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 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
*rcu_state_p
;
114 LIST_HEAD(rcu_struct_flavors
);
116 /* Dump rcu_node combining tree at boot to verify correct setup. */
117 static bool dump_tree
;
118 module_param(dump_tree
, bool, 0444);
119 /* Control rcu_node-tree auto-balancing at boot time. */
120 static bool rcu_fanout_exact
;
121 module_param(rcu_fanout_exact
, bool, 0444);
122 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
123 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
124 module_param(rcu_fanout_leaf
, int, 0444);
125 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
126 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
133 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
136 * The rcu_scheduler_active variable transitions from zero to one just
137 * before the first task is spawned. So when this variable is zero, RCU
138 * can assume that there is but one task, allowing RCU to (for example)
139 * optimize synchronize_sched() to a simple barrier(). When this variable
140 * is one, RCU must actually do all the hard work required to detect real
141 * grace periods. This variable is also used to suppress boot-time false
142 * positives from lockdep-RCU error checking.
144 int rcu_scheduler_active __read_mostly
;
145 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
148 * The rcu_scheduler_fully_active variable transitions from zero to one
149 * during the early_initcall() processing, which is after the scheduler
150 * is capable of creating new tasks. So RCU processing (for example,
151 * creating tasks for RCU priority boosting) must be delayed until after
152 * rcu_scheduler_fully_active transitions from zero to one. We also
153 * currently delay invocation of any RCU callbacks until after this point.
155 * It might later prove better for people registering RCU callbacks during
156 * early boot to take responsibility for these callbacks, but one step at
159 static int rcu_scheduler_fully_active __read_mostly
;
161 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
162 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
163 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
164 static void invoke_rcu_core(void);
165 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
167 /* rcuc/rcub kthread realtime priority */
168 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
169 module_param(kthread_prio
, int, 0644);
171 /* Delay in jiffies for grace-period initialization delays, debug only. */
173 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
174 static int gp_preinit_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY
;
175 module_param(gp_preinit_delay
, int, 0644);
176 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
177 static const int gp_preinit_delay
;
178 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
180 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
181 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
182 module_param(gp_init_delay
, int, 0644);
183 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
184 static const int gp_init_delay
;
185 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
187 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
188 static int gp_cleanup_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY
;
189 module_param(gp_cleanup_delay
, int, 0644);
190 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
191 static const int gp_cleanup_delay
;
192 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
195 * Number of grace periods between delays, normalized by the duration of
196 * the delay. The longer the the delay, the more the grace periods between
197 * each delay. The reason for this normalization is that it means that,
198 * for non-zero delays, the overall slowdown of grace periods is constant
199 * regardless of the duration of the delay. This arrangement balances
200 * the need for long delays to increase some race probabilities with the
201 * need for fast grace periods to increase other race probabilities.
203 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
206 * Track the rcutorture test sequence number and the update version
207 * number within a given test. The rcutorture_testseq is incremented
208 * on every rcutorture module load and unload, so has an odd value
209 * when a test is running. The rcutorture_vernum is set to zero
210 * when rcutorture starts and is incremented on each rcutorture update.
211 * These variables enable correlating rcutorture output with the
212 * RCU tracing information.
214 unsigned long rcutorture_testseq
;
215 unsigned long rcutorture_vernum
;
218 * Compute the mask of online CPUs for the specified rcu_node structure.
219 * This will not be stable unless the rcu_node structure's ->lock is
220 * held, but the bit corresponding to the current CPU will be stable
223 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
225 return READ_ONCE(rnp
->qsmaskinitnext
);
229 * Return true if an RCU grace period is in progress. The READ_ONCE()s
230 * permit this function to be invoked without holding the root rcu_node
231 * structure's ->lock, but of course results can be subject to change.
233 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
235 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
239 * Note a quiescent state. Because we do not need to know
240 * how many quiescent states passed, just if there was at least
241 * one since the start of the grace period, this just sets a flag.
242 * The caller must have disabled preemption.
244 void rcu_sched_qs(void)
246 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
247 trace_rcu_grace_period(TPS("rcu_sched"),
248 __this_cpu_read(rcu_sched_data
.gpnum
),
250 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
256 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
257 trace_rcu_grace_period(TPS("rcu_bh"),
258 __this_cpu_read(rcu_bh_data
.gpnum
),
260 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
264 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
266 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
267 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
268 .dynticks
= ATOMIC_INIT(1),
269 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
270 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
271 .dynticks_idle
= ATOMIC_INIT(1),
272 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
275 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
276 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
279 * Let the RCU core know that this CPU has gone through the scheduler,
280 * which is a quiescent state. This is called when the need for a
281 * quiescent state is urgent, so we burn an atomic operation and full
282 * memory barriers to let the RCU core know about it, regardless of what
283 * this CPU might (or might not) do in the near future.
285 * We inform the RCU core by emulating a zero-duration dyntick-idle
286 * period, which we in turn do by incrementing the ->dynticks counter
289 static void rcu_momentary_dyntick_idle(void)
292 struct rcu_data
*rdp
;
293 struct rcu_dynticks
*rdtp
;
295 struct rcu_state
*rsp
;
297 local_irq_save(flags
);
300 * Yes, we can lose flag-setting operations. This is OK, because
301 * the flag will be set again after some delay.
303 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
304 raw_cpu_write(rcu_sched_qs_mask
, 0);
306 /* Find the flavor that needs a quiescent state. */
307 for_each_rcu_flavor(rsp
) {
308 rdp
= raw_cpu_ptr(rsp
->rda
);
309 if (!(resched_mask
& rsp
->flavor_mask
))
311 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
312 if (READ_ONCE(rdp
->mynode
->completed
) !=
313 READ_ONCE(rdp
->cond_resched_completed
))
317 * Pretend to be momentarily idle for the quiescent state.
318 * This allows the grace-period kthread to record the
319 * quiescent state, with no need for this CPU to do anything
322 rdtp
= this_cpu_ptr(&rcu_dynticks
);
323 smp_mb__before_atomic(); /* Earlier stuff before QS. */
324 atomic_add(2, &rdtp
->dynticks
); /* QS. */
325 smp_mb__after_atomic(); /* Later stuff after QS. */
328 local_irq_restore(flags
);
332 * Note a context switch. This is a quiescent state for RCU-sched,
333 * and requires special handling for preemptible RCU.
334 * The caller must have disabled preemption.
336 void rcu_note_context_switch(void)
338 trace_rcu_utilization(TPS("Start context switch"));
340 rcu_preempt_note_context_switch();
341 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
342 rcu_momentary_dyntick_idle();
343 trace_rcu_utilization(TPS("End context switch"));
345 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
348 * Register a quiescent state for all RCU flavors. If there is an
349 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
350 * dyntick-idle quiescent state visible to other CPUs (but only for those
351 * RCU flavors in desperate need of a quiescent state, which will normally
352 * be none of them). Either way, do a lightweight quiescent state for
355 void rcu_all_qs(void)
357 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
358 rcu_momentary_dyntick_idle();
359 this_cpu_inc(rcu_qs_ctr
);
361 EXPORT_SYMBOL_GPL(rcu_all_qs
);
363 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
364 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
365 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
367 module_param(blimit
, long, 0444);
368 module_param(qhimark
, long, 0444);
369 module_param(qlowmark
, long, 0444);
371 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
372 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
374 module_param(jiffies_till_first_fqs
, ulong
, 0644);
375 module_param(jiffies_till_next_fqs
, ulong
, 0644);
378 * How long the grace period must be before we start recruiting
379 * quiescent-state help from rcu_note_context_switch().
381 static ulong jiffies_till_sched_qs
= HZ
/ 20;
382 module_param(jiffies_till_sched_qs
, ulong
, 0644);
384 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
385 struct rcu_data
*rdp
);
386 static void force_qs_rnp(struct rcu_state
*rsp
,
387 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
388 unsigned long *maxj
),
389 bool *isidle
, unsigned long *maxj
);
390 static void force_quiescent_state(struct rcu_state
*rsp
);
391 static int rcu_pending(void);
394 * Return the number of RCU batches started thus far for debug & stats.
396 unsigned long rcu_batches_started(void)
398 return rcu_state_p
->gpnum
;
400 EXPORT_SYMBOL_GPL(rcu_batches_started
);
403 * Return the number of RCU-sched batches started thus far for debug & stats.
405 unsigned long rcu_batches_started_sched(void)
407 return rcu_sched_state
.gpnum
;
409 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
412 * Return the number of RCU BH batches started thus far for debug & stats.
414 unsigned long rcu_batches_started_bh(void)
416 return rcu_bh_state
.gpnum
;
418 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
421 * Return the number of RCU batches completed thus far for debug & stats.
423 unsigned long rcu_batches_completed(void)
425 return rcu_state_p
->completed
;
427 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
430 * Return the number of RCU-sched batches completed thus far for debug & stats.
432 unsigned long rcu_batches_completed_sched(void)
434 return rcu_sched_state
.completed
;
436 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
439 * Return the number of RCU BH batches completed thus far for debug & stats.
441 unsigned long rcu_batches_completed_bh(void)
443 return rcu_bh_state
.completed
;
445 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
448 * Force a quiescent state.
450 void rcu_force_quiescent_state(void)
452 force_quiescent_state(rcu_state_p
);
454 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
457 * Force a quiescent state for RCU BH.
459 void rcu_bh_force_quiescent_state(void)
461 force_quiescent_state(&rcu_bh_state
);
463 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
466 * Force a quiescent state for RCU-sched.
468 void rcu_sched_force_quiescent_state(void)
470 force_quiescent_state(&rcu_sched_state
);
472 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
475 * Show the state of the grace-period kthreads.
477 void show_rcu_gp_kthreads(void)
479 struct rcu_state
*rsp
;
481 for_each_rcu_flavor(rsp
) {
482 pr_info("%s: wait state: %d ->state: %#lx\n",
483 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
484 /* sched_show_task(rsp->gp_kthread); */
487 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
490 * Record the number of times rcutorture tests have been initiated and
491 * terminated. This information allows the debugfs tracing stats to be
492 * correlated to the rcutorture messages, even when the rcutorture module
493 * is being repeatedly loaded and unloaded. In other words, we cannot
494 * store this state in rcutorture itself.
496 void rcutorture_record_test_transition(void)
498 rcutorture_testseq
++;
499 rcutorture_vernum
= 0;
501 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
504 * Send along grace-period-related data for rcutorture diagnostics.
506 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
507 unsigned long *gpnum
, unsigned long *completed
)
509 struct rcu_state
*rsp
= NULL
;
518 case RCU_SCHED_FLAVOR
:
519 rsp
= &rcu_sched_state
;
525 *flags
= READ_ONCE(rsp
->gp_flags
);
526 *gpnum
= READ_ONCE(rsp
->gpnum
);
527 *completed
= READ_ONCE(rsp
->completed
);
534 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
537 * Record the number of writer passes through the current rcutorture test.
538 * This is also used to correlate debugfs tracing stats with the rcutorture
541 void rcutorture_record_progress(unsigned long vernum
)
545 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
548 * Does the CPU have callbacks ready to be invoked?
551 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
553 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
554 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
558 * Return the root node of the specified rcu_state structure.
560 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
562 return &rsp
->node
[0];
566 * Is there any need for future grace periods?
567 * Interrupts must be disabled. If the caller does not hold the root
568 * rnp_node structure's ->lock, the results are advisory only.
570 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
572 struct rcu_node
*rnp
= rcu_get_root(rsp
);
573 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
574 int *fp
= &rnp
->need_future_gp
[idx
];
576 return READ_ONCE(*fp
);
580 * Does the current CPU require a not-yet-started grace period?
581 * The caller must have disabled interrupts to prevent races with
582 * normal callback registry.
585 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
589 if (rcu_gp_in_progress(rsp
))
590 return 0; /* No, a grace period is already in progress. */
591 if (rcu_future_needs_gp(rsp
))
592 return 1; /* Yes, a no-CBs CPU needs one. */
593 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
594 return 0; /* No, this is a no-CBs (or offline) CPU. */
595 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
596 return 1; /* Yes, this CPU has newly registered callbacks. */
597 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
598 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
599 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
600 rdp
->nxtcompleted
[i
]))
601 return 1; /* Yes, CBs for future grace period. */
602 return 0; /* No grace period needed. */
606 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
608 * If the new value of the ->dynticks_nesting counter now is zero,
609 * we really have entered idle, and must do the appropriate accounting.
610 * The caller must have disabled interrupts.
612 static void rcu_eqs_enter_common(long long oldval
, bool user
)
614 struct rcu_state
*rsp
;
615 struct rcu_data
*rdp
;
616 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
618 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
619 if (!user
&& !is_idle_task(current
)) {
620 struct task_struct
*idle __maybe_unused
=
621 idle_task(smp_processor_id());
623 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
624 ftrace_dump(DUMP_ORIG
);
625 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
626 current
->pid
, current
->comm
,
627 idle
->pid
, idle
->comm
); /* must be idle task! */
629 for_each_rcu_flavor(rsp
) {
630 rdp
= this_cpu_ptr(rsp
->rda
);
631 do_nocb_deferred_wakeup(rdp
);
633 rcu_prepare_for_idle();
634 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
635 smp_mb__before_atomic(); /* See above. */
636 atomic_inc(&rdtp
->dynticks
);
637 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
638 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
639 rcu_dynticks_task_enter();
642 * It is illegal to enter an extended quiescent state while
643 * in an RCU read-side critical section.
645 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
646 "Illegal idle entry in RCU read-side critical section.");
647 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
648 "Illegal idle entry in RCU-bh read-side critical section.");
649 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
650 "Illegal idle entry in RCU-sched read-side critical section.");
654 * Enter an RCU extended quiescent state, which can be either the
655 * idle loop or adaptive-tickless usermode execution.
657 static void rcu_eqs_enter(bool user
)
660 struct rcu_dynticks
*rdtp
;
662 rdtp
= this_cpu_ptr(&rcu_dynticks
);
663 oldval
= rdtp
->dynticks_nesting
;
664 WARN_ON_ONCE((oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
665 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
666 rdtp
->dynticks_nesting
= 0;
667 rcu_eqs_enter_common(oldval
, user
);
669 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
674 * rcu_idle_enter - inform RCU that current CPU is entering idle
676 * Enter idle mode, in other words, -leave- the mode in which RCU
677 * read-side critical sections can occur. (Though RCU read-side
678 * critical sections can occur in irq handlers in idle, a possibility
679 * handled by irq_enter() and irq_exit().)
681 * We crowbar the ->dynticks_nesting field to zero to allow for
682 * the possibility of usermode upcalls having messed up our count
683 * of interrupt nesting level during the prior busy period.
685 void rcu_idle_enter(void)
689 local_irq_save(flags
);
690 rcu_eqs_enter(false);
691 rcu_sysidle_enter(0);
692 local_irq_restore(flags
);
694 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
696 #ifdef CONFIG_RCU_USER_QS
698 * rcu_user_enter - inform RCU that we are resuming userspace.
700 * Enter RCU idle mode right before resuming userspace. No use of RCU
701 * is permitted between this call and rcu_user_exit(). This way the
702 * CPU doesn't need to maintain the tick for RCU maintenance purposes
703 * when the CPU runs in userspace.
705 void rcu_user_enter(void)
709 #endif /* CONFIG_RCU_USER_QS */
712 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
714 * Exit from an interrupt handler, which might possibly result in entering
715 * idle mode, in other words, leaving the mode in which read-side critical
716 * sections can occur.
718 * This code assumes that the idle loop never does anything that might
719 * result in unbalanced calls to irq_enter() and irq_exit(). If your
720 * architecture violates this assumption, RCU will give you what you
721 * deserve, good and hard. But very infrequently and irreproducibly.
723 * Use things like work queues to work around this limitation.
725 * You have been warned.
727 void rcu_irq_exit(void)
731 struct rcu_dynticks
*rdtp
;
733 local_irq_save(flags
);
734 rdtp
= this_cpu_ptr(&rcu_dynticks
);
735 oldval
= rdtp
->dynticks_nesting
;
736 rdtp
->dynticks_nesting
--;
737 WARN_ON_ONCE(rdtp
->dynticks_nesting
< 0);
738 if (rdtp
->dynticks_nesting
)
739 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
741 rcu_eqs_enter_common(oldval
, true);
742 rcu_sysidle_enter(1);
743 local_irq_restore(flags
);
747 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
749 * If the new value of the ->dynticks_nesting counter was previously zero,
750 * we really have exited idle, and must do the appropriate accounting.
751 * The caller must have disabled interrupts.
753 static void rcu_eqs_exit_common(long long oldval
, int user
)
755 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
757 rcu_dynticks_task_exit();
758 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
759 atomic_inc(&rdtp
->dynticks
);
760 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
761 smp_mb__after_atomic(); /* See above. */
762 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
763 rcu_cleanup_after_idle();
764 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
765 if (!user
&& !is_idle_task(current
)) {
766 struct task_struct
*idle __maybe_unused
=
767 idle_task(smp_processor_id());
769 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
770 oldval
, rdtp
->dynticks_nesting
);
771 ftrace_dump(DUMP_ORIG
);
772 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
773 current
->pid
, current
->comm
,
774 idle
->pid
, idle
->comm
); /* must be idle task! */
779 * Exit an RCU extended quiescent state, which can be either the
780 * idle loop or adaptive-tickless usermode execution.
782 static void rcu_eqs_exit(bool user
)
784 struct rcu_dynticks
*rdtp
;
787 rdtp
= this_cpu_ptr(&rcu_dynticks
);
788 oldval
= rdtp
->dynticks_nesting
;
789 WARN_ON_ONCE(oldval
< 0);
790 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
791 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
793 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
794 rcu_eqs_exit_common(oldval
, user
);
799 * rcu_idle_exit - inform RCU that current CPU is leaving idle
801 * Exit idle mode, in other words, -enter- the mode in which RCU
802 * read-side critical sections can occur.
804 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
805 * allow for the possibility of usermode upcalls messing up our count
806 * of interrupt nesting level during the busy period that is just
809 void rcu_idle_exit(void)
813 local_irq_save(flags
);
816 local_irq_restore(flags
);
818 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
820 #ifdef CONFIG_RCU_USER_QS
822 * rcu_user_exit - inform RCU that we are exiting userspace.
824 * Exit RCU idle mode while entering the kernel because it can
825 * run a RCU read side critical section anytime.
827 void rcu_user_exit(void)
831 #endif /* CONFIG_RCU_USER_QS */
834 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
836 * Enter an interrupt handler, which might possibly result in exiting
837 * idle mode, in other words, entering the mode in which read-side critical
838 * sections can occur.
840 * Note that the Linux kernel is fully capable of entering an interrupt
841 * handler that it never exits, for example when doing upcalls to
842 * user mode! This code assumes that the idle loop never does upcalls to
843 * user mode. If your architecture does do upcalls from the idle loop (or
844 * does anything else that results in unbalanced calls to the irq_enter()
845 * and irq_exit() functions), RCU will give you what you deserve, good
846 * and hard. But very infrequently and irreproducibly.
848 * Use things like work queues to work around this limitation.
850 * You have been warned.
852 void rcu_irq_enter(void)
855 struct rcu_dynticks
*rdtp
;
858 local_irq_save(flags
);
859 rdtp
= this_cpu_ptr(&rcu_dynticks
);
860 oldval
= rdtp
->dynticks_nesting
;
861 rdtp
->dynticks_nesting
++;
862 WARN_ON_ONCE(rdtp
->dynticks_nesting
== 0);
864 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
866 rcu_eqs_exit_common(oldval
, true);
868 local_irq_restore(flags
);
872 * rcu_nmi_enter - inform RCU of entry to NMI context
874 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
875 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
876 * that the CPU is active. This implementation permits nested NMIs, as
877 * long as the nesting level does not overflow an int. (You will probably
878 * run out of stack space first.)
880 void rcu_nmi_enter(void)
882 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
885 /* Complain about underflow. */
886 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
889 * If idle from RCU viewpoint, atomically increment ->dynticks
890 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
891 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
892 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
893 * to be in the outermost NMI handler that interrupted an RCU-idle
894 * period (observation due to Andy Lutomirski).
896 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
897 smp_mb__before_atomic(); /* Force delay from prior write. */
898 atomic_inc(&rdtp
->dynticks
);
899 /* atomic_inc() before later RCU read-side crit sects */
900 smp_mb__after_atomic(); /* See above. */
901 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
904 rdtp
->dynticks_nmi_nesting
+= incby
;
909 * rcu_nmi_exit - inform RCU of exit from NMI context
911 * If we are returning from the outermost NMI handler that interrupted an
912 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
913 * to let the RCU grace-period handling know that the CPU is back to
916 void rcu_nmi_exit(void)
918 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
921 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
922 * (We are exiting an NMI handler, so RCU better be paying attention
925 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
926 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
929 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
930 * leave it in non-RCU-idle state.
932 if (rdtp
->dynticks_nmi_nesting
!= 1) {
933 rdtp
->dynticks_nmi_nesting
-= 2;
937 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
938 rdtp
->dynticks_nmi_nesting
= 0;
939 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
940 smp_mb__before_atomic(); /* See above. */
941 atomic_inc(&rdtp
->dynticks
);
942 smp_mb__after_atomic(); /* Force delay to next write. */
943 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
947 * __rcu_is_watching - are RCU read-side critical sections safe?
949 * Return true if RCU is watching the running CPU, which means that
950 * this CPU can safely enter RCU read-side critical sections. Unlike
951 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
952 * least disabled preemption.
954 bool notrace
__rcu_is_watching(void)
956 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
960 * rcu_is_watching - see if RCU thinks that the current CPU is idle
962 * If the current CPU is in its idle loop and is neither in an interrupt
963 * or NMI handler, return true.
965 bool notrace
rcu_is_watching(void)
970 ret
= __rcu_is_watching();
974 EXPORT_SYMBOL_GPL(rcu_is_watching
);
976 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
979 * Is the current CPU online? Disable preemption to avoid false positives
980 * that could otherwise happen due to the current CPU number being sampled,
981 * this task being preempted, its old CPU being taken offline, resuming
982 * on some other CPU, then determining that its old CPU is now offline.
983 * It is OK to use RCU on an offline processor during initial boot, hence
984 * the check for rcu_scheduler_fully_active. Note also that it is OK
985 * for a CPU coming online to use RCU for one jiffy prior to marking itself
986 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
987 * offline to continue to use RCU for one jiffy after marking itself
988 * offline in the cpu_online_mask. This leniency is necessary given the
989 * non-atomic nature of the online and offline processing, for example,
990 * the fact that a CPU enters the scheduler after completing the CPU_DYING
993 * This is also why RCU internally marks CPUs online during the
994 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
996 * Disable checking if in an NMI handler because we cannot safely report
997 * errors from NMI handlers anyway.
999 bool rcu_lockdep_current_cpu_online(void)
1001 struct rcu_data
*rdp
;
1002 struct rcu_node
*rnp
;
1008 rdp
= this_cpu_ptr(&rcu_sched_data
);
1010 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
1011 !rcu_scheduler_fully_active
;
1015 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
1017 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1020 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1022 * If the current CPU is idle or running at a first-level (not nested)
1023 * interrupt from idle, return true. The caller must have at least
1024 * disabled preemption.
1026 static int rcu_is_cpu_rrupt_from_idle(void)
1028 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1032 * Snapshot the specified CPU's dynticks counter so that we can later
1033 * credit them with an implicit quiescent state. Return 1 if this CPU
1034 * is in dynticks idle mode, which is an extended quiescent state.
1036 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1037 bool *isidle
, unsigned long *maxj
)
1039 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1040 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1041 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1042 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1045 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1046 rdp
->mynode
->gpnum
))
1047 WRITE_ONCE(rdp
->gpwrap
, true);
1053 * Return true if the specified CPU has passed through a quiescent
1054 * state by virtue of being in or having passed through an dynticks
1055 * idle state since the last call to dyntick_save_progress_counter()
1056 * for this same CPU, or by virtue of having been offline.
1058 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1059 bool *isidle
, unsigned long *maxj
)
1065 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1066 snap
= (unsigned int)rdp
->dynticks_snap
;
1069 * If the CPU passed through or entered a dynticks idle phase with
1070 * no active irq/NMI handlers, then we can safely pretend that the CPU
1071 * already acknowledged the request to pass through a quiescent
1072 * state. Either way, that CPU cannot possibly be in an RCU
1073 * read-side critical section that started before the beginning
1074 * of the current RCU grace period.
1076 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1077 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1078 rdp
->dynticks_fqs
++;
1083 * Check for the CPU being offline, but only if the grace period
1084 * is old enough. We don't need to worry about the CPU changing
1085 * state: If we see it offline even once, it has been through a
1088 * The reason for insisting that the grace period be at least
1089 * one jiffy old is that CPUs that are not quite online and that
1090 * have just gone offline can still execute RCU read-side critical
1093 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1094 return 0; /* Grace period is not old enough. */
1096 if (cpu_is_offline(rdp
->cpu
)) {
1097 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1103 * A CPU running for an extended time within the kernel can
1104 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1105 * even context-switching back and forth between a pair of
1106 * in-kernel CPU-bound tasks cannot advance grace periods.
1107 * So if the grace period is old enough, make the CPU pay attention.
1108 * Note that the unsynchronized assignments to the per-CPU
1109 * rcu_sched_qs_mask variable are safe. Yes, setting of
1110 * bits can be lost, but they will be set again on the next
1111 * force-quiescent-state pass. So lost bit sets do not result
1112 * in incorrect behavior, merely in a grace period lasting
1113 * a few jiffies longer than it might otherwise. Because
1114 * there are at most four threads involved, and because the
1115 * updates are only once every few jiffies, the probability of
1116 * lossage (and thus of slight grace-period extension) is
1119 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1120 * is set too high, we override with half of the RCU CPU stall
1123 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1124 if (ULONG_CMP_GE(jiffies
,
1125 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1126 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1127 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1128 WRITE_ONCE(rdp
->cond_resched_completed
,
1129 READ_ONCE(rdp
->mynode
->completed
));
1130 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1132 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1133 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1134 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1135 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1136 /* Time to beat on that CPU again! */
1137 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1138 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1145 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1147 unsigned long j
= jiffies
;
1151 smp_wmb(); /* Record start time before stall time. */
1152 j1
= rcu_jiffies_till_stall_check();
1153 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1154 rsp
->jiffies_resched
= j
+ j1
/ 2;
1155 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1159 * Complain about starvation of grace-period kthread.
1161 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1167 gpa
= READ_ONCE(rsp
->gp_activity
);
1168 if (j
- gpa
> 2 * HZ
)
1169 pr_err("%s kthread starved for %ld jiffies!\n",
1170 rsp
->name
, j
- gpa
);
1174 * Dump stacks of all tasks running on stalled CPUs.
1176 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1179 unsigned long flags
;
1180 struct rcu_node
*rnp
;
1182 rcu_for_each_leaf_node(rsp
, rnp
) {
1183 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1184 if (rnp
->qsmask
!= 0) {
1185 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1186 if (rnp
->qsmask
& (1UL << cpu
))
1187 dump_cpu_task(rnp
->grplo
+ cpu
);
1189 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1193 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1197 unsigned long flags
;
1201 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1204 /* Only let one CPU complain about others per time interval. */
1206 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1207 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1208 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1209 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1212 WRITE_ONCE(rsp
->jiffies_stall
,
1213 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1214 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1217 * OK, time to rat on our buddy...
1218 * See Documentation/RCU/stallwarn.txt for info on how to debug
1219 * RCU CPU stall warnings.
1221 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1223 print_cpu_stall_info_begin();
1224 rcu_for_each_leaf_node(rsp
, rnp
) {
1225 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1226 ndetected
+= rcu_print_task_stall(rnp
);
1227 if (rnp
->qsmask
!= 0) {
1228 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1229 if (rnp
->qsmask
& (1UL << cpu
)) {
1230 print_cpu_stall_info(rsp
,
1235 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1238 print_cpu_stall_info_end();
1239 for_each_possible_cpu(cpu
)
1240 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1241 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1242 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1243 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1245 rcu_dump_cpu_stacks(rsp
);
1247 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1248 READ_ONCE(rsp
->completed
) == gpnum
) {
1249 pr_err("INFO: Stall ended before state dump start\n");
1252 gpa
= READ_ONCE(rsp
->gp_activity
);
1253 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1254 rsp
->name
, j
- gpa
, j
, gpa
,
1255 jiffies_till_next_fqs
,
1256 rcu_get_root(rsp
)->qsmask
);
1257 /* In this case, the current CPU might be at fault. */
1258 sched_show_task(current
);
1262 /* Complain about tasks blocking the grace period. */
1263 rcu_print_detail_task_stall(rsp
);
1265 rcu_check_gp_kthread_starvation(rsp
);
1267 force_quiescent_state(rsp
); /* Kick them all. */
1270 static void print_cpu_stall(struct rcu_state
*rsp
)
1273 unsigned long flags
;
1274 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1278 * OK, time to rat on ourselves...
1279 * See Documentation/RCU/stallwarn.txt for info on how to debug
1280 * RCU CPU stall warnings.
1282 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1283 print_cpu_stall_info_begin();
1284 print_cpu_stall_info(rsp
, smp_processor_id());
1285 print_cpu_stall_info_end();
1286 for_each_possible_cpu(cpu
)
1287 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1288 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1289 jiffies
- rsp
->gp_start
,
1290 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1292 rcu_check_gp_kthread_starvation(rsp
);
1294 rcu_dump_cpu_stacks(rsp
);
1296 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1297 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1298 WRITE_ONCE(rsp
->jiffies_stall
,
1299 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1300 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1303 * Attempt to revive the RCU machinery by forcing a context switch.
1305 * A context switch would normally allow the RCU state machine to make
1306 * progress and it could be we're stuck in kernel space without context
1307 * switches for an entirely unreasonable amount of time.
1309 resched_cpu(smp_processor_id());
1312 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1314 unsigned long completed
;
1315 unsigned long gpnum
;
1319 struct rcu_node
*rnp
;
1321 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1326 * Lots of memory barriers to reject false positives.
1328 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1329 * then rsp->gp_start, and finally rsp->completed. These values
1330 * are updated in the opposite order with memory barriers (or
1331 * equivalent) during grace-period initialization and cleanup.
1332 * Now, a false positive can occur if we get an new value of
1333 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1334 * the memory barriers, the only way that this can happen is if one
1335 * grace period ends and another starts between these two fetches.
1336 * Detect this by comparing rsp->completed with the previous fetch
1339 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1340 * and rsp->gp_start suffice to forestall false positives.
1342 gpnum
= READ_ONCE(rsp
->gpnum
);
1343 smp_rmb(); /* Pick up ->gpnum first... */
1344 js
= READ_ONCE(rsp
->jiffies_stall
);
1345 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1346 gps
= READ_ONCE(rsp
->gp_start
);
1347 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1348 completed
= READ_ONCE(rsp
->completed
);
1349 if (ULONG_CMP_GE(completed
, gpnum
) ||
1350 ULONG_CMP_LT(j
, js
) ||
1351 ULONG_CMP_GE(gps
, js
))
1352 return; /* No stall or GP completed since entering function. */
1354 if (rcu_gp_in_progress(rsp
) &&
1355 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1357 /* We haven't checked in, so go dump stack. */
1358 print_cpu_stall(rsp
);
1360 } else if (rcu_gp_in_progress(rsp
) &&
1361 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1363 /* They had a few time units to dump stack, so complain. */
1364 print_other_cpu_stall(rsp
, gpnum
);
1369 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1371 * Set the stall-warning timeout way off into the future, thus preventing
1372 * any RCU CPU stall-warning messages from appearing in the current set of
1373 * RCU grace periods.
1375 * The caller must disable hard irqs.
1377 void rcu_cpu_stall_reset(void)
1379 struct rcu_state
*rsp
;
1381 for_each_rcu_flavor(rsp
)
1382 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1386 * Initialize the specified rcu_data structure's default callback list
1387 * to empty. The default callback list is the one that is not used by
1388 * no-callbacks CPUs.
1390 static void init_default_callback_list(struct rcu_data
*rdp
)
1394 rdp
->nxtlist
= NULL
;
1395 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1396 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1400 * Initialize the specified rcu_data structure's callback list to empty.
1402 static void init_callback_list(struct rcu_data
*rdp
)
1404 if (init_nocb_callback_list(rdp
))
1406 init_default_callback_list(rdp
);
1410 * Determine the value that ->completed will have at the end of the
1411 * next subsequent grace period. This is used to tag callbacks so that
1412 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1413 * been dyntick-idle for an extended period with callbacks under the
1414 * influence of RCU_FAST_NO_HZ.
1416 * The caller must hold rnp->lock with interrupts disabled.
1418 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1419 struct rcu_node
*rnp
)
1422 * If RCU is idle, we just wait for the next grace period.
1423 * But we can only be sure that RCU is idle if we are looking
1424 * at the root rcu_node structure -- otherwise, a new grace
1425 * period might have started, but just not yet gotten around
1426 * to initializing the current non-root rcu_node structure.
1428 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1429 return rnp
->completed
+ 1;
1432 * Otherwise, wait for a possible partial grace period and
1433 * then the subsequent full grace period.
1435 return rnp
->completed
+ 2;
1439 * Trace-event helper function for rcu_start_future_gp() and
1440 * rcu_nocb_wait_gp().
1442 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1443 unsigned long c
, const char *s
)
1445 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1446 rnp
->completed
, c
, rnp
->level
,
1447 rnp
->grplo
, rnp
->grphi
, s
);
1451 * Start some future grace period, as needed to handle newly arrived
1452 * callbacks. The required future grace periods are recorded in each
1453 * rcu_node structure's ->need_future_gp field. Returns true if there
1454 * is reason to awaken the grace-period kthread.
1456 * The caller must hold the specified rcu_node structure's ->lock.
1458 static bool __maybe_unused
1459 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1460 unsigned long *c_out
)
1465 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1468 * Pick up grace-period number for new callbacks. If this
1469 * grace period is already marked as needed, return to the caller.
1471 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1472 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1473 if (rnp
->need_future_gp
[c
& 0x1]) {
1474 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1479 * If either this rcu_node structure or the root rcu_node structure
1480 * believe that a grace period is in progress, then we must wait
1481 * for the one following, which is in "c". Because our request
1482 * will be noticed at the end of the current grace period, we don't
1483 * need to explicitly start one. We only do the lockless check
1484 * of rnp_root's fields if the current rcu_node structure thinks
1485 * there is no grace period in flight, and because we hold rnp->lock,
1486 * the only possible change is when rnp_root's two fields are
1487 * equal, in which case rnp_root->gpnum might be concurrently
1488 * incremented. But that is OK, as it will just result in our
1489 * doing some extra useless work.
1491 if (rnp
->gpnum
!= rnp
->completed
||
1492 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1493 rnp
->need_future_gp
[c
& 0x1]++;
1494 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1499 * There might be no grace period in progress. If we don't already
1500 * hold it, acquire the root rcu_node structure's lock in order to
1501 * start one (if needed).
1503 if (rnp
!= rnp_root
) {
1504 raw_spin_lock(&rnp_root
->lock
);
1505 smp_mb__after_unlock_lock();
1509 * Get a new grace-period number. If there really is no grace
1510 * period in progress, it will be smaller than the one we obtained
1511 * earlier. Adjust callbacks as needed. Note that even no-CBs
1512 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1514 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1515 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1516 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1517 rdp
->nxtcompleted
[i
] = c
;
1520 * If the needed for the required grace period is already
1521 * recorded, trace and leave.
1523 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1524 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1528 /* Record the need for the future grace period. */
1529 rnp_root
->need_future_gp
[c
& 0x1]++;
1531 /* If a grace period is not already in progress, start one. */
1532 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1533 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1535 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1536 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1539 if (rnp
!= rnp_root
)
1540 raw_spin_unlock(&rnp_root
->lock
);
1548 * Clean up any old requests for the just-ended grace period. Also return
1549 * whether any additional grace periods have been requested. Also invoke
1550 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1551 * waiting for this grace period to complete.
1553 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1555 int c
= rnp
->completed
;
1557 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1559 rcu_nocb_gp_cleanup(rsp
, rnp
);
1560 rnp
->need_future_gp
[c
& 0x1] = 0;
1561 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1562 trace_rcu_future_gp(rnp
, rdp
, c
,
1563 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1568 * Awaken the grace-period kthread for the specified flavor of RCU.
1569 * Don't do a self-awaken, and don't bother awakening when there is
1570 * nothing for the grace-period kthread to do (as in several CPUs
1571 * raced to awaken, and we lost), and finally don't try to awaken
1572 * a kthread that has not yet been created.
1574 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1576 if (current
== rsp
->gp_kthread
||
1577 !READ_ONCE(rsp
->gp_flags
) ||
1580 wake_up(&rsp
->gp_wq
);
1584 * If there is room, assign a ->completed number to any callbacks on
1585 * this CPU that have not already been assigned. Also accelerate any
1586 * callbacks that were previously assigned a ->completed number that has
1587 * since proven to be too conservative, which can happen if callbacks get
1588 * assigned a ->completed number while RCU is idle, but with reference to
1589 * a non-root rcu_node structure. This function is idempotent, so it does
1590 * not hurt to call it repeatedly. Returns an flag saying that we should
1591 * awaken the RCU grace-period kthread.
1593 * The caller must hold rnp->lock with interrupts disabled.
1595 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1596 struct rcu_data
*rdp
)
1602 /* If the CPU has no callbacks, nothing to do. */
1603 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1607 * Starting from the sublist containing the callbacks most
1608 * recently assigned a ->completed number and working down, find the
1609 * first sublist that is not assignable to an upcoming grace period.
1610 * Such a sublist has something in it (first two tests) and has
1611 * a ->completed number assigned that will complete sooner than
1612 * the ->completed number for newly arrived callbacks (last test).
1614 * The key point is that any later sublist can be assigned the
1615 * same ->completed number as the newly arrived callbacks, which
1616 * means that the callbacks in any of these later sublist can be
1617 * grouped into a single sublist, whether or not they have already
1618 * been assigned a ->completed number.
1620 c
= rcu_cbs_completed(rsp
, rnp
);
1621 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1622 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1623 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1627 * If there are no sublist for unassigned callbacks, leave.
1628 * At the same time, advance "i" one sublist, so that "i" will
1629 * index into the sublist where all the remaining callbacks should
1632 if (++i
>= RCU_NEXT_TAIL
)
1636 * Assign all subsequent callbacks' ->completed number to the next
1637 * full grace period and group them all in the sublist initially
1640 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1641 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1642 rdp
->nxtcompleted
[i
] = c
;
1644 /* Record any needed additional grace periods. */
1645 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1647 /* Trace depending on how much we were able to accelerate. */
1648 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1649 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1651 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1656 * Move any callbacks whose grace period has completed to the
1657 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1658 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1659 * sublist. This function is idempotent, so it does not hurt to
1660 * invoke it repeatedly. As long as it is not invoked -too- often...
1661 * Returns true if the RCU grace-period kthread needs to be awakened.
1663 * The caller must hold rnp->lock with interrupts disabled.
1665 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1666 struct rcu_data
*rdp
)
1670 /* If the CPU has no callbacks, nothing to do. */
1671 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1675 * Find all callbacks whose ->completed numbers indicate that they
1676 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1678 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1679 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1681 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1683 /* Clean up any sublist tail pointers that were misordered above. */
1684 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1685 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1687 /* Copy down callbacks to fill in empty sublists. */
1688 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1689 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1691 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1692 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1695 /* Classify any remaining callbacks. */
1696 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1700 * Update CPU-local rcu_data state to record the beginnings and ends of
1701 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1702 * structure corresponding to the current CPU, and must have irqs disabled.
1703 * Returns true if the grace-period kthread needs to be awakened.
1705 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1706 struct rcu_data
*rdp
)
1710 /* Handle the ends of any preceding grace periods first. */
1711 if (rdp
->completed
== rnp
->completed
&&
1712 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1714 /* No grace period end, so just accelerate recent callbacks. */
1715 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1719 /* Advance callbacks. */
1720 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1722 /* Remember that we saw this grace-period completion. */
1723 rdp
->completed
= rnp
->completed
;
1724 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1727 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1729 * If the current grace period is waiting for this CPU,
1730 * set up to detect a quiescent state, otherwise don't
1731 * go looking for one.
1733 rdp
->gpnum
= rnp
->gpnum
;
1734 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1735 rdp
->passed_quiesce
= 0;
1736 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1737 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1738 zero_cpu_stall_ticks(rdp
);
1739 WRITE_ONCE(rdp
->gpwrap
, false);
1744 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1746 unsigned long flags
;
1748 struct rcu_node
*rnp
;
1750 local_irq_save(flags
);
1752 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1753 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1754 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1755 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1756 local_irq_restore(flags
);
1759 smp_mb__after_unlock_lock();
1760 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1761 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1763 rcu_gp_kthread_wake(rsp
);
1766 static void rcu_gp_slow(struct rcu_state
*rsp
, int delay
)
1769 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1770 schedule_timeout_uninterruptible(delay
);
1774 * Initialize a new grace period. Return 0 if no grace period required.
1776 static int rcu_gp_init(struct rcu_state
*rsp
)
1778 unsigned long oldmask
;
1779 struct rcu_data
*rdp
;
1780 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1782 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1783 raw_spin_lock_irq(&rnp
->lock
);
1784 smp_mb__after_unlock_lock();
1785 if (!READ_ONCE(rsp
->gp_flags
)) {
1786 /* Spurious wakeup, tell caller to go back to sleep. */
1787 raw_spin_unlock_irq(&rnp
->lock
);
1790 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1792 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1794 * Grace period already in progress, don't start another.
1795 * Not supposed to be able to happen.
1797 raw_spin_unlock_irq(&rnp
->lock
);
1801 /* Advance to a new grace period and initialize state. */
1802 record_gp_stall_check_time(rsp
);
1803 /* Record GP times before starting GP, hence smp_store_release(). */
1804 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1805 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1806 raw_spin_unlock_irq(&rnp
->lock
);
1809 * Apply per-leaf buffered online and offline operations to the
1810 * rcu_node tree. Note that this new grace period need not wait
1811 * for subsequent online CPUs, and that quiescent-state forcing
1812 * will handle subsequent offline CPUs.
1814 rcu_for_each_leaf_node(rsp
, rnp
) {
1815 rcu_gp_slow(rsp
, gp_preinit_delay
);
1816 raw_spin_lock_irq(&rnp
->lock
);
1817 smp_mb__after_unlock_lock();
1818 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1819 !rnp
->wait_blkd_tasks
) {
1820 /* Nothing to do on this leaf rcu_node structure. */
1821 raw_spin_unlock_irq(&rnp
->lock
);
1825 /* Record old state, apply changes to ->qsmaskinit field. */
1826 oldmask
= rnp
->qsmaskinit
;
1827 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1829 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1830 if (!oldmask
!= !rnp
->qsmaskinit
) {
1831 if (!oldmask
) /* First online CPU for this rcu_node. */
1832 rcu_init_new_rnp(rnp
);
1833 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1834 rnp
->wait_blkd_tasks
= true;
1835 else /* Last offline CPU and can propagate. */
1836 rcu_cleanup_dead_rnp(rnp
);
1840 * If all waited-on tasks from prior grace period are
1841 * done, and if all this rcu_node structure's CPUs are
1842 * still offline, propagate up the rcu_node tree and
1843 * clear ->wait_blkd_tasks. Otherwise, if one of this
1844 * rcu_node structure's CPUs has since come back online,
1845 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1846 * checks for this, so just call it unconditionally).
1848 if (rnp
->wait_blkd_tasks
&&
1849 (!rcu_preempt_has_tasks(rnp
) ||
1851 rnp
->wait_blkd_tasks
= false;
1852 rcu_cleanup_dead_rnp(rnp
);
1855 raw_spin_unlock_irq(&rnp
->lock
);
1859 * Set the quiescent-state-needed bits in all the rcu_node
1860 * structures for all currently online CPUs in breadth-first order,
1861 * starting from the root rcu_node structure, relying on the layout
1862 * of the tree within the rsp->node[] array. Note that other CPUs
1863 * will access only the leaves of the hierarchy, thus seeing that no
1864 * grace period is in progress, at least until the corresponding
1865 * leaf node has been initialized. In addition, we have excluded
1866 * CPU-hotplug operations.
1868 * The grace period cannot complete until the initialization
1869 * process finishes, because this kthread handles both.
1871 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1872 rcu_gp_slow(rsp
, gp_init_delay
);
1873 raw_spin_lock_irq(&rnp
->lock
);
1874 smp_mb__after_unlock_lock();
1875 rdp
= this_cpu_ptr(rsp
->rda
);
1876 rcu_preempt_check_blocked_tasks(rnp
);
1877 rnp
->qsmask
= rnp
->qsmaskinit
;
1878 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1879 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1880 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1881 if (rnp
== rdp
->mynode
)
1882 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1883 rcu_preempt_boost_start_gp(rnp
);
1884 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1885 rnp
->level
, rnp
->grplo
,
1886 rnp
->grphi
, rnp
->qsmask
);
1887 raw_spin_unlock_irq(&rnp
->lock
);
1888 cond_resched_rcu_qs();
1889 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1896 * Do one round of quiescent-state forcing.
1898 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1900 int fqs_state
= fqs_state_in
;
1901 bool isidle
= false;
1903 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1905 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1907 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1908 /* Collect dyntick-idle snapshots. */
1909 if (is_sysidle_rcu_state(rsp
)) {
1911 maxj
= jiffies
- ULONG_MAX
/ 4;
1913 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1915 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1916 fqs_state
= RCU_FORCE_QS
;
1918 /* Handle dyntick-idle and offline CPUs. */
1920 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1922 /* Clear flag to prevent immediate re-entry. */
1923 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1924 raw_spin_lock_irq(&rnp
->lock
);
1925 smp_mb__after_unlock_lock();
1926 WRITE_ONCE(rsp
->gp_flags
,
1927 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
1928 raw_spin_unlock_irq(&rnp
->lock
);
1934 * Clean up after the old grace period.
1936 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1938 unsigned long gp_duration
;
1939 bool needgp
= false;
1941 struct rcu_data
*rdp
;
1942 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1944 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1945 raw_spin_lock_irq(&rnp
->lock
);
1946 smp_mb__after_unlock_lock();
1947 gp_duration
= jiffies
- rsp
->gp_start
;
1948 if (gp_duration
> rsp
->gp_max
)
1949 rsp
->gp_max
= gp_duration
;
1952 * We know the grace period is complete, but to everyone else
1953 * it appears to still be ongoing. But it is also the case
1954 * that to everyone else it looks like there is nothing that
1955 * they can do to advance the grace period. It is therefore
1956 * safe for us to drop the lock in order to mark the grace
1957 * period as completed in all of the rcu_node structures.
1959 raw_spin_unlock_irq(&rnp
->lock
);
1962 * Propagate new ->completed value to rcu_node structures so
1963 * that other CPUs don't have to wait until the start of the next
1964 * grace period to process their callbacks. This also avoids
1965 * some nasty RCU grace-period initialization races by forcing
1966 * the end of the current grace period to be completely recorded in
1967 * all of the rcu_node structures before the beginning of the next
1968 * grace period is recorded in any of the rcu_node structures.
1970 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1971 raw_spin_lock_irq(&rnp
->lock
);
1972 smp_mb__after_unlock_lock();
1973 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
1974 WARN_ON_ONCE(rnp
->qsmask
);
1975 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
1976 rdp
= this_cpu_ptr(rsp
->rda
);
1977 if (rnp
== rdp
->mynode
)
1978 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1979 /* smp_mb() provided by prior unlock-lock pair. */
1980 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1981 raw_spin_unlock_irq(&rnp
->lock
);
1982 cond_resched_rcu_qs();
1983 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1984 rcu_gp_slow(rsp
, gp_cleanup_delay
);
1986 rnp
= rcu_get_root(rsp
);
1987 raw_spin_lock_irq(&rnp
->lock
);
1988 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1989 rcu_nocb_gp_set(rnp
, nocb
);
1991 /* Declare grace period done. */
1992 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
1993 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
1994 rsp
->fqs_state
= RCU_GP_IDLE
;
1995 rdp
= this_cpu_ptr(rsp
->rda
);
1996 /* Advance CBs to reduce false positives below. */
1997 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
1998 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
1999 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2000 trace_rcu_grace_period(rsp
->name
,
2001 READ_ONCE(rsp
->gpnum
),
2004 raw_spin_unlock_irq(&rnp
->lock
);
2008 * Body of kthread that handles grace periods.
2010 static int __noreturn
rcu_gp_kthread(void *arg
)
2016 struct rcu_state
*rsp
= arg
;
2017 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2019 rcu_bind_gp_kthread();
2022 /* Handle grace-period start. */
2024 trace_rcu_grace_period(rsp
->name
,
2025 READ_ONCE(rsp
->gpnum
),
2027 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2028 wait_event_interruptible(rsp
->gp_wq
,
2029 READ_ONCE(rsp
->gp_flags
) &
2031 /* Locking provides needed memory barrier. */
2032 if (rcu_gp_init(rsp
))
2034 cond_resched_rcu_qs();
2035 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2036 WARN_ON(signal_pending(current
));
2037 trace_rcu_grace_period(rsp
->name
,
2038 READ_ONCE(rsp
->gpnum
),
2042 /* Handle quiescent-state forcing. */
2043 fqs_state
= RCU_SAVE_DYNTICK
;
2044 j
= jiffies_till_first_fqs
;
2047 jiffies_till_first_fqs
= HZ
;
2052 rsp
->jiffies_force_qs
= jiffies
+ j
;
2053 trace_rcu_grace_period(rsp
->name
,
2054 READ_ONCE(rsp
->gpnum
),
2056 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2057 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
2058 ((gf
= READ_ONCE(rsp
->gp_flags
)) &
2060 (!READ_ONCE(rnp
->qsmask
) &&
2061 !rcu_preempt_blocked_readers_cgp(rnp
)),
2063 /* Locking provides needed memory barriers. */
2064 /* If grace period done, leave loop. */
2065 if (!READ_ONCE(rnp
->qsmask
) &&
2066 !rcu_preempt_blocked_readers_cgp(rnp
))
2068 /* If time for quiescent-state forcing, do it. */
2069 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2070 (gf
& RCU_GP_FLAG_FQS
)) {
2071 trace_rcu_grace_period(rsp
->name
,
2072 READ_ONCE(rsp
->gpnum
),
2074 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
2075 trace_rcu_grace_period(rsp
->name
,
2076 READ_ONCE(rsp
->gpnum
),
2078 cond_resched_rcu_qs();
2079 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2081 /* Deal with stray signal. */
2082 cond_resched_rcu_qs();
2083 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2084 WARN_ON(signal_pending(current
));
2085 trace_rcu_grace_period(rsp
->name
,
2086 READ_ONCE(rsp
->gpnum
),
2089 j
= jiffies_till_next_fqs
;
2092 jiffies_till_next_fqs
= HZ
;
2095 jiffies_till_next_fqs
= 1;
2099 /* Handle grace-period end. */
2100 rcu_gp_cleanup(rsp
);
2105 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2106 * in preparation for detecting the next grace period. The caller must hold
2107 * the root node's ->lock and hard irqs must be disabled.
2109 * Note that it is legal for a dying CPU (which is marked as offline) to
2110 * invoke this function. This can happen when the dying CPU reports its
2113 * Returns true if the grace-period kthread must be awakened.
2116 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2117 struct rcu_data
*rdp
)
2119 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2121 * Either we have not yet spawned the grace-period
2122 * task, this CPU does not need another grace period,
2123 * or a grace period is already in progress.
2124 * Either way, don't start a new grace period.
2128 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2129 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2133 * We can't do wakeups while holding the rnp->lock, as that
2134 * could cause possible deadlocks with the rq->lock. Defer
2135 * the wakeup to our caller.
2141 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2142 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2143 * is invoked indirectly from rcu_advance_cbs(), which would result in
2144 * endless recursion -- or would do so if it wasn't for the self-deadlock
2145 * that is encountered beforehand.
2147 * Returns true if the grace-period kthread needs to be awakened.
2149 static bool rcu_start_gp(struct rcu_state
*rsp
)
2151 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2152 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2156 * If there is no grace period in progress right now, any
2157 * callbacks we have up to this point will be satisfied by the
2158 * next grace period. Also, advancing the callbacks reduces the
2159 * probability of false positives from cpu_needs_another_gp()
2160 * resulting in pointless grace periods. So, advance callbacks
2161 * then start the grace period!
2163 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2164 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2169 * Report a full set of quiescent states to the specified rcu_state
2170 * data structure. This involves cleaning up after the prior grace
2171 * period and letting rcu_start_gp() start up the next grace period
2172 * if one is needed. Note that the caller must hold rnp->lock, which
2173 * is released before return.
2175 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2176 __releases(rcu_get_root(rsp
)->lock
)
2178 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2179 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2180 rcu_gp_kthread_wake(rsp
);
2184 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2185 * Allows quiescent states for a group of CPUs to be reported at one go
2186 * to the specified rcu_node structure, though all the CPUs in the group
2187 * must be represented by the same rcu_node structure (which need not be a
2188 * leaf rcu_node structure, though it often will be). The gps parameter
2189 * is the grace-period snapshot, which means that the quiescent states
2190 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2191 * must be held upon entry, and it is released before return.
2194 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2195 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2196 __releases(rnp
->lock
)
2198 unsigned long oldmask
= 0;
2199 struct rcu_node
*rnp_c
;
2201 /* Walk up the rcu_node hierarchy. */
2203 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2206 * Our bit has already been cleared, or the
2207 * relevant grace period is already over, so done.
2209 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2212 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2213 rnp
->qsmask
&= ~mask
;
2214 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2215 mask
, rnp
->qsmask
, rnp
->level
,
2216 rnp
->grplo
, rnp
->grphi
,
2218 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2220 /* Other bits still set at this level, so done. */
2221 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2224 mask
= rnp
->grpmask
;
2225 if (rnp
->parent
== NULL
) {
2227 /* No more levels. Exit loop holding root lock. */
2231 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2234 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2235 smp_mb__after_unlock_lock();
2236 oldmask
= rnp_c
->qsmask
;
2240 * Get here if we are the last CPU to pass through a quiescent
2241 * state for this grace period. Invoke rcu_report_qs_rsp()
2242 * to clean up and start the next grace period if one is needed.
2244 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2248 * Record a quiescent state for all tasks that were previously queued
2249 * on the specified rcu_node structure and that were blocking the current
2250 * RCU grace period. The caller must hold the specified rnp->lock with
2251 * irqs disabled, and this lock is released upon return, but irqs remain
2254 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2255 struct rcu_node
*rnp
, unsigned long flags
)
2256 __releases(rnp
->lock
)
2260 struct rcu_node
*rnp_p
;
2262 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2263 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2264 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2265 return; /* Still need more quiescent states! */
2268 rnp_p
= rnp
->parent
;
2269 if (rnp_p
== NULL
) {
2271 * Only one rcu_node structure in the tree, so don't
2272 * try to report up to its nonexistent parent!
2274 rcu_report_qs_rsp(rsp
, flags
);
2278 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2280 mask
= rnp
->grpmask
;
2281 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2282 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
2283 smp_mb__after_unlock_lock();
2284 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2288 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2289 * structure. This must be either called from the specified CPU, or
2290 * called when the specified CPU is known to be offline (and when it is
2291 * also known that no other CPU is concurrently trying to help the offline
2292 * CPU). The lastcomp argument is used to make sure we are still in the
2293 * grace period of interest. We don't want to end the current grace period
2294 * based on quiescent states detected in an earlier grace period!
2297 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2299 unsigned long flags
;
2302 struct rcu_node
*rnp
;
2305 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2306 smp_mb__after_unlock_lock();
2307 if ((rdp
->passed_quiesce
== 0 &&
2308 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2309 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2313 * The grace period in which this quiescent state was
2314 * recorded has ended, so don't report it upwards.
2315 * We will instead need a new quiescent state that lies
2316 * within the current grace period.
2318 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2319 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2320 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2323 mask
= rdp
->grpmask
;
2324 if ((rnp
->qsmask
& mask
) == 0) {
2325 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2327 rdp
->qs_pending
= 0;
2330 * This GP can't end until cpu checks in, so all of our
2331 * callbacks can be processed during the next GP.
2333 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2335 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2336 /* ^^^ Released rnp->lock */
2338 rcu_gp_kthread_wake(rsp
);
2343 * Check to see if there is a new grace period of which this CPU
2344 * is not yet aware, and if so, set up local rcu_data state for it.
2345 * Otherwise, see if this CPU has just passed through its first
2346 * quiescent state for this grace period, and record that fact if so.
2349 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2351 /* Check for grace-period ends and beginnings. */
2352 note_gp_changes(rsp
, rdp
);
2355 * Does this CPU still need to do its part for current grace period?
2356 * If no, return and let the other CPUs do their part as well.
2358 if (!rdp
->qs_pending
)
2362 * Was there a quiescent state since the beginning of the grace
2363 * period? If no, then exit and wait for the next call.
2365 if (!rdp
->passed_quiesce
&&
2366 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2370 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2373 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2376 #ifdef CONFIG_HOTPLUG_CPU
2379 * Send the specified CPU's RCU callbacks to the orphanage. The
2380 * specified CPU must be offline, and the caller must hold the
2384 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2385 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2387 /* No-CBs CPUs do not have orphanable callbacks. */
2388 if (rcu_is_nocb_cpu(rdp
->cpu
))
2392 * Orphan the callbacks. First adjust the counts. This is safe
2393 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2394 * cannot be running now. Thus no memory barrier is required.
2396 if (rdp
->nxtlist
!= NULL
) {
2397 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2398 rsp
->qlen
+= rdp
->qlen
;
2399 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2401 WRITE_ONCE(rdp
->qlen
, 0);
2405 * Next, move those callbacks still needing a grace period to
2406 * the orphanage, where some other CPU will pick them up.
2407 * Some of the callbacks might have gone partway through a grace
2408 * period, but that is too bad. They get to start over because we
2409 * cannot assume that grace periods are synchronized across CPUs.
2410 * We don't bother updating the ->nxttail[] array yet, instead
2411 * we just reset the whole thing later on.
2413 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2414 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2415 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2416 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2420 * Then move the ready-to-invoke callbacks to the orphanage,
2421 * where some other CPU will pick them up. These will not be
2422 * required to pass though another grace period: They are done.
2424 if (rdp
->nxtlist
!= NULL
) {
2425 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2426 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2430 * Finally, initialize the rcu_data structure's list to empty and
2431 * disallow further callbacks on this CPU.
2433 init_callback_list(rdp
);
2434 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2438 * Adopt the RCU callbacks from the specified rcu_state structure's
2439 * orphanage. The caller must hold the ->orphan_lock.
2441 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2444 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2446 /* No-CBs CPUs are handled specially. */
2447 if (rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2450 /* Do the accounting first. */
2451 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2452 rdp
->qlen
+= rsp
->qlen
;
2453 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2454 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2455 rcu_idle_count_callbacks_posted();
2460 * We do not need a memory barrier here because the only way we
2461 * can get here if there is an rcu_barrier() in flight is if
2462 * we are the task doing the rcu_barrier().
2465 /* First adopt the ready-to-invoke callbacks. */
2466 if (rsp
->orphan_donelist
!= NULL
) {
2467 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2468 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2469 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2470 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2471 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2472 rsp
->orphan_donelist
= NULL
;
2473 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2476 /* And then adopt the callbacks that still need a grace period. */
2477 if (rsp
->orphan_nxtlist
!= NULL
) {
2478 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2479 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2480 rsp
->orphan_nxtlist
= NULL
;
2481 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2486 * Trace the fact that this CPU is going offline.
2488 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2490 RCU_TRACE(unsigned long mask
);
2491 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2492 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2494 RCU_TRACE(mask
= rdp
->grpmask
);
2495 trace_rcu_grace_period(rsp
->name
,
2496 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2501 * All CPUs for the specified rcu_node structure have gone offline,
2502 * and all tasks that were preempted within an RCU read-side critical
2503 * section while running on one of those CPUs have since exited their RCU
2504 * read-side critical section. Some other CPU is reporting this fact with
2505 * the specified rcu_node structure's ->lock held and interrupts disabled.
2506 * This function therefore goes up the tree of rcu_node structures,
2507 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2508 * the leaf rcu_node structure's ->qsmaskinit field has already been
2511 * This function does check that the specified rcu_node structure has
2512 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2513 * prematurely. That said, invoking it after the fact will cost you
2514 * a needless lock acquisition. So once it has done its work, don't
2517 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2520 struct rcu_node
*rnp
= rnp_leaf
;
2522 if (rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2525 mask
= rnp
->grpmask
;
2529 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2530 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2531 rnp
->qsmaskinit
&= ~mask
;
2532 rnp
->qsmask
&= ~mask
;
2533 if (rnp
->qsmaskinit
) {
2534 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2537 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2542 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2543 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2546 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2548 unsigned long flags
;
2550 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2551 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2553 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2554 mask
= rdp
->grpmask
;
2555 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2556 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2557 rnp
->qsmaskinitnext
&= ~mask
;
2558 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2562 * The CPU has been completely removed, and some other CPU is reporting
2563 * this fact from process context. Do the remainder of the cleanup,
2564 * including orphaning the outgoing CPU's RCU callbacks, and also
2565 * adopting them. There can only be one CPU hotplug operation at a time,
2566 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2568 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2570 unsigned long flags
;
2571 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2572 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2574 /* Adjust any no-longer-needed kthreads. */
2575 rcu_boost_kthread_setaffinity(rnp
, -1);
2577 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2578 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2579 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2580 rcu_adopt_orphan_cbs(rsp
, flags
);
2581 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2583 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2584 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2585 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2588 #else /* #ifdef CONFIG_HOTPLUG_CPU */
2590 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2594 static void __maybe_unused
rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2598 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2602 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2606 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2609 * Invoke any RCU callbacks that have made it to the end of their grace
2610 * period. Thottle as specified by rdp->blimit.
2612 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2614 unsigned long flags
;
2615 struct rcu_head
*next
, *list
, **tail
;
2616 long bl
, count
, count_lazy
;
2619 /* If no callbacks are ready, just return. */
2620 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2621 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2622 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2623 need_resched(), is_idle_task(current
),
2624 rcu_is_callbacks_kthread());
2629 * Extract the list of ready callbacks, disabling to prevent
2630 * races with call_rcu() from interrupt handlers.
2632 local_irq_save(flags
);
2633 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2635 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2636 list
= rdp
->nxtlist
;
2637 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2638 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2639 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2640 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2641 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2642 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2643 local_irq_restore(flags
);
2645 /* Invoke callbacks. */
2646 count
= count_lazy
= 0;
2650 debug_rcu_head_unqueue(list
);
2651 if (__rcu_reclaim(rsp
->name
, list
))
2654 /* Stop only if limit reached and CPU has something to do. */
2655 if (++count
>= bl
&&
2657 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2661 local_irq_save(flags
);
2662 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2663 is_idle_task(current
),
2664 rcu_is_callbacks_kthread());
2666 /* Update count, and requeue any remaining callbacks. */
2668 *tail
= rdp
->nxtlist
;
2669 rdp
->nxtlist
= list
;
2670 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2671 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2672 rdp
->nxttail
[i
] = tail
;
2676 smp_mb(); /* List handling before counting for rcu_barrier(). */
2677 rdp
->qlen_lazy
-= count_lazy
;
2678 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2679 rdp
->n_cbs_invoked
+= count
;
2681 /* Reinstate batch limit if we have worked down the excess. */
2682 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2683 rdp
->blimit
= blimit
;
2685 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2686 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2687 rdp
->qlen_last_fqs_check
= 0;
2688 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2689 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2690 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2691 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2693 local_irq_restore(flags
);
2695 /* Re-invoke RCU core processing if there are callbacks remaining. */
2696 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2701 * Check to see if this CPU is in a non-context-switch quiescent state
2702 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2703 * Also schedule RCU core processing.
2705 * This function must be called from hardirq context. It is normally
2706 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2707 * false, there is no point in invoking rcu_check_callbacks().
2709 void rcu_check_callbacks(int user
)
2711 trace_rcu_utilization(TPS("Start scheduler-tick"));
2712 increment_cpu_stall_ticks();
2713 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2716 * Get here if this CPU took its interrupt from user
2717 * mode or from the idle loop, and if this is not a
2718 * nested interrupt. In this case, the CPU is in
2719 * a quiescent state, so note it.
2721 * No memory barrier is required here because both
2722 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2723 * variables that other CPUs neither access nor modify,
2724 * at least not while the corresponding CPU is online.
2730 } else if (!in_softirq()) {
2733 * Get here if this CPU did not take its interrupt from
2734 * softirq, in other words, if it is not interrupting
2735 * a rcu_bh read-side critical section. This is an _bh
2736 * critical section, so note it.
2741 rcu_preempt_check_callbacks();
2745 rcu_note_voluntary_context_switch(current
);
2746 trace_rcu_utilization(TPS("End scheduler-tick"));
2750 * Scan the leaf rcu_node structures, processing dyntick state for any that
2751 * have not yet encountered a quiescent state, using the function specified.
2752 * Also initiate boosting for any threads blocked on the root rcu_node.
2754 * The caller must have suppressed start of new grace periods.
2756 static void force_qs_rnp(struct rcu_state
*rsp
,
2757 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2758 unsigned long *maxj
),
2759 bool *isidle
, unsigned long *maxj
)
2763 unsigned long flags
;
2765 struct rcu_node
*rnp
;
2767 rcu_for_each_leaf_node(rsp
, rnp
) {
2768 cond_resched_rcu_qs();
2770 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2771 smp_mb__after_unlock_lock();
2772 if (!rcu_gp_in_progress(rsp
)) {
2773 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2776 if (rnp
->qsmask
== 0) {
2777 if (rcu_state_p
== &rcu_sched_state
||
2778 rsp
!= rcu_state_p
||
2779 rcu_preempt_blocked_readers_cgp(rnp
)) {
2781 * No point in scanning bits because they
2782 * are all zero. But we might need to
2783 * priority-boost blocked readers.
2785 rcu_initiate_boost(rnp
, flags
);
2786 /* rcu_initiate_boost() releases rnp->lock */
2790 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2792 * Race between grace-period
2793 * initialization and task exiting RCU
2794 * read-side critical section: Report.
2796 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2797 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2803 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2804 if ((rnp
->qsmask
& bit
) != 0) {
2805 if ((rnp
->qsmaskinit
& bit
) == 0)
2806 *isidle
= false; /* Pending hotplug. */
2807 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2812 /* Idle/offline CPUs, report (releases rnp->lock. */
2813 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2815 /* Nothing to do here, so just drop the lock. */
2816 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2822 * Force quiescent states on reluctant CPUs, and also detect which
2823 * CPUs are in dyntick-idle mode.
2825 static void force_quiescent_state(struct rcu_state
*rsp
)
2827 unsigned long flags
;
2829 struct rcu_node
*rnp
;
2830 struct rcu_node
*rnp_old
= NULL
;
2832 /* Funnel through hierarchy to reduce memory contention. */
2833 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2834 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2835 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2836 !raw_spin_trylock(&rnp
->fqslock
);
2837 if (rnp_old
!= NULL
)
2838 raw_spin_unlock(&rnp_old
->fqslock
);
2840 rsp
->n_force_qs_lh
++;
2845 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2847 /* Reached the root of the rcu_node tree, acquire lock. */
2848 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2849 smp_mb__after_unlock_lock();
2850 raw_spin_unlock(&rnp_old
->fqslock
);
2851 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2852 rsp
->n_force_qs_lh
++;
2853 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2854 return; /* Someone beat us to it. */
2856 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2857 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2858 rcu_gp_kthread_wake(rsp
);
2862 * This does the RCU core processing work for the specified rcu_state
2863 * and rcu_data structures. This may be called only from the CPU to
2864 * whom the rdp belongs.
2867 __rcu_process_callbacks(struct rcu_state
*rsp
)
2869 unsigned long flags
;
2871 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2873 WARN_ON_ONCE(rdp
->beenonline
== 0);
2875 /* Update RCU state based on any recent quiescent states. */
2876 rcu_check_quiescent_state(rsp
, rdp
);
2878 /* Does this CPU require a not-yet-started grace period? */
2879 local_irq_save(flags
);
2880 if (cpu_needs_another_gp(rsp
, rdp
)) {
2881 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2882 needwake
= rcu_start_gp(rsp
);
2883 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2885 rcu_gp_kthread_wake(rsp
);
2887 local_irq_restore(flags
);
2890 /* If there are callbacks ready, invoke them. */
2891 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2892 invoke_rcu_callbacks(rsp
, rdp
);
2894 /* Do any needed deferred wakeups of rcuo kthreads. */
2895 do_nocb_deferred_wakeup(rdp
);
2899 * Do RCU core processing for the current CPU.
2901 static void rcu_process_callbacks(struct softirq_action
*unused
)
2903 struct rcu_state
*rsp
;
2905 if (cpu_is_offline(smp_processor_id()))
2907 trace_rcu_utilization(TPS("Start RCU core"));
2908 for_each_rcu_flavor(rsp
)
2909 __rcu_process_callbacks(rsp
);
2910 trace_rcu_utilization(TPS("End RCU core"));
2914 * Schedule RCU callback invocation. If the specified type of RCU
2915 * does not support RCU priority boosting, just do a direct call,
2916 * otherwise wake up the per-CPU kernel kthread. Note that because we
2917 * are running on the current CPU with softirqs disabled, the
2918 * rcu_cpu_kthread_task cannot disappear out from under us.
2920 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2922 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2924 if (likely(!rsp
->boost
)) {
2925 rcu_do_batch(rsp
, rdp
);
2928 invoke_rcu_callbacks_kthread();
2931 static void invoke_rcu_core(void)
2933 if (cpu_online(smp_processor_id()))
2934 raise_softirq(RCU_SOFTIRQ
);
2938 * Handle any core-RCU processing required by a call_rcu() invocation.
2940 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2941 struct rcu_head
*head
, unsigned long flags
)
2946 * If called from an extended quiescent state, invoke the RCU
2947 * core in order to force a re-evaluation of RCU's idleness.
2949 if (!rcu_is_watching())
2952 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2953 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2957 * Force the grace period if too many callbacks or too long waiting.
2958 * Enforce hysteresis, and don't invoke force_quiescent_state()
2959 * if some other CPU has recently done so. Also, don't bother
2960 * invoking force_quiescent_state() if the newly enqueued callback
2961 * is the only one waiting for a grace period to complete.
2963 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2965 /* Are we ignoring a completed grace period? */
2966 note_gp_changes(rsp
, rdp
);
2968 /* Start a new grace period if one not already started. */
2969 if (!rcu_gp_in_progress(rsp
)) {
2970 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2972 raw_spin_lock(&rnp_root
->lock
);
2973 smp_mb__after_unlock_lock();
2974 needwake
= rcu_start_gp(rsp
);
2975 raw_spin_unlock(&rnp_root
->lock
);
2977 rcu_gp_kthread_wake(rsp
);
2979 /* Give the grace period a kick. */
2980 rdp
->blimit
= LONG_MAX
;
2981 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2982 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2983 force_quiescent_state(rsp
);
2984 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2985 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2991 * RCU callback function to leak a callback.
2993 static void rcu_leak_callback(struct rcu_head
*rhp
)
2998 * Helper function for call_rcu() and friends. The cpu argument will
2999 * normally be -1, indicating "currently running CPU". It may specify
3000 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3001 * is expected to specify a CPU.
3004 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
3005 struct rcu_state
*rsp
, int cpu
, bool lazy
)
3007 unsigned long flags
;
3008 struct rcu_data
*rdp
;
3010 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
3011 if (debug_rcu_head_queue(head
)) {
3012 /* Probable double call_rcu(), so leak the callback. */
3013 WRITE_ONCE(head
->func
, rcu_leak_callback
);
3014 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3021 * Opportunistically note grace-period endings and beginnings.
3022 * Note that we might see a beginning right after we see an
3023 * end, but never vice versa, since this CPU has to pass through
3024 * a quiescent state betweentimes.
3026 local_irq_save(flags
);
3027 rdp
= this_cpu_ptr(rsp
->rda
);
3029 /* Add the callback to our list. */
3030 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3034 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3035 if (likely(rdp
->mynode
)) {
3036 /* Post-boot, so this should be for a no-CBs CPU. */
3037 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3038 WARN_ON_ONCE(offline
);
3039 /* Offline CPU, _call_rcu() illegal, leak callback. */
3040 local_irq_restore(flags
);
3044 * Very early boot, before rcu_init(). Initialize if needed
3045 * and then drop through to queue the callback.
3048 WARN_ON_ONCE(!rcu_is_watching());
3049 if (!likely(rdp
->nxtlist
))
3050 init_default_callback_list(rdp
);
3052 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3056 rcu_idle_count_callbacks_posted();
3057 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3058 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3059 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3061 if (__is_kfree_rcu_offset((unsigned long)func
))
3062 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3063 rdp
->qlen_lazy
, rdp
->qlen
);
3065 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3067 /* Go handle any RCU core processing required. */
3068 __call_rcu_core(rsp
, rdp
, head
, flags
);
3069 local_irq_restore(flags
);
3073 * Queue an RCU-sched callback for invocation after a grace period.
3075 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3077 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3079 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3082 * Queue an RCU callback for invocation after a quicker grace period.
3084 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3086 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3088 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3091 * Queue an RCU callback for lazy invocation after a grace period.
3092 * This will likely be later named something like "call_rcu_lazy()",
3093 * but this change will require some way of tagging the lazy RCU
3094 * callbacks in the list of pending callbacks. Until then, this
3095 * function may only be called from __kfree_rcu().
3097 void kfree_call_rcu(struct rcu_head
*head
,
3098 void (*func
)(struct rcu_head
*rcu
))
3100 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3102 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3105 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3106 * any blocking grace-period wait automatically implies a grace period
3107 * if there is only one CPU online at any point time during execution
3108 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3109 * occasionally incorrectly indicate that there are multiple CPUs online
3110 * when there was in fact only one the whole time, as this just adds
3111 * some overhead: RCU still operates correctly.
3113 static inline int rcu_blocking_is_gp(void)
3117 might_sleep(); /* Check for RCU read-side critical section. */
3119 ret
= num_online_cpus() <= 1;
3125 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3127 * Control will return to the caller some time after a full rcu-sched
3128 * grace period has elapsed, in other words after all currently executing
3129 * rcu-sched read-side critical sections have completed. These read-side
3130 * critical sections are delimited by rcu_read_lock_sched() and
3131 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3132 * local_irq_disable(), and so on may be used in place of
3133 * rcu_read_lock_sched().
3135 * This means that all preempt_disable code sequences, including NMI and
3136 * non-threaded hardware-interrupt handlers, in progress on entry will
3137 * have completed before this primitive returns. However, this does not
3138 * guarantee that softirq handlers will have completed, since in some
3139 * kernels, these handlers can run in process context, and can block.
3141 * Note that this guarantee implies further memory-ordering guarantees.
3142 * On systems with more than one CPU, when synchronize_sched() returns,
3143 * each CPU is guaranteed to have executed a full memory barrier since the
3144 * end of its last RCU-sched read-side critical section whose beginning
3145 * preceded the call to synchronize_sched(). In addition, each CPU having
3146 * an RCU read-side critical section that extends beyond the return from
3147 * synchronize_sched() is guaranteed to have executed a full memory barrier
3148 * after the beginning of synchronize_sched() and before the beginning of
3149 * that RCU read-side critical section. Note that these guarantees include
3150 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3151 * that are executing in the kernel.
3153 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3154 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3155 * to have executed a full memory barrier during the execution of
3156 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3157 * again only if the system has more than one CPU).
3159 * This primitive provides the guarantees made by the (now removed)
3160 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3161 * guarantees that rcu_read_lock() sections will have completed.
3162 * In "classic RCU", these two guarantees happen to be one and
3163 * the same, but can differ in realtime RCU implementations.
3165 void synchronize_sched(void)
3167 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3168 !lock_is_held(&rcu_lock_map
) &&
3169 !lock_is_held(&rcu_sched_lock_map
),
3170 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3171 if (rcu_blocking_is_gp())
3173 if (rcu_gp_is_expedited())
3174 synchronize_sched_expedited();
3176 wait_rcu_gp(call_rcu_sched
);
3178 EXPORT_SYMBOL_GPL(synchronize_sched
);
3181 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3183 * Control will return to the caller some time after a full rcu_bh grace
3184 * period has elapsed, in other words after all currently executing rcu_bh
3185 * read-side critical sections have completed. RCU read-side critical
3186 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3187 * and may be nested.
3189 * See the description of synchronize_sched() for more detailed information
3190 * on memory ordering guarantees.
3192 void synchronize_rcu_bh(void)
3194 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3195 !lock_is_held(&rcu_lock_map
) &&
3196 !lock_is_held(&rcu_sched_lock_map
),
3197 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3198 if (rcu_blocking_is_gp())
3200 if (rcu_gp_is_expedited())
3201 synchronize_rcu_bh_expedited();
3203 wait_rcu_gp(call_rcu_bh
);
3205 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3208 * get_state_synchronize_rcu - Snapshot current RCU state
3210 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3211 * to determine whether or not a full grace period has elapsed in the
3214 unsigned long get_state_synchronize_rcu(void)
3217 * Any prior manipulation of RCU-protected data must happen
3218 * before the load from ->gpnum.
3223 * Make sure this load happens before the purportedly
3224 * time-consuming work between get_state_synchronize_rcu()
3225 * and cond_synchronize_rcu().
3227 return smp_load_acquire(&rcu_state_p
->gpnum
);
3229 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3232 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3234 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3236 * If a full RCU grace period has elapsed since the earlier call to
3237 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3238 * synchronize_rcu() to wait for a full grace period.
3240 * Yes, this function does not take counter wrap into account. But
3241 * counter wrap is harmless. If the counter wraps, we have waited for
3242 * more than 2 billion grace periods (and way more on a 64-bit system!),
3243 * so waiting for one additional grace period should be just fine.
3245 void cond_synchronize_rcu(unsigned long oldstate
)
3247 unsigned long newstate
;
3250 * Ensure that this load happens before any RCU-destructive
3251 * actions the caller might carry out after we return.
3253 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3254 if (ULONG_CMP_GE(oldstate
, newstate
))
3257 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3259 static int synchronize_sched_expedited_cpu_stop(void *data
)
3262 * There must be a full memory barrier on each affected CPU
3263 * between the time that try_stop_cpus() is called and the
3264 * time that it returns.
3266 * In the current initial implementation of cpu_stop, the
3267 * above condition is already met when the control reaches
3268 * this point and the following smp_mb() is not strictly
3269 * necessary. Do smp_mb() anyway for documentation and
3270 * robustness against future implementation changes.
3272 smp_mb(); /* See above comment block. */
3277 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3279 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3280 * approach to force the grace period to end quickly. This consumes
3281 * significant time on all CPUs and is unfriendly to real-time workloads,
3282 * so is thus not recommended for any sort of common-case code. In fact,
3283 * if you are using synchronize_sched_expedited() in a loop, please
3284 * restructure your code to batch your updates, and then use a single
3285 * synchronize_sched() instead.
3287 * This implementation can be thought of as an application of ticket
3288 * locking to RCU, with sync_sched_expedited_started and
3289 * sync_sched_expedited_done taking on the roles of the halves
3290 * of the ticket-lock word. Each task atomically increments
3291 * sync_sched_expedited_started upon entry, snapshotting the old value,
3292 * then attempts to stop all the CPUs. If this succeeds, then each
3293 * CPU will have executed a context switch, resulting in an RCU-sched
3294 * grace period. We are then done, so we use atomic_cmpxchg() to
3295 * update sync_sched_expedited_done to match our snapshot -- but
3296 * only if someone else has not already advanced past our snapshot.
3298 * On the other hand, if try_stop_cpus() fails, we check the value
3299 * of sync_sched_expedited_done. If it has advanced past our
3300 * initial snapshot, then someone else must have forced a grace period
3301 * some time after we took our snapshot. In this case, our work is
3302 * done for us, and we can simply return. Otherwise, we try again,
3303 * but keep our initial snapshot for purposes of checking for someone
3304 * doing our work for us.
3306 * If we fail too many times in a row, we fall back to synchronize_sched().
3308 void synchronize_sched_expedited(void)
3313 long firstsnap
, s
, snap
;
3315 struct rcu_state
*rsp
= &rcu_sched_state
;
3318 * If we are in danger of counter wrap, just do synchronize_sched().
3319 * By allowing sync_sched_expedited_started to advance no more than
3320 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3321 * that more than 3.5 billion CPUs would be required to force a
3322 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3323 * course be required on a 64-bit system.
3325 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3326 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3328 synchronize_sched();
3329 atomic_long_inc(&rsp
->expedited_wrap
);
3334 * Take a ticket. Note that atomic_inc_return() implies a
3335 * full memory barrier.
3337 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3339 if (!try_get_online_cpus()) {
3340 /* CPU hotplug operation in flight, fall back to normal GP. */
3341 wait_rcu_gp(call_rcu_sched
);
3342 atomic_long_inc(&rsp
->expedited_normal
);
3345 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3347 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3348 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3350 cpumask_copy(cm
, cpu_online_mask
);
3351 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3352 for_each_cpu(cpu
, cm
) {
3353 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3355 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3356 cpumask_clear_cpu(cpu
, cm
);
3358 if (cpumask_weight(cm
) == 0)
3363 * Each pass through the following loop attempts to force a
3364 * context switch on each CPU.
3366 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3367 synchronize_sched_expedited_cpu_stop
,
3370 atomic_long_inc(&rsp
->expedited_tryfail
);
3372 /* Check to see if someone else did our work for us. */
3373 s
= atomic_long_read(&rsp
->expedited_done
);
3374 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3375 /* ensure test happens before caller kfree */
3376 smp_mb__before_atomic(); /* ^^^ */
3377 atomic_long_inc(&rsp
->expedited_workdone1
);
3378 free_cpumask_var(cm
);
3382 /* No joy, try again later. Or just synchronize_sched(). */
3383 if (trycount
++ < 10) {
3384 udelay(trycount
* num_online_cpus());
3386 wait_rcu_gp(call_rcu_sched
);
3387 atomic_long_inc(&rsp
->expedited_normal
);
3388 free_cpumask_var(cm
);
3392 /* Recheck to see if someone else did our work for us. */
3393 s
= atomic_long_read(&rsp
->expedited_done
);
3394 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3395 /* ensure test happens before caller kfree */
3396 smp_mb__before_atomic(); /* ^^^ */
3397 atomic_long_inc(&rsp
->expedited_workdone2
);
3398 free_cpumask_var(cm
);
3403 * Refetching sync_sched_expedited_started allows later
3404 * callers to piggyback on our grace period. We retry
3405 * after they started, so our grace period works for them,
3406 * and they started after our first try, so their grace
3407 * period works for us.
3409 if (!try_get_online_cpus()) {
3410 /* CPU hotplug operation in flight, use normal GP. */
3411 wait_rcu_gp(call_rcu_sched
);
3412 atomic_long_inc(&rsp
->expedited_normal
);
3413 free_cpumask_var(cm
);
3416 snap
= atomic_long_read(&rsp
->expedited_start
);
3417 smp_mb(); /* ensure read is before try_stop_cpus(). */
3419 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3422 free_cpumask_var(cm
);
3425 * Everyone up to our most recent fetch is covered by our grace
3426 * period. Update the counter, but only if our work is still
3427 * relevant -- which it won't be if someone who started later
3428 * than we did already did their update.
3431 atomic_long_inc(&rsp
->expedited_done_tries
);
3432 s
= atomic_long_read(&rsp
->expedited_done
);
3433 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3434 /* ensure test happens before caller kfree */
3435 smp_mb__before_atomic(); /* ^^^ */
3436 atomic_long_inc(&rsp
->expedited_done_lost
);
3439 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3440 atomic_long_inc(&rsp
->expedited_done_exit
);
3444 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3447 * Check to see if there is any immediate RCU-related work to be done
3448 * by the current CPU, for the specified type of RCU, returning 1 if so.
3449 * The checks are in order of increasing expense: checks that can be
3450 * carried out against CPU-local state are performed first. However,
3451 * we must check for CPU stalls first, else we might not get a chance.
3453 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3455 struct rcu_node
*rnp
= rdp
->mynode
;
3457 rdp
->n_rcu_pending
++;
3459 /* Check for CPU stalls, if enabled. */
3460 check_cpu_stall(rsp
, rdp
);
3462 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3463 if (rcu_nohz_full_cpu(rsp
))
3466 /* Is the RCU core waiting for a quiescent state from this CPU? */
3467 if (rcu_scheduler_fully_active
&&
3468 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3469 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3470 rdp
->n_rp_qs_pending
++;
3471 } else if (rdp
->qs_pending
&&
3472 (rdp
->passed_quiesce
||
3473 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3474 rdp
->n_rp_report_qs
++;
3478 /* Does this CPU have callbacks ready to invoke? */
3479 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3480 rdp
->n_rp_cb_ready
++;
3484 /* Has RCU gone idle with this CPU needing another grace period? */
3485 if (cpu_needs_another_gp(rsp
, rdp
)) {
3486 rdp
->n_rp_cpu_needs_gp
++;
3490 /* Has another RCU grace period completed? */
3491 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3492 rdp
->n_rp_gp_completed
++;
3496 /* Has a new RCU grace period started? */
3497 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3498 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3499 rdp
->n_rp_gp_started
++;
3503 /* Does this CPU need a deferred NOCB wakeup? */
3504 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3505 rdp
->n_rp_nocb_defer_wakeup
++;
3510 rdp
->n_rp_need_nothing
++;
3515 * Check to see if there is any immediate RCU-related work to be done
3516 * by the current CPU, returning 1 if so. This function is part of the
3517 * RCU implementation; it is -not- an exported member of the RCU API.
3519 static int rcu_pending(void)
3521 struct rcu_state
*rsp
;
3523 for_each_rcu_flavor(rsp
)
3524 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3530 * Return true if the specified CPU has any callback. If all_lazy is
3531 * non-NULL, store an indication of whether all callbacks are lazy.
3532 * (If there are no callbacks, all of them are deemed to be lazy.)
3534 static int __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3538 struct rcu_data
*rdp
;
3539 struct rcu_state
*rsp
;
3541 for_each_rcu_flavor(rsp
) {
3542 rdp
= this_cpu_ptr(rsp
->rda
);
3546 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3557 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3558 * the compiler is expected to optimize this away.
3560 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3561 int cpu
, unsigned long done
)
3563 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3564 atomic_read(&rsp
->barrier_cpu_count
), done
);
3568 * RCU callback function for _rcu_barrier(). If we are last, wake
3569 * up the task executing _rcu_barrier().
3571 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3573 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3574 struct rcu_state
*rsp
= rdp
->rsp
;
3576 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3577 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3578 complete(&rsp
->barrier_completion
);
3580 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3585 * Called with preemption disabled, and from cross-cpu IRQ context.
3587 static void rcu_barrier_func(void *type
)
3589 struct rcu_state
*rsp
= type
;
3590 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3592 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3593 atomic_inc(&rsp
->barrier_cpu_count
);
3594 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3598 * Orchestrate the specified type of RCU barrier, waiting for all
3599 * RCU callbacks of the specified type to complete.
3601 static void _rcu_barrier(struct rcu_state
*rsp
)
3604 struct rcu_data
*rdp
;
3605 unsigned long snap
= READ_ONCE(rsp
->n_barrier_done
);
3606 unsigned long snap_done
;
3608 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3610 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3611 mutex_lock(&rsp
->barrier_mutex
);
3614 * Ensure that all prior references, including to ->n_barrier_done,
3615 * are ordered before the _rcu_barrier() machinery.
3617 smp_mb(); /* See above block comment. */
3620 * Recheck ->n_barrier_done to see if others did our work for us.
3621 * This means checking ->n_barrier_done for an even-to-odd-to-even
3622 * transition. The "if" expression below therefore rounds the old
3623 * value up to the next even number and adds two before comparing.
3625 snap_done
= rsp
->n_barrier_done
;
3626 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3629 * If the value in snap is odd, we needed to wait for the current
3630 * rcu_barrier() to complete, then wait for the next one, in other
3631 * words, we need the value of snap_done to be three larger than
3632 * the value of snap. On the other hand, if the value in snap is
3633 * even, we only had to wait for the next rcu_barrier() to complete,
3634 * in other words, we need the value of snap_done to be only two
3635 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3636 * this for us (thank you, Linus!).
3638 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3639 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3640 smp_mb(); /* caller's subsequent code after above check. */
3641 mutex_unlock(&rsp
->barrier_mutex
);
3646 * Increment ->n_barrier_done to avoid duplicate work. Use
3647 * WRITE_ONCE() to prevent the compiler from speculating
3648 * the increment to precede the early-exit check.
3650 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3651 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3652 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3653 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3656 * Initialize the count to one rather than to zero in order to
3657 * avoid a too-soon return to zero in case of a short grace period
3658 * (or preemption of this task). Exclude CPU-hotplug operations
3659 * to ensure that no offline CPU has callbacks queued.
3661 init_completion(&rsp
->barrier_completion
);
3662 atomic_set(&rsp
->barrier_cpu_count
, 1);
3666 * Force each CPU with callbacks to register a new callback.
3667 * When that callback is invoked, we will know that all of the
3668 * corresponding CPU's preceding callbacks have been invoked.
3670 for_each_possible_cpu(cpu
) {
3671 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3673 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3674 if (rcu_is_nocb_cpu(cpu
)) {
3675 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3676 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3677 rsp
->n_barrier_done
);
3679 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3680 rsp
->n_barrier_done
);
3681 smp_mb__before_atomic();
3682 atomic_inc(&rsp
->barrier_cpu_count
);
3683 __call_rcu(&rdp
->barrier_head
,
3684 rcu_barrier_callback
, rsp
, cpu
, 0);
3686 } else if (READ_ONCE(rdp
->qlen
)) {
3687 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3688 rsp
->n_barrier_done
);
3689 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3691 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3692 rsp
->n_barrier_done
);
3698 * Now that we have an rcu_barrier_callback() callback on each
3699 * CPU, and thus each counted, remove the initial count.
3701 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3702 complete(&rsp
->barrier_completion
);
3704 /* Increment ->n_barrier_done to prevent duplicate work. */
3705 smp_mb(); /* Keep increment after above mechanism. */
3706 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3707 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3708 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3709 smp_mb(); /* Keep increment before caller's subsequent code. */
3711 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3712 wait_for_completion(&rsp
->barrier_completion
);
3714 /* Other rcu_barrier() invocations can now safely proceed. */
3715 mutex_unlock(&rsp
->barrier_mutex
);
3719 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3721 void rcu_barrier_bh(void)
3723 _rcu_barrier(&rcu_bh_state
);
3725 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3728 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3730 void rcu_barrier_sched(void)
3732 _rcu_barrier(&rcu_sched_state
);
3734 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3737 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3738 * first CPU in a given leaf rcu_node structure coming online. The caller
3739 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3742 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3745 struct rcu_node
*rnp
= rnp_leaf
;
3748 mask
= rnp
->grpmask
;
3752 raw_spin_lock(&rnp
->lock
); /* Interrupts already disabled. */
3753 rnp
->qsmaskinit
|= mask
;
3754 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
3759 * Do boot-time initialization of a CPU's per-CPU RCU data.
3762 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3764 unsigned long flags
;
3765 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3766 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3768 /* Set up local state, ensuring consistent view of global state. */
3769 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3770 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3771 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3772 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3773 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3776 rcu_boot_init_nocb_percpu_data(rdp
);
3777 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3781 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3782 * offline event can be happening at a given time. Note also that we
3783 * can accept some slop in the rsp->completed access due to the fact
3784 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3787 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3789 unsigned long flags
;
3791 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3792 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3794 /* Set up local state, ensuring consistent view of global state. */
3795 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3796 rdp
->beenonline
= 1; /* We have now been online. */
3797 rdp
->qlen_last_fqs_check
= 0;
3798 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3799 rdp
->blimit
= blimit
;
3801 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3802 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3803 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3804 atomic_set(&rdp
->dynticks
->dynticks
,
3805 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3806 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3809 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3810 * propagation up the rcu_node tree will happen at the beginning
3811 * of the next grace period.
3814 mask
= rdp
->grpmask
;
3815 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3816 smp_mb__after_unlock_lock();
3817 rnp
->qsmaskinitnext
|= mask
;
3818 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3819 rdp
->completed
= rnp
->completed
;
3820 rdp
->passed_quiesce
= false;
3821 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
3822 rdp
->qs_pending
= false;
3823 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3824 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3827 static void rcu_prepare_cpu(int cpu
)
3829 struct rcu_state
*rsp
;
3831 for_each_rcu_flavor(rsp
)
3832 rcu_init_percpu_data(cpu
, rsp
);
3836 * Handle CPU online/offline notification events.
3838 int rcu_cpu_notify(struct notifier_block
*self
,
3839 unsigned long action
, void *hcpu
)
3841 long cpu
= (long)hcpu
;
3842 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3843 struct rcu_node
*rnp
= rdp
->mynode
;
3844 struct rcu_state
*rsp
;
3847 case CPU_UP_PREPARE
:
3848 case CPU_UP_PREPARE_FROZEN
:
3849 rcu_prepare_cpu(cpu
);
3850 rcu_prepare_kthreads(cpu
);
3851 rcu_spawn_all_nocb_kthreads(cpu
);
3854 case CPU_DOWN_FAILED
:
3855 rcu_boost_kthread_setaffinity(rnp
, -1);
3857 case CPU_DOWN_PREPARE
:
3858 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3861 case CPU_DYING_FROZEN
:
3862 for_each_rcu_flavor(rsp
)
3863 rcu_cleanup_dying_cpu(rsp
);
3865 case CPU_DYING_IDLE
:
3866 for_each_rcu_flavor(rsp
) {
3867 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3871 case CPU_DEAD_FROZEN
:
3872 case CPU_UP_CANCELED
:
3873 case CPU_UP_CANCELED_FROZEN
:
3874 for_each_rcu_flavor(rsp
) {
3875 rcu_cleanup_dead_cpu(cpu
, rsp
);
3876 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3885 static int rcu_pm_notify(struct notifier_block
*self
,
3886 unsigned long action
, void *hcpu
)
3889 case PM_HIBERNATION_PREPARE
:
3890 case PM_SUSPEND_PREPARE
:
3891 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3894 case PM_POST_HIBERNATION
:
3895 case PM_POST_SUSPEND
:
3896 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3897 rcu_unexpedite_gp();
3906 * Spawn the kthreads that handle each RCU flavor's grace periods.
3908 static int __init
rcu_spawn_gp_kthread(void)
3910 unsigned long flags
;
3911 int kthread_prio_in
= kthread_prio
;
3912 struct rcu_node
*rnp
;
3913 struct rcu_state
*rsp
;
3914 struct sched_param sp
;
3915 struct task_struct
*t
;
3917 /* Force priority into range. */
3918 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3920 else if (kthread_prio
< 0)
3922 else if (kthread_prio
> 99)
3924 if (kthread_prio
!= kthread_prio_in
)
3925 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3926 kthread_prio
, kthread_prio_in
);
3928 rcu_scheduler_fully_active
= 1;
3929 for_each_rcu_flavor(rsp
) {
3930 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3932 rnp
= rcu_get_root(rsp
);
3933 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3934 rsp
->gp_kthread
= t
;
3936 sp
.sched_priority
= kthread_prio
;
3937 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3940 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3942 rcu_spawn_nocb_kthreads();
3943 rcu_spawn_boost_kthreads();
3946 early_initcall(rcu_spawn_gp_kthread
);
3949 * This function is invoked towards the end of the scheduler's initialization
3950 * process. Before this is called, the idle task might contain
3951 * RCU read-side critical sections (during which time, this idle
3952 * task is booting the system). After this function is called, the
3953 * idle tasks are prohibited from containing RCU read-side critical
3954 * sections. This function also enables RCU lockdep checking.
3956 void rcu_scheduler_starting(void)
3958 WARN_ON(num_online_cpus() != 1);
3959 WARN_ON(nr_context_switches() > 0);
3960 rcu_scheduler_active
= 1;
3964 * Compute the per-level fanout, either using the exact fanout specified
3965 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
3967 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3971 if (rcu_fanout_exact
) {
3972 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3973 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3974 rsp
->levelspread
[i
] = RCU_FANOUT
;
3980 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3981 ccur
= rsp
->levelcnt
[i
];
3982 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3989 * Helper function for rcu_init() that initializes one rcu_state structure.
3991 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3992 struct rcu_data __percpu
*rda
)
3994 static const char * const buf
[] = {
3998 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3999 static const char * const fqs
[] = {
4003 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
4004 static u8 fl_mask
= 0x1;
4008 struct rcu_node
*rnp
;
4010 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4012 /* Silence gcc 4.8 false positive about array index out of range. */
4013 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4014 panic("rcu_init_one: rcu_num_lvls out of range");
4016 /* Initialize the level-tracking arrays. */
4018 for (i
= 0; i
< rcu_num_lvls
; i
++)
4019 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
4020 for (i
= 1; i
< rcu_num_lvls
; i
++)
4021 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
4022 rcu_init_levelspread(rsp
);
4023 rsp
->flavor_mask
= fl_mask
;
4026 /* Initialize the elements themselves, starting from the leaves. */
4028 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4029 cpustride
*= rsp
->levelspread
[i
];
4030 rnp
= rsp
->level
[i
];
4031 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
4032 raw_spin_lock_init(&rnp
->lock
);
4033 lockdep_set_class_and_name(&rnp
->lock
,
4034 &rcu_node_class
[i
], buf
[i
]);
4035 raw_spin_lock_init(&rnp
->fqslock
);
4036 lockdep_set_class_and_name(&rnp
->fqslock
,
4037 &rcu_fqs_class
[i
], fqs
[i
]);
4038 rnp
->gpnum
= rsp
->gpnum
;
4039 rnp
->completed
= rsp
->completed
;
4041 rnp
->qsmaskinit
= 0;
4042 rnp
->grplo
= j
* cpustride
;
4043 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4044 if (rnp
->grphi
>= nr_cpu_ids
)
4045 rnp
->grphi
= nr_cpu_ids
- 1;
4051 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
4052 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4053 rnp
->parent
= rsp
->level
[i
- 1] +
4054 j
/ rsp
->levelspread
[i
- 1];
4057 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4058 rcu_init_one_nocb(rnp
);
4062 init_waitqueue_head(&rsp
->gp_wq
);
4063 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4064 for_each_possible_cpu(i
) {
4065 while (i
> rnp
->grphi
)
4067 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4068 rcu_boot_init_percpu_data(i
, rsp
);
4070 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4074 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4075 * replace the definitions in tree.h because those are needed to size
4076 * the ->node array in the rcu_state structure.
4078 static void __init
rcu_init_geometry(void)
4084 int rcu_capacity
[MAX_RCU_LVLS
+ 1];
4087 * Initialize any unspecified boot parameters.
4088 * The default values of jiffies_till_first_fqs and
4089 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4090 * value, which is a function of HZ, then adding one for each
4091 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4093 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4094 if (jiffies_till_first_fqs
== ULONG_MAX
)
4095 jiffies_till_first_fqs
= d
;
4096 if (jiffies_till_next_fqs
== ULONG_MAX
)
4097 jiffies_till_next_fqs
= d
;
4099 /* If the compile-time values are accurate, just leave. */
4100 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4101 nr_cpu_ids
== NR_CPUS
)
4103 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4104 rcu_fanout_leaf
, nr_cpu_ids
);
4107 * Compute number of nodes that can be handled an rcu_node tree
4108 * with the given number of levels. Setting rcu_capacity[0] makes
4109 * some of the arithmetic easier.
4111 rcu_capacity
[0] = 1;
4112 rcu_capacity
[1] = rcu_fanout_leaf
;
4113 for (i
= 2; i
<= MAX_RCU_LVLS
; i
++)
4114 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4117 * The boot-time rcu_fanout_leaf parameter is only permitted
4118 * to increase the leaf-level fanout, not decrease it. Of course,
4119 * the leaf-level fanout cannot exceed the number of bits in
4120 * the rcu_node masks. Finally, the tree must be able to accommodate
4121 * the configured number of CPUs. Complain and fall back to the
4122 * compile-time values if these limits are exceeded.
4124 if (rcu_fanout_leaf
< RCU_FANOUT_LEAF
||
4125 rcu_fanout_leaf
> sizeof(unsigned long) * 8 ||
4126 n
> rcu_capacity
[MAX_RCU_LVLS
]) {
4131 /* Calculate the number of rcu_nodes at each level of the tree. */
4132 for (i
= 1; i
<= MAX_RCU_LVLS
; i
++)
4133 if (n
<= rcu_capacity
[i
]) {
4134 for (j
= 0; j
<= i
; j
++)
4136 DIV_ROUND_UP(n
, rcu_capacity
[i
- j
]);
4138 for (j
= i
+ 1; j
<= MAX_RCU_LVLS
; j
++)
4143 /* Calculate the total number of rcu_node structures. */
4145 for (i
= 0; i
<= MAX_RCU_LVLS
; i
++)
4146 rcu_num_nodes
+= num_rcu_lvl
[i
];
4151 * Dump out the structure of the rcu_node combining tree associated
4152 * with the rcu_state structure referenced by rsp.
4154 static void __init
rcu_dump_rcu_node_tree(struct rcu_state
*rsp
)
4157 struct rcu_node
*rnp
;
4159 pr_info("rcu_node tree layout dump\n");
4161 rcu_for_each_node_breadth_first(rsp
, rnp
) {
4162 if (rnp
->level
!= level
) {
4167 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4172 void __init
rcu_init(void)
4176 rcu_early_boot_tests();
4178 rcu_bootup_announce();
4179 rcu_init_geometry();
4180 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
4181 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
4183 rcu_dump_rcu_node_tree(&rcu_sched_state
);
4184 __rcu_init_preempt();
4185 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4188 * We don't need protection against CPU-hotplug here because
4189 * this is called early in boot, before either interrupts
4190 * or the scheduler are operational.
4192 cpu_notifier(rcu_cpu_notify
, 0);
4193 pm_notifier(rcu_pm_notify
, 0);
4194 for_each_online_cpu(cpu
)
4195 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4198 #include "tree_plugin.h"