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/trace_events.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
72 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
95 struct rcu_state sname##_state = { \
96 .level = { &sname##_state.node[0] }, \
97 .rda = &sname##_data, \
99 .fqs_state = RCU_GP_IDLE, \
100 .gpnum = 0UL - 300UL, \
101 .completed = 0UL - 300UL, \
102 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
103 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
104 .orphan_donetail = &sname##_state.orphan_donelist, \
105 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
106 .name = RCU_STATE_NAME(sname), \
110 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
111 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
113 static struct rcu_state
*const rcu_state_p
;
114 static struct rcu_data __percpu
*const rcu_data_p
;
115 LIST_HEAD(rcu_struct_flavors
);
117 /* Dump rcu_node combining tree at boot to verify correct setup. */
118 static bool dump_tree
;
119 module_param(dump_tree
, bool, 0444);
120 /* Control rcu_node-tree auto-balancing at boot time. */
121 static bool rcu_fanout_exact
;
122 module_param(rcu_fanout_exact
, bool, 0444);
123 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
124 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
125 module_param(rcu_fanout_leaf
, int, 0444);
126 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
127 static int num_rcu_lvl
[] = { /* Number of rcu_nodes at specified level. */
134 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
137 * The rcu_scheduler_active variable transitions from zero to one just
138 * before the first task is spawned. So when this variable is zero, RCU
139 * can assume that there is but one task, allowing RCU to (for example)
140 * optimize synchronize_sched() to a simple barrier(). When this variable
141 * is one, RCU must actually do all the hard work required to detect real
142 * grace periods. This variable is also used to suppress boot-time false
143 * positives from lockdep-RCU error checking.
145 int rcu_scheduler_active __read_mostly
;
146 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
149 * The rcu_scheduler_fully_active variable transitions from zero to one
150 * during the early_initcall() processing, which is after the scheduler
151 * is capable of creating new tasks. So RCU processing (for example,
152 * creating tasks for RCU priority boosting) must be delayed until after
153 * rcu_scheduler_fully_active transitions from zero to one. We also
154 * currently delay invocation of any RCU callbacks until after this point.
156 * It might later prove better for people registering RCU callbacks during
157 * early boot to take responsibility for these callbacks, but one step at
160 static int rcu_scheduler_fully_active __read_mostly
;
162 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
163 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
164 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
165 static void invoke_rcu_core(void);
166 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
168 /* rcuc/rcub kthread realtime priority */
169 #ifdef CONFIG_RCU_KTHREAD_PRIO
170 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
171 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
172 static int kthread_prio
= IS_ENABLED(CONFIG_RCU_BOOST
) ? 1 : 0;
173 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
174 module_param(kthread_prio
, int, 0644);
176 /* Delay in jiffies for grace-period initialization delays, debug only. */
178 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
179 static int gp_preinit_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY
;
180 module_param(gp_preinit_delay
, int, 0644);
181 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
182 static const int gp_preinit_delay
;
183 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
185 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
186 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
187 module_param(gp_init_delay
, int, 0644);
188 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
189 static const int gp_init_delay
;
190 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
192 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
193 static int gp_cleanup_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY
;
194 module_param(gp_cleanup_delay
, int, 0644);
195 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
196 static const int gp_cleanup_delay
;
197 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
200 * Number of grace periods between delays, normalized by the duration of
201 * the delay. The longer the the delay, the more the grace periods between
202 * each delay. The reason for this normalization is that it means that,
203 * for non-zero delays, the overall slowdown of grace periods is constant
204 * regardless of the duration of the delay. This arrangement balances
205 * the need for long delays to increase some race probabilities with the
206 * need for fast grace periods to increase other race probabilities.
208 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
211 * Track the rcutorture test sequence number and the update version
212 * number within a given test. The rcutorture_testseq is incremented
213 * on every rcutorture module load and unload, so has an odd value
214 * when a test is running. The rcutorture_vernum is set to zero
215 * when rcutorture starts and is incremented on each rcutorture update.
216 * These variables enable correlating rcutorture output with the
217 * RCU tracing information.
219 unsigned long rcutorture_testseq
;
220 unsigned long rcutorture_vernum
;
223 * Compute the mask of online CPUs for the specified rcu_node structure.
224 * This will not be stable unless the rcu_node structure's ->lock is
225 * held, but the bit corresponding to the current CPU will be stable
228 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
230 return READ_ONCE(rnp
->qsmaskinitnext
);
234 * Return true if an RCU grace period is in progress. The READ_ONCE()s
235 * permit this function to be invoked without holding the root rcu_node
236 * structure's ->lock, but of course results can be subject to change.
238 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
240 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
244 * Note a quiescent state. Because we do not need to know
245 * how many quiescent states passed, just if there was at least
246 * one since the start of the grace period, this just sets a flag.
247 * The caller must have disabled preemption.
249 void rcu_sched_qs(void)
251 if (!__this_cpu_read(rcu_sched_data
.passed_quiesce
)) {
252 trace_rcu_grace_period(TPS("rcu_sched"),
253 __this_cpu_read(rcu_sched_data
.gpnum
),
255 __this_cpu_write(rcu_sched_data
.passed_quiesce
, 1);
261 if (!__this_cpu_read(rcu_bh_data
.passed_quiesce
)) {
262 trace_rcu_grace_period(TPS("rcu_bh"),
263 __this_cpu_read(rcu_bh_data
.gpnum
),
265 __this_cpu_write(rcu_bh_data
.passed_quiesce
, 1);
269 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
271 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
272 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
273 .dynticks
= ATOMIC_INIT(1),
274 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
275 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
276 .dynticks_idle
= ATOMIC_INIT(1),
277 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
280 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
281 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
284 * Let the RCU core know that this CPU has gone through the scheduler,
285 * which is a quiescent state. This is called when the need for a
286 * quiescent state is urgent, so we burn an atomic operation and full
287 * memory barriers to let the RCU core know about it, regardless of what
288 * this CPU might (or might not) do in the near future.
290 * We inform the RCU core by emulating a zero-duration dyntick-idle
291 * period, which we in turn do by incrementing the ->dynticks counter
294 static void rcu_momentary_dyntick_idle(void)
297 struct rcu_data
*rdp
;
298 struct rcu_dynticks
*rdtp
;
300 struct rcu_state
*rsp
;
302 local_irq_save(flags
);
305 * Yes, we can lose flag-setting operations. This is OK, because
306 * the flag will be set again after some delay.
308 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
309 raw_cpu_write(rcu_sched_qs_mask
, 0);
311 /* Find the flavor that needs a quiescent state. */
312 for_each_rcu_flavor(rsp
) {
313 rdp
= raw_cpu_ptr(rsp
->rda
);
314 if (!(resched_mask
& rsp
->flavor_mask
))
316 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
317 if (READ_ONCE(rdp
->mynode
->completed
) !=
318 READ_ONCE(rdp
->cond_resched_completed
))
322 * Pretend to be momentarily idle for the quiescent state.
323 * This allows the grace-period kthread to record the
324 * quiescent state, with no need for this CPU to do anything
327 rdtp
= this_cpu_ptr(&rcu_dynticks
);
328 smp_mb__before_atomic(); /* Earlier stuff before QS. */
329 atomic_add(2, &rdtp
->dynticks
); /* QS. */
330 smp_mb__after_atomic(); /* Later stuff after QS. */
333 local_irq_restore(flags
);
337 * Note a context switch. This is a quiescent state for RCU-sched,
338 * and requires special handling for preemptible RCU.
339 * The caller must have disabled preemption.
341 void rcu_note_context_switch(void)
343 trace_rcu_utilization(TPS("Start context switch"));
345 rcu_preempt_note_context_switch();
346 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
347 rcu_momentary_dyntick_idle();
348 trace_rcu_utilization(TPS("End context switch"));
350 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
353 * Register a quiescent state for all RCU flavors. If there is an
354 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
355 * dyntick-idle quiescent state visible to other CPUs (but only for those
356 * RCU flavors in desperate need of a quiescent state, which will normally
357 * be none of them). Either way, do a lightweight quiescent state for
360 void rcu_all_qs(void)
362 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
363 rcu_momentary_dyntick_idle();
364 this_cpu_inc(rcu_qs_ctr
);
366 EXPORT_SYMBOL_GPL(rcu_all_qs
);
368 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
369 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
370 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
372 module_param(blimit
, long, 0444);
373 module_param(qhimark
, long, 0444);
374 module_param(qlowmark
, long, 0444);
376 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
377 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
379 module_param(jiffies_till_first_fqs
, ulong
, 0644);
380 module_param(jiffies_till_next_fqs
, ulong
, 0644);
383 * How long the grace period must be before we start recruiting
384 * quiescent-state help from rcu_note_context_switch().
386 static ulong jiffies_till_sched_qs
= HZ
/ 20;
387 module_param(jiffies_till_sched_qs
, ulong
, 0644);
389 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
390 struct rcu_data
*rdp
);
391 static void force_qs_rnp(struct rcu_state
*rsp
,
392 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
393 unsigned long *maxj
),
394 bool *isidle
, unsigned long *maxj
);
395 static void force_quiescent_state(struct rcu_state
*rsp
);
396 static int rcu_pending(void);
399 * Return the number of RCU batches started thus far for debug & stats.
401 unsigned long rcu_batches_started(void)
403 return rcu_state_p
->gpnum
;
405 EXPORT_SYMBOL_GPL(rcu_batches_started
);
408 * Return the number of RCU-sched batches started thus far for debug & stats.
410 unsigned long rcu_batches_started_sched(void)
412 return rcu_sched_state
.gpnum
;
414 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
417 * Return the number of RCU BH batches started thus far for debug & stats.
419 unsigned long rcu_batches_started_bh(void)
421 return rcu_bh_state
.gpnum
;
423 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
426 * Return the number of RCU batches completed thus far for debug & stats.
428 unsigned long rcu_batches_completed(void)
430 return rcu_state_p
->completed
;
432 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
435 * Return the number of RCU-sched batches completed thus far for debug & stats.
437 unsigned long rcu_batches_completed_sched(void)
439 return rcu_sched_state
.completed
;
441 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
444 * Return the number of RCU BH batches completed thus far for debug & stats.
446 unsigned long rcu_batches_completed_bh(void)
448 return rcu_bh_state
.completed
;
450 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
453 * Force a quiescent state.
455 void rcu_force_quiescent_state(void)
457 force_quiescent_state(rcu_state_p
);
459 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
462 * Force a quiescent state for RCU BH.
464 void rcu_bh_force_quiescent_state(void)
466 force_quiescent_state(&rcu_bh_state
);
468 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
471 * Force a quiescent state for RCU-sched.
473 void rcu_sched_force_quiescent_state(void)
475 force_quiescent_state(&rcu_sched_state
);
477 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
480 * Show the state of the grace-period kthreads.
482 void show_rcu_gp_kthreads(void)
484 struct rcu_state
*rsp
;
486 for_each_rcu_flavor(rsp
) {
487 pr_info("%s: wait state: %d ->state: %#lx\n",
488 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
489 /* sched_show_task(rsp->gp_kthread); */
492 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
495 * Record the number of times rcutorture tests have been initiated and
496 * terminated. This information allows the debugfs tracing stats to be
497 * correlated to the rcutorture messages, even when the rcutorture module
498 * is being repeatedly loaded and unloaded. In other words, we cannot
499 * store this state in rcutorture itself.
501 void rcutorture_record_test_transition(void)
503 rcutorture_testseq
++;
504 rcutorture_vernum
= 0;
506 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
509 * Send along grace-period-related data for rcutorture diagnostics.
511 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
512 unsigned long *gpnum
, unsigned long *completed
)
514 struct rcu_state
*rsp
= NULL
;
523 case RCU_SCHED_FLAVOR
:
524 rsp
= &rcu_sched_state
;
530 *flags
= READ_ONCE(rsp
->gp_flags
);
531 *gpnum
= READ_ONCE(rsp
->gpnum
);
532 *completed
= READ_ONCE(rsp
->completed
);
539 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
542 * Record the number of writer passes through the current rcutorture test.
543 * This is also used to correlate debugfs tracing stats with the rcutorture
546 void rcutorture_record_progress(unsigned long vernum
)
550 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
553 * Does the CPU have callbacks ready to be invoked?
556 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
558 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
559 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
563 * Return the root node of the specified rcu_state structure.
565 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
567 return &rsp
->node
[0];
571 * Is there any need for future grace periods?
572 * Interrupts must be disabled. If the caller does not hold the root
573 * rnp_node structure's ->lock, the results are advisory only.
575 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
577 struct rcu_node
*rnp
= rcu_get_root(rsp
);
578 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
579 int *fp
= &rnp
->need_future_gp
[idx
];
581 return READ_ONCE(*fp
);
585 * Does the current CPU require a not-yet-started grace period?
586 * The caller must have disabled interrupts to prevent races with
587 * normal callback registry.
590 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
594 if (rcu_gp_in_progress(rsp
))
595 return 0; /* No, a grace period is already in progress. */
596 if (rcu_future_needs_gp(rsp
))
597 return 1; /* Yes, a no-CBs CPU needs one. */
598 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
599 return 0; /* No, this is a no-CBs (or offline) CPU. */
600 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
601 return 1; /* Yes, this CPU has newly registered callbacks. */
602 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
603 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
604 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
605 rdp
->nxtcompleted
[i
]))
606 return 1; /* Yes, CBs for future grace period. */
607 return 0; /* No grace period needed. */
611 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
613 * If the new value of the ->dynticks_nesting counter now is zero,
614 * we really have entered idle, and must do the appropriate accounting.
615 * The caller must have disabled interrupts.
617 static void rcu_eqs_enter_common(long long oldval
, bool user
)
619 struct rcu_state
*rsp
;
620 struct rcu_data
*rdp
;
621 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
623 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
624 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
625 !user
&& !is_idle_task(current
)) {
626 struct task_struct
*idle __maybe_unused
=
627 idle_task(smp_processor_id());
629 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
630 ftrace_dump(DUMP_ORIG
);
631 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
632 current
->pid
, current
->comm
,
633 idle
->pid
, idle
->comm
); /* must be idle task! */
635 for_each_rcu_flavor(rsp
) {
636 rdp
= this_cpu_ptr(rsp
->rda
);
637 do_nocb_deferred_wakeup(rdp
);
639 rcu_prepare_for_idle();
640 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
641 smp_mb__before_atomic(); /* See above. */
642 atomic_inc(&rdtp
->dynticks
);
643 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
644 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
645 atomic_read(&rdtp
->dynticks
) & 0x1);
646 rcu_dynticks_task_enter();
649 * It is illegal to enter an extended quiescent state while
650 * in an RCU read-side critical section.
652 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map
),
653 "Illegal idle entry in RCU read-side critical section.");
654 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
),
655 "Illegal idle entry in RCU-bh read-side critical section.");
656 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map
),
657 "Illegal idle entry in RCU-sched read-side critical section.");
661 * Enter an RCU extended quiescent state, which can be either the
662 * idle loop or adaptive-tickless usermode execution.
664 static void rcu_eqs_enter(bool user
)
667 struct rcu_dynticks
*rdtp
;
669 rdtp
= this_cpu_ptr(&rcu_dynticks
);
670 oldval
= rdtp
->dynticks_nesting
;
671 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
672 (oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
673 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
674 rdtp
->dynticks_nesting
= 0;
675 rcu_eqs_enter_common(oldval
, user
);
677 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
682 * rcu_idle_enter - inform RCU that current CPU is entering idle
684 * Enter idle mode, in other words, -leave- the mode in which RCU
685 * read-side critical sections can occur. (Though RCU read-side
686 * critical sections can occur in irq handlers in idle, a possibility
687 * handled by irq_enter() and irq_exit().)
689 * We crowbar the ->dynticks_nesting field to zero to allow for
690 * the possibility of usermode upcalls having messed up our count
691 * of interrupt nesting level during the prior busy period.
693 void rcu_idle_enter(void)
697 local_irq_save(flags
);
698 rcu_eqs_enter(false);
699 rcu_sysidle_enter(0);
700 local_irq_restore(flags
);
702 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
704 #ifdef CONFIG_RCU_USER_QS
706 * rcu_user_enter - inform RCU that we are resuming userspace.
708 * Enter RCU idle mode right before resuming userspace. No use of RCU
709 * is permitted between this call and rcu_user_exit(). This way the
710 * CPU doesn't need to maintain the tick for RCU maintenance purposes
711 * when the CPU runs in userspace.
713 void rcu_user_enter(void)
717 #endif /* CONFIG_RCU_USER_QS */
720 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
722 * Exit from an interrupt handler, which might possibly result in entering
723 * idle mode, in other words, leaving the mode in which read-side critical
724 * sections can occur.
726 * This code assumes that the idle loop never does anything that might
727 * result in unbalanced calls to irq_enter() and irq_exit(). If your
728 * architecture violates this assumption, RCU will give you what you
729 * deserve, good and hard. But very infrequently and irreproducibly.
731 * Use things like work queues to work around this limitation.
733 * You have been warned.
735 void rcu_irq_exit(void)
739 struct rcu_dynticks
*rdtp
;
741 local_irq_save(flags
);
742 rdtp
= this_cpu_ptr(&rcu_dynticks
);
743 oldval
= rdtp
->dynticks_nesting
;
744 rdtp
->dynticks_nesting
--;
745 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
746 rdtp
->dynticks_nesting
< 0);
747 if (rdtp
->dynticks_nesting
)
748 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
750 rcu_eqs_enter_common(oldval
, true);
751 rcu_sysidle_enter(1);
752 local_irq_restore(flags
);
756 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
758 * If the new value of the ->dynticks_nesting counter was previously zero,
759 * we really have exited idle, and must do the appropriate accounting.
760 * The caller must have disabled interrupts.
762 static void rcu_eqs_exit_common(long long oldval
, int user
)
764 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
766 rcu_dynticks_task_exit();
767 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
768 atomic_inc(&rdtp
->dynticks
);
769 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
770 smp_mb__after_atomic(); /* See above. */
771 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
772 !(atomic_read(&rdtp
->dynticks
) & 0x1));
773 rcu_cleanup_after_idle();
774 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
775 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
776 !user
&& !is_idle_task(current
)) {
777 struct task_struct
*idle __maybe_unused
=
778 idle_task(smp_processor_id());
780 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
781 oldval
, rdtp
->dynticks_nesting
);
782 ftrace_dump(DUMP_ORIG
);
783 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
784 current
->pid
, current
->comm
,
785 idle
->pid
, idle
->comm
); /* must be idle task! */
790 * Exit an RCU extended quiescent state, which can be either the
791 * idle loop or adaptive-tickless usermode execution.
793 static void rcu_eqs_exit(bool user
)
795 struct rcu_dynticks
*rdtp
;
798 rdtp
= this_cpu_ptr(&rcu_dynticks
);
799 oldval
= rdtp
->dynticks_nesting
;
800 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) && oldval
< 0);
801 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
802 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
804 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
805 rcu_eqs_exit_common(oldval
, user
);
810 * rcu_idle_exit - inform RCU that current CPU is leaving idle
812 * Exit idle mode, in other words, -enter- the mode in which RCU
813 * read-side critical sections can occur.
815 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
816 * allow for the possibility of usermode upcalls messing up our count
817 * of interrupt nesting level during the busy period that is just
820 void rcu_idle_exit(void)
824 local_irq_save(flags
);
827 local_irq_restore(flags
);
829 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
831 #ifdef CONFIG_RCU_USER_QS
833 * rcu_user_exit - inform RCU that we are exiting userspace.
835 * Exit RCU idle mode while entering the kernel because it can
836 * run a RCU read side critical section anytime.
838 void rcu_user_exit(void)
842 #endif /* CONFIG_RCU_USER_QS */
845 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
847 * Enter an interrupt handler, which might possibly result in exiting
848 * idle mode, in other words, entering the mode in which read-side critical
849 * sections can occur.
851 * Note that the Linux kernel is fully capable of entering an interrupt
852 * handler that it never exits, for example when doing upcalls to
853 * user mode! This code assumes that the idle loop never does upcalls to
854 * user mode. If your architecture does do upcalls from the idle loop (or
855 * does anything else that results in unbalanced calls to the irq_enter()
856 * and irq_exit() functions), RCU will give you what you deserve, good
857 * and hard. But very infrequently and irreproducibly.
859 * Use things like work queues to work around this limitation.
861 * You have been warned.
863 void rcu_irq_enter(void)
866 struct rcu_dynticks
*rdtp
;
869 local_irq_save(flags
);
870 rdtp
= this_cpu_ptr(&rcu_dynticks
);
871 oldval
= rdtp
->dynticks_nesting
;
872 rdtp
->dynticks_nesting
++;
873 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
874 rdtp
->dynticks_nesting
== 0);
876 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
878 rcu_eqs_exit_common(oldval
, true);
880 local_irq_restore(flags
);
884 * rcu_nmi_enter - inform RCU of entry to NMI context
886 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
887 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
888 * that the CPU is active. This implementation permits nested NMIs, as
889 * long as the nesting level does not overflow an int. (You will probably
890 * run out of stack space first.)
892 void rcu_nmi_enter(void)
894 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
897 /* Complain about underflow. */
898 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
901 * If idle from RCU viewpoint, atomically increment ->dynticks
902 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
903 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
904 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
905 * to be in the outermost NMI handler that interrupted an RCU-idle
906 * period (observation due to Andy Lutomirski).
908 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
909 smp_mb__before_atomic(); /* Force delay from prior write. */
910 atomic_inc(&rdtp
->dynticks
);
911 /* atomic_inc() before later RCU read-side crit sects */
912 smp_mb__after_atomic(); /* See above. */
913 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
916 rdtp
->dynticks_nmi_nesting
+= incby
;
921 * rcu_nmi_exit - inform RCU of exit from NMI context
923 * If we are returning from the outermost NMI handler that interrupted an
924 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
925 * to let the RCU grace-period handling know that the CPU is back to
928 void rcu_nmi_exit(void)
930 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
933 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
934 * (We are exiting an NMI handler, so RCU better be paying attention
937 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
938 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
941 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
942 * leave it in non-RCU-idle state.
944 if (rdtp
->dynticks_nmi_nesting
!= 1) {
945 rdtp
->dynticks_nmi_nesting
-= 2;
949 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
950 rdtp
->dynticks_nmi_nesting
= 0;
951 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
952 smp_mb__before_atomic(); /* See above. */
953 atomic_inc(&rdtp
->dynticks
);
954 smp_mb__after_atomic(); /* Force delay to next write. */
955 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
959 * __rcu_is_watching - are RCU read-side critical sections safe?
961 * Return true if RCU is watching the running CPU, which means that
962 * this CPU can safely enter RCU read-side critical sections. Unlike
963 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
964 * least disabled preemption.
966 bool notrace
__rcu_is_watching(void)
968 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
972 * rcu_is_watching - see if RCU thinks that the current CPU is idle
974 * If the current CPU is in its idle loop and is neither in an interrupt
975 * or NMI handler, return true.
977 bool notrace
rcu_is_watching(void)
982 ret
= __rcu_is_watching();
986 EXPORT_SYMBOL_GPL(rcu_is_watching
);
988 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
991 * Is the current CPU online? Disable preemption to avoid false positives
992 * that could otherwise happen due to the current CPU number being sampled,
993 * this task being preempted, its old CPU being taken offline, resuming
994 * on some other CPU, then determining that its old CPU is now offline.
995 * It is OK to use RCU on an offline processor during initial boot, hence
996 * the check for rcu_scheduler_fully_active. Note also that it is OK
997 * for a CPU coming online to use RCU for one jiffy prior to marking itself
998 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
999 * offline to continue to use RCU for one jiffy after marking itself
1000 * offline in the cpu_online_mask. This leniency is necessary given the
1001 * non-atomic nature of the online and offline processing, for example,
1002 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1005 * This is also why RCU internally marks CPUs online during the
1006 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1008 * Disable checking if in an NMI handler because we cannot safely report
1009 * errors from NMI handlers anyway.
1011 bool rcu_lockdep_current_cpu_online(void)
1013 struct rcu_data
*rdp
;
1014 struct rcu_node
*rnp
;
1020 rdp
= this_cpu_ptr(&rcu_sched_data
);
1022 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
1023 !rcu_scheduler_fully_active
;
1027 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
1029 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1032 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1034 * If the current CPU is idle or running at a first-level (not nested)
1035 * interrupt from idle, return true. The caller must have at least
1036 * disabled preemption.
1038 static int rcu_is_cpu_rrupt_from_idle(void)
1040 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1044 * Snapshot the specified CPU's dynticks counter so that we can later
1045 * credit them with an implicit quiescent state. Return 1 if this CPU
1046 * is in dynticks idle mode, which is an extended quiescent state.
1048 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1049 bool *isidle
, unsigned long *maxj
)
1051 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1052 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1053 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1054 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1057 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1058 rdp
->mynode
->gpnum
))
1059 WRITE_ONCE(rdp
->gpwrap
, true);
1065 * Return true if the specified CPU has passed through a quiescent
1066 * state by virtue of being in or having passed through an dynticks
1067 * idle state since the last call to dyntick_save_progress_counter()
1068 * for this same CPU, or by virtue of having been offline.
1070 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1071 bool *isidle
, unsigned long *maxj
)
1077 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1078 snap
= (unsigned int)rdp
->dynticks_snap
;
1081 * If the CPU passed through or entered a dynticks idle phase with
1082 * no active irq/NMI handlers, then we can safely pretend that the CPU
1083 * already acknowledged the request to pass through a quiescent
1084 * state. Either way, that CPU cannot possibly be in an RCU
1085 * read-side critical section that started before the beginning
1086 * of the current RCU grace period.
1088 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1089 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1090 rdp
->dynticks_fqs
++;
1095 * Check for the CPU being offline, but only if the grace period
1096 * is old enough. We don't need to worry about the CPU changing
1097 * state: If we see it offline even once, it has been through a
1100 * The reason for insisting that the grace period be at least
1101 * one jiffy old is that CPUs that are not quite online and that
1102 * have just gone offline can still execute RCU read-side critical
1105 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1106 return 0; /* Grace period is not old enough. */
1108 if (cpu_is_offline(rdp
->cpu
)) {
1109 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1115 * A CPU running for an extended time within the kernel can
1116 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1117 * even context-switching back and forth between a pair of
1118 * in-kernel CPU-bound tasks cannot advance grace periods.
1119 * So if the grace period is old enough, make the CPU pay attention.
1120 * Note that the unsynchronized assignments to the per-CPU
1121 * rcu_sched_qs_mask variable are safe. Yes, setting of
1122 * bits can be lost, but they will be set again on the next
1123 * force-quiescent-state pass. So lost bit sets do not result
1124 * in incorrect behavior, merely in a grace period lasting
1125 * a few jiffies longer than it might otherwise. Because
1126 * there are at most four threads involved, and because the
1127 * updates are only once every few jiffies, the probability of
1128 * lossage (and thus of slight grace-period extension) is
1131 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1132 * is set too high, we override with half of the RCU CPU stall
1135 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1136 if (ULONG_CMP_GE(jiffies
,
1137 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1138 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1139 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1140 WRITE_ONCE(rdp
->cond_resched_completed
,
1141 READ_ONCE(rdp
->mynode
->completed
));
1142 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1144 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1145 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1146 rdp
->rsp
->jiffies_resched
+= 5; /* Enable beating. */
1147 } else if (ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1148 /* Time to beat on that CPU again! */
1149 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1150 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1157 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1159 unsigned long j
= jiffies
;
1163 smp_wmb(); /* Record start time before stall time. */
1164 j1
= rcu_jiffies_till_stall_check();
1165 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1166 rsp
->jiffies_resched
= j
+ j1
/ 2;
1167 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1171 * Complain about starvation of grace-period kthread.
1173 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1179 gpa
= READ_ONCE(rsp
->gp_activity
);
1180 if (j
- gpa
> 2 * HZ
)
1181 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n",
1183 rsp
->gpnum
, rsp
->completed
,
1184 rsp
->gp_flags
, rsp
->gp_state
,
1185 rsp
->gp_kthread
? rsp
->gp_kthread
->state
: 0);
1189 * Dump stacks of all tasks running on stalled CPUs.
1191 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1194 unsigned long flags
;
1195 struct rcu_node
*rnp
;
1197 rcu_for_each_leaf_node(rsp
, rnp
) {
1198 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1199 if (rnp
->qsmask
!= 0) {
1200 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1201 if (rnp
->qsmask
& (1UL << cpu
))
1202 dump_cpu_task(rnp
->grplo
+ cpu
);
1204 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1208 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1212 unsigned long flags
;
1216 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1219 /* Only let one CPU complain about others per time interval. */
1221 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1222 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1223 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1224 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1227 WRITE_ONCE(rsp
->jiffies_stall
,
1228 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1229 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1232 * OK, time to rat on our buddy...
1233 * See Documentation/RCU/stallwarn.txt for info on how to debug
1234 * RCU CPU stall warnings.
1236 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1238 print_cpu_stall_info_begin();
1239 rcu_for_each_leaf_node(rsp
, rnp
) {
1240 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1241 ndetected
+= rcu_print_task_stall(rnp
);
1242 if (rnp
->qsmask
!= 0) {
1243 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1244 if (rnp
->qsmask
& (1UL << cpu
)) {
1245 print_cpu_stall_info(rsp
,
1250 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1253 print_cpu_stall_info_end();
1254 for_each_possible_cpu(cpu
)
1255 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1256 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1257 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1258 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1260 rcu_dump_cpu_stacks(rsp
);
1262 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1263 READ_ONCE(rsp
->completed
) == gpnum
) {
1264 pr_err("INFO: Stall ended before state dump start\n");
1267 gpa
= READ_ONCE(rsp
->gp_activity
);
1268 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1269 rsp
->name
, j
- gpa
, j
, gpa
,
1270 jiffies_till_next_fqs
,
1271 rcu_get_root(rsp
)->qsmask
);
1272 /* In this case, the current CPU might be at fault. */
1273 sched_show_task(current
);
1277 /* Complain about tasks blocking the grace period. */
1278 rcu_print_detail_task_stall(rsp
);
1280 rcu_check_gp_kthread_starvation(rsp
);
1282 force_quiescent_state(rsp
); /* Kick them all. */
1285 static void print_cpu_stall(struct rcu_state
*rsp
)
1288 unsigned long flags
;
1289 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1293 * OK, time to rat on ourselves...
1294 * See Documentation/RCU/stallwarn.txt for info on how to debug
1295 * RCU CPU stall warnings.
1297 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1298 print_cpu_stall_info_begin();
1299 print_cpu_stall_info(rsp
, smp_processor_id());
1300 print_cpu_stall_info_end();
1301 for_each_possible_cpu(cpu
)
1302 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1303 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1304 jiffies
- rsp
->gp_start
,
1305 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1307 rcu_check_gp_kthread_starvation(rsp
);
1309 rcu_dump_cpu_stacks(rsp
);
1311 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1312 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1313 WRITE_ONCE(rsp
->jiffies_stall
,
1314 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1315 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1318 * Attempt to revive the RCU machinery by forcing a context switch.
1320 * A context switch would normally allow the RCU state machine to make
1321 * progress and it could be we're stuck in kernel space without context
1322 * switches for an entirely unreasonable amount of time.
1324 resched_cpu(smp_processor_id());
1327 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1329 unsigned long completed
;
1330 unsigned long gpnum
;
1334 struct rcu_node
*rnp
;
1336 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1341 * Lots of memory barriers to reject false positives.
1343 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1344 * then rsp->gp_start, and finally rsp->completed. These values
1345 * are updated in the opposite order with memory barriers (or
1346 * equivalent) during grace-period initialization and cleanup.
1347 * Now, a false positive can occur if we get an new value of
1348 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1349 * the memory barriers, the only way that this can happen is if one
1350 * grace period ends and another starts between these two fetches.
1351 * Detect this by comparing rsp->completed with the previous fetch
1354 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1355 * and rsp->gp_start suffice to forestall false positives.
1357 gpnum
= READ_ONCE(rsp
->gpnum
);
1358 smp_rmb(); /* Pick up ->gpnum first... */
1359 js
= READ_ONCE(rsp
->jiffies_stall
);
1360 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1361 gps
= READ_ONCE(rsp
->gp_start
);
1362 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1363 completed
= READ_ONCE(rsp
->completed
);
1364 if (ULONG_CMP_GE(completed
, gpnum
) ||
1365 ULONG_CMP_LT(j
, js
) ||
1366 ULONG_CMP_GE(gps
, js
))
1367 return; /* No stall or GP completed since entering function. */
1369 if (rcu_gp_in_progress(rsp
) &&
1370 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1372 /* We haven't checked in, so go dump stack. */
1373 print_cpu_stall(rsp
);
1375 } else if (rcu_gp_in_progress(rsp
) &&
1376 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1378 /* They had a few time units to dump stack, so complain. */
1379 print_other_cpu_stall(rsp
, gpnum
);
1384 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1386 * Set the stall-warning timeout way off into the future, thus preventing
1387 * any RCU CPU stall-warning messages from appearing in the current set of
1388 * RCU grace periods.
1390 * The caller must disable hard irqs.
1392 void rcu_cpu_stall_reset(void)
1394 struct rcu_state
*rsp
;
1396 for_each_rcu_flavor(rsp
)
1397 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1401 * Initialize the specified rcu_data structure's default callback list
1402 * to empty. The default callback list is the one that is not used by
1403 * no-callbacks CPUs.
1405 static void init_default_callback_list(struct rcu_data
*rdp
)
1409 rdp
->nxtlist
= NULL
;
1410 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1411 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1415 * Initialize the specified rcu_data structure's callback list to empty.
1417 static void init_callback_list(struct rcu_data
*rdp
)
1419 if (init_nocb_callback_list(rdp
))
1421 init_default_callback_list(rdp
);
1425 * Determine the value that ->completed will have at the end of the
1426 * next subsequent grace period. This is used to tag callbacks so that
1427 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1428 * been dyntick-idle for an extended period with callbacks under the
1429 * influence of RCU_FAST_NO_HZ.
1431 * The caller must hold rnp->lock with interrupts disabled.
1433 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1434 struct rcu_node
*rnp
)
1437 * If RCU is idle, we just wait for the next grace period.
1438 * But we can only be sure that RCU is idle if we are looking
1439 * at the root rcu_node structure -- otherwise, a new grace
1440 * period might have started, but just not yet gotten around
1441 * to initializing the current non-root rcu_node structure.
1443 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1444 return rnp
->completed
+ 1;
1447 * Otherwise, wait for a possible partial grace period and
1448 * then the subsequent full grace period.
1450 return rnp
->completed
+ 2;
1454 * Trace-event helper function for rcu_start_future_gp() and
1455 * rcu_nocb_wait_gp().
1457 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1458 unsigned long c
, const char *s
)
1460 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1461 rnp
->completed
, c
, rnp
->level
,
1462 rnp
->grplo
, rnp
->grphi
, s
);
1466 * Start some future grace period, as needed to handle newly arrived
1467 * callbacks. The required future grace periods are recorded in each
1468 * rcu_node structure's ->need_future_gp field. Returns true if there
1469 * is reason to awaken the grace-period kthread.
1471 * The caller must hold the specified rcu_node structure's ->lock.
1473 static bool __maybe_unused
1474 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1475 unsigned long *c_out
)
1480 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1483 * Pick up grace-period number for new callbacks. If this
1484 * grace period is already marked as needed, return to the caller.
1486 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1487 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1488 if (rnp
->need_future_gp
[c
& 0x1]) {
1489 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1494 * If either this rcu_node structure or the root rcu_node structure
1495 * believe that a grace period is in progress, then we must wait
1496 * for the one following, which is in "c". Because our request
1497 * will be noticed at the end of the current grace period, we don't
1498 * need to explicitly start one. We only do the lockless check
1499 * of rnp_root's fields if the current rcu_node structure thinks
1500 * there is no grace period in flight, and because we hold rnp->lock,
1501 * the only possible change is when rnp_root's two fields are
1502 * equal, in which case rnp_root->gpnum might be concurrently
1503 * incremented. But that is OK, as it will just result in our
1504 * doing some extra useless work.
1506 if (rnp
->gpnum
!= rnp
->completed
||
1507 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1508 rnp
->need_future_gp
[c
& 0x1]++;
1509 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1514 * There might be no grace period in progress. If we don't already
1515 * hold it, acquire the root rcu_node structure's lock in order to
1516 * start one (if needed).
1518 if (rnp
!= rnp_root
) {
1519 raw_spin_lock(&rnp_root
->lock
);
1520 smp_mb__after_unlock_lock();
1524 * Get a new grace-period number. If there really is no grace
1525 * period in progress, it will be smaller than the one we obtained
1526 * earlier. Adjust callbacks as needed. Note that even no-CBs
1527 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1529 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1530 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1531 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1532 rdp
->nxtcompleted
[i
] = c
;
1535 * If the needed for the required grace period is already
1536 * recorded, trace and leave.
1538 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1539 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1543 /* Record the need for the future grace period. */
1544 rnp_root
->need_future_gp
[c
& 0x1]++;
1546 /* If a grace period is not already in progress, start one. */
1547 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1548 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1550 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1551 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1554 if (rnp
!= rnp_root
)
1555 raw_spin_unlock(&rnp_root
->lock
);
1563 * Clean up any old requests for the just-ended grace period. Also return
1564 * whether any additional grace periods have been requested. Also invoke
1565 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1566 * waiting for this grace period to complete.
1568 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1570 int c
= rnp
->completed
;
1572 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1574 rcu_nocb_gp_cleanup(rsp
, rnp
);
1575 rnp
->need_future_gp
[c
& 0x1] = 0;
1576 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1577 trace_rcu_future_gp(rnp
, rdp
, c
,
1578 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1583 * Awaken the grace-period kthread for the specified flavor of RCU.
1584 * Don't do a self-awaken, and don't bother awakening when there is
1585 * nothing for the grace-period kthread to do (as in several CPUs
1586 * raced to awaken, and we lost), and finally don't try to awaken
1587 * a kthread that has not yet been created.
1589 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1591 if (current
== rsp
->gp_kthread
||
1592 !READ_ONCE(rsp
->gp_flags
) ||
1595 wake_up(&rsp
->gp_wq
);
1599 * If there is room, assign a ->completed number to any callbacks on
1600 * this CPU that have not already been assigned. Also accelerate any
1601 * callbacks that were previously assigned a ->completed number that has
1602 * since proven to be too conservative, which can happen if callbacks get
1603 * assigned a ->completed number while RCU is idle, but with reference to
1604 * a non-root rcu_node structure. This function is idempotent, so it does
1605 * not hurt to call it repeatedly. Returns an flag saying that we should
1606 * awaken the RCU grace-period kthread.
1608 * The caller must hold rnp->lock with interrupts disabled.
1610 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1611 struct rcu_data
*rdp
)
1617 /* If the CPU has no callbacks, nothing to do. */
1618 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1622 * Starting from the sublist containing the callbacks most
1623 * recently assigned a ->completed number and working down, find the
1624 * first sublist that is not assignable to an upcoming grace period.
1625 * Such a sublist has something in it (first two tests) and has
1626 * a ->completed number assigned that will complete sooner than
1627 * the ->completed number for newly arrived callbacks (last test).
1629 * The key point is that any later sublist can be assigned the
1630 * same ->completed number as the newly arrived callbacks, which
1631 * means that the callbacks in any of these later sublist can be
1632 * grouped into a single sublist, whether or not they have already
1633 * been assigned a ->completed number.
1635 c
= rcu_cbs_completed(rsp
, rnp
);
1636 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1637 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1638 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1642 * If there are no sublist for unassigned callbacks, leave.
1643 * At the same time, advance "i" one sublist, so that "i" will
1644 * index into the sublist where all the remaining callbacks should
1647 if (++i
>= RCU_NEXT_TAIL
)
1651 * Assign all subsequent callbacks' ->completed number to the next
1652 * full grace period and group them all in the sublist initially
1655 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1656 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1657 rdp
->nxtcompleted
[i
] = c
;
1659 /* Record any needed additional grace periods. */
1660 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1662 /* Trace depending on how much we were able to accelerate. */
1663 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1664 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1666 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1671 * Move any callbacks whose grace period has completed to the
1672 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1673 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1674 * sublist. This function is idempotent, so it does not hurt to
1675 * invoke it repeatedly. As long as it is not invoked -too- often...
1676 * Returns true if the RCU grace-period kthread needs to be awakened.
1678 * The caller must hold rnp->lock with interrupts disabled.
1680 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1681 struct rcu_data
*rdp
)
1685 /* If the CPU has no callbacks, nothing to do. */
1686 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1690 * Find all callbacks whose ->completed numbers indicate that they
1691 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1693 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1694 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1696 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1698 /* Clean up any sublist tail pointers that were misordered above. */
1699 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1700 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1702 /* Copy down callbacks to fill in empty sublists. */
1703 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1704 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1706 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1707 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1710 /* Classify any remaining callbacks. */
1711 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1715 * Update CPU-local rcu_data state to record the beginnings and ends of
1716 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1717 * structure corresponding to the current CPU, and must have irqs disabled.
1718 * Returns true if the grace-period kthread needs to be awakened.
1720 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1721 struct rcu_data
*rdp
)
1725 /* Handle the ends of any preceding grace periods first. */
1726 if (rdp
->completed
== rnp
->completed
&&
1727 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1729 /* No grace period end, so just accelerate recent callbacks. */
1730 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1734 /* Advance callbacks. */
1735 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1737 /* Remember that we saw this grace-period completion. */
1738 rdp
->completed
= rnp
->completed
;
1739 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1742 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1744 * If the current grace period is waiting for this CPU,
1745 * set up to detect a quiescent state, otherwise don't
1746 * go looking for one.
1748 rdp
->gpnum
= rnp
->gpnum
;
1749 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1750 rdp
->passed_quiesce
= 0;
1751 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1752 rdp
->qs_pending
= !!(rnp
->qsmask
& rdp
->grpmask
);
1753 zero_cpu_stall_ticks(rdp
);
1754 WRITE_ONCE(rdp
->gpwrap
, false);
1759 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1761 unsigned long flags
;
1763 struct rcu_node
*rnp
;
1765 local_irq_save(flags
);
1767 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1768 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1769 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1770 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
1771 local_irq_restore(flags
);
1774 smp_mb__after_unlock_lock();
1775 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1776 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1778 rcu_gp_kthread_wake(rsp
);
1781 static void rcu_gp_slow(struct rcu_state
*rsp
, int delay
)
1784 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1785 schedule_timeout_uninterruptible(delay
);
1789 * Initialize a new grace period. Return 0 if no grace period required.
1791 static int rcu_gp_init(struct rcu_state
*rsp
)
1793 unsigned long oldmask
;
1794 struct rcu_data
*rdp
;
1795 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1797 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1798 raw_spin_lock_irq(&rnp
->lock
);
1799 smp_mb__after_unlock_lock();
1800 if (!READ_ONCE(rsp
->gp_flags
)) {
1801 /* Spurious wakeup, tell caller to go back to sleep. */
1802 raw_spin_unlock_irq(&rnp
->lock
);
1805 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1807 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1809 * Grace period already in progress, don't start another.
1810 * Not supposed to be able to happen.
1812 raw_spin_unlock_irq(&rnp
->lock
);
1816 /* Advance to a new grace period and initialize state. */
1817 record_gp_stall_check_time(rsp
);
1818 /* Record GP times before starting GP, hence smp_store_release(). */
1819 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1820 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1821 raw_spin_unlock_irq(&rnp
->lock
);
1824 * Apply per-leaf buffered online and offline operations to the
1825 * rcu_node tree. Note that this new grace period need not wait
1826 * for subsequent online CPUs, and that quiescent-state forcing
1827 * will handle subsequent offline CPUs.
1829 rcu_for_each_leaf_node(rsp
, rnp
) {
1830 rcu_gp_slow(rsp
, gp_preinit_delay
);
1831 raw_spin_lock_irq(&rnp
->lock
);
1832 smp_mb__after_unlock_lock();
1833 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1834 !rnp
->wait_blkd_tasks
) {
1835 /* Nothing to do on this leaf rcu_node structure. */
1836 raw_spin_unlock_irq(&rnp
->lock
);
1840 /* Record old state, apply changes to ->qsmaskinit field. */
1841 oldmask
= rnp
->qsmaskinit
;
1842 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1844 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1845 if (!oldmask
!= !rnp
->qsmaskinit
) {
1846 if (!oldmask
) /* First online CPU for this rcu_node. */
1847 rcu_init_new_rnp(rnp
);
1848 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1849 rnp
->wait_blkd_tasks
= true;
1850 else /* Last offline CPU and can propagate. */
1851 rcu_cleanup_dead_rnp(rnp
);
1855 * If all waited-on tasks from prior grace period are
1856 * done, and if all this rcu_node structure's CPUs are
1857 * still offline, propagate up the rcu_node tree and
1858 * clear ->wait_blkd_tasks. Otherwise, if one of this
1859 * rcu_node structure's CPUs has since come back online,
1860 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1861 * checks for this, so just call it unconditionally).
1863 if (rnp
->wait_blkd_tasks
&&
1864 (!rcu_preempt_has_tasks(rnp
) ||
1866 rnp
->wait_blkd_tasks
= false;
1867 rcu_cleanup_dead_rnp(rnp
);
1870 raw_spin_unlock_irq(&rnp
->lock
);
1874 * Set the quiescent-state-needed bits in all the rcu_node
1875 * structures for all currently online CPUs in breadth-first order,
1876 * starting from the root rcu_node structure, relying on the layout
1877 * of the tree within the rsp->node[] array. Note that other CPUs
1878 * will access only the leaves of the hierarchy, thus seeing that no
1879 * grace period is in progress, at least until the corresponding
1880 * leaf node has been initialized. In addition, we have excluded
1881 * CPU-hotplug operations.
1883 * The grace period cannot complete until the initialization
1884 * process finishes, because this kthread handles both.
1886 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1887 rcu_gp_slow(rsp
, gp_init_delay
);
1888 raw_spin_lock_irq(&rnp
->lock
);
1889 smp_mb__after_unlock_lock();
1890 rdp
= this_cpu_ptr(rsp
->rda
);
1891 rcu_preempt_check_blocked_tasks(rnp
);
1892 rnp
->qsmask
= rnp
->qsmaskinit
;
1893 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1894 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1895 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1896 if (rnp
== rdp
->mynode
)
1897 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1898 rcu_preempt_boost_start_gp(rnp
);
1899 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1900 rnp
->level
, rnp
->grplo
,
1901 rnp
->grphi
, rnp
->qsmask
);
1902 raw_spin_unlock_irq(&rnp
->lock
);
1903 cond_resched_rcu_qs();
1904 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1911 * Do one round of quiescent-state forcing.
1913 static int rcu_gp_fqs(struct rcu_state
*rsp
, int fqs_state_in
)
1915 int fqs_state
= fqs_state_in
;
1916 bool isidle
= false;
1918 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1920 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1922 if (fqs_state
== RCU_SAVE_DYNTICK
) {
1923 /* Collect dyntick-idle snapshots. */
1924 if (is_sysidle_rcu_state(rsp
)) {
1926 maxj
= jiffies
- ULONG_MAX
/ 4;
1928 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1930 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1931 fqs_state
= RCU_FORCE_QS
;
1933 /* Handle dyntick-idle and offline CPUs. */
1935 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1937 /* Clear flag to prevent immediate re-entry. */
1938 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1939 raw_spin_lock_irq(&rnp
->lock
);
1940 smp_mb__after_unlock_lock();
1941 WRITE_ONCE(rsp
->gp_flags
,
1942 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
1943 raw_spin_unlock_irq(&rnp
->lock
);
1949 * Clean up after the old grace period.
1951 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
1953 unsigned long gp_duration
;
1954 bool needgp
= false;
1956 struct rcu_data
*rdp
;
1957 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1959 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1960 raw_spin_lock_irq(&rnp
->lock
);
1961 smp_mb__after_unlock_lock();
1962 gp_duration
= jiffies
- rsp
->gp_start
;
1963 if (gp_duration
> rsp
->gp_max
)
1964 rsp
->gp_max
= gp_duration
;
1967 * We know the grace period is complete, but to everyone else
1968 * it appears to still be ongoing. But it is also the case
1969 * that to everyone else it looks like there is nothing that
1970 * they can do to advance the grace period. It is therefore
1971 * safe for us to drop the lock in order to mark the grace
1972 * period as completed in all of the rcu_node structures.
1974 raw_spin_unlock_irq(&rnp
->lock
);
1977 * Propagate new ->completed value to rcu_node structures so
1978 * that other CPUs don't have to wait until the start of the next
1979 * grace period to process their callbacks. This also avoids
1980 * some nasty RCU grace-period initialization races by forcing
1981 * the end of the current grace period to be completely recorded in
1982 * all of the rcu_node structures before the beginning of the next
1983 * grace period is recorded in any of the rcu_node structures.
1985 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1986 raw_spin_lock_irq(&rnp
->lock
);
1987 smp_mb__after_unlock_lock();
1988 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
1989 WARN_ON_ONCE(rnp
->qsmask
);
1990 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
1991 rdp
= this_cpu_ptr(rsp
->rda
);
1992 if (rnp
== rdp
->mynode
)
1993 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
1994 /* smp_mb() provided by prior unlock-lock pair. */
1995 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
1996 raw_spin_unlock_irq(&rnp
->lock
);
1997 cond_resched_rcu_qs();
1998 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1999 rcu_gp_slow(rsp
, gp_cleanup_delay
);
2001 rnp
= rcu_get_root(rsp
);
2002 raw_spin_lock_irq(&rnp
->lock
);
2003 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
2004 rcu_nocb_gp_set(rnp
, nocb
);
2006 /* Declare grace period done. */
2007 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
2008 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
2009 rsp
->fqs_state
= RCU_GP_IDLE
;
2010 rdp
= this_cpu_ptr(rsp
->rda
);
2011 /* Advance CBs to reduce false positives below. */
2012 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
2013 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
2014 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2015 trace_rcu_grace_period(rsp
->name
,
2016 READ_ONCE(rsp
->gpnum
),
2019 raw_spin_unlock_irq(&rnp
->lock
);
2023 * Body of kthread that handles grace periods.
2025 static int __noreturn
rcu_gp_kthread(void *arg
)
2031 struct rcu_state
*rsp
= arg
;
2032 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2034 rcu_bind_gp_kthread();
2037 /* Handle grace-period start. */
2039 trace_rcu_grace_period(rsp
->name
,
2040 READ_ONCE(rsp
->gpnum
),
2042 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2043 wait_event_interruptible(rsp
->gp_wq
,
2044 READ_ONCE(rsp
->gp_flags
) &
2046 rsp
->gp_state
= RCU_GP_DONE_GPS
;
2047 /* Locking provides needed memory barrier. */
2048 if (rcu_gp_init(rsp
))
2050 cond_resched_rcu_qs();
2051 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2052 WARN_ON(signal_pending(current
));
2053 trace_rcu_grace_period(rsp
->name
,
2054 READ_ONCE(rsp
->gpnum
),
2058 /* Handle quiescent-state forcing. */
2059 fqs_state
= RCU_SAVE_DYNTICK
;
2060 j
= jiffies_till_first_fqs
;
2063 jiffies_till_first_fqs
= HZ
;
2068 rsp
->jiffies_force_qs
= jiffies
+ j
;
2069 trace_rcu_grace_period(rsp
->name
,
2070 READ_ONCE(rsp
->gpnum
),
2072 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2073 ret
= wait_event_interruptible_timeout(rsp
->gp_wq
,
2074 ((gf
= READ_ONCE(rsp
->gp_flags
)) &
2076 (!READ_ONCE(rnp
->qsmask
) &&
2077 !rcu_preempt_blocked_readers_cgp(rnp
)),
2079 rsp
->gp_state
= RCU_GP_DONE_FQS
;
2080 /* Locking provides needed memory barriers. */
2081 /* If grace period done, leave loop. */
2082 if (!READ_ONCE(rnp
->qsmask
) &&
2083 !rcu_preempt_blocked_readers_cgp(rnp
))
2085 /* If time for quiescent-state forcing, do it. */
2086 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2087 (gf
& RCU_GP_FLAG_FQS
)) {
2088 trace_rcu_grace_period(rsp
->name
,
2089 READ_ONCE(rsp
->gpnum
),
2091 fqs_state
= rcu_gp_fqs(rsp
, fqs_state
);
2092 trace_rcu_grace_period(rsp
->name
,
2093 READ_ONCE(rsp
->gpnum
),
2095 cond_resched_rcu_qs();
2096 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2098 /* Deal with stray signal. */
2099 cond_resched_rcu_qs();
2100 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2101 WARN_ON(signal_pending(current
));
2102 trace_rcu_grace_period(rsp
->name
,
2103 READ_ONCE(rsp
->gpnum
),
2106 j
= jiffies_till_next_fqs
;
2109 jiffies_till_next_fqs
= HZ
;
2112 jiffies_till_next_fqs
= 1;
2116 /* Handle grace-period end. */
2117 rsp
->gp_state
= RCU_GP_CLEANUP
;
2118 rcu_gp_cleanup(rsp
);
2119 rsp
->gp_state
= RCU_GP_CLEANED
;
2124 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2125 * in preparation for detecting the next grace period. The caller must hold
2126 * the root node's ->lock and hard irqs must be disabled.
2128 * Note that it is legal for a dying CPU (which is marked as offline) to
2129 * invoke this function. This can happen when the dying CPU reports its
2132 * Returns true if the grace-period kthread must be awakened.
2135 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2136 struct rcu_data
*rdp
)
2138 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2140 * Either we have not yet spawned the grace-period
2141 * task, this CPU does not need another grace period,
2142 * or a grace period is already in progress.
2143 * Either way, don't start a new grace period.
2147 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2148 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2152 * We can't do wakeups while holding the rnp->lock, as that
2153 * could cause possible deadlocks with the rq->lock. Defer
2154 * the wakeup to our caller.
2160 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2161 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2162 * is invoked indirectly from rcu_advance_cbs(), which would result in
2163 * endless recursion -- or would do so if it wasn't for the self-deadlock
2164 * that is encountered beforehand.
2166 * Returns true if the grace-period kthread needs to be awakened.
2168 static bool rcu_start_gp(struct rcu_state
*rsp
)
2170 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2171 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2175 * If there is no grace period in progress right now, any
2176 * callbacks we have up to this point will be satisfied by the
2177 * next grace period. Also, advancing the callbacks reduces the
2178 * probability of false positives from cpu_needs_another_gp()
2179 * resulting in pointless grace periods. So, advance callbacks
2180 * then start the grace period!
2182 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2183 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2188 * Report a full set of quiescent states to the specified rcu_state
2189 * data structure. This involves cleaning up after the prior grace
2190 * period and letting rcu_start_gp() start up the next grace period
2191 * if one is needed. Note that the caller must hold rnp->lock, which
2192 * is released before return.
2194 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2195 __releases(rcu_get_root(rsp
)->lock
)
2197 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2198 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2199 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2200 rcu_gp_kthread_wake(rsp
);
2204 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2205 * Allows quiescent states for a group of CPUs to be reported at one go
2206 * to the specified rcu_node structure, though all the CPUs in the group
2207 * must be represented by the same rcu_node structure (which need not be a
2208 * leaf rcu_node structure, though it often will be). The gps parameter
2209 * is the grace-period snapshot, which means that the quiescent states
2210 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2211 * must be held upon entry, and it is released before return.
2214 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2215 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2216 __releases(rnp
->lock
)
2218 unsigned long oldmask
= 0;
2219 struct rcu_node
*rnp_c
;
2221 /* Walk up the rcu_node hierarchy. */
2223 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2226 * Our bit has already been cleared, or the
2227 * relevant grace period is already over, so done.
2229 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2232 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2233 rnp
->qsmask
&= ~mask
;
2234 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2235 mask
, rnp
->qsmask
, rnp
->level
,
2236 rnp
->grplo
, rnp
->grphi
,
2238 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2240 /* Other bits still set at this level, so done. */
2241 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2244 mask
= rnp
->grpmask
;
2245 if (rnp
->parent
== NULL
) {
2247 /* No more levels. Exit loop holding root lock. */
2251 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2254 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2255 smp_mb__after_unlock_lock();
2256 oldmask
= rnp_c
->qsmask
;
2260 * Get here if we are the last CPU to pass through a quiescent
2261 * state for this grace period. Invoke rcu_report_qs_rsp()
2262 * to clean up and start the next grace period if one is needed.
2264 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2268 * Record a quiescent state for all tasks that were previously queued
2269 * on the specified rcu_node structure and that were blocking the current
2270 * RCU grace period. The caller must hold the specified rnp->lock with
2271 * irqs disabled, and this lock is released upon return, but irqs remain
2274 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2275 struct rcu_node
*rnp
, unsigned long flags
)
2276 __releases(rnp
->lock
)
2280 struct rcu_node
*rnp_p
;
2282 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2283 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2284 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2285 return; /* Still need more quiescent states! */
2288 rnp_p
= rnp
->parent
;
2289 if (rnp_p
== NULL
) {
2291 * Only one rcu_node structure in the tree, so don't
2292 * try to report up to its nonexistent parent!
2294 rcu_report_qs_rsp(rsp
, flags
);
2298 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2300 mask
= rnp
->grpmask
;
2301 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2302 raw_spin_lock(&rnp_p
->lock
); /* irqs already disabled. */
2303 smp_mb__after_unlock_lock();
2304 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2308 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2309 * structure. This must be either called from the specified CPU, or
2310 * called when the specified CPU is known to be offline (and when it is
2311 * also known that no other CPU is concurrently trying to help the offline
2312 * CPU). The lastcomp argument is used to make sure we are still in the
2313 * grace period of interest. We don't want to end the current grace period
2314 * based on quiescent states detected in an earlier grace period!
2317 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2319 unsigned long flags
;
2322 struct rcu_node
*rnp
;
2325 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2326 smp_mb__after_unlock_lock();
2327 if ((rdp
->passed_quiesce
== 0 &&
2328 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2329 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2333 * The grace period in which this quiescent state was
2334 * recorded has ended, so don't report it upwards.
2335 * We will instead need a new quiescent state that lies
2336 * within the current grace period.
2338 rdp
->passed_quiesce
= 0; /* need qs for new gp. */
2339 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2340 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2343 mask
= rdp
->grpmask
;
2344 if ((rnp
->qsmask
& mask
) == 0) {
2345 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2347 rdp
->qs_pending
= 0;
2350 * This GP can't end until cpu checks in, so all of our
2351 * callbacks can be processed during the next GP.
2353 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2355 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2356 /* ^^^ Released rnp->lock */
2358 rcu_gp_kthread_wake(rsp
);
2363 * Check to see if there is a new grace period of which this CPU
2364 * is not yet aware, and if so, set up local rcu_data state for it.
2365 * Otherwise, see if this CPU has just passed through its first
2366 * quiescent state for this grace period, and record that fact if so.
2369 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2371 /* Check for grace-period ends and beginnings. */
2372 note_gp_changes(rsp
, rdp
);
2375 * Does this CPU still need to do its part for current grace period?
2376 * If no, return and let the other CPUs do their part as well.
2378 if (!rdp
->qs_pending
)
2382 * Was there a quiescent state since the beginning of the grace
2383 * period? If no, then exit and wait for the next call.
2385 if (!rdp
->passed_quiesce
&&
2386 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2390 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2393 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2397 * Send the specified CPU's RCU callbacks to the orphanage. The
2398 * specified CPU must be offline, and the caller must hold the
2402 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2403 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2405 /* No-CBs CPUs do not have orphanable callbacks. */
2406 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) || rcu_is_nocb_cpu(rdp
->cpu
))
2410 * Orphan the callbacks. First adjust the counts. This is safe
2411 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2412 * cannot be running now. Thus no memory barrier is required.
2414 if (rdp
->nxtlist
!= NULL
) {
2415 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2416 rsp
->qlen
+= rdp
->qlen
;
2417 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2419 WRITE_ONCE(rdp
->qlen
, 0);
2423 * Next, move those callbacks still needing a grace period to
2424 * the orphanage, where some other CPU will pick them up.
2425 * Some of the callbacks might have gone partway through a grace
2426 * period, but that is too bad. They get to start over because we
2427 * cannot assume that grace periods are synchronized across CPUs.
2428 * We don't bother updating the ->nxttail[] array yet, instead
2429 * we just reset the whole thing later on.
2431 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2432 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2433 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2434 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2438 * Then move the ready-to-invoke callbacks to the orphanage,
2439 * where some other CPU will pick them up. These will not be
2440 * required to pass though another grace period: They are done.
2442 if (rdp
->nxtlist
!= NULL
) {
2443 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2444 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2448 * Finally, initialize the rcu_data structure's list to empty and
2449 * disallow further callbacks on this CPU.
2451 init_callback_list(rdp
);
2452 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2456 * Adopt the RCU callbacks from the specified rcu_state structure's
2457 * orphanage. The caller must hold the ->orphan_lock.
2459 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2462 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2464 /* No-CBs CPUs are handled specially. */
2465 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2466 rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2469 /* Do the accounting first. */
2470 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2471 rdp
->qlen
+= rsp
->qlen
;
2472 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2473 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2474 rcu_idle_count_callbacks_posted();
2479 * We do not need a memory barrier here because the only way we
2480 * can get here if there is an rcu_barrier() in flight is if
2481 * we are the task doing the rcu_barrier().
2484 /* First adopt the ready-to-invoke callbacks. */
2485 if (rsp
->orphan_donelist
!= NULL
) {
2486 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2487 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2488 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2489 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2490 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2491 rsp
->orphan_donelist
= NULL
;
2492 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2495 /* And then adopt the callbacks that still need a grace period. */
2496 if (rsp
->orphan_nxtlist
!= NULL
) {
2497 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2498 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2499 rsp
->orphan_nxtlist
= NULL
;
2500 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2505 * Trace the fact that this CPU is going offline.
2507 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2509 RCU_TRACE(unsigned long mask
);
2510 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2511 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2513 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2516 RCU_TRACE(mask
= rdp
->grpmask
);
2517 trace_rcu_grace_period(rsp
->name
,
2518 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2523 * All CPUs for the specified rcu_node structure have gone offline,
2524 * and all tasks that were preempted within an RCU read-side critical
2525 * section while running on one of those CPUs have since exited their RCU
2526 * read-side critical section. Some other CPU is reporting this fact with
2527 * the specified rcu_node structure's ->lock held and interrupts disabled.
2528 * This function therefore goes up the tree of rcu_node structures,
2529 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2530 * the leaf rcu_node structure's ->qsmaskinit field has already been
2533 * This function does check that the specified rcu_node structure has
2534 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2535 * prematurely. That said, invoking it after the fact will cost you
2536 * a needless lock acquisition. So once it has done its work, don't
2539 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2542 struct rcu_node
*rnp
= rnp_leaf
;
2544 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2545 rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2548 mask
= rnp
->grpmask
;
2552 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
2553 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2554 rnp
->qsmaskinit
&= ~mask
;
2555 rnp
->qsmask
&= ~mask
;
2556 if (rnp
->qsmaskinit
) {
2557 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2560 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
2565 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2566 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2569 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
2571 unsigned long flags
;
2573 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2574 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2576 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2579 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2580 mask
= rdp
->grpmask
;
2581 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2582 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2583 rnp
->qsmaskinitnext
&= ~mask
;
2584 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2588 * The CPU has been completely removed, and some other CPU is reporting
2589 * this fact from process context. Do the remainder of the cleanup,
2590 * including orphaning the outgoing CPU's RCU callbacks, and also
2591 * adopting them. There can only be one CPU hotplug operation at a time,
2592 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2594 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2596 unsigned long flags
;
2597 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2598 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2600 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2603 /* Adjust any no-longer-needed kthreads. */
2604 rcu_boost_kthread_setaffinity(rnp
, -1);
2606 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2607 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2608 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2609 rcu_adopt_orphan_cbs(rsp
, flags
);
2610 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2612 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2613 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2614 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2618 * Invoke any RCU callbacks that have made it to the end of their grace
2619 * period. Thottle as specified by rdp->blimit.
2621 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2623 unsigned long flags
;
2624 struct rcu_head
*next
, *list
, **tail
;
2625 long bl
, count
, count_lazy
;
2628 /* If no callbacks are ready, just return. */
2629 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2630 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2631 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2632 need_resched(), is_idle_task(current
),
2633 rcu_is_callbacks_kthread());
2638 * Extract the list of ready callbacks, disabling to prevent
2639 * races with call_rcu() from interrupt handlers.
2641 local_irq_save(flags
);
2642 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2644 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2645 list
= rdp
->nxtlist
;
2646 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2647 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2648 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2649 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2650 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2651 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2652 local_irq_restore(flags
);
2654 /* Invoke callbacks. */
2655 count
= count_lazy
= 0;
2659 debug_rcu_head_unqueue(list
);
2660 if (__rcu_reclaim(rsp
->name
, list
))
2663 /* Stop only if limit reached and CPU has something to do. */
2664 if (++count
>= bl
&&
2666 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2670 local_irq_save(flags
);
2671 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2672 is_idle_task(current
),
2673 rcu_is_callbacks_kthread());
2675 /* Update count, and requeue any remaining callbacks. */
2677 *tail
= rdp
->nxtlist
;
2678 rdp
->nxtlist
= list
;
2679 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2680 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2681 rdp
->nxttail
[i
] = tail
;
2685 smp_mb(); /* List handling before counting for rcu_barrier(). */
2686 rdp
->qlen_lazy
-= count_lazy
;
2687 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2688 rdp
->n_cbs_invoked
+= count
;
2690 /* Reinstate batch limit if we have worked down the excess. */
2691 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2692 rdp
->blimit
= blimit
;
2694 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2695 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2696 rdp
->qlen_last_fqs_check
= 0;
2697 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2698 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2699 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2700 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2702 local_irq_restore(flags
);
2704 /* Re-invoke RCU core processing if there are callbacks remaining. */
2705 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2710 * Check to see if this CPU is in a non-context-switch quiescent state
2711 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2712 * Also schedule RCU core processing.
2714 * This function must be called from hardirq context. It is normally
2715 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2716 * false, there is no point in invoking rcu_check_callbacks().
2718 void rcu_check_callbacks(int user
)
2720 trace_rcu_utilization(TPS("Start scheduler-tick"));
2721 increment_cpu_stall_ticks();
2722 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2725 * Get here if this CPU took its interrupt from user
2726 * mode or from the idle loop, and if this is not a
2727 * nested interrupt. In this case, the CPU is in
2728 * a quiescent state, so note it.
2730 * No memory barrier is required here because both
2731 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2732 * variables that other CPUs neither access nor modify,
2733 * at least not while the corresponding CPU is online.
2739 } else if (!in_softirq()) {
2742 * Get here if this CPU did not take its interrupt from
2743 * softirq, in other words, if it is not interrupting
2744 * a rcu_bh read-side critical section. This is an _bh
2745 * critical section, so note it.
2750 rcu_preempt_check_callbacks();
2754 rcu_note_voluntary_context_switch(current
);
2755 trace_rcu_utilization(TPS("End scheduler-tick"));
2759 * Scan the leaf rcu_node structures, processing dyntick state for any that
2760 * have not yet encountered a quiescent state, using the function specified.
2761 * Also initiate boosting for any threads blocked on the root rcu_node.
2763 * The caller must have suppressed start of new grace periods.
2765 static void force_qs_rnp(struct rcu_state
*rsp
,
2766 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2767 unsigned long *maxj
),
2768 bool *isidle
, unsigned long *maxj
)
2772 unsigned long flags
;
2774 struct rcu_node
*rnp
;
2776 rcu_for_each_leaf_node(rsp
, rnp
) {
2777 cond_resched_rcu_qs();
2779 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
2780 smp_mb__after_unlock_lock();
2781 if (rnp
->qsmask
== 0) {
2782 if (rcu_state_p
== &rcu_sched_state
||
2783 rsp
!= rcu_state_p
||
2784 rcu_preempt_blocked_readers_cgp(rnp
)) {
2786 * No point in scanning bits because they
2787 * are all zero. But we might need to
2788 * priority-boost blocked readers.
2790 rcu_initiate_boost(rnp
, flags
);
2791 /* rcu_initiate_boost() releases rnp->lock */
2795 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2797 * Race between grace-period
2798 * initialization and task exiting RCU
2799 * read-side critical section: Report.
2801 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2802 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2808 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2809 if ((rnp
->qsmask
& bit
) != 0) {
2810 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2815 /* Idle/offline CPUs, report (releases rnp->lock. */
2816 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2818 /* Nothing to do here, so just drop the lock. */
2819 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
2825 * Force quiescent states on reluctant CPUs, and also detect which
2826 * CPUs are in dyntick-idle mode.
2828 static void force_quiescent_state(struct rcu_state
*rsp
)
2830 unsigned long flags
;
2832 struct rcu_node
*rnp
;
2833 struct rcu_node
*rnp_old
= NULL
;
2835 /* Funnel through hierarchy to reduce memory contention. */
2836 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2837 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2838 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2839 !raw_spin_trylock(&rnp
->fqslock
);
2840 if (rnp_old
!= NULL
)
2841 raw_spin_unlock(&rnp_old
->fqslock
);
2843 rsp
->n_force_qs_lh
++;
2848 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2850 /* Reached the root of the rcu_node tree, acquire lock. */
2851 raw_spin_lock_irqsave(&rnp_old
->lock
, flags
);
2852 smp_mb__after_unlock_lock();
2853 raw_spin_unlock(&rnp_old
->fqslock
);
2854 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2855 rsp
->n_force_qs_lh
++;
2856 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2857 return; /* Someone beat us to it. */
2859 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2860 raw_spin_unlock_irqrestore(&rnp_old
->lock
, flags
);
2861 rcu_gp_kthread_wake(rsp
);
2865 * This does the RCU core processing work for the specified rcu_state
2866 * and rcu_data structures. This may be called only from the CPU to
2867 * whom the rdp belongs.
2870 __rcu_process_callbacks(struct rcu_state
*rsp
)
2872 unsigned long flags
;
2874 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2876 WARN_ON_ONCE(rdp
->beenonline
== 0);
2878 /* Update RCU state based on any recent quiescent states. */
2879 rcu_check_quiescent_state(rsp
, rdp
);
2881 /* Does this CPU require a not-yet-started grace period? */
2882 local_irq_save(flags
);
2883 if (cpu_needs_another_gp(rsp
, rdp
)) {
2884 raw_spin_lock(&rcu_get_root(rsp
)->lock
); /* irqs disabled. */
2885 needwake
= rcu_start_gp(rsp
);
2886 raw_spin_unlock_irqrestore(&rcu_get_root(rsp
)->lock
, flags
);
2888 rcu_gp_kthread_wake(rsp
);
2890 local_irq_restore(flags
);
2893 /* If there are callbacks ready, invoke them. */
2894 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2895 invoke_rcu_callbacks(rsp
, rdp
);
2897 /* Do any needed deferred wakeups of rcuo kthreads. */
2898 do_nocb_deferred_wakeup(rdp
);
2902 * Do RCU core processing for the current CPU.
2904 static void rcu_process_callbacks(struct softirq_action
*unused
)
2906 struct rcu_state
*rsp
;
2908 if (cpu_is_offline(smp_processor_id()))
2910 trace_rcu_utilization(TPS("Start RCU core"));
2911 for_each_rcu_flavor(rsp
)
2912 __rcu_process_callbacks(rsp
);
2913 trace_rcu_utilization(TPS("End RCU core"));
2917 * Schedule RCU callback invocation. If the specified type of RCU
2918 * does not support RCU priority boosting, just do a direct call,
2919 * otherwise wake up the per-CPU kernel kthread. Note that because we
2920 * are running on the current CPU with softirqs disabled, the
2921 * rcu_cpu_kthread_task cannot disappear out from under us.
2923 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2925 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2927 if (likely(!rsp
->boost
)) {
2928 rcu_do_batch(rsp
, rdp
);
2931 invoke_rcu_callbacks_kthread();
2934 static void invoke_rcu_core(void)
2936 if (cpu_online(smp_processor_id()))
2937 raise_softirq(RCU_SOFTIRQ
);
2941 * Handle any core-RCU processing required by a call_rcu() invocation.
2943 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2944 struct rcu_head
*head
, unsigned long flags
)
2949 * If called from an extended quiescent state, invoke the RCU
2950 * core in order to force a re-evaluation of RCU's idleness.
2952 if (!rcu_is_watching())
2955 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2956 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2960 * Force the grace period if too many callbacks or too long waiting.
2961 * Enforce hysteresis, and don't invoke force_quiescent_state()
2962 * if some other CPU has recently done so. Also, don't bother
2963 * invoking force_quiescent_state() if the newly enqueued callback
2964 * is the only one waiting for a grace period to complete.
2966 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2968 /* Are we ignoring a completed grace period? */
2969 note_gp_changes(rsp
, rdp
);
2971 /* Start a new grace period if one not already started. */
2972 if (!rcu_gp_in_progress(rsp
)) {
2973 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2975 raw_spin_lock(&rnp_root
->lock
);
2976 smp_mb__after_unlock_lock();
2977 needwake
= rcu_start_gp(rsp
);
2978 raw_spin_unlock(&rnp_root
->lock
);
2980 rcu_gp_kthread_wake(rsp
);
2982 /* Give the grace period a kick. */
2983 rdp
->blimit
= LONG_MAX
;
2984 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
2985 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
2986 force_quiescent_state(rsp
);
2987 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2988 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2994 * RCU callback function to leak a callback.
2996 static void rcu_leak_callback(struct rcu_head
*rhp
)
3001 * Helper function for call_rcu() and friends. The cpu argument will
3002 * normally be -1, indicating "currently running CPU". It may specify
3003 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3004 * is expected to specify a CPU.
3007 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
3008 struct rcu_state
*rsp
, int cpu
, bool lazy
)
3010 unsigned long flags
;
3011 struct rcu_data
*rdp
;
3013 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
3014 if (debug_rcu_head_queue(head
)) {
3015 /* Probable double call_rcu(), so leak the callback. */
3016 WRITE_ONCE(head
->func
, rcu_leak_callback
);
3017 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3024 * Opportunistically note grace-period endings and beginnings.
3025 * Note that we might see a beginning right after we see an
3026 * end, but never vice versa, since this CPU has to pass through
3027 * a quiescent state betweentimes.
3029 local_irq_save(flags
);
3030 rdp
= this_cpu_ptr(rsp
->rda
);
3032 /* Add the callback to our list. */
3033 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3037 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3038 if (likely(rdp
->mynode
)) {
3039 /* Post-boot, so this should be for a no-CBs CPU. */
3040 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3041 WARN_ON_ONCE(offline
);
3042 /* Offline CPU, _call_rcu() illegal, leak callback. */
3043 local_irq_restore(flags
);
3047 * Very early boot, before rcu_init(). Initialize if needed
3048 * and then drop through to queue the callback.
3051 WARN_ON_ONCE(!rcu_is_watching());
3052 if (!likely(rdp
->nxtlist
))
3053 init_default_callback_list(rdp
);
3055 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3059 rcu_idle_count_callbacks_posted();
3060 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3061 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3062 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3064 if (__is_kfree_rcu_offset((unsigned long)func
))
3065 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3066 rdp
->qlen_lazy
, rdp
->qlen
);
3068 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3070 /* Go handle any RCU core processing required. */
3071 __call_rcu_core(rsp
, rdp
, head
, flags
);
3072 local_irq_restore(flags
);
3076 * Queue an RCU-sched callback for invocation after a grace period.
3078 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3080 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3082 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3085 * Queue an RCU callback for invocation after a quicker grace period.
3087 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
3089 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3091 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3094 * Queue an RCU callback for lazy invocation after a grace period.
3095 * This will likely be later named something like "call_rcu_lazy()",
3096 * but this change will require some way of tagging the lazy RCU
3097 * callbacks in the list of pending callbacks. Until then, this
3098 * function may only be called from __kfree_rcu().
3100 void kfree_call_rcu(struct rcu_head
*head
,
3101 void (*func
)(struct rcu_head
*rcu
))
3103 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3105 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3108 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3109 * any blocking grace-period wait automatically implies a grace period
3110 * if there is only one CPU online at any point time during execution
3111 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3112 * occasionally incorrectly indicate that there are multiple CPUs online
3113 * when there was in fact only one the whole time, as this just adds
3114 * some overhead: RCU still operates correctly.
3116 static inline int rcu_blocking_is_gp(void)
3120 might_sleep(); /* Check for RCU read-side critical section. */
3122 ret
= num_online_cpus() <= 1;
3128 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3130 * Control will return to the caller some time after a full rcu-sched
3131 * grace period has elapsed, in other words after all currently executing
3132 * rcu-sched read-side critical sections have completed. These read-side
3133 * critical sections are delimited by rcu_read_lock_sched() and
3134 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3135 * local_irq_disable(), and so on may be used in place of
3136 * rcu_read_lock_sched().
3138 * This means that all preempt_disable code sequences, including NMI and
3139 * non-threaded hardware-interrupt handlers, in progress on entry will
3140 * have completed before this primitive returns. However, this does not
3141 * guarantee that softirq handlers will have completed, since in some
3142 * kernels, these handlers can run in process context, and can block.
3144 * Note that this guarantee implies further memory-ordering guarantees.
3145 * On systems with more than one CPU, when synchronize_sched() returns,
3146 * each CPU is guaranteed to have executed a full memory barrier since the
3147 * end of its last RCU-sched read-side critical section whose beginning
3148 * preceded the call to synchronize_sched(). In addition, each CPU having
3149 * an RCU read-side critical section that extends beyond the return from
3150 * synchronize_sched() is guaranteed to have executed a full memory barrier
3151 * after the beginning of synchronize_sched() and before the beginning of
3152 * that RCU read-side critical section. Note that these guarantees include
3153 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3154 * that are executing in the kernel.
3156 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3157 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3158 * to have executed a full memory barrier during the execution of
3159 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3160 * again only if the system has more than one CPU).
3162 * This primitive provides the guarantees made by the (now removed)
3163 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3164 * guarantees that rcu_read_lock() sections will have completed.
3165 * In "classic RCU", these two guarantees happen to be one and
3166 * the same, but can differ in realtime RCU implementations.
3168 void synchronize_sched(void)
3170 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3171 !lock_is_held(&rcu_lock_map
) &&
3172 !lock_is_held(&rcu_sched_lock_map
),
3173 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3174 if (rcu_blocking_is_gp())
3176 if (rcu_gp_is_expedited())
3177 synchronize_sched_expedited();
3179 wait_rcu_gp(call_rcu_sched
);
3181 EXPORT_SYMBOL_GPL(synchronize_sched
);
3184 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3186 * Control will return to the caller some time after a full rcu_bh grace
3187 * period has elapsed, in other words after all currently executing rcu_bh
3188 * read-side critical sections have completed. RCU read-side critical
3189 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3190 * and may be nested.
3192 * See the description of synchronize_sched() for more detailed information
3193 * on memory ordering guarantees.
3195 void synchronize_rcu_bh(void)
3197 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map
) &&
3198 !lock_is_held(&rcu_lock_map
) &&
3199 !lock_is_held(&rcu_sched_lock_map
),
3200 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3201 if (rcu_blocking_is_gp())
3203 if (rcu_gp_is_expedited())
3204 synchronize_rcu_bh_expedited();
3206 wait_rcu_gp(call_rcu_bh
);
3208 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3211 * get_state_synchronize_rcu - Snapshot current RCU state
3213 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3214 * to determine whether or not a full grace period has elapsed in the
3217 unsigned long get_state_synchronize_rcu(void)
3220 * Any prior manipulation of RCU-protected data must happen
3221 * before the load from ->gpnum.
3226 * Make sure this load happens before the purportedly
3227 * time-consuming work between get_state_synchronize_rcu()
3228 * and cond_synchronize_rcu().
3230 return smp_load_acquire(&rcu_state_p
->gpnum
);
3232 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3235 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3237 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3239 * If a full RCU grace period has elapsed since the earlier call to
3240 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3241 * synchronize_rcu() to wait for a full grace period.
3243 * Yes, this function does not take counter wrap into account. But
3244 * counter wrap is harmless. If the counter wraps, we have waited for
3245 * more than 2 billion grace periods (and way more on a 64-bit system!),
3246 * so waiting for one additional grace period should be just fine.
3248 void cond_synchronize_rcu(unsigned long oldstate
)
3250 unsigned long newstate
;
3253 * Ensure that this load happens before any RCU-destructive
3254 * actions the caller might carry out after we return.
3256 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3257 if (ULONG_CMP_GE(oldstate
, newstate
))
3260 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3262 static int synchronize_sched_expedited_cpu_stop(void *data
)
3265 * There must be a full memory barrier on each affected CPU
3266 * between the time that try_stop_cpus() is called and the
3267 * time that it returns.
3269 * In the current initial implementation of cpu_stop, the
3270 * above condition is already met when the control reaches
3271 * this point and the following smp_mb() is not strictly
3272 * necessary. Do smp_mb() anyway for documentation and
3273 * robustness against future implementation changes.
3275 smp_mb(); /* See above comment block. */
3280 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3282 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3283 * approach to force the grace period to end quickly. This consumes
3284 * significant time on all CPUs and is unfriendly to real-time workloads,
3285 * so is thus not recommended for any sort of common-case code. In fact,
3286 * if you are using synchronize_sched_expedited() in a loop, please
3287 * restructure your code to batch your updates, and then use a single
3288 * synchronize_sched() instead.
3290 * This implementation can be thought of as an application of ticket
3291 * locking to RCU, with sync_sched_expedited_started and
3292 * sync_sched_expedited_done taking on the roles of the halves
3293 * of the ticket-lock word. Each task atomically increments
3294 * sync_sched_expedited_started upon entry, snapshotting the old value,
3295 * then attempts to stop all the CPUs. If this succeeds, then each
3296 * CPU will have executed a context switch, resulting in an RCU-sched
3297 * grace period. We are then done, so we use atomic_cmpxchg() to
3298 * update sync_sched_expedited_done to match our snapshot -- but
3299 * only if someone else has not already advanced past our snapshot.
3301 * On the other hand, if try_stop_cpus() fails, we check the value
3302 * of sync_sched_expedited_done. If it has advanced past our
3303 * initial snapshot, then someone else must have forced a grace period
3304 * some time after we took our snapshot. In this case, our work is
3305 * done for us, and we can simply return. Otherwise, we try again,
3306 * but keep our initial snapshot for purposes of checking for someone
3307 * doing our work for us.
3309 * If we fail too many times in a row, we fall back to synchronize_sched().
3311 void synchronize_sched_expedited(void)
3316 long firstsnap
, s
, snap
;
3318 struct rcu_state
*rsp
= &rcu_sched_state
;
3321 * If we are in danger of counter wrap, just do synchronize_sched().
3322 * By allowing sync_sched_expedited_started to advance no more than
3323 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3324 * that more than 3.5 billion CPUs would be required to force a
3325 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3326 * course be required on a 64-bit system.
3328 if (ULONG_CMP_GE((ulong
)atomic_long_read(&rsp
->expedited_start
),
3329 (ulong
)atomic_long_read(&rsp
->expedited_done
) +
3331 wait_rcu_gp(call_rcu_sched
);
3332 atomic_long_inc(&rsp
->expedited_wrap
);
3337 * Take a ticket. Note that atomic_inc_return() implies a
3338 * full memory barrier.
3340 snap
= atomic_long_inc_return(&rsp
->expedited_start
);
3342 if (!try_get_online_cpus()) {
3343 /* CPU hotplug operation in flight, fall back to normal GP. */
3344 wait_rcu_gp(call_rcu_sched
);
3345 atomic_long_inc(&rsp
->expedited_normal
);
3348 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3350 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3351 cma
= zalloc_cpumask_var(&cm
, GFP_KERNEL
);
3353 cpumask_copy(cm
, cpu_online_mask
);
3354 cpumask_clear_cpu(raw_smp_processor_id(), cm
);
3355 for_each_cpu(cpu
, cm
) {
3356 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3358 if (!(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3359 cpumask_clear_cpu(cpu
, cm
);
3361 if (cpumask_weight(cm
) == 0)
3366 * Each pass through the following loop attempts to force a
3367 * context switch on each CPU.
3369 while (try_stop_cpus(cma
? cm
: cpu_online_mask
,
3370 synchronize_sched_expedited_cpu_stop
,
3373 atomic_long_inc(&rsp
->expedited_tryfail
);
3375 /* Check to see if someone else did our work for us. */
3376 s
= atomic_long_read(&rsp
->expedited_done
);
3377 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3378 /* ensure test happens before caller kfree */
3379 smp_mb__before_atomic(); /* ^^^ */
3380 atomic_long_inc(&rsp
->expedited_workdone1
);
3381 free_cpumask_var(cm
);
3385 /* No joy, try again later. Or just synchronize_sched(). */
3386 if (trycount
++ < 10) {
3387 udelay(trycount
* num_online_cpus());
3389 wait_rcu_gp(call_rcu_sched
);
3390 atomic_long_inc(&rsp
->expedited_normal
);
3391 free_cpumask_var(cm
);
3395 /* Recheck to see if someone else did our work for us. */
3396 s
= atomic_long_read(&rsp
->expedited_done
);
3397 if (ULONG_CMP_GE((ulong
)s
, (ulong
)firstsnap
)) {
3398 /* ensure test happens before caller kfree */
3399 smp_mb__before_atomic(); /* ^^^ */
3400 atomic_long_inc(&rsp
->expedited_workdone2
);
3401 free_cpumask_var(cm
);
3406 * Refetching sync_sched_expedited_started allows later
3407 * callers to piggyback on our grace period. We retry
3408 * after they started, so our grace period works for them,
3409 * and they started after our first try, so their grace
3410 * period works for us.
3412 if (!try_get_online_cpus()) {
3413 /* CPU hotplug operation in flight, use normal GP. */
3414 wait_rcu_gp(call_rcu_sched
);
3415 atomic_long_inc(&rsp
->expedited_normal
);
3416 free_cpumask_var(cm
);
3419 snap
= atomic_long_read(&rsp
->expedited_start
);
3420 smp_mb(); /* ensure read is before try_stop_cpus(). */
3422 atomic_long_inc(&rsp
->expedited_stoppedcpus
);
3425 free_cpumask_var(cm
);
3428 * Everyone up to our most recent fetch is covered by our grace
3429 * period. Update the counter, but only if our work is still
3430 * relevant -- which it won't be if someone who started later
3431 * than we did already did their update.
3434 atomic_long_inc(&rsp
->expedited_done_tries
);
3435 s
= atomic_long_read(&rsp
->expedited_done
);
3436 if (ULONG_CMP_GE((ulong
)s
, (ulong
)snap
)) {
3437 /* ensure test happens before caller kfree */
3438 smp_mb__before_atomic(); /* ^^^ */
3439 atomic_long_inc(&rsp
->expedited_done_lost
);
3442 } while (atomic_long_cmpxchg(&rsp
->expedited_done
, s
, snap
) != s
);
3443 atomic_long_inc(&rsp
->expedited_done_exit
);
3447 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3450 * Check to see if there is any immediate RCU-related work to be done
3451 * by the current CPU, for the specified type of RCU, returning 1 if so.
3452 * The checks are in order of increasing expense: checks that can be
3453 * carried out against CPU-local state are performed first. However,
3454 * we must check for CPU stalls first, else we might not get a chance.
3456 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3458 struct rcu_node
*rnp
= rdp
->mynode
;
3460 rdp
->n_rcu_pending
++;
3462 /* Check for CPU stalls, if enabled. */
3463 check_cpu_stall(rsp
, rdp
);
3465 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3466 if (rcu_nohz_full_cpu(rsp
))
3469 /* Is the RCU core waiting for a quiescent state from this CPU? */
3470 if (rcu_scheduler_fully_active
&&
3471 rdp
->qs_pending
&& !rdp
->passed_quiesce
&&
3472 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3473 rdp
->n_rp_qs_pending
++;
3474 } else if (rdp
->qs_pending
&&
3475 (rdp
->passed_quiesce
||
3476 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3477 rdp
->n_rp_report_qs
++;
3481 /* Does this CPU have callbacks ready to invoke? */
3482 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3483 rdp
->n_rp_cb_ready
++;
3487 /* Has RCU gone idle with this CPU needing another grace period? */
3488 if (cpu_needs_another_gp(rsp
, rdp
)) {
3489 rdp
->n_rp_cpu_needs_gp
++;
3493 /* Has another RCU grace period completed? */
3494 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3495 rdp
->n_rp_gp_completed
++;
3499 /* Has a new RCU grace period started? */
3500 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3501 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3502 rdp
->n_rp_gp_started
++;
3506 /* Does this CPU need a deferred NOCB wakeup? */
3507 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3508 rdp
->n_rp_nocb_defer_wakeup
++;
3513 rdp
->n_rp_need_nothing
++;
3518 * Check to see if there is any immediate RCU-related work to be done
3519 * by the current CPU, returning 1 if so. This function is part of the
3520 * RCU implementation; it is -not- an exported member of the RCU API.
3522 static int rcu_pending(void)
3524 struct rcu_state
*rsp
;
3526 for_each_rcu_flavor(rsp
)
3527 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3533 * Return true if the specified CPU has any callback. If all_lazy is
3534 * non-NULL, store an indication of whether all callbacks are lazy.
3535 * (If there are no callbacks, all of them are deemed to be lazy.)
3537 static bool __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3541 struct rcu_data
*rdp
;
3542 struct rcu_state
*rsp
;
3544 for_each_rcu_flavor(rsp
) {
3545 rdp
= this_cpu_ptr(rsp
->rda
);
3549 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3560 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3561 * the compiler is expected to optimize this away.
3563 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3564 int cpu
, unsigned long done
)
3566 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3567 atomic_read(&rsp
->barrier_cpu_count
), done
);
3571 * RCU callback function for _rcu_barrier(). If we are last, wake
3572 * up the task executing _rcu_barrier().
3574 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3576 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3577 struct rcu_state
*rsp
= rdp
->rsp
;
3579 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3580 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->n_barrier_done
);
3581 complete(&rsp
->barrier_completion
);
3583 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->n_barrier_done
);
3588 * Called with preemption disabled, and from cross-cpu IRQ context.
3590 static void rcu_barrier_func(void *type
)
3592 struct rcu_state
*rsp
= type
;
3593 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3595 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->n_barrier_done
);
3596 atomic_inc(&rsp
->barrier_cpu_count
);
3597 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3601 * Orchestrate the specified type of RCU barrier, waiting for all
3602 * RCU callbacks of the specified type to complete.
3604 static void _rcu_barrier(struct rcu_state
*rsp
)
3607 struct rcu_data
*rdp
;
3608 unsigned long snap
= READ_ONCE(rsp
->n_barrier_done
);
3609 unsigned long snap_done
;
3611 _rcu_barrier_trace(rsp
, "Begin", -1, snap
);
3613 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3614 mutex_lock(&rsp
->barrier_mutex
);
3617 * Ensure that all prior references, including to ->n_barrier_done,
3618 * are ordered before the _rcu_barrier() machinery.
3620 smp_mb(); /* See above block comment. */
3623 * Recheck ->n_barrier_done to see if others did our work for us.
3624 * This means checking ->n_barrier_done for an even-to-odd-to-even
3625 * transition. The "if" expression below therefore rounds the old
3626 * value up to the next even number and adds two before comparing.
3628 snap_done
= rsp
->n_barrier_done
;
3629 _rcu_barrier_trace(rsp
, "Check", -1, snap_done
);
3632 * If the value in snap is odd, we needed to wait for the current
3633 * rcu_barrier() to complete, then wait for the next one, in other
3634 * words, we need the value of snap_done to be three larger than
3635 * the value of snap. On the other hand, if the value in snap is
3636 * even, we only had to wait for the next rcu_barrier() to complete,
3637 * in other words, we need the value of snap_done to be only two
3638 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3639 * this for us (thank you, Linus!).
3641 if (ULONG_CMP_GE(snap_done
, (snap
+ 3) & ~0x1)) {
3642 _rcu_barrier_trace(rsp
, "EarlyExit", -1, snap_done
);
3643 smp_mb(); /* caller's subsequent code after above check. */
3644 mutex_unlock(&rsp
->barrier_mutex
);
3649 * Increment ->n_barrier_done to avoid duplicate work. Use
3650 * WRITE_ONCE() to prevent the compiler from speculating
3651 * the increment to precede the early-exit check.
3653 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3654 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 1);
3655 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->n_barrier_done
);
3656 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3659 * Initialize the count to one rather than to zero in order to
3660 * avoid a too-soon return to zero in case of a short grace period
3661 * (or preemption of this task). Exclude CPU-hotplug operations
3662 * to ensure that no offline CPU has callbacks queued.
3664 init_completion(&rsp
->barrier_completion
);
3665 atomic_set(&rsp
->barrier_cpu_count
, 1);
3669 * Force each CPU with callbacks to register a new callback.
3670 * When that callback is invoked, we will know that all of the
3671 * corresponding CPU's preceding callbacks have been invoked.
3673 for_each_possible_cpu(cpu
) {
3674 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3676 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3677 if (rcu_is_nocb_cpu(cpu
)) {
3678 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3679 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3680 rsp
->n_barrier_done
);
3682 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3683 rsp
->n_barrier_done
);
3684 smp_mb__before_atomic();
3685 atomic_inc(&rsp
->barrier_cpu_count
);
3686 __call_rcu(&rdp
->barrier_head
,
3687 rcu_barrier_callback
, rsp
, cpu
, 0);
3689 } else if (READ_ONCE(rdp
->qlen
)) {
3690 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3691 rsp
->n_barrier_done
);
3692 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3694 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3695 rsp
->n_barrier_done
);
3701 * Now that we have an rcu_barrier_callback() callback on each
3702 * CPU, and thus each counted, remove the initial count.
3704 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3705 complete(&rsp
->barrier_completion
);
3707 /* Increment ->n_barrier_done to prevent duplicate work. */
3708 smp_mb(); /* Keep increment after above mechanism. */
3709 WRITE_ONCE(rsp
->n_barrier_done
, rsp
->n_barrier_done
+ 1);
3710 WARN_ON_ONCE((rsp
->n_barrier_done
& 0x1) != 0);
3711 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->n_barrier_done
);
3712 smp_mb(); /* Keep increment before caller's subsequent code. */
3714 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3715 wait_for_completion(&rsp
->barrier_completion
);
3717 /* Other rcu_barrier() invocations can now safely proceed. */
3718 mutex_unlock(&rsp
->barrier_mutex
);
3722 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3724 void rcu_barrier_bh(void)
3726 _rcu_barrier(&rcu_bh_state
);
3728 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3731 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3733 void rcu_barrier_sched(void)
3735 _rcu_barrier(&rcu_sched_state
);
3737 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3740 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3741 * first CPU in a given leaf rcu_node structure coming online. The caller
3742 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3745 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3748 struct rcu_node
*rnp
= rnp_leaf
;
3751 mask
= rnp
->grpmask
;
3755 raw_spin_lock(&rnp
->lock
); /* Interrupts already disabled. */
3756 rnp
->qsmaskinit
|= mask
;
3757 raw_spin_unlock(&rnp
->lock
); /* Interrupts remain disabled. */
3762 * Do boot-time initialization of a CPU's per-CPU RCU data.
3765 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3767 unsigned long flags
;
3768 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3769 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3771 /* Set up local state, ensuring consistent view of global state. */
3772 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3773 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
3774 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3775 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3776 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3779 rcu_boot_init_nocb_percpu_data(rdp
);
3780 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3784 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3785 * offline event can be happening at a given time. Note also that we
3786 * can accept some slop in the rsp->completed access due to the fact
3787 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3790 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3792 unsigned long flags
;
3794 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3795 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3797 /* Set up local state, ensuring consistent view of global state. */
3798 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3799 rdp
->beenonline
= 1; /* We have now been online. */
3800 rdp
->qlen_last_fqs_check
= 0;
3801 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3802 rdp
->blimit
= blimit
;
3804 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3805 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3806 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3807 atomic_set(&rdp
->dynticks
->dynticks
,
3808 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3809 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
3812 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3813 * propagation up the rcu_node tree will happen at the beginning
3814 * of the next grace period.
3817 mask
= rdp
->grpmask
;
3818 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
3819 smp_mb__after_unlock_lock();
3820 rnp
->qsmaskinitnext
|= mask
;
3821 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3822 rdp
->completed
= rnp
->completed
;
3823 rdp
->passed_quiesce
= false;
3824 rdp
->rcu_qs_ctr_snap
= per_cpu(rcu_qs_ctr
, cpu
);
3825 rdp
->qs_pending
= false;
3826 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3827 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3830 static void rcu_prepare_cpu(int cpu
)
3832 struct rcu_state
*rsp
;
3834 for_each_rcu_flavor(rsp
)
3835 rcu_init_percpu_data(cpu
, rsp
);
3839 * Handle CPU online/offline notification events.
3841 int rcu_cpu_notify(struct notifier_block
*self
,
3842 unsigned long action
, void *hcpu
)
3844 long cpu
= (long)hcpu
;
3845 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3846 struct rcu_node
*rnp
= rdp
->mynode
;
3847 struct rcu_state
*rsp
;
3850 case CPU_UP_PREPARE
:
3851 case CPU_UP_PREPARE_FROZEN
:
3852 rcu_prepare_cpu(cpu
);
3853 rcu_prepare_kthreads(cpu
);
3854 rcu_spawn_all_nocb_kthreads(cpu
);
3857 case CPU_DOWN_FAILED
:
3858 rcu_boost_kthread_setaffinity(rnp
, -1);
3860 case CPU_DOWN_PREPARE
:
3861 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3864 case CPU_DYING_FROZEN
:
3865 for_each_rcu_flavor(rsp
)
3866 rcu_cleanup_dying_cpu(rsp
);
3868 case CPU_DYING_IDLE
:
3869 for_each_rcu_flavor(rsp
) {
3870 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3874 case CPU_DEAD_FROZEN
:
3875 case CPU_UP_CANCELED
:
3876 case CPU_UP_CANCELED_FROZEN
:
3877 for_each_rcu_flavor(rsp
) {
3878 rcu_cleanup_dead_cpu(cpu
, rsp
);
3879 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3888 static int rcu_pm_notify(struct notifier_block
*self
,
3889 unsigned long action
, void *hcpu
)
3892 case PM_HIBERNATION_PREPARE
:
3893 case PM_SUSPEND_PREPARE
:
3894 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3897 case PM_POST_HIBERNATION
:
3898 case PM_POST_SUSPEND
:
3899 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3900 rcu_unexpedite_gp();
3909 * Spawn the kthreads that handle each RCU flavor's grace periods.
3911 static int __init
rcu_spawn_gp_kthread(void)
3913 unsigned long flags
;
3914 int kthread_prio_in
= kthread_prio
;
3915 struct rcu_node
*rnp
;
3916 struct rcu_state
*rsp
;
3917 struct sched_param sp
;
3918 struct task_struct
*t
;
3920 /* Force priority into range. */
3921 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3923 else if (kthread_prio
< 0)
3925 else if (kthread_prio
> 99)
3927 if (kthread_prio
!= kthread_prio_in
)
3928 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3929 kthread_prio
, kthread_prio_in
);
3931 rcu_scheduler_fully_active
= 1;
3932 for_each_rcu_flavor(rsp
) {
3933 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3935 rnp
= rcu_get_root(rsp
);
3936 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
3937 rsp
->gp_kthread
= t
;
3939 sp
.sched_priority
= kthread_prio
;
3940 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3943 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
3945 rcu_spawn_nocb_kthreads();
3946 rcu_spawn_boost_kthreads();
3949 early_initcall(rcu_spawn_gp_kthread
);
3952 * This function is invoked towards the end of the scheduler's initialization
3953 * process. Before this is called, the idle task might contain
3954 * RCU read-side critical sections (during which time, this idle
3955 * task is booting the system). After this function is called, the
3956 * idle tasks are prohibited from containing RCU read-side critical
3957 * sections. This function also enables RCU lockdep checking.
3959 void rcu_scheduler_starting(void)
3961 WARN_ON(num_online_cpus() != 1);
3962 WARN_ON(nr_context_switches() > 0);
3963 rcu_scheduler_active
= 1;
3967 * Compute the per-level fanout, either using the exact fanout specified
3968 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
3970 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
3974 if (rcu_fanout_exact
) {
3975 rsp
->levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3976 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3977 rsp
->levelspread
[i
] = RCU_FANOUT
;
3983 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3984 ccur
= rsp
->levelcnt
[i
];
3985 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
3992 * Helper function for rcu_init() that initializes one rcu_state structure.
3994 static void __init
rcu_init_one(struct rcu_state
*rsp
,
3995 struct rcu_data __percpu
*rda
)
3997 static const char * const buf
[] = {
4001 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
4002 static const char * const fqs
[] = {
4006 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
4007 static u8 fl_mask
= 0x1;
4011 struct rcu_node
*rnp
;
4013 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4015 /* Silence gcc 4.8 false positive about array index out of range. */
4016 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4017 panic("rcu_init_one: rcu_num_lvls out of range");
4019 /* Initialize the level-tracking arrays. */
4021 for (i
= 0; i
< rcu_num_lvls
; i
++)
4022 rsp
->levelcnt
[i
] = num_rcu_lvl
[i
];
4023 for (i
= 1; i
< rcu_num_lvls
; i
++)
4024 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
4025 rcu_init_levelspread(rsp
);
4026 rsp
->flavor_mask
= fl_mask
;
4029 /* Initialize the elements themselves, starting from the leaves. */
4031 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4032 cpustride
*= rsp
->levelspread
[i
];
4033 rnp
= rsp
->level
[i
];
4034 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
4035 raw_spin_lock_init(&rnp
->lock
);
4036 lockdep_set_class_and_name(&rnp
->lock
,
4037 &rcu_node_class
[i
], buf
[i
]);
4038 raw_spin_lock_init(&rnp
->fqslock
);
4039 lockdep_set_class_and_name(&rnp
->fqslock
,
4040 &rcu_fqs_class
[i
], fqs
[i
]);
4041 rnp
->gpnum
= rsp
->gpnum
;
4042 rnp
->completed
= rsp
->completed
;
4044 rnp
->qsmaskinit
= 0;
4045 rnp
->grplo
= j
* cpustride
;
4046 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4047 if (rnp
->grphi
>= nr_cpu_ids
)
4048 rnp
->grphi
= nr_cpu_ids
- 1;
4054 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
4055 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4056 rnp
->parent
= rsp
->level
[i
- 1] +
4057 j
/ rsp
->levelspread
[i
- 1];
4060 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4061 rcu_init_one_nocb(rnp
);
4065 init_waitqueue_head(&rsp
->gp_wq
);
4066 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4067 for_each_possible_cpu(i
) {
4068 while (i
> rnp
->grphi
)
4070 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4071 rcu_boot_init_percpu_data(i
, rsp
);
4073 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4077 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4078 * replace the definitions in tree.h because those are needed to size
4079 * the ->node array in the rcu_state structure.
4081 static void __init
rcu_init_geometry(void)
4085 int rcu_capacity
[MAX_RCU_LVLS
];
4088 * Initialize any unspecified boot parameters.
4089 * The default values of jiffies_till_first_fqs and
4090 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4091 * value, which is a function of HZ, then adding one for each
4092 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4094 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4095 if (jiffies_till_first_fqs
== ULONG_MAX
)
4096 jiffies_till_first_fqs
= d
;
4097 if (jiffies_till_next_fqs
== ULONG_MAX
)
4098 jiffies_till_next_fqs
= d
;
4100 /* If the compile-time values are accurate, just leave. */
4101 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4102 nr_cpu_ids
== NR_CPUS
)
4104 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4105 rcu_fanout_leaf
, nr_cpu_ids
);
4108 * The boot-time rcu_fanout_leaf parameter is only permitted
4109 * to increase the leaf-level fanout, not decrease it. Of course,
4110 * the leaf-level fanout cannot exceed the number of bits in
4111 * the rcu_node masks. Complain and fall back to the compile-
4112 * time values if these limits are exceeded.
4114 if (rcu_fanout_leaf
< RCU_FANOUT_LEAF
||
4115 rcu_fanout_leaf
> sizeof(unsigned long) * 8) {
4121 * Compute number of nodes that can be handled an rcu_node tree
4122 * with the given number of levels.
4124 rcu_capacity
[0] = rcu_fanout_leaf
;
4125 for (i
= 1; i
< MAX_RCU_LVLS
; i
++)
4126 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4129 * The tree must be able to accommodate the configured number of CPUs.
4130 * If this limit is exceeded than we have a serious problem elsewhere.
4132 if (nr_cpu_ids
> rcu_capacity
[MAX_RCU_LVLS
- 1])
4133 panic("rcu_init_geometry: rcu_capacity[] is too small");
4135 /* Calculate the number of levels in the tree. */
4136 for (i
= 0; nr_cpu_ids
> rcu_capacity
[i
]; i
++) {
4138 rcu_num_lvls
= i
+ 1;
4140 /* Calculate the number of rcu_nodes at each level of the tree. */
4141 for (i
= 0; i
< rcu_num_lvls
; i
++) {
4142 int cap
= rcu_capacity
[(rcu_num_lvls
- 1) - i
];
4143 num_rcu_lvl
[i
] = DIV_ROUND_UP(nr_cpu_ids
, cap
);
4146 /* Calculate the total number of rcu_node structures. */
4148 for (i
= 0; i
< rcu_num_lvls
; i
++)
4149 rcu_num_nodes
+= num_rcu_lvl
[i
];
4153 * Dump out the structure of the rcu_node combining tree associated
4154 * with the rcu_state structure referenced by rsp.
4156 static void __init
rcu_dump_rcu_node_tree(struct rcu_state
*rsp
)
4159 struct rcu_node
*rnp
;
4161 pr_info("rcu_node tree layout dump\n");
4163 rcu_for_each_node_breadth_first(rsp
, rnp
) {
4164 if (rnp
->level
!= level
) {
4169 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4174 void __init
rcu_init(void)
4178 rcu_early_boot_tests();
4180 rcu_bootup_announce();
4181 rcu_init_geometry();
4182 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
4183 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
4185 rcu_dump_rcu_node_tree(&rcu_sched_state
);
4186 __rcu_init_preempt();
4187 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4190 * We don't need protection against CPU-hotplug here because
4191 * this is called early in boot, before either interrupts
4192 * or the scheduler are operational.
4194 cpu_notifier(rcu_cpu_notify
, 0);
4195 pm_notifier(rcu_pm_notify
, 0);
4196 for_each_online_cpu(cpu
)
4197 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4200 #include "tree_plugin.h"