_gp
[c
& 0x1]++;
1465 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1470 * There might be no grace period in progress. If we don't already
1471 * hold it, acquire the root rcu_node structure's lock in order to
1472 * start one (if needed).
1474 if (rnp
!= rnp_root
) {
1475 raw_spin_lock(&rnp_root
->lock
);
1476 smp_mb__after_unlock_lock();
1480 * Get a new grace-period number. If there really is no grace
1481 * period in progress, it will be smaller than the one we obtained
1482 * earlier. Adjust callbacks as needed. Note that even no-CBs
1483 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1485 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1486 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1487 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1488 rdp
->nxtcompleted
[i
] = c
;
1491 * If the needed for the required grace period is already
1492 * recorded, trace and leave.
1494 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1495 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1499 /* Record the need for the future grace period. */
1500 rnp_root
->need_future_gp
[c
& 0x1]++;
1502 /* If a grace period is not already in progress, start one. */
1503 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1504 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1506 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1507 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1510 if (rnp
!= rnp_root
)
1511 raw_spin_unlock(&rnp_root
->lock
);
1519 * Clean up any old requests for the just-ended grace period. Also return
1520 * whether any additional grace periods have been requested. Also invoke
1521 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1522 * waiting for this grace period to complete.
1524 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1526 int c
= rnp
->completed
;
1528 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1530 rcu_nocb_gp_cleanup(rsp
, rnp
);
1531 rnp
->need_future_gp
[c
& 0x1] = 0;
1532 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1533 trace_rcu_future_gp(rnp
, rdp
, c
,
1534 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1539 * Awaken the grace-period kthread for the specified flavor of RCU.
1540 * Don't do a self-awaken, and don't bother awakening when there is
1541 * nothing for the grace-period kthread to do (as in several CPUs
1542 * raced to awaken, and we lost), and finally don't try to awaken
1543 * a kthread that has not yet been created.
1545 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1547 if (current
== rsp
->gp_kthread
||
1548 !READ_ONCE(rsp
->gp_flags
) ||
1551 wake_up(&rsp
->gp_wq
);
1555 * If there is room, assign a ->completed number to any callbacks on
1556 * this CPU that have not already been assigned. Also accelerate any
1557 * callbacks that were previously assigned a ->completed number that has
1558 * since proven to be too conservative, which can happen if callbacks get
1559 * assigned a ->completed number while RCU is idle, but with reference to
1560 * a non-root rcu_node structure. This function is idempotent, so it does
1561 * not hurt to call it repeatedly. Returns an flag saying that we should
1562 * awaken the RCU grace-period kthread.
1564 * The caller must hold rnp->lock with interrupts disabled.
1566 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1567 struct rcu_data
*rdp
)
1573 /* If the CPU has no callbacks, nothing to do. */
1574 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1578 * Starting from the sublist containing the callbacks most
1579 * recently assigned a ->completed number and working down, find the
1580 * first sublist that is not assignable to an upcoming grace period.
1581 * Such a sublist has something in it (first two tests) and has
1582 * a ->completed number assigned that will complete sooner than
1583 * the ->completed number for newly arrived callbacks (last test).
1585 * The key point is that any later sublist can be assigned the
1586 * same ->completed number as the newly arrived callbacks, which
1587 * means that the callbacks in any of these later sublist can be
1588 * grouped into a single sublist, whether or not they have already
1589 * been assigned a ->completed number.
1591 c
= rcu_cbs_completed(rsp
, rnp
);
1592 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1593 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1594 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1598 * If there are no sublist for unassigned callbacks, leave.
1599 * At the same time, advance "i" one sublist, so that "i" will
1600 * index into the sublist where all the remaining callbacks should
1603 if (++i
>= RCU_NEXT_TAIL
)
1607 * Assign all subsequent callbacks' ->completed number to the next
1608 * full grace period and group them all in the sublist initially
1611 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1612 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1613 rdp
->nxtcompleted
[i
] = c
;
1615 /* Record any needed additional grace periods. */
1616 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1618 /* Trace depending on how much we were able to accelerate. */
1619 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1620 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1622 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1627 * Move any callbacks whose grace period has completed to the
1628 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1629 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1630 * sublist. This function is idempotent, so it does not hurt to
1631 * invoke it repeatedly. As long as it is not invoked -too- often...
1632 * Returns true if the RCU grace-period kthread needs to be awakened.
1634 * The caller must hold rnp->lock with interrupts disabled.
1636 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1637 struct rcu_data
*rdp
)
1641 /* If the CPU has no callbacks, nothing to do. */
1642 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1646 * Find all callbacks whose ->completed numbers indicate that they
1647 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1649 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1650 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1652 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1654 /* Clean up any sublist tail pointers that were misordered above. */
1655 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1656 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1658 /* Copy down callbacks to fill in empty sublists. */
1659 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1660 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1662 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1663 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1666 /* Classify any remaining callbacks. */
1667 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1671 * Update CPU-local rcu_data state to record the beginnings and ends of
1672 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1673 * structure corresponding to the current CPU, and must have irqs disabled.
1674 * Returns true if the grace-period kthread needs to be awakened.
1676 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1677 struct rcu_data
*rdp
)
1681 /* Handle the ends of any preceding grace periods first. */
1682 if (rdp
->completed
== rnp
->completed
&&
1683 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1685 /* No grace period end, so just accelerate recent callbacks. */
1686 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1690 /* Advance callbacks. */
1691 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1693 /* Remember that we saw this grace-period completion. */
1694 rdp
->completed
= rnp
->completed
;
1695 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1698 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1700 * If the current grace period is waiting for this CPU,
1701 * set up to detect a quiescent state, otherwise don't
1702 * go looking for one.
1704 rdp
->gpnum
= rnp
->gpnum
;
1705 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1706 rdp
->passed_quiesce
= 0;
1707 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1708 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1709 zero_cpu_stall_ticks(rdp
);
1710 WRITE_ONCE(rdp
->gpwrap
, false);
1715 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1717 unsigned long flags
;
1719 struct rcu_node
*rnp
;
1721 local_irq_save(flags
);
1723 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1724 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1725 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1726 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1727 local_irq_restore(flags
);
1730 smp_mb__after_unlock_lock();
1731 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1732 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1734 rcu_gp_kthread_wake(rsp
);
1738 * Initialize a new grace period. Return 0 if no grace period required.
1740 static int rcu_gp_init(struct rcu_state
*rsp
)
1742 unsigned long oldmask
;
1743 struct rcu_data
*rdp
;
1744 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1746 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1747 raw_spin_lock_irq(&rnp
->lock
);
1748 smp_mb__after_unlock_lock();
1749 if (!READ_ONCE(rsp
->gp_flags
)) {
1750 /* Spurious wakeup, tell caller to go back to sleep. */
1751 raw_spin_unlock_irq(&rnp
->lock
);
1754 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1756 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1758 * Grace period already in progress, don't start another.
1759 * Not supposed to be able to happen.
1761 raw_spin_unlock_irq(&rnp
->lock
);
1765 /* Advance to a new grace period and initialize state. */
1766 record_gp_stall_check_time(rsp
);
1767 /* Record GP times before starting GP, hence smp_store_release(). */
1768 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1769 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1770 raw_spin_unlock_irq(&rnp
->lock
);
1773 * Apply per-leaf buffered online and offline operations to the
1774 * rcu_node tree. Note that this new grace period need not wait
1775 * for subsequent online CPUs, and that quiescent-state forcing
1776 * will handle subsequent offline CPUs.
1778 rcu_for_each_leaf_node(rsp
, rnp
) {
1779 raw_spin_lock_irq(&rnp
->lock
);
1780 smp_mb__after_unlock_lock();
1781 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1782 !rnp
->wait_blkd_tasks
) {
1783 /* Nothing to do on this leaf rcu_node structure. */
1784 raw_spin_unlock_irq(&rnp
->lock
);
1788 /* Record old state, apply changes to ->qsmaskinit field. */
1789 oldmask
= rnp
->qsmaskinit
;
1790 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1792 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1793 if (!oldmask
!= !rnp
->qsmaskinit
) {
1794 if (!oldmask
) /* First online CPU for this rcu_node. */
1795 rcu_init_new_rnp(rnp
);
1796 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1797 rnp
->wait_blkd_tasks
= true;
1798 else /* Last offline CPU and can propagate. */
1799 rcu_cleanup_dead_rnp(rnp
);
1803 * If all waited-on tasks from prior grace period are
1804 * done, and if all this rcu_node structure's CPUs are
1805 * still offline, propagate up the rcu_node tree and
1806 * clear ->wait_blkd_tasks. Otherwise, if one of this
1807 * rcu_node structure's CPUs has since come back online,
1808 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1809 * checks for this, so just call it unconditionally).
1811 if (rnp
->wait_blkd_tasks
&&
1812 (!rcu_preempt_has_tasks(rnp
) ||
1814 rnp
->wait_blkd_tasks
= false;
1815 rcu_cleanup_dead_rnp(rnp
);
1818 raw_spin_unlock_irq(&rnp
->lock
);
1822 * Set the quiescent-state-needed bits in all the rcu_node
1823 * structures for all currently online CPUs in breadth-first order,
1824 * starting from the root rcu_node structure, relying on the layout
1825 * of the tree within the rsp->node[] array. Note that other CPUs
1826 * will access only the leaves of the hierarchy, thus seeing that no
1827 * grace period is in progress, at least until the corresponding
1828 * leaf node has been initialized. In addition, we have excluded
1829 * CPU-hotplug operations.
1831 * The grace period cannot complete until the initialization
1832 * process finishes, because this kthread handles both.
1834 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1835 raw_spin_lock_irq(&rnp
->lock
);
1836 smp_mb__after_unlock_lock();
1837 rdp
= this_cpu_ptr(rsp
->rda
);
1838 rcu_preempt_check_blocked_tasks(rnp
);
1839 rnp
->qsmask
= rnp
->qsmaskinit
;
1840 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1841 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1842 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1843 if (rnp
== rdp
->mynode
)
1844 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1845 rcu_preempt_boost_start_gp(rnp
);
1846 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1847 rnp
->level
, rnp
->grplo
,
1848 rnp
->grphi
, rnp
->qsmask
);
1849 raw_spin_unlock_irq(&rnp
->lock
);
1850 cond_resched_rcu_qs();
1851 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1852 if (gp_init_delay
> 0 &&
1853 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
)))
1854 schedule_timeout_uninterruptible(gp_init_delay
);
1861 * Do one round of quiescent-state forcing.
1863 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1865 int fqs_state
= fqs_state_in
;
1866 bool isidle
= false;
1868 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1870 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1872 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1873 /* Collect dyntick-idle snapshots. */
1874 if (is_sysidle_rcu_state(rsp
)) {
1876 maxj
= jiffies
- ULONG_MAX
/ 4;
1878 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1880 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1881 fqs_state
= RCU_FORCE_QS
;
1883 /* Handle dyntick-idle and offline CPUs. */
1885 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1887 /* Clear flag to prevent immediate re-entry. */
1888 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1889 raw_spin_lock_irq(&rnp
->lock
);
1890 smp_mb__after_unlock_lock();
1891 WRITE_ONCE(rsp
->gp_flags
,
1892 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
1893 raw_spin_unlock_irq(&rnp
->lock
);
1899 * Clean up after the old grace period.
1901 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1903 unsigned long gp_duration
;
1904 bool needgp
= false;
1906 struct rcu_data
*rdp
;
1907 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1909 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1910 raw_spin_lock_irq(&rnp
->lock
);
1911 smp_mb__after_unlock_lock();
1912 gp_duration
= jiffies
- rsp
->gp_start
;
1913 if (gp_duration
> rsp
->gp_max
)
1914 rsp
->gp_max
= gp_duration
;
1917 * We know the grace period is complete, but to everyone else
1918 * it appears to still be ongoing. But it is also the case
1919 * that to everyone else it looks like there is nothing that
1920 * they can do to advance the grace period. It is therefore
1921 * safe for us to drop the lock in order to mark the grace
1922 * period as completed in all of the rcu_node structures.
1924 raw_spin_unlock_irq(&rnp
->lock
);
1927 * Propagate new ->completed value to rcu_node structures so
1928 * that other CPUs don't have to wait until the start of the next
1929 * grace period to process their callbacks. This also avoids
1930 * some nasty RCU grace-period initialization races by forcing
1931 * the end of the current grace period to be completely recorded in
1932 * all of the rcu_node structures before the beginning of the next
1933 * grace period is recorded in any of the rcu_node structures.
1935 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1936 raw_spin_lock_irq(&rnp
->lock
);
1937 smp_mb__after_unlock_lock();
1938 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
1939 WARN_ON_ONCE(rnp
->qsmask
);
1940 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
1941 rdp
= this_cpu_ptr(rsp
->rda
);
1942 if (rnp
== rdp
->mynode
)
1943 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1944 /* smp_mb() provided by prior unlock-lock pair. */
1945 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1946 raw_spin_unlock_irq(&rnp
->lock
);
1947 cond_resched_rcu_qs();
1948 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1950 rnp
= rcu_get_root(rsp
);
1951 raw_spin_lock_irq(&rnp
->lock
);
1952 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1953 rcu_nocb_gp_set(rnp
, nocb
);
1955 /* Declare grace period done. */
1956 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
1957 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1958 rsp
->fqs_state
= RCU_GP_IDLE
;
1959 rdp
= this_cpu_ptr(rsp
->rda
);
1960 /* Advance CBs to reduce false positives below. */
1961 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1962 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1963 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
1964 trace_rcu_grace_period(rsp
->name
,
1965 READ_ONCE(rsp
->gpnum
),
1968 raw_spin_unlock_irq(&rnp
->lock
);
1972 * Body of kthread that handles grace periods.
1974 static int __noreturn
rcu_gp_kthread(void *arg
)
1980 struct rcu_state
*rsp
= arg
;
1981 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1983 rcu_bind_gp_kthread();
1986 /* Handle grace-period start. */
1988 trace_rcu_grace_period(rsp
->name
,
1989 READ_ONCE(rsp
->gpnum
),
1991 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
1992 wait_event_interruptible(rsp
->gp_wq
,
1993 READ_ONCE(rsp
->gp_flags
) &
1995 /* Locking provides needed memory barrier. */
1996 if (rcu_gp_init(rsp
))
1998 cond_resched_rcu_qs();
1999 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2000 WARN_ON(signal_pending(current
));
2001 trace_rcu_grace_period(rsp
->name
,
2002 READ_ONCE(rsp
->gpnum
),
2006 /* Handle quiescent-state forcing. */
2007 fqs_state
= RCU_SAVE_DYNTICK
;
2008 j
= jiffies_till_first_fqs
;
2011 jiffies_till_first_fqs
= HZ
;
2016 rsp
->jiffies_force_qs
= jiffies
+ j
;
2017 trace_rcu_grace_period(rsp
->name
,
2018 READ_ONCE(rsp
->gpnum
),
2020 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2021 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
2022 ((gf
= READ_ONCE(rsp
->gp_flags
)) &
2024 (!READ_ONCE(rnp
->qsmask
) &&
2025 !rcu_preempt_blocked_readers_cgp(rnp
)),
2027 /* Locking provides needed memory barriers. */
2028 /* If grace period done, leave loop. */
2029 if (!READ_ONCE(rnp
->qsmask
) &&
2030 !rcu_preempt_blocked_readers_cgp(rnp
))
2032 /* If time for quiescent-state forcing, do it. */
2033 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2034 (gf
& RCU_GP_FLAG_FQS
)) {
2035 trace_rcu_grace_period(rsp
->name
,
2036 READ_ONCE(rsp
->gpnum
),
2038 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
2039 trace_rcu_grace_period(rsp
->name
,
2040 READ_ONCE(rsp
->gpnum
),
2042 cond_resched_rcu_qs();
2043 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2045 /* Deal with stray signal. */
2046 cond_resched_rcu_qs();
2047 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2048 WARN_ON(signal_pending(current
));
2049 trace_rcu_grace_period(rsp
->name
,
2050 READ_ONCE(rsp
->gpnum
),
2053 j
= jiffies_till_next_fqs
;
2056 jiffies_till_next_fqs
= HZ
;
2059 jiffies_till_next_fqs
= 1;
2063 /* Handle grace-period end. */
2064 rcu_gp_cleanup(rsp
);
2069 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2070 * in preparation for detecting the next grace period. The caller must hold
2071 * the root node's ->lock and hard irqs must be disabled.
2073 * Note that it is legal for a dying CPU (which is marked as offline) to
2074 * invoke this function. This can happen when the dying CPU reports its
2077 * Returns true if the grace-period kthread must be awakened.
2080 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2081 struct rcu_data
*rdp
)
2083 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2085 * Either we have not yet spawned the grace-period
2086 * task, this CPU does not need another grace period,
2087 * or a grace period is already in progress.
2088 * Either way, don't start a new grace period.
2092 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2093 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2097 * We can't do wakeups while holding the rnp->lock, as that
2098 * could cause possible deadlocks with the rq->lock. Defer
2099 * the wakeup to our caller.
2105 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2106 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2107 * is invoked indirectly from rcu_advance_cbs(), which would result in
2108 * endless recursion -- or would do so if it wasn't for the self-deadlock
2109 * that is encountered beforehand.
2111 * Returns true if the grace-period kthread needs to be awakened.
2113 static bool rcu_start_gp(struct rcu_state
*rsp
)
2115 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2116 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2120 * If there is no grace period in progress right now, any
2121 * callbacks we have up to this point will be satisfied by the
2122 * next grace period. Also, advancing the callbacks reduces the
2123 * probability of false positives from cpu_needs_another_gp()
2124 * resulting in pointless grace periods. So, advance callbacks
2125 * then start the grace period!
2127 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2128 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2133 * Report a full set of quiescent states to the specified rcu_state
2134 * data structure. This involves cleaning up after the prior grace
2135 * period and letting rcu_start_gp() start up the next grace period
2136 * if one is needed. Note that the caller must hold rnp->lock, which
2137 * is released before return.
2139 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2140 __releases(rcu_get_root(rsp
)->lock
)
2142 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2143 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2144 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2145 rcu_gp_kthread_wake(rsp
);
2149 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2150 * Allows quiescent states for a group of CPUs to be reported at one go
2151 * to the specified rcu_node structure, though all the CPUs in the group
2152 * must be represented by the same rcu_node structure (which need not be a
2153 * leaf rcu_node structure, though it often will be). The gps parameter
2154 * is the grace-period snapshot, which means that the quiescent states
2155 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2156 * must be held upon entry, and it is released before return.
2159 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2160 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2161 __releases(rnp
->lock
)
2163 unsigned long oldmask
= 0;
2164 struct rcu_node
*rnp_c
;
2166 /* Walk up the rcu_node hierarchy. */
2168 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2171 * Our bit has already been cleared, or the
2172 * relevant grace period is already over, so done.
2174 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2177 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2178 rnp
->qsmask
&= ~mask
;
2179 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2180 mask
, rnp
->qsmask
, rnp
->level
,
2181 rnp
->grplo
, rnp
->grphi
,
2183 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2185 /* Other bits still set at this level, so done. */
2186 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2189 mask
= rnp
->grpmask
;
2190 if (rnp
->parent
== NULL
) {
2192 /* No more levels. Exit loop holding root lock. */
2196 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2199 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2200 smp_mb__after_unlock_lock();
2201 oldmask
= rnp_c
->qsmask
;
2205 * Get here if we are the last CPU to pass through a quiescent
2206 * state for this grace period. Invoke rcu_report_qs_rsp()
2207 * to clean up and start the next grace period if one is needed.
2209 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2213 * Record a quiescent state for all tasks that were previously queued
2214 * on the specified rcu_node structure and that were blocking the current
2215 * RCU grace period. The caller must hold the specified rnp->lock with
2216 * irqs disabled, and this lock is released upon return, but irqs remain
2219 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2220 struct rcu_node
*rnp
, unsigned long flags
)
2221 __releases(rnp
->lock
)
2225 struct rcu_node
*rnp_p
;
2227 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2228 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2229 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2230 return; /* Still need more quiescent states! */
2233 rnp_p
= rnp
->parent
;
2234 if (rnp_p
== NULL
) {
2236 * Only one rcu_node structure in the tree, so don't
2237 * try to report up to its nonexistent parent!
2239 rcu_report_qs_rsp(rsp
, flags
);
2243 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2245 mask
= rnp
->grpmask
;
2246 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2247 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
2248 smp_mb__after_unlock_lock();
2249 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2253 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2254 * structure. This must be either called from the specified CPU, or
2255 * called when the specified CPU is known to be offline (and when it is
2256 * also known that no other CPU is concurrently trying to help the offline
2257 * CPU). The lastcomp argument is used to make sure we are still in the
2258 * grace period of interest. We don't want to end the current grace period
2259 * based on quiescent states detected in an earlier grace period!
2262 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2264 unsigned long flags
;
2267 struct rcu_node
*rnp
;
2270 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2271 smp_mb__after_unlock_lock();
2272 if ((rdp
->passed_quiesce
== 0 &&
2273 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2274 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2278 * The grace period in which this quiescent state was
2279 * recorded has ended, so don't report it upwards.
2280 * We will instead need a new quiescent state that lies
2281 * within the current grace period.
2283 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2284 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2285 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2288 mask
= rdp
->grpmask
;
2289 if ((rnp
->qsmask
& mask
) == 0) {
2290 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2292 rdp
->qs_pending
= 0;
2295 * This GP can't end until cpu checks in, so all of our
2296 * callbacks can be processed during the next GP.
2298 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2300 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2301 /* ^^^ Released rnp->lock */
2303 rcu_gp_kthread_wake(rsp
);
2308 * Check to see if there is a new grace period of which this CPU
2309 * is not yet aware, and if so, set up local rcu_data state for it.
2310 * Otherwise, see if this CPU has just passed through its first
2311 * quiescent state for this grace period, and record that fact if so.
2314 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2316 /* Check for grace-period ends and beginnings. */
2317 note_gp_changes(rsp
, rdp
);
2320 * Does this CPU still need to do its part for current grace period?
2321 * If no, return and let the other CPUs do their part as well.
2323 if (!rdp
->qs_pending
)
2327 * Was there a quiescent state since the beginning of the grace
2328 * period? If no, then exit and wait for the next call.
2330 if (!rdp
->passed_quiesce
&&
2331 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2335 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2338 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2341 #ifdef CONFIG_HOTPLUG_CPU
2344 * Send the specified CPU's RCU callbacks to the orphanage. The
2345 * specified CPU must be offline, and the caller must hold the
2349 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2350 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2352 /* No-CBs CPUs do not have orphanable callbacks. */
2353 if (rcu_is_nocb_cpu(rdp
->cpu
))
2357 * Orphan the callbacks. First adjust the counts. This is safe
2358 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2359 * cannot be running now. Thus no memory barrier is required.
2361 if (rdp
->nxtlist
!= NULL
) {
2362 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2363 rsp
->qlen
+= rdp
->qlen
;
2364 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2366 WRITE_ONCE(rdp
->qlen
, 0);
2370 * Next, move those callbacks still needing a grace period to
2371 * the orphanage, where some other CPU will pick them up.
2372 * Some of the callbacks might have gone partway through a grace
2373 * period, but that is too bad. They get to start over because we
2374 * cannot assume that grace periods are synchronized across CPUs.
2375 * We don't bother updating the ->nxttail[] array yet, instead
2376 * we just reset the whole thing later on.
2378 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2379 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2380 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2381 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2385 * Then move the ready-to-invoke callbacks to the orphanage,
2386 * where some other CPU will pick them up. These will not be
2387 * required to pass though another grace period: They are done.
2389 if (rdp
->nxtlist
!= NULL
) {
2390 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2391 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2395 * Finally, initialize the rcu_data structure's list to empty and
2396 * disallow further callbacks on this CPU.
2398 init_callback_list(rdp
);
2399 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2403 * Adopt the RCU callbacks from the specified rcu_state structure's
2404 * orphanage. The caller must hold the ->orphan_lock.
2406 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2409 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2411 /* No-CBs CPUs are handled specially. */
2412 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2415 /* Do the accounting first. */
2416 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2417 rdp
->qlen
+= rsp
->qlen
;
2418 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2419 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2420 rcu_idle_count_callbacks_posted();
2425 * We do not need a memory barrier here because the only way we
2426 * can get here if there is an rcu_barrier() in flight is if
2427 * we are the task doing the rcu_barrier().
2430 /* First adopt the ready-to-invoke callbacks. */
2431 if (rsp
->orphan_donelist
!= NULL
) {
2432 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2433 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2434 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2435 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2436 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2437 rsp
->orphan_donelist
= NULL
;
2438 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2441 /* And then adopt the callbacks that still need a grace period. */
2442 if (rsp
->orphan_nxtlist
!= NULL
) {
2443 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2444 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2445 rsp
->orphan_nxtlist
= NULL
;
2446 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2451 * Trace the fact that this CPU is going offline.
2453 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2455 RCU_TRACE(unsigned long mask
);
2456 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2457 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2459 RCU_TRACE(mask
= rdp
->grpmask
);
2460 trace_rcu_grace_period(rsp
->name
,
2461 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2466 * All CPUs for the specified rcu_node structure have gone offline,
2467 * and all tasks that were preempted within an RCU read-side critical
2468 * section while running on one of those CPUs have since exited their RCU
2469 * read-side critical section. Some other CPU is reporting this fact with
2470 * the specified rcu_node structure's ->lock held and interrupts disabled.
2471 * This function therefore goes up the tree of rcu_node structures,
2472 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2473 * the leaf rcu_node structure's ->qsmaskinit field has already been
2476 * This function does check that the specified rcu_node structure has
2477 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2478 * prematurely. That said, invoking it after the fact will cost you
2479 * a needless lock acquisition. So once it has done its work, don't
2482 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2485 struct rcu_node
*rnp
= rnp_leaf
;
2487 if (rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2490 mask
= rnp
->grpmask
;
2494 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2495 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2496 rnp
->qsmaskinit
&= ~mask
;
2497 rnp
->qsmask
&= ~mask
;
2498 if (rnp
->qsmaskinit
) {
2499 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2502 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2507 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2508 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2511 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2513 unsigned long flags
;
2515 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2516 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2518 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2519 mask
= rdp
->grpmask
;
2520 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2521 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2522 rnp
->qsmaskinitnext
&= ~mask
;
2523 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2527 * The CPU has been completely removed, and some other CPU is reporting
2528 * this fact from process context. Do the remainder of the cleanup,
2529 * including orphaning the outgoing CPU's RCU callbacks, and also
2530 * adopting them. There can only be one CPU hotplug operation at a time,
2531 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2533 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2535 unsigned long flags
;
2536 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2537 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2539 /* Adjust any no-longer-needed kthreads. */
2540 rcu_boost_kthread_setaffinity(rnp
, -1);
2542 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2543 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2544 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2545 rcu_adopt_orphan_cbs(rsp
, flags
);
2546 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2548 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2549 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2550 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2553 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2555 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2559 static void __maybe_unused
rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2563 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2567 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2571 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2574 * Invoke any RCU callbacks that have made it to the end of their grace
2575 * period. Thottle as specified by rdp->blimit.
2577 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2579 unsigned long flags
;
2580 struct rcu_head
*next
, *list
, **tail
;
2581 long bl
, count
, count_lazy
;
2584 /* If no callbacks are ready, just return. */
2585 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2586 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2587 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2588 need_resched(), is_idle_task(current
),
2589 rcu_is_callbacks_kthread());
2594 * Extract the list of ready callbacks, disabling to prevent
2595 * races with call_rcu() from interrupt handlers.
2597 local_irq_save(flags
);
2598 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2600 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2601 list
= rdp
->nxtlist
;
2602 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2603 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2604 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2605 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2606 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2607 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2608 local_irq_restore(flags
);
2610 /* Invoke callbacks. */
2611 count
= count_lazy
= 0;
2615 debug_rcu_head_unqueue(list
);
2616 if (__rcu_reclaim(rsp
->name
, list
))
2619 /* Stop only if limit reached and CPU has something to do. */
2620 if (++count
>= bl
&&
2622 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2626 local_irq_save(flags
);
2627 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2628 is_idle_task(current
),
2629 rcu_is_callbacks_kthread());
2631 /* Update count, and requeue any remaining callbacks. */
2633 *tail
= rdp
->nxtlist
;
2634 rdp
->nxtlist
= list
;
2635 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2636 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2637 rdp
->nxttail
[i
] = tail
;
2641 smp_mb(); /* List handling before counting for rcu_barrier(). */
2642 rdp
->qlen_lazy
-= count_lazy
;
2643 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2644 rdp
->n_cbs_invoked
+= count
;
2646 /* Reinstate batch limit if we have worked down the excess. */
2647 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2648 rdp
->blimit
= blimit
;
2650 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2651 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2652 rdp
->qlen_last_fqs_check
= 0;
2653 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2654 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2655 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2656 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2658 local_irq_restore(flags
);
2660 /* Re-invoke RCU core processing if there are callbacks remaining. */
2661 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2666 * Check to see if this CPU is in a non-context-switch quiescent state
2667 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2668 * Also schedule RCU core processing.
2670 * This function must be called from hardirq context. It is normally
2671 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2672 * false, there is no point in invoking rcu_check_callbacks().
2674 void rcu_check_callbacks(int user
)
2676 trace_rcu_utilization(TPS("Start scheduler-tick"));
2677 increment_cpu_stall_ticks();
2678 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2681 * Get here if this CPU took its interrupt from user
2682 * mode or from the idle loop, and if this is not a
2683 * nested interrupt. In this case, the CPU is in
2684 * a quiescent state, so note it.
2686 * No memory barrier is required here because both
2687 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2688 * variables that other CPUs neither access nor modify,
2689 * at least not while the corresponding CPU is online.
2695 } else if (!in_softirq()) {
2698 * Get here if this CPU did not take its interrupt from
2699 * softirq, in other words, if it is not interrupting
2700 * a rcu_bh read-side critical section. This is an _bh
2701 * critical section, so note it.
2706 rcu_preempt_check_callbacks();
2710 rcu_note_voluntary_context_switch(current
);
2711 trace_rcu_utilization(TPS("End scheduler-tick"));
2715 * Scan the leaf rcu_node structures, processing dyntick state for any that
2716 * have not yet encountered a quiescent state, using the function specified.
2717 * Also initiate boosting for any threads blocked on the root rcu_node.
2719 * The caller must have suppressed start of new grace periods.
2721 static void force_qs_rnp(struct rcu_state
*rsp
,
2722 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2723 unsigned long *maxj
),
2724 bool *isidle
, unsigned long *maxj
)
2728 unsigned long flags
;
2730 struct rcu_node
*rnp
;
2732 rcu_for_each_leaf_node(rsp
, rnp
) {
2733 cond_resched_rcu_qs();
2735 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2736 smp_mb__after_unlock_lock();
2737 if (!rcu_gp_in_progress(rsp
)) {
2738 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2741 if (rnp
->qsmask
== 0) {
2742 if (rcu_state_p
== &rcu_sched_state
||
2743 rsp
!= rcu_state_p
||
2744 rcu_preempt_blocked_readers_cgp(rnp
)) {
2746 * No point in scanning bits because they
2747 * are all zero. But we might need to
2748 * priority-boost blocked readers.
2750 rcu_initiate_boost(rnp
, flags
);
2751 /* rcu_initiate_boost() releases rnp->lock */
2755 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2757 * Race between grace-period
2758 * initialization and task exiting RCU
2759 * read-side critical section: Report.
2761 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2762 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2768 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2769 if ((rnp
->qsmask
& bit
) != 0) {
2770 if ((rnp
->qsmaskinit
& bit
) == 0)
2771 *isidle
= false; /* Pending hotplug. */
2772 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2777 /* Idle/offline CPUs, report (releases rnp->lock. */
2778 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2780 /* Nothing to do here, so just drop the lock. */
2781 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2787 * Force quiescent states on reluctant CPUs, and also detect which
2788 * CPUs are in dyntick-idle mode.
2790 static void force_quiescent_state(struct rcu_state
*rsp
)
2792 unsigned long flags
;
2794 struct rcu_node
*rnp
;
2795 struct rcu_node
*rnp_old
= NULL
;
2797 /* Funnel through hierarchy to reduce memory contention. */
2798 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2799 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2800 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2801 !raw_spin_trylock(&rnp
->fqslock
);
2802 if (rnp_old
!= NULL
)
2803 raw_spin_unlock(&rnp_old
->fqslock
);
2805 rsp
->n_force_qs_lh
++;
2810 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2812 /* Reached the root of the rcu_node tree, acquire lock. */
2813 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2814 smp_mb__after_unlock_lock();
2815 raw_spin_unlock(&rnp_old
->fqslock
);
2816 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2817 rsp
->n_force_qs_lh
++;
2818 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2819 return; /* Someone beat us to it. */
2821 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2822 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2823 rcu_gp_kthread_wake(rsp
);
2827 * This does the RCU core processing work for the specified rcu_state
2828 * and rcu_data structures. This may be called only from the CPU to
2829 * whom the rdp belongs.
2832 __rcu_process_callbacks(struct rcu_state
*rsp
)
2834 unsigned long flags
;
2836 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2838 WARN_ON_ONCE(rdp
->beenonline
== 0);
2840 /* Update RCU state based on any recent quiescent states. */
2841 rcu_check_quiescent_state(rsp
, rdp
);
2843 /* Does this CPU require a not-yet-started grace period? */
2844 local_irq_save(flags
);
2845 if (cpu_needs_another_gp(rsp
, rdp
)) {
2846 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2847 needwake
= rcu_start_gp(rsp
);
2848 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2850 rcu_gp_kthread_wake(rsp
);
2852 local_irq_restore(flags
);
2855 /* If there are callbacks ready, invoke them. */
2856 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2857 invoke_rcu_callbacks(rsp
, rdp
);
2859 /* Do any needed deferred wakeups of rcuo kthreads. */
2860 do_nocb_deferred_wakeup(rdp
);
2864 * Do RCU core processing for the current CPU.
2866 static void rcu_process_callbacks(struct softirq_action
*unused
)
2868 struct rcu_state
*rsp
;
2870 if (cpu_is_offline(smp_processor_id()))
2872 trace_rcu_utilization(TPS("Start RCU core"));
2873 for_each_rcu_flavor(rsp
)
2874 __rcu_process_callbacks(rsp
);
2875 trace_rcu_utilization(TPS("End RCU core"));
2879 * Schedule RCU callback invocation. If the specified type of RCU
2880 * does not support RCU priority boosting, just do a direct call,
2881 * otherwise wake up the per-CPU kernel kthread. Note that because we
2882 * are running on the current CPU with softirqs disabled, the
2883 * rcu_cpu_kthread_task cannot disappear out from under us.
2885 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2887 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2889 if (likely(!rsp
->boost
)) {
2890 rcu_do_batch(rsp
, rdp
);
2893 invoke_rcu_callbacks_kthread();
2896 static void invoke_rcu_core(void)
2898 if (cpu_online(smp_processor_id()))
2899 raise_softirq(RCU_SOFTIRQ
);
2903 * Handle any core-RCU processing required by a call_rcu() invocation.
2905 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2906 struct rcu_head
*head
, unsigned long flags
)
2911 * If called from an extended quiescent state, invoke the RCU
2912 * core in order to force a re-evaluation of RCU's idleness.
2914 if (!rcu_is_watching())
2917 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2918 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2922 * Force the grace period if too many callbacks or too long waiting.
2923 * Enforce hysteresis, and don't invoke force_quiescent_state()
2924 * if some other CPU has recently done so. Also, don't bother
2925 * invoking force_quiescent_state() if the newly enqueued callback
2926 * is the only one waiting for a grace period to complete.
2928 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2930 /* Are we ignoring a completed grace period? */
2931 note_gp_changes(rsp
, rdp
);
2933 /* Start a new grace period if one not already started. */
2934 if (!rcu_gp_in_progress(rsp
)) {
2935 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2937 raw_spin_lock(&rnp_root
->lock
);
2938 smp_mb__after_unlock_lock();
2939 needwake
= rcu_start_gp(rsp
);
2940 raw_spin_unlock(&rnp_root
->lock
);
2942 rcu_gp_kthread_wake(rsp
);
2944 /* Give the grace period a kick. */
2945 rdp
->blimit
= LONG_MAX
;
2946 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2947 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2948 force_quiescent_state(rsp
);
2949 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2950 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2956 * RCU callback function to leak a callback.
2958 static void rcu_leak_callback(struct rcu_head
*rhp
)
2963 * Helper function for call_rcu() and friends. The cpu argument will
2964 * normally be -1, indicating "currently running CPU". It may specify
2965 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2966 * is expected to specify a CPU.
2969 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
2970 struct rcu_state
*rsp
, int cpu
, bool lazy
)
2972 unsigned long flags
;
2973 struct rcu_data
*rdp
;
2975 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
2976 if (debug_rcu_head_queue(head
)) {
2977 /* Probable double call_rcu(), so leak the callback. */
2978 WRITE_ONCE(head
->func
, rcu_leak_callback
);
2979 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2986 * Opportunistically note grace-period endings and beginnings.
2987 * Note that we might see a beginning right after we see an
2988 * end, but never vice versa, since this CPU has to pass through
2989 * a quiescent state betweentimes.
2991 local_irq_save(flags
);
2992 rdp
= this_cpu_ptr(rsp
->rda
);
2994 /* Add the callback to our list. */
2995 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
2999 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3000 if (likely(rdp
->mynode
)) {
3001 /* Post-boot, so this should be for a no-CBs CPU. */
3002 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3003 WARN_ON_ONCE(offline
);
3004 /* Offline CPU, _call_rcu() illegal, leak callback. */
3005 local_irq_restore(flags
);
3009 * Very early boot, before rcu_init(). Initialize if needed
3010 * and then drop through to queue the callback.
3013 WARN_ON_ONCE(!rcu_is_watching());
3014 if (!likely(rdp
->nxtlist
))
3015 init_default_callback_list(rdp
);
3017 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3021 rcu_idle_count_callbacks_posted();
3022 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3023 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3024 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3026 if (__is_kfree_rcu_offset((unsigned long)func
))
3027 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3028 rdp
->qlen_lazy
, rdp
->qlen
);
3030 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3032 /* Go handle any RCU core processing required. */
3033 __call_rcu_core(rsp
, rdp
, head
, flags
);
3034 local_irq_restore(flags
);
3038 * Queue an RCU-sched callback for invocation after a grace period.
3040 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3042 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3044 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3047 * Queue an RCU callback for invocation after a quicker grace period.
3049 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3051 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3053 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3056 * Queue an RCU callback for lazy invocation after a grace period.
3057 * This will likely be later named something like "call_rcu_lazy()",
3058 * but this change will require some way of tagging the lazy RCU
3059 * callbacks in the list of pending callbacks. Until then, this
3060 * function may only be called from __kfree_rcu().
3062 void kfree_call_rcu(struct rcu_head
*head
,
3063 void (*func
)(struct rcu_head
*rcu
))
3065 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3067 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3070 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3071 * any blocking grace-period wait automatically implies a grace period
3072 * if there is only one CPU online at any point time during execution
3073 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3074 * occasionally incorrectly indicate that there are multiple CPUs online
3075 * when there was in fact only one the whole time, as this just adds
3076 * some overhead: RCU still operates correctly.
3078 static inline int rcu_blocking_is_gp(void)
3082 might_sleep(); /* Check for RCU read-side critical section. */
3084 ret
= num_online_cpus() <= 1;
3090 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3092 * Control will return to the caller some time after a full rcu-sched
3093 * grace period has elapsed, in other words after all currently executing
3094 * rcu-sched read-side critical sections have completed. These read-side
3095 * critical sections are delimited by rcu_read_lock_sched() and
3096 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3097 * local_irq_disable(), and so on may be used in place of
3098 * rcu_read_lock_sched().
3100 * This means that all preempt_disable code sequences, including NMI and
3101 * non-threaded hardware-interrupt handlers, in progress on entry will
3102 * have completed before this primitive returns. However, this does not
3103 * guarantee that softirq handlers will have completed, since in some
3104 * kernels, these handlers can run in process context, and can block.
3106 * Note that this guarantee implies further memory-ordering guarantees.
3107 * On systems with more than one CPU, when synchronize_sched() returns,
3108 * each CPU is guaranteed to have executed a full memory barrier since the
3109 * end of its last RCU-sched read-side critical section whose beginning
3110 * preceded the call to synchronize_sched(). In addition, each CPU having
3111 * an RCU read-side critical section that extends beyond the return from
3112 * synchronize_sched() is guaranteed to have executed a full memory barrier
3113 * after the beginning of synchronize_sched() and before the beginning of
3114 * that RCU read-side critical section. Note that these guarantees include
3115 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3116 * that are executing in the kernel.
3118 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3119 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3120 * to have executed a full memory barrier during the execution of
3121 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3122 * again only if the system has more than one CPU).
3124 * This primitive provides the guarantees made by the (now removed)
3125 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3126 * guarantees that rcu_read_lock() sections will have completed.
3127 * In "classic RCU", these two guarantees happen to be one and
3128 * the same, but can differ in realtime RCU implementations.
3130 void synchronize_sched(void)
3132 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3133 !lock_is_held(&rcu_lock_map
) &&
3134 !lock_is_held(&rcu_sched_lock_map
),
3135 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3136 if (rcu_blocking_is_gp())
3138 if (rcu_gp_is_expedited())
3139 synchronize_sched_expedited();
3141 wait_rcu_gp(call_rcu_sched
);
3143 EXPORT_SYMBOL_GPL(synchronize_sched
);
3146 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3148 * Control will return to the caller some time after a full rcu_bh grace
3149 * period has elapsed, in other words after all currently executing rcu_bh
3150 * read-side critical sections have completed. RCU read-side critical
3151 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3152 * and may be nested.
3154 * See the description of synchronize_sched() for more detailed information
3155 * on memory ordering guarantees.
3157 void synchronize_rcu_bh(void)
3159 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3160 !lock_is_held(&rcu_lock_map
) &&
3161 !lock_is_held(&rcu_sched_lock_map
),
3162 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3163 if (rcu_blocking_is_gp())
3165 if (rcu_gp_is_expedited())
3166 synchronize_rcu_bh_expedited();
3168 wait_rcu_gp(call_rcu_bh
);
3170 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3173 * get_state_synchronize_rcu - Snapshot current RCU state
3175 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3176 * to determine whether or not a full grace period has elapsed in the
3179 unsigned long get_state_synchronize_rcu(void)
3182 * Any prior manipulation of RCU-protected data must happen
3183 * before the load from ->gpnum.
3188 * Make sure this load happens before the purportedly
3189 * time-consuming work between get_state_synchronize_rcu()
3190 * and cond_synchronize_rcu().
3192 return smp_load_acquire(&rcu_state_p
->gpnum
);
3194 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3197 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3199 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3201 * If a full RCU grace period has elapsed since the earlier call to
3202 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3203 * synchronize_rcu() to wait for a full grace period.
3205 * Yes, this function does not take counter wrap into account. But
3206 * counter wrap is harmless. If the counter wraps, we have waited for
3207 * more than 2 billion grace periods (and way more on a 64-bit system!),
3208 * so waiting for one additional grace period should be just fine.
3210 void cond_synchronize_rcu(unsigned long oldstate
)
3212 unsigned long newstate
;
3215 * Ensure that this load happens before any RCU-destructive
3216 * actions the caller might carry out after we return.
3218 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3219 if (ULONG_CMP_GE(oldstate
, newstate
))
3222 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3224 static int synchronize_sched_expedited_cpu_stop(void *data
)
3227 * There must be a full memory barrier on each affected CPU
3228 * between the time that try_stop_cpus() is called and the
3229 * time that it returns.
3231 * In the current initial implementation of cpu_stop, the
3232 * above condition is already met when the control reaches
3233 * this point and the following smp_mb() is not strictly
3234 * necessary. Do smp_mb() anyway for documentation and
3235 * robustness against future implementation changes.
3237 smp_mb(); /* See above comment block. */
3242 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3244 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3245 * approach to force the grace period to end quickly. This consumes
3246 * significant time on all CPUs and is unfriendly to real-time workloads,
3247 * so is thus not recommended for any sort of common-case code. In fact,
3248 * if you are using synchronize_sched_expedited() in a loop, please
3249 * restructure your code to batch your updates, and then use a single
3250 * synchronize_sched() instead.
3252 * This implementation can be thought of as an application of ticket
3253 * locking to RCU, with sync_sched_expedited_started and
3254 * sync_sched_expedited_done taking on the roles of the halves
3255 * of the ticket-lock word. Each task atomically increments
3256 * sync_sched_expedited_started upon entry, snapshotting the old value,
3257 * then attempts to stop all the CPUs. If this succeeds, then each
3258 * CPU will have executed a context switch, resulting in an RCU-sched
3259 * grace period. We are then done, so we use atomic_cmpxchg() to
3260 * update sync_sched_expedited_done to match our snapshot -- but
3261 * only if someone else has not already advanced past our snapshot.
3263 * On the other hand, if try_stop_cpus() fails, we check the value
3264 * of sync_sched_expedited_done. If it has advanced past our
3265 * initial snapshot, then someone else must have forced a grace period
3266 * some time after we took our snapshot. In this case, our work is
3267 * done for us, and we can simply return. Otherwise, we try again,
3268 * but keep our initial snapshot for purposes of checking for someone
3269 * doing our work for us.
3271 * If we fail too many times in a row, we fall back to synchronize_sched().
3273 void synchronize_sched_expedited(void)
3278 long firstsnap
, s
, snap
;
3280 struct rcu_state
*rsp
= &rcu_sched_state
;
3283 * If we are in danger of counter wrap, just do synchronize_sched().
3284 * By allowing sync_sched_expedited_started to advance no more than
3285 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3286 * that more than 3.5 billion CPUs would be required to force a
3287 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3288 * course be required on a 64-bit system.
3290 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3291 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3293 wait_rcu_gp(call_rcu_sched
);
3294 atomic_long_inc(&rsp
->expedited_wrap
);
3299 * Take a ticket. Note that atomic_inc_return() implies a
3300 * full memory barrier.
3302 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3304 if (!try_get_online_cpus()) {
3305 /* CPU hotplug operation in flight, fall back to normal GP. */
3306 wait_rcu_gp(call_rcu_sched
);
3307 atomic_long_inc(&rsp
->expedited_normal
);
3310 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3312 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3313 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3315 cpumask_copy(cm
, cpu_online_mask
);
3316 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3317 for_each_cpu(cpu
, cm
) {
3318 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3320 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3321 cpumask_clear_cpu(cpu
, cm
);
3323 if (cpumask_weight(cm
) == 0)
3328 * Each pass through the following loop attempts to force a
3329 * context switch on each CPU.
3331 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3332 synchronize_sched_expedited_cpu_stop
,
3335 atomic_long_inc(&rsp
->expedited_tryfail
);
3337 /* Check to see if someone else did our work for us. */
3338 s
= atomic_long_read(&rsp
->expedited_done
);
3339 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3340 /* ensure test happens before caller kfree */
3341 smp_mb__before_atomic(); /* ^^^ */
3342 atomic_long_inc(&rsp
->expedited_workdone1
);
3343 free_cpumask_var(cm
);
3347 /* No joy, try again later. Or just synchronize_sched(). */
3348 if (trycount
++ < 10) {
3349 udelay(trycount
* num_online_cpus());
3351 wait_rcu_gp(call_rcu_sched
);
3352 atomic_long_inc(&rsp
->expedited_normal
);
3353 free_cpumask_var(cm
);
3357 /* Recheck to see if someone else did our work for us. */
3358 s
= atomic_long_read(&rsp
->expedited_done
);
3359 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3360 /* ensure test happens before caller kfree */
3361 smp_mb__before_atomic(); /* ^^^ */
3362 atomic_long_inc(&rsp
->expedited_workdone2
);
3363 free_cpumask_var(cm
);
3368 * Refetching sync_sched_expedited_started allows later
3369 * callers to piggyback on our grace period. We retry
3370 * after they started, so our grace period works for them,
3371 * and they started after our first try, so their grace
3372 * period works for us.
3374 if (!try_get_online_cpus()) {
3375 /* CPU hotplug operation in flight, use normal GP. */
3376 wait_rcu_gp(call_rcu_sched
);
3377 atomic_long_inc(&rsp
->expedited_normal
);
3378 free_cpumask_var(cm
);
3381 snap
= atomic_long_read(&rsp
->expedited_start
);
3382 smp_mb(); /* ensure read is before try_stop_cpus(). */
3384 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3387 free_cpumask_var(cm
);
3390 * Everyone up to our most recent fetch is covered by our grace
3391 * period. Update the counter, but only if our work is still
3392 * relevant -- which it won't be if someone who started later
3393 * than we did already did their update.
3396 atomic_long_inc(&rsp
->expedited_done_tries
);
3397 s
= atomic_long_read(&rsp
->expedited_done
);
3398 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3399 /* ensure test happens before caller kfree */
3400 smp_mb__before_atomic(); /* ^^^ */
3401 atomic_long_inc(&rsp
->expedited_done_lost
);
3404 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3405 atomic_long_inc(&rsp
->expedited_done_exit
);
3409 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3412 * Check to see if there is any immediate RCU-related work to be done
3413 * by the current CPU, for the specified type of RCU, returning 1 if so.
3414 * The checks are in order of increasing expense: checks that can be
3415 * carried out against CPU-local state are performed first. However,
3416 * we must check for CPU stalls first, else we might not get a chance.
3418 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3420 struct rcu_node
*rnp
= rdp
->mynode
;
3422 rdp
->n_rcu_pending
++;
3424 /* Check for CPU stalls, if enabled. */
3425 check_cpu_stall(rsp
, rdp
);
3427 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3428 if (rcu_nohz_full_cpu(rsp
))
3431 /* Is the RCU core waiting for a quiescent state from this CPU? */
3432 if (rcu_scheduler_fully_active
&&
3433 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3434 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3435 rdp
->n_rp_qs_pending
++;
3436 } else if (rdp
->qs_pending
&&
3437 (rdp
->passed_quiesce
||
3438 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3439 rdp
->n_rp_report_qs
++;
3443 /* Does this CPU have callbacks ready to invoke? */
3444 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3445 rdp
->n_rp_cb_ready
++;
3449 /* Has RCU gone idle with this CPU needing another grace period? */
3450 if (cpu_needs_another_gp(rsp
, rdp
)) {
3451 rdp
->n_rp_cpu_needs_gp
++;
3455 /* Has another RCU grace period completed? */
3456 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3457 rdp
->n_rp_gp_completed
++;
3461 /* Has a new RCU grace period started? */
3462 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3463 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3464 rdp
->n_rp_gp_started
++;
3468 /* Does this CPU need a deferred NOCB wakeup? */
3469 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3470 rdp
->n_rp_nocb_defer_wakeup
++;
3475 rdp
->n_rp_need_nothing
++;
3480 * Check to see if there is any immediate RCU-related work to be done
3481 * by the current CPU, returning 1 if so. This function is part of the
3482 * RCU implementation; it is -not- an exported member of the RCU API.
3484 static int rcu_pending(void)
3486 struct rcu_state
*rsp
;
3488 for_each_rcu_flavor(rsp
)
3489 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3495 * Return true if the specified CPU has any callback. If all_lazy is
3496 * non-NULL, store an indication of whether all callbacks are lazy.
3497 * (If there are no callbacks, all of them are deemed to be lazy.)
3499 static bool __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3503 struct rcu_data
*rdp
;
3504 struct rcu_state
*rsp
;
3506 for_each_rcu_flavor(rsp
) {
3507 rdp
= this_cpu_ptr(rsp
->rda
);
3511 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3522 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3523 * the compiler is expected to optimize this away.
3525 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3526 int cpu
, unsigned long done
)
3528 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3529 atomic_read(&rsp
->barrier_cpu_count
), done
);
3533 * RCU callback function for _rcu_barrier(). If we are last, wake
3534 * up the task executing _rcu_barrier().
3536 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3538 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3539 struct rcu_state
*rsp
= rdp
->rsp
;
3541 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3542 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3543 complete(&rsp
->barrier_completion
);
3545 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3550 * Called with preemption disabled, and from cross-cpu IRQ context.
3552 static void rcu_barrier_func(void *type
)
3554 struct rcu_state
*rsp
= type
;
3555 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3557 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3558 atomic_inc(&rsp
->barrier_cpu_count
);
3559 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3563 * Orchestrate the specified type of RCU barrier, waiting for all
3564 * RCU callbacks of the specified type to complete.
3566 static void _rcu_barrier(struct rcu_state
*rsp
)
3569 struct rcu_data
*rdp
;
3570 unsigned long snap
= READ_ONCE(rsp
->n_barrier_done
);
3571 unsigned long snap_done
;
3573 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3575 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3576 mutex_lock(&rsp
->barrier_mutex
);
3579 * Ensure that all prior references, including to ->n_barrier_done,
3580 * are ordered before the _rcu_barrier() machinery.
3582 smp_mb(); /* See above block comment. */
3585 * Recheck ->n_barrier_done to see if others did our work for us.
3586 * This means checking ->n_barrier_done for an even-to-odd-to-even
3587 * transition. The "if" expression below therefore rounds the old
3588 * value up to the next even number and adds two before comparing.
3590 snap_done
= rsp
->n_barrier_done
;
3591 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3594 * If the value in snap is odd, we needed to wait for the current
3595 * rcu_barrier() to complete, then wait for the next one, in other
3596 * words, we need the value of snap_done to be three larger than
3597 * the value of snap. On the other hand, if the value in snap is
3598 * even, we only had to wait for the next rcu_barrier() to complete,
3599 * in other words, we need the value of snap_done to be only two
3600 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3601 * this for us (thank you, Linus!).
3603 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3604 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3605 smp_mb(); /* caller's subsequent code after above check. */
3606 mutex_unlock(&rsp
->barrier_mutex
);
3611 * Increment ->n_barrier_done to avoid duplicate work. Use
3612 * WRITE_ONCE() to prevent the compiler from speculating
3613 * the increment to precede the early-exit check.
3615 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3616 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3617 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3618 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3621 * Initialize the count to one rather than to zero in order to
3622 * avoid a too-soon return to zero in case of a short grace period
3623 * (or preemption of this task). Exclude CPU-hotplug operations
3624 * to ensure that no offline CPU has callbacks queued.
3626 init_completion(&rsp
->barrier_completion
);
3627 atomic_set(&rsp
->barrier_cpu_count
, 1);
3631 * Force each CPU with callbacks to register a new callback.
3632 * When that callback is invoked, we will know that all of the
3633 * corresponding CPU's preceding callbacks have been invoked.
3635 for_each_possible_cpu(cpu
) {
3636 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3638 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3639 if (rcu_is_nocb_cpu(cpu
)) {
3640 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3641 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3642 rsp
->n_barrier_done
);
3644 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3645 rsp
->n_barrier_done
);
3646 smp_mb__before_atomic();
3647 atomic_inc(&rsp
->barrier_cpu_count
);
3648 __call_rcu(&rdp
->barrier_head
,
3649 rcu_barrier_callback
, rsp
, cpu
, 0);
3651 } else if (READ_ONCE(rdp
->qlen
)) {
3652 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3653 rsp
->n_barrier_done
);
3654 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3656 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3657 rsp
->n_barrier_done
);
3663 * Now that we have an rcu_barrier_callback() callback on each
3664 * CPU, and thus each counted, remove the initial count.
3666 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3667 complete(&rsp
->barrier_completion
);
3669 /* Increment ->n_barrier_done to prevent duplicate work. */
3670 smp_mb(); /* Keep increment after above mechanism. */
3671 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3672 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3673 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3674 smp_mb(); /* Keep increment before caller's subsequent code. */
3676 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3677 wait_for_completion(&rsp
->barrier_completion
);
3679 /* Other rcu_barrier() invocations can now safely proceed. */
3680 mutex_unlock(&rsp
->barrier_mutex
);
3684 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3686 void rcu_barrier_bh(void)
3688 _rcu_barrier(&rcu_bh_state
);
3690 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3693 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3695 void rcu_barrier_sched(void)
3697 _rcu_barrier(&rcu_sched_state
);
3699 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3702 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3703 * first CPU in a given leaf rcu_node structure coming online. The caller
3704 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3707 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3710 struct rcu_node
*rnp
= rnp_leaf
;
3713 mask
= rnp
->grpmask
;
3717 raw_spin_lock(&rnp
->lock
); /* Interrupts already disabled. */
3718 rnp
->qsmaskinit
|= mask
;
3719 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
3724 * Do boot-time initialization of a CPU's per-CPU RCU data.
3727 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3729 unsigned long flags
;
3730 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3731 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3733 /* Set up local state, ensuring consistent view of global state. */
3734 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3735 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3736 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3737 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3738 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3741 rcu_boot_init_nocb_percpu_data(rdp
);
3742 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3746 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3747 * offline event can be happening at a given time. Note also that we
3748 * can accept some slop in the rsp->completed access due to the fact
3749 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3752 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3754 unsigned long flags
;
3756 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3757 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3759 /* Set up local state, ensuring consistent view of global state. */
3760 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3761 rdp
->beenonline
= 1; /* We have now been online. */
3762 rdp
->qlen_last_fqs_check
= 0;
3763 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3764 rdp
->blimit
= blimit
;
3766 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3767 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3768 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3769 atomic_set(&rdp
->dynticks
->dynticks
,
3770 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3771 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3774 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3775 * propagation up the rcu_node tree will happen at the beginning
3776 * of the next grace period.
3779 mask
= rdp
->grpmask
;
3780 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3781 smp_mb__after_unlock_lock();
3782 rnp
->qsmaskinitnext
|= mask
;
3783 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3784 rdp
->completed
= rnp
->completed
;
3785 rdp
->passed_quiesce
= false;
3786 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3787 rdp
->qs_pending
= false;
3788 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3789 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3792 static void rcu_prepare_cpu(int cpu
)
3794 struct rcu_state
*rsp
;
3796 for_each_rcu_flavor(rsp
)
3797 rcu_init_percpu_data(cpu
, rsp
);
3801 * Handle CPU online/offline notification events.
3803 int rcu_cpu_notify(struct notifier_block
*self
,
3804 unsigned long action
, void *hcpu
)
3806 long cpu
= (long)hcpu
;
3807 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3808 struct rcu_node
*rnp
= rdp
->mynode
;
3809 struct rcu_state
*rsp
;
3812 case CPU_UP_PREPARE
:
3813 case CPU_UP_PREPARE_FROZEN
:
3814 rcu_prepare_cpu(cpu
);
3815 rcu_prepare_kthreads(cpu
);
3816 rcu_spawn_all_nocb_kthreads(cpu
);
3819 case CPU_DOWN_FAILED
:
3820 rcu_boost_kthread_setaffinity(rnp
, -1);
3822 case CPU_DOWN_PREPARE
:
3823 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3826 case CPU_DYING_FROZEN
:
3827 for_each_rcu_flavor(rsp
)
3828 rcu_cleanup_dying_cpu(rsp
);
3830 case CPU_DYING_IDLE
:
3831 for_each_rcu_flavor(rsp
) {
3832 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3836 case CPU_DEAD_FROZEN
:
3837 case CPU_UP_CANCELED
:
3838 case CPU_UP_CANCELED_FROZEN
:
3839 for_each_rcu_flavor(rsp
) {
3840 rcu_cleanup_dead_cpu(cpu
, rsp
);
3841 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3850 static int rcu_pm_notify(struct notifier_block
*self
,
3851 unsigned long action
, void *hcpu
)
3854 case PM_HIBERNATION_PREPARE
:
3855 case PM_SUSPEND_PREPARE
:
3856 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3859 case PM_POST_HIBERNATION
:
3860 case PM_POST_SUSPEND
:
3861 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3862 rcu_unexpedite_gp();
3871 * Spawn the kthreads that handle each RCU flavor's grace periods.
3873 static int __init
rcu_spawn_gp_kthread(void)
3875 unsigned long flags
;
3876 int kthread_prio_in
= kthread_prio
;
3877 struct rcu_node
*rnp
;
3878 struct rcu_state
*rsp
;
3879 struct sched_param sp
;
3880 struct task_struct
*t
;
3882 /* Force priority into range. */
3883 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3885 else if (kthread_prio
< 0)
3887 else if (kthread_prio
> 99)
3889 if (kthread_prio
!= kthread_prio_in
)
3890 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3891 kthread_prio
, kthread_prio_in
);
3893 rcu_scheduler_fully_active
= 1;
3894 for_each_rcu_flavor(rsp
) {
3895 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3897 rnp
= rcu_get_root(rsp
);
3898 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3899 rsp
->gp_kthread
= t
;
3901 sp
.sched_priority
= kthread_prio
;
3902 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3905 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3907 rcu_spawn_nocb_kthreads();
3908 rcu_spawn_boost_kthreads();
3911 early_initcall(rcu_spawn_gp_kthread
);
3914 * This function is invoked towards the end of the scheduler's initialization
3915 * process. Before this is called, the idle task might contain
3916 * RCU read-side critical sections (during which time, this idle
3917 * task is booting the system). After this function is called, the
3918 * idle tasks are prohibited from containing RCU read-side critical
3919 * sections. This function also enables RCU lockdep checking.
3921 void rcu_scheduler_starting(void)
3923 WARN_ON(num_online_cpus() != 1);
3924 WARN_ON(nr_context_switches() > 0);
3925 rcu_scheduler_active
= 1;
3929 * Compute the per-level fanout, either using the exact fanout specified
3930 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3932 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3936 if (IS_ENABLED(CONFIG_RCU_FANOUT_EXACT
)) {
3937 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3938 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3939 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
3945 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3946 ccur
= rsp
->levelcnt
[i
];
3947 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3954 * Helper function for rcu_init() that initializes one rcu_state structure.
3956 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3957 struct rcu_data __percpu
*rda
)
3959 static const char * const buf
[] = {
3963 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3964 static const char * const fqs
[] = {
3968 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3969 static u8 fl_mask
= 0x1;
3973 struct rcu_node
*rnp
;
3975 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
3977 /* Silence gcc 4.8 warning about array index out of range. */
3978 if (rcu_num_lvls
> RCU_NUM_LVLS
)
3979 panic("rcu_init_one: rcu_num_lvls overflow");
3981 /* Initialize the level-tracking arrays. */
3983 for (i
= 0; i
< rcu_num_lvls
; i
++)
3984 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
3985 for (i
= 1; i
< rcu_num_lvls
; i
++)
3986 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
3987 rcu_init_levelspread(rsp
);
3988 rsp
->flavor_mask
= fl_mask
;
3991 /* Initialize the elements themselves, starting from the leaves. */
3993 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3994 cpustride
*= rsp
->levelspread
[i
];
3995 rnp
= rsp
->level
[i
];
3996 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
3997 raw_spin_lock_init(&rnp
->lock
);
3998 lockdep_set_class_and_name(&rnp
->lock
,
3999 &rcu_node_class
[i
], buf
[i
]);
4000 raw_spin_lock_init(&rnp
->fqslock
);
4001 lockdep_set_class_and_name(&rnp
->fqslock
,
4002 &rcu_fqs_class
[i
], fqs
[i
]);
4003 rnp
->gpnum
= rsp
->gpnum
;
4004 rnp
->completed
= rsp
->completed
;
4006 rnp
->qsmaskinit
= 0;
4007 rnp
->grplo
= j
* cpustride
;
4008 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4009 if (rnp
->grphi
>= nr_cpu_ids
)
4010 rnp
->grphi
= nr_cpu_ids
- 1;
4016 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
4017 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4018 rnp
->parent
= rsp
->level
[i
- 1] +
4019 j
/ rsp
->levelspread
[i
- 1];
4022 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4023 rcu_init_one_nocb(rnp
);
4027 init_waitqueue_head(&rsp
->gp_wq
);
4028 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4029 for_each_possible_cpu(i
) {
4030 while (i
> rnp
->grphi
)
4032 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4033 rcu_boot_init_percpu_data(i
, rsp
);
4035 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4039 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4040 * replace the definitions in tree.h because those are needed to size
4041 * the ->node array in the rcu_state structure.
4043 static void __init
rcu_init_geometry(void)
4049 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
4052 * Initialize any unspecified boot parameters.
4053 * The default values of jiffies_till_first_fqs and
4054 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4055 * value, which is a function of HZ, then adding one for each
4056 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4058 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4059 if (jiffies_till_first_fqs
== ULONG_MAX
)
4060 jiffies_till_first_fqs
= d
;
4061 if (jiffies_till_next_fqs
== ULONG_MAX
)
4062 jiffies_till_next_fqs
= d
;
4064 /* If the compile-time values are accurate, just leave. */
4065 if (rcu_fanout_leaf
== CONFIG_RCU_FANOUT_LEAF
&&
4066 nr_cpu_ids
== NR_CPUS
)
4068 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4069 rcu_fanout_leaf
, nr_cpu_ids
);
4072 * Compute number of nodes that can be handled an rcu_node tree
4073 * with the given number of levels. Setting rcu_capacity[0] makes
4074 * some of the arithmetic easier.
4076 rcu_capacity
[0] = 1;
4077 rcu_capacity
[1] = rcu_fanout_leaf
;
4078 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
4079 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * CONFIG_RCU_FANOUT
;
4082 * The boot-time rcu_fanout_leaf parameter is only permitted
4083 * to increase the leaf-level fanout, not decrease it. Of course,
4084 * the leaf-level fanout cannot exceed the number of bits in
4085 * the rcu_node masks. Finally, the tree must be able to accommodate
4086 * the configured number of CPUs. Complain and fall back to the
4087 * compile-time values if these limits are exceeded.
4089 if (rcu_fanout_leaf
< CONFIG_RCU_FANOUT_LEAF
||
4090 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
4091 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
4096 /* Calculate the number of rcu_nodes at each level of the tree. */
4097 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
4098 if (n
<= rcu_capacity
[i
]) {
4099 for (j
= 0; j
<= i
; j
++)
4101 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
4103 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
4108 /* Calculate the total number of rcu_node structures. */
4110 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
4111 rcu_num_nodes
+= num_rcu_lvl
[i
];
4115 void __init
rcu_init(void)
4119 rcu_early_boot_tests();
4121 rcu_bootup_announce();
4122 rcu_init_geometry();
4123 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
4124 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
4125 __rcu_init_preempt();
4126 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4129 * We don't need protection against CPU-hotplug here because
4130 * this is called early in boot, before either interrupts
4131 * or the scheduler are operational.
4133 cpu_notifier(rcu_cpu_notify
, 0);
4134 pm_notifier(rcu_pm_notify
, 0);
4135 for_each_online_cpu(cpu
)
4136 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4139 #include "tree_plugin.h"