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. */
72 * In order to export the rcu_state name to the tracing tools, it
73 * needs to be added in the __tracepoint_string section.
74 * This requires defining a separate variable tp_<sname>_varname
75 * that points to the string being used, and this will allow
76 * the tracing userspace tools to be able to decipher the string
77 * address to the matching string.
80 # define DEFINE_RCU_TPS(sname) \
81 static char sname##_varname[] = #sname; \
82 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
83 # define RCU_STATE_NAME(sname) sname##_varname
85 # define DEFINE_RCU_TPS(sname)
86 # define RCU_STATE_NAME(sname) __stringify(sname)
89 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
90 DEFINE_RCU_TPS(sname) \
91 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
92 struct rcu_state sname##_state = { \
93 .level = { &sname##_state.node[0] }, \
94 .rda = &sname##_data, \
96 .gp_state = RCU_GP_IDLE, \
97 .gpnum = 0UL - 300UL, \
98 .completed = 0UL - 300UL, \
99 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
100 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
101 .orphan_donetail = &sname##_state.orphan_donelist, \
102 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
103 .name = RCU_STATE_NAME(sname), \
105 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
108 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
109 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
111 static struct rcu_state
*const rcu_state_p
;
112 LIST_HEAD(rcu_struct_flavors
);
114 /* Dump rcu_node combining tree at boot to verify correct setup. */
115 static bool dump_tree
;
116 module_param(dump_tree
, bool, 0444);
117 /* Control rcu_node-tree auto-balancing at boot time. */
118 static bool rcu_fanout_exact
;
119 module_param(rcu_fanout_exact
, bool, 0444);
120 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
121 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
122 module_param(rcu_fanout_leaf
, int, 0444);
123 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
124 /* Number of rcu_nodes at specified level. */
125 static int num_rcu_lvl
[] = NUM_RCU_LVL_INIT
;
126 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
129 * The rcu_scheduler_active variable transitions from zero to one just
130 * before the first task is spawned. So when this variable is zero, RCU
131 * can assume that there is but one task, allowing RCU to (for example)
132 * optimize synchronize_sched() to a simple barrier(). When this variable
133 * is one, RCU must actually do all the hard work required to detect real
134 * grace periods. This variable is also used to suppress boot-time false
135 * positives from lockdep-RCU error checking.
137 int rcu_scheduler_active __read_mostly
;
138 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
141 * The rcu_scheduler_fully_active variable transitions from zero to one
142 * during the early_initcall() processing, which is after the scheduler
143 * is capable of creating new tasks. So RCU processing (for example,
144 * creating tasks for RCU priority boosting) must be delayed until after
145 * rcu_scheduler_fully_active transitions from zero to one. We also
146 * currently delay invocation of any RCU callbacks until after this point.
148 * It might later prove better for people registering RCU callbacks during
149 * early boot to take responsibility for these callbacks, but one step at
152 static int rcu_scheduler_fully_active __read_mostly
;
154 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
155 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
156 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
157 static void invoke_rcu_core(void);
158 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
159 static void rcu_report_exp_rdp(struct rcu_state
*rsp
,
160 struct rcu_data
*rdp
, bool wake
);
162 /* rcuc/rcub kthread realtime priority */
163 #ifdef CONFIG_RCU_KTHREAD_PRIO
164 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
165 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
166 static int kthread_prio
= IS_ENABLED(CONFIG_RCU_BOOST
) ? 1 : 0;
167 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
168 module_param(kthread_prio
, int, 0644);
170 /* Delay in jiffies for grace-period initialization delays, debug only. */
172 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
173 static int gp_preinit_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY
;
174 module_param(gp_preinit_delay
, int, 0644);
175 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
176 static const int gp_preinit_delay
;
177 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
179 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
180 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
181 module_param(gp_init_delay
, int, 0644);
182 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
183 static const int gp_init_delay
;
184 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
186 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
187 static int gp_cleanup_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY
;
188 module_param(gp_cleanup_delay
, int, 0644);
189 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
190 static const int gp_cleanup_delay
;
191 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
194 * Number of grace periods between delays, normalized by the duration of
195 * the delay. The longer the the delay, the more the grace periods between
196 * each delay. The reason for this normalization is that it means that,
197 * for non-zero delays, the overall slowdown of grace periods is constant
198 * regardless of the duration of the delay. This arrangement balances
199 * the need for long delays to increase some race probabilities with the
200 * need for fast grace periods to increase other race probabilities.
202 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
205 * Track the rcutorture test sequence number and the update version
206 * number within a given test. The rcutorture_testseq is incremented
207 * on every rcutorture module load and unload, so has an odd value
208 * when a test is running. The rcutorture_vernum is set to zero
209 * when rcutorture starts and is incremented on each rcutorture update.
210 * These variables enable correlating rcutorture output with the
211 * RCU tracing information.
213 unsigned long rcutorture_testseq
;
214 unsigned long rcutorture_vernum
;
217 * Compute the mask of online CPUs for the specified rcu_node structure.
218 * This will not be stable unless the rcu_node structure's ->lock is
219 * held, but the bit corresponding to the current CPU will be stable
222 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
224 return READ_ONCE(rnp
->qsmaskinitnext
);
228 * Return true if an RCU grace period is in progress. The READ_ONCE()s
229 * permit this function to be invoked without holding the root rcu_node
230 * structure's ->lock, but of course results can be subject to change.
232 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
234 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
238 * Note a quiescent state. Because we do not need to know
239 * how many quiescent states passed, just if there was at least
240 * one since the start of the grace period, this just sets a flag.
241 * The caller must have disabled preemption.
243 void rcu_sched_qs(void)
245 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.s
))
247 trace_rcu_grace_period(TPS("rcu_sched"),
248 __this_cpu_read(rcu_sched_data
.gpnum
),
250 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.norm
, false);
251 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
253 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, false);
254 rcu_report_exp_rdp(&rcu_sched_state
,
255 this_cpu_ptr(&rcu_sched_data
), true);
260 if (__this_cpu_read(rcu_bh_data
.cpu_no_qs
.s
)) {
261 trace_rcu_grace_period(TPS("rcu_bh"),
262 __this_cpu_read(rcu_bh_data
.gpnum
),
264 __this_cpu_write(rcu_bh_data
.cpu_no_qs
.b
.norm
, false);
268 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
270 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
271 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
272 .dynticks
= ATOMIC_INIT(1),
273 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
274 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
275 .dynticks_idle
= ATOMIC_INIT(1),
276 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
279 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
280 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
283 * Let the RCU core know that this CPU has gone through the scheduler,
284 * which is a quiescent state. This is called when the need for a
285 * quiescent state is urgent, so we burn an atomic operation and full
286 * memory barriers to let the RCU core know about it, regardless of what
287 * this CPU might (or might not) do in the near future.
289 * We inform the RCU core by emulating a zero-duration dyntick-idle
290 * period, which we in turn do by incrementing the ->dynticks counter
293 * The caller must have disabled interrupts.
295 static void rcu_momentary_dyntick_idle(void)
297 struct rcu_data
*rdp
;
298 struct rcu_dynticks
*rdtp
;
300 struct rcu_state
*rsp
;
303 * Yes, we can lose flag-setting operations. This is OK, because
304 * the flag will be set again after some delay.
306 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
307 raw_cpu_write(rcu_sched_qs_mask
, 0);
309 /* Find the flavor that needs a quiescent state. */
310 for_each_rcu_flavor(rsp
) {
311 rdp
= raw_cpu_ptr(rsp
->rda
);
312 if (!(resched_mask
& rsp
->flavor_mask
))
314 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
315 if (READ_ONCE(rdp
->mynode
->completed
) !=
316 READ_ONCE(rdp
->cond_resched_completed
))
320 * Pretend to be momentarily idle for the quiescent state.
321 * This allows the grace-period kthread to record the
322 * quiescent state, with no need for this CPU to do anything
325 rdtp
= this_cpu_ptr(&rcu_dynticks
);
326 smp_mb__before_atomic(); /* Earlier stuff before QS. */
327 atomic_add(2, &rdtp
->dynticks
); /* QS. */
328 smp_mb__after_atomic(); /* Later stuff after QS. */
334 * Note a context switch. This is a quiescent state for RCU-sched,
335 * and requires special handling for preemptible RCU.
336 * The caller must have disabled interrupts.
338 void rcu_note_context_switch(void)
340 barrier(); /* Avoid RCU read-side critical sections leaking down. */
341 trace_rcu_utilization(TPS("Start context switch"));
343 rcu_preempt_note_context_switch();
344 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
345 rcu_momentary_dyntick_idle();
346 trace_rcu_utilization(TPS("End context switch"));
347 barrier(); /* Avoid RCU read-side critical sections leaking up. */
349 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
352 * Register a quiescent state for all RCU flavors. If there is an
353 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
354 * dyntick-idle quiescent state visible to other CPUs (but only for those
355 * RCU flavors in desperate need of a quiescent state, which will normally
356 * be none of them). Either way, do a lightweight quiescent state for
359 * The barrier() calls are redundant in the common case when this is
360 * called externally, but just in case this is called from within this
364 void rcu_all_qs(void)
368 barrier(); /* Avoid RCU read-side critical sections leaking down. */
369 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
))) {
370 local_irq_save(flags
);
371 rcu_momentary_dyntick_idle();
372 local_irq_restore(flags
);
374 if (unlikely(raw_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))) {
376 * Yes, we just checked a per-CPU variable with preemption
377 * enabled, so we might be migrated to some other CPU at
378 * this point. That is OK because in that case, the
379 * migration will supply the needed quiescent state.
380 * We might end up needlessly disabling preemption and
381 * invoking rcu_sched_qs() on the destination CPU, but
382 * the probability and cost are both quite low, so this
383 * should not be a problem in practice.
389 this_cpu_inc(rcu_qs_ctr
);
390 barrier(); /* Avoid RCU read-side critical sections leaking up. */
392 EXPORT_SYMBOL_GPL(rcu_all_qs
);
394 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
395 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
396 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
398 module_param(blimit
, long, 0444);
399 module_param(qhimark
, long, 0444);
400 module_param(qlowmark
, long, 0444);
402 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
403 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
405 module_param(jiffies_till_first_fqs
, ulong
, 0644);
406 module_param(jiffies_till_next_fqs
, ulong
, 0644);
409 * How long the grace period must be before we start recruiting
410 * quiescent-state help from rcu_note_context_switch().
412 static ulong jiffies_till_sched_qs
= HZ
/ 20;
413 module_param(jiffies_till_sched_qs
, ulong
, 0644);
415 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
416 struct rcu_data
*rdp
);
417 static void force_qs_rnp(struct rcu_state
*rsp
,
418 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
419 unsigned long *maxj
),
420 bool *isidle
, unsigned long *maxj
);
421 static void force_quiescent_state(struct rcu_state
*rsp
);
422 static int rcu_pending(void);
425 * Return the number of RCU batches started thus far for debug & stats.
427 unsigned long rcu_batches_started(void)
429 return rcu_state_p
->gpnum
;
431 EXPORT_SYMBOL_GPL(rcu_batches_started
);
434 * Return the number of RCU-sched batches started thus far for debug & stats.
436 unsigned long rcu_batches_started_sched(void)
438 return rcu_sched_state
.gpnum
;
440 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
443 * Return the number of RCU BH batches started thus far for debug & stats.
445 unsigned long rcu_batches_started_bh(void)
447 return rcu_bh_state
.gpnum
;
449 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
452 * Return the number of RCU batches completed thus far for debug & stats.
454 unsigned long rcu_batches_completed(void)
456 return rcu_state_p
->completed
;
458 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
461 * Return the number of RCU-sched batches completed thus far for debug & stats.
463 unsigned long rcu_batches_completed_sched(void)
465 return rcu_sched_state
.completed
;
467 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
470 * Return the number of RCU BH batches completed thus far for debug & stats.
472 unsigned long rcu_batches_completed_bh(void)
474 return rcu_bh_state
.completed
;
476 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
479 * Force a quiescent state.
481 void rcu_force_quiescent_state(void)
483 force_quiescent_state(rcu_state_p
);
485 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
488 * Force a quiescent state for RCU BH.
490 void rcu_bh_force_quiescent_state(void)
492 force_quiescent_state(&rcu_bh_state
);
494 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
497 * Force a quiescent state for RCU-sched.
499 void rcu_sched_force_quiescent_state(void)
501 force_quiescent_state(&rcu_sched_state
);
503 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
506 * Show the state of the grace-period kthreads.
508 void show_rcu_gp_kthreads(void)
510 struct rcu_state
*rsp
;
512 for_each_rcu_flavor(rsp
) {
513 pr_info("%s: wait state: %d ->state: %#lx\n",
514 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
515 /* sched_show_task(rsp->gp_kthread); */
518 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
521 * Record the number of times rcutorture tests have been initiated and
522 * terminated. This information allows the debugfs tracing stats to be
523 * correlated to the rcutorture messages, even when the rcutorture module
524 * is being repeatedly loaded and unloaded. In other words, we cannot
525 * store this state in rcutorture itself.
527 void rcutorture_record_test_transition(void)
529 rcutorture_testseq
++;
530 rcutorture_vernum
= 0;
532 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
535 * Send along grace-period-related data for rcutorture diagnostics.
537 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
538 unsigned long *gpnum
, unsigned long *completed
)
540 struct rcu_state
*rsp
= NULL
;
549 case RCU_SCHED_FLAVOR
:
550 rsp
= &rcu_sched_state
;
556 *flags
= READ_ONCE(rsp
->gp_flags
);
557 *gpnum
= READ_ONCE(rsp
->gpnum
);
558 *completed
= READ_ONCE(rsp
->completed
);
565 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
568 * Record the number of writer passes through the current rcutorture test.
569 * This is also used to correlate debugfs tracing stats with the rcutorture
572 void rcutorture_record_progress(unsigned long vernum
)
576 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
579 * Does the CPU have callbacks ready to be invoked?
582 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
584 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
585 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
589 * Return the root node of the specified rcu_state structure.
591 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
593 return &rsp
->node
[0];
597 * Is there any need for future grace periods?
598 * Interrupts must be disabled. If the caller does not hold the root
599 * rnp_node structure's ->lock, the results are advisory only.
601 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
603 struct rcu_node
*rnp
= rcu_get_root(rsp
);
604 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
605 int *fp
= &rnp
->need_future_gp
[idx
];
607 return READ_ONCE(*fp
);
611 * Does the current CPU require a not-yet-started grace period?
612 * The caller must have disabled interrupts to prevent races with
613 * normal callback registry.
616 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
620 if (rcu_gp_in_progress(rsp
))
621 return false; /* No, a grace period is already in progress. */
622 if (rcu_future_needs_gp(rsp
))
623 return true; /* Yes, a no-CBs CPU needs one. */
624 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
625 return false; /* No, this is a no-CBs (or offline) CPU. */
626 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
627 return true; /* Yes, CPU has newly registered callbacks. */
628 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
629 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
630 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
631 rdp
->nxtcompleted
[i
]))
632 return true; /* Yes, CBs for future grace period. */
633 return false; /* No grace period needed. */
637 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
639 * If the new value of the ->dynticks_nesting counter now is zero,
640 * we really have entered idle, and must do the appropriate accounting.
641 * The caller must have disabled interrupts.
643 static void rcu_eqs_enter_common(long long oldval
, bool user
)
645 struct rcu_state
*rsp
;
646 struct rcu_data
*rdp
;
647 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
649 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
650 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
651 !user
&& !is_idle_task(current
)) {
652 struct task_struct
*idle __maybe_unused
=
653 idle_task(smp_processor_id());
655 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
656 rcu_ftrace_dump(DUMP_ORIG
);
657 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
658 current
->pid
, current
->comm
,
659 idle
->pid
, idle
->comm
); /* must be idle task! */
661 for_each_rcu_flavor(rsp
) {
662 rdp
= this_cpu_ptr(rsp
->rda
);
663 do_nocb_deferred_wakeup(rdp
);
665 rcu_prepare_for_idle();
666 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
667 smp_mb__before_atomic(); /* See above. */
668 atomic_inc(&rdtp
->dynticks
);
669 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
670 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
671 atomic_read(&rdtp
->dynticks
) & 0x1);
672 rcu_dynticks_task_enter();
675 * It is illegal to enter an extended quiescent state while
676 * in an RCU read-side critical section.
678 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map
),
679 "Illegal idle entry in RCU read-side critical section.");
680 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
),
681 "Illegal idle entry in RCU-bh read-side critical section.");
682 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map
),
683 "Illegal idle entry in RCU-sched read-side critical section.");
687 * Enter an RCU extended quiescent state, which can be either the
688 * idle loop or adaptive-tickless usermode execution.
690 static void rcu_eqs_enter(bool user
)
693 struct rcu_dynticks
*rdtp
;
695 rdtp
= this_cpu_ptr(&rcu_dynticks
);
696 oldval
= rdtp
->dynticks_nesting
;
697 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
698 (oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
699 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
700 rdtp
->dynticks_nesting
= 0;
701 rcu_eqs_enter_common(oldval
, user
);
703 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
708 * rcu_idle_enter - inform RCU that current CPU is entering idle
710 * Enter idle mode, in other words, -leave- the mode in which RCU
711 * read-side critical sections can occur. (Though RCU read-side
712 * critical sections can occur in irq handlers in idle, a possibility
713 * handled by irq_enter() and irq_exit().)
715 * We crowbar the ->dynticks_nesting field to zero to allow for
716 * the possibility of usermode upcalls having messed up our count
717 * of interrupt nesting level during the prior busy period.
719 void rcu_idle_enter(void)
723 local_irq_save(flags
);
724 rcu_eqs_enter(false);
725 rcu_sysidle_enter(0);
726 local_irq_restore(flags
);
728 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
730 #ifdef CONFIG_NO_HZ_FULL
732 * rcu_user_enter - inform RCU that we are resuming userspace.
734 * Enter RCU idle mode right before resuming userspace. No use of RCU
735 * is permitted between this call and rcu_user_exit(). This way the
736 * CPU doesn't need to maintain the tick for RCU maintenance purposes
737 * when the CPU runs in userspace.
739 void rcu_user_enter(void)
743 #endif /* CONFIG_NO_HZ_FULL */
746 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
748 * Exit from an interrupt handler, which might possibly result in entering
749 * idle mode, in other words, leaving the mode in which read-side critical
750 * sections can occur. The caller must have disabled interrupts.
752 * This code assumes that the idle loop never does anything that might
753 * result in unbalanced calls to irq_enter() and irq_exit(). If your
754 * architecture violates this assumption, RCU will give you what you
755 * deserve, good and hard. But very infrequently and irreproducibly.
757 * Use things like work queues to work around this limitation.
759 * You have been warned.
761 void rcu_irq_exit(void)
764 struct rcu_dynticks
*rdtp
;
766 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
767 rdtp
= this_cpu_ptr(&rcu_dynticks
);
768 oldval
= rdtp
->dynticks_nesting
;
769 rdtp
->dynticks_nesting
--;
770 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
771 rdtp
->dynticks_nesting
< 0);
772 if (rdtp
->dynticks_nesting
)
773 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
775 rcu_eqs_enter_common(oldval
, true);
776 rcu_sysidle_enter(1);
780 * Wrapper for rcu_irq_exit() where interrupts are enabled.
782 void rcu_irq_exit_irqson(void)
786 local_irq_save(flags
);
788 local_irq_restore(flags
);
792 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
794 * If the new value of the ->dynticks_nesting counter was previously zero,
795 * we really have exited idle, and must do the appropriate accounting.
796 * The caller must have disabled interrupts.
798 static void rcu_eqs_exit_common(long long oldval
, int user
)
800 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
802 rcu_dynticks_task_exit();
803 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
804 atomic_inc(&rdtp
->dynticks
);
805 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
806 smp_mb__after_atomic(); /* See above. */
807 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
808 !(atomic_read(&rdtp
->dynticks
) & 0x1));
809 rcu_cleanup_after_idle();
810 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
811 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
812 !user
&& !is_idle_task(current
)) {
813 struct task_struct
*idle __maybe_unused
=
814 idle_task(smp_processor_id());
816 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
817 oldval
, rdtp
->dynticks_nesting
);
818 rcu_ftrace_dump(DUMP_ORIG
);
819 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
820 current
->pid
, current
->comm
,
821 idle
->pid
, idle
->comm
); /* must be idle task! */
826 * Exit an RCU extended quiescent state, which can be either the
827 * idle loop or adaptive-tickless usermode execution.
829 static void rcu_eqs_exit(bool user
)
831 struct rcu_dynticks
*rdtp
;
834 rdtp
= this_cpu_ptr(&rcu_dynticks
);
835 oldval
= rdtp
->dynticks_nesting
;
836 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) && oldval
< 0);
837 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
838 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
840 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
841 rcu_eqs_exit_common(oldval
, user
);
846 * rcu_idle_exit - inform RCU that current CPU is leaving idle
848 * Exit idle mode, in other words, -enter- the mode in which RCU
849 * read-side critical sections can occur.
851 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
852 * allow for the possibility of usermode upcalls messing up our count
853 * of interrupt nesting level during the busy period that is just
856 void rcu_idle_exit(void)
860 local_irq_save(flags
);
863 local_irq_restore(flags
);
865 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
867 #ifdef CONFIG_NO_HZ_FULL
869 * rcu_user_exit - inform RCU that we are exiting userspace.
871 * Exit RCU idle mode while entering the kernel because it can
872 * run a RCU read side critical section anytime.
874 void rcu_user_exit(void)
878 #endif /* CONFIG_NO_HZ_FULL */
881 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
883 * Enter an interrupt handler, which might possibly result in exiting
884 * idle mode, in other words, entering the mode in which read-side critical
885 * sections can occur. The caller must have disabled interrupts.
887 * Note that the Linux kernel is fully capable of entering an interrupt
888 * handler that it never exits, for example when doing upcalls to
889 * user mode! This code assumes that the idle loop never does upcalls to
890 * user mode. If your architecture does do upcalls from the idle loop (or
891 * does anything else that results in unbalanced calls to the irq_enter()
892 * and irq_exit() functions), RCU will give you what you deserve, good
893 * and hard. But very infrequently and irreproducibly.
895 * Use things like work queues to work around this limitation.
897 * You have been warned.
899 void rcu_irq_enter(void)
901 struct rcu_dynticks
*rdtp
;
904 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
905 rdtp
= this_cpu_ptr(&rcu_dynticks
);
906 oldval
= rdtp
->dynticks_nesting
;
907 rdtp
->dynticks_nesting
++;
908 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
909 rdtp
->dynticks_nesting
== 0);
911 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
913 rcu_eqs_exit_common(oldval
, true);
918 * Wrapper for rcu_irq_enter() where interrupts are enabled.
920 void rcu_irq_enter_irqson(void)
924 local_irq_save(flags
);
926 local_irq_restore(flags
);
930 * rcu_nmi_enter - inform RCU of entry to NMI context
932 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
933 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
934 * that the CPU is active. This implementation permits nested NMIs, as
935 * long as the nesting level does not overflow an int. (You will probably
936 * run out of stack space first.)
938 void rcu_nmi_enter(void)
940 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
943 /* Complain about underflow. */
944 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
947 * If idle from RCU viewpoint, atomically increment ->dynticks
948 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
949 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
950 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
951 * to be in the outermost NMI handler that interrupted an RCU-idle
952 * period (observation due to Andy Lutomirski).
954 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
955 smp_mb__before_atomic(); /* Force delay from prior write. */
956 atomic_inc(&rdtp
->dynticks
);
957 /* atomic_inc() before later RCU read-side crit sects */
958 smp_mb__after_atomic(); /* See above. */
959 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
962 rdtp
->dynticks_nmi_nesting
+= incby
;
967 * rcu_nmi_exit - inform RCU of exit from NMI context
969 * If we are returning from the outermost NMI handler that interrupted an
970 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
971 * to let the RCU grace-period handling know that the CPU is back to
974 void rcu_nmi_exit(void)
976 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
979 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
980 * (We are exiting an NMI handler, so RCU better be paying attention
983 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
984 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
987 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
988 * leave it in non-RCU-idle state.
990 if (rdtp
->dynticks_nmi_nesting
!= 1) {
991 rdtp
->dynticks_nmi_nesting
-= 2;
995 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
996 rdtp
->dynticks_nmi_nesting
= 0;
997 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
998 smp_mb__before_atomic(); /* See above. */
999 atomic_inc(&rdtp
->dynticks
);
1000 smp_mb__after_atomic(); /* Force delay to next write. */
1001 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
1005 * __rcu_is_watching - are RCU read-side critical sections safe?
1007 * Return true if RCU is watching the running CPU, which means that
1008 * this CPU can safely enter RCU read-side critical sections. Unlike
1009 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1010 * least disabled preemption.
1012 bool notrace
__rcu_is_watching(void)
1014 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
1018 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1020 * If the current CPU is in its idle loop and is neither in an interrupt
1021 * or NMI handler, return true.
1023 bool notrace
rcu_is_watching(void)
1027 preempt_disable_notrace();
1028 ret
= __rcu_is_watching();
1029 preempt_enable_notrace();
1032 EXPORT_SYMBOL_GPL(rcu_is_watching
);
1034 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1037 * Is the current CPU online? Disable preemption to avoid false positives
1038 * that could otherwise happen due to the current CPU number being sampled,
1039 * this task being preempted, its old CPU being taken offline, resuming
1040 * on some other CPU, then determining that its old CPU is now offline.
1041 * It is OK to use RCU on an offline processor during initial boot, hence
1042 * the check for rcu_scheduler_fully_active. Note also that it is OK
1043 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1044 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1045 * offline to continue to use RCU for one jiffy after marking itself
1046 * offline in the cpu_online_mask. This leniency is necessary given the
1047 * non-atomic nature of the online and offline processing, for example,
1048 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1051 * This is also why RCU internally marks CPUs online during the
1052 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1054 * Disable checking if in an NMI handler because we cannot safely report
1055 * errors from NMI handlers anyway.
1057 bool rcu_lockdep_current_cpu_online(void)
1059 struct rcu_data
*rdp
;
1060 struct rcu_node
*rnp
;
1066 rdp
= this_cpu_ptr(&rcu_sched_data
);
1068 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
1069 !rcu_scheduler_fully_active
;
1073 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
1075 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1078 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1080 * If the current CPU is idle or running at a first-level (not nested)
1081 * interrupt from idle, return true. The caller must have at least
1082 * disabled preemption.
1084 static int rcu_is_cpu_rrupt_from_idle(void)
1086 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1090 * Snapshot the specified CPU's dynticks counter so that we can later
1091 * credit them with an implicit quiescent state. Return 1 if this CPU
1092 * is in dynticks idle mode, which is an extended quiescent state.
1094 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1095 bool *isidle
, unsigned long *maxj
)
1097 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1098 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1099 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1100 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1101 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1102 rdp
->mynode
->gpnum
))
1103 WRITE_ONCE(rdp
->gpwrap
, true);
1110 * Return true if the specified CPU has passed through a quiescent
1111 * state by virtue of being in or having passed through an dynticks
1112 * idle state since the last call to dyntick_save_progress_counter()
1113 * for this same CPU, or by virtue of having been offline.
1115 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1116 bool *isidle
, unsigned long *maxj
)
1122 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1123 snap
= (unsigned int)rdp
->dynticks_snap
;
1126 * If the CPU passed through or entered a dynticks idle phase with
1127 * no active irq/NMI handlers, then we can safely pretend that the CPU
1128 * already acknowledged the request to pass through a quiescent
1129 * state. Either way, that CPU cannot possibly be in an RCU
1130 * read-side critical section that started before the beginning
1131 * of the current RCU grace period.
1133 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1134 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1135 rdp
->dynticks_fqs
++;
1140 * Check for the CPU being offline, but only if the grace period
1141 * is old enough. We don't need to worry about the CPU changing
1142 * state: If we see it offline even once, it has been through a
1145 * The reason for insisting that the grace period be at least
1146 * one jiffy old is that CPUs that are not quite online and that
1147 * have just gone offline can still execute RCU read-side critical
1150 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1151 return 0; /* Grace period is not old enough. */
1153 if (cpu_is_offline(rdp
->cpu
)) {
1154 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1160 * A CPU running for an extended time within the kernel can
1161 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1162 * even context-switching back and forth between a pair of
1163 * in-kernel CPU-bound tasks cannot advance grace periods.
1164 * So if the grace period is old enough, make the CPU pay attention.
1165 * Note that the unsynchronized assignments to the per-CPU
1166 * rcu_sched_qs_mask variable are safe. Yes, setting of
1167 * bits can be lost, but they will be set again on the next
1168 * force-quiescent-state pass. So lost bit sets do not result
1169 * in incorrect behavior, merely in a grace period lasting
1170 * a few jiffies longer than it might otherwise. Because
1171 * there are at most four threads involved, and because the
1172 * updates are only once every few jiffies, the probability of
1173 * lossage (and thus of slight grace-period extension) is
1176 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1177 * is set too high, we override with half of the RCU CPU stall
1180 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1181 if (ULONG_CMP_GE(jiffies
,
1182 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1183 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1184 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1185 WRITE_ONCE(rdp
->cond_resched_completed
,
1186 READ_ONCE(rdp
->mynode
->completed
));
1187 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1189 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1191 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1194 /* And if it has been a really long time, kick the CPU as well. */
1195 if (ULONG_CMP_GE(jiffies
,
1196 rdp
->rsp
->gp_start
+ 2 * jiffies_till_sched_qs
) ||
1197 ULONG_CMP_GE(jiffies
, rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
))
1198 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1203 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1205 unsigned long j
= jiffies
;
1209 smp_wmb(); /* Record start time before stall time. */
1210 j1
= rcu_jiffies_till_stall_check();
1211 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1212 rsp
->jiffies_resched
= j
+ j1
/ 2;
1213 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1217 * Convert a ->gp_state value to a character string.
1219 static const char *gp_state_getname(short gs
)
1221 if (gs
< 0 || gs
>= ARRAY_SIZE(gp_state_names
))
1223 return gp_state_names
[gs
];
1227 * Complain about starvation of grace-period kthread.
1229 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1235 gpa
= READ_ONCE(rsp
->gp_activity
);
1236 if (j
- gpa
> 2 * HZ
) {
1237 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1239 rsp
->gpnum
, rsp
->completed
,
1241 gp_state_getname(rsp
->gp_state
), rsp
->gp_state
,
1242 rsp
->gp_kthread
? rsp
->gp_kthread
->state
: ~0);
1243 if (rsp
->gp_kthread
)
1244 sched_show_task(rsp
->gp_kthread
);
1249 * Dump stacks of all tasks running on stalled CPUs.
1251 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1254 unsigned long flags
;
1255 struct rcu_node
*rnp
;
1257 rcu_for_each_leaf_node(rsp
, rnp
) {
1258 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1259 if (rnp
->qsmask
!= 0) {
1260 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1261 if (rnp
->qsmask
& (1UL << cpu
))
1262 dump_cpu_task(rnp
->grplo
+ cpu
);
1264 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1268 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1272 unsigned long flags
;
1276 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1279 /* Only let one CPU complain about others per time interval. */
1281 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1282 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1283 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1284 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1287 WRITE_ONCE(rsp
->jiffies_stall
,
1288 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1289 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1292 * OK, time to rat on our buddy...
1293 * See Documentation/RCU/stallwarn.txt for info on how to debug
1294 * RCU CPU stall warnings.
1296 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1298 print_cpu_stall_info_begin();
1299 rcu_for_each_leaf_node(rsp
, rnp
) {
1300 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1301 ndetected
+= rcu_print_task_stall(rnp
);
1302 if (rnp
->qsmask
!= 0) {
1303 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1304 if (rnp
->qsmask
& (1UL << cpu
)) {
1305 print_cpu_stall_info(rsp
,
1310 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1313 print_cpu_stall_info_end();
1314 for_each_possible_cpu(cpu
)
1315 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1316 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1317 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1318 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1320 rcu_dump_cpu_stacks(rsp
);
1322 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1323 READ_ONCE(rsp
->completed
) == gpnum
) {
1324 pr_err("INFO: Stall ended before state dump start\n");
1327 gpa
= READ_ONCE(rsp
->gp_activity
);
1328 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1329 rsp
->name
, j
- gpa
, j
, gpa
,
1330 jiffies_till_next_fqs
,
1331 rcu_get_root(rsp
)->qsmask
);
1332 /* In this case, the current CPU might be at fault. */
1333 sched_show_task(current
);
1337 /* Complain about tasks blocking the grace period. */
1338 rcu_print_detail_task_stall(rsp
);
1340 rcu_check_gp_kthread_starvation(rsp
);
1342 force_quiescent_state(rsp
); /* Kick them all. */
1345 static void print_cpu_stall(struct rcu_state
*rsp
)
1348 unsigned long flags
;
1349 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1353 * OK, time to rat on ourselves...
1354 * See Documentation/RCU/stallwarn.txt for info on how to debug
1355 * RCU CPU stall warnings.
1357 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1358 print_cpu_stall_info_begin();
1359 print_cpu_stall_info(rsp
, smp_processor_id());
1360 print_cpu_stall_info_end();
1361 for_each_possible_cpu(cpu
)
1362 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1363 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1364 jiffies
- rsp
->gp_start
,
1365 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1367 rcu_check_gp_kthread_starvation(rsp
);
1369 rcu_dump_cpu_stacks(rsp
);
1371 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1372 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1373 WRITE_ONCE(rsp
->jiffies_stall
,
1374 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1375 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1378 * Attempt to revive the RCU machinery by forcing a context switch.
1380 * A context switch would normally allow the RCU state machine to make
1381 * progress and it could be we're stuck in kernel space without context
1382 * switches for an entirely unreasonable amount of time.
1384 resched_cpu(smp_processor_id());
1387 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1389 unsigned long completed
;
1390 unsigned long gpnum
;
1394 struct rcu_node
*rnp
;
1396 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1401 * Lots of memory barriers to reject false positives.
1403 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1404 * then rsp->gp_start, and finally rsp->completed. These values
1405 * are updated in the opposite order with memory barriers (or
1406 * equivalent) during grace-period initialization and cleanup.
1407 * Now, a false positive can occur if we get an new value of
1408 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1409 * the memory barriers, the only way that this can happen is if one
1410 * grace period ends and another starts between these two fetches.
1411 * Detect this by comparing rsp->completed with the previous fetch
1414 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1415 * and rsp->gp_start suffice to forestall false positives.
1417 gpnum
= READ_ONCE(rsp
->gpnum
);
1418 smp_rmb(); /* Pick up ->gpnum first... */
1419 js
= READ_ONCE(rsp
->jiffies_stall
);
1420 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1421 gps
= READ_ONCE(rsp
->gp_start
);
1422 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1423 completed
= READ_ONCE(rsp
->completed
);
1424 if (ULONG_CMP_GE(completed
, gpnum
) ||
1425 ULONG_CMP_LT(j
, js
) ||
1426 ULONG_CMP_GE(gps
, js
))
1427 return; /* No stall or GP completed since entering function. */
1429 if (rcu_gp_in_progress(rsp
) &&
1430 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1432 /* We haven't checked in, so go dump stack. */
1433 print_cpu_stall(rsp
);
1435 } else if (rcu_gp_in_progress(rsp
) &&
1436 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1438 /* They had a few time units to dump stack, so complain. */
1439 print_other_cpu_stall(rsp
, gpnum
);
1444 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1446 * Set the stall-warning timeout way off into the future, thus preventing
1447 * any RCU CPU stall-warning messages from appearing in the current set of
1448 * RCU grace periods.
1450 * The caller must disable hard irqs.
1452 void rcu_cpu_stall_reset(void)
1454 struct rcu_state
*rsp
;
1456 for_each_rcu_flavor(rsp
)
1457 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1461 * Initialize the specified rcu_data structure's default callback list
1462 * to empty. The default callback list is the one that is not used by
1463 * no-callbacks CPUs.
1465 static void init_default_callback_list(struct rcu_data
*rdp
)
1469 rdp
->nxtlist
= NULL
;
1470 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1471 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1475 * Initialize the specified rcu_data structure's callback list to empty.
1477 static void init_callback_list(struct rcu_data
*rdp
)
1479 if (init_nocb_callback_list(rdp
))
1481 init_default_callback_list(rdp
);
1485 * Determine the value that ->completed will have at the end of the
1486 * next subsequent grace period. This is used to tag callbacks so that
1487 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1488 * been dyntick-idle for an extended period with callbacks under the
1489 * influence of RCU_FAST_NO_HZ.
1491 * The caller must hold rnp->lock with interrupts disabled.
1493 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1494 struct rcu_node
*rnp
)
1497 * If RCU is idle, we just wait for the next grace period.
1498 * But we can only be sure that RCU is idle if we are looking
1499 * at the root rcu_node structure -- otherwise, a new grace
1500 * period might have started, but just not yet gotten around
1501 * to initializing the current non-root rcu_node structure.
1503 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1504 return rnp
->completed
+ 1;
1507 * Otherwise, wait for a possible partial grace period and
1508 * then the subsequent full grace period.
1510 return rnp
->completed
+ 2;
1514 * Trace-event helper function for rcu_start_future_gp() and
1515 * rcu_nocb_wait_gp().
1517 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1518 unsigned long c
, const char *s
)
1520 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1521 rnp
->completed
, c
, rnp
->level
,
1522 rnp
->grplo
, rnp
->grphi
, s
);
1526 * Start some future grace period, as needed to handle newly arrived
1527 * callbacks. The required future grace periods are recorded in each
1528 * rcu_node structure's ->need_future_gp field. Returns true if there
1529 * is reason to awaken the grace-period kthread.
1531 * The caller must hold the specified rcu_node structure's ->lock.
1533 static bool __maybe_unused
1534 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1535 unsigned long *c_out
)
1540 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1543 * Pick up grace-period number for new callbacks. If this
1544 * grace period is already marked as needed, return to the caller.
1546 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1547 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1548 if (rnp
->need_future_gp
[c
& 0x1]) {
1549 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1554 * If either this rcu_node structure or the root rcu_node structure
1555 * believe that a grace period is in progress, then we must wait
1556 * for the one following, which is in "c". Because our request
1557 * will be noticed at the end of the current grace period, we don't
1558 * need to explicitly start one. We only do the lockless check
1559 * of rnp_root's fields if the current rcu_node structure thinks
1560 * there is no grace period in flight, and because we hold rnp->lock,
1561 * the only possible change is when rnp_root's two fields are
1562 * equal, in which case rnp_root->gpnum might be concurrently
1563 * incremented. But that is OK, as it will just result in our
1564 * doing some extra useless work.
1566 if (rnp
->gpnum
!= rnp
->completed
||
1567 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1568 rnp
->need_future_gp
[c
& 0x1]++;
1569 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1574 * There might be no grace period in progress. If we don't already
1575 * hold it, acquire the root rcu_node structure's lock in order to
1576 * start one (if needed).
1578 if (rnp
!= rnp_root
)
1579 raw_spin_lock_rcu_node(rnp_root
);
1582 * Get a new grace-period number. If there really is no grace
1583 * period in progress, it will be smaller than the one we obtained
1584 * earlier. Adjust callbacks as needed. Note that even no-CBs
1585 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1587 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1588 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1589 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1590 rdp
->nxtcompleted
[i
] = c
;
1593 * If the needed for the required grace period is already
1594 * recorded, trace and leave.
1596 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1597 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1601 /* Record the need for the future grace period. */
1602 rnp_root
->need_future_gp
[c
& 0x1]++;
1604 /* If a grace period is not already in progress, start one. */
1605 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1606 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1608 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1609 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1612 if (rnp
!= rnp_root
)
1613 raw_spin_unlock_rcu_node(rnp_root
);
1621 * Clean up any old requests for the just-ended grace period. Also return
1622 * whether any additional grace periods have been requested. Also invoke
1623 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1624 * waiting for this grace period to complete.
1626 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1628 int c
= rnp
->completed
;
1630 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1632 rnp
->need_future_gp
[c
& 0x1] = 0;
1633 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1634 trace_rcu_future_gp(rnp
, rdp
, c
,
1635 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1640 * Awaken the grace-period kthread for the specified flavor of RCU.
1641 * Don't do a self-awaken, and don't bother awakening when there is
1642 * nothing for the grace-period kthread to do (as in several CPUs
1643 * raced to awaken, and we lost), and finally don't try to awaken
1644 * a kthread that has not yet been created.
1646 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1648 if (current
== rsp
->gp_kthread
||
1649 !READ_ONCE(rsp
->gp_flags
) ||
1652 swake_up(&rsp
->gp_wq
);
1656 * If there is room, assign a ->completed number to any callbacks on
1657 * this CPU that have not already been assigned. Also accelerate any
1658 * callbacks that were previously assigned a ->completed number that has
1659 * since proven to be too conservative, which can happen if callbacks get
1660 * assigned a ->completed number while RCU is idle, but with reference to
1661 * a non-root rcu_node structure. This function is idempotent, so it does
1662 * not hurt to call it repeatedly. Returns an flag saying that we should
1663 * awaken the RCU grace-period kthread.
1665 * The caller must hold rnp->lock with interrupts disabled.
1667 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1668 struct rcu_data
*rdp
)
1674 /* If the CPU has no callbacks, nothing to do. */
1675 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1679 * Starting from the sublist containing the callbacks most
1680 * recently assigned a ->completed number and working down, find the
1681 * first sublist that is not assignable to an upcoming grace period.
1682 * Such a sublist has something in it (first two tests) and has
1683 * a ->completed number assigned that will complete sooner than
1684 * the ->completed number for newly arrived callbacks (last test).
1686 * The key point is that any later sublist can be assigned the
1687 * same ->completed number as the newly arrived callbacks, which
1688 * means that the callbacks in any of these later sublist can be
1689 * grouped into a single sublist, whether or not they have already
1690 * been assigned a ->completed number.
1692 c
= rcu_cbs_completed(rsp
, rnp
);
1693 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1694 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1695 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1699 * If there are no sublist for unassigned callbacks, leave.
1700 * At the same time, advance "i" one sublist, so that "i" will
1701 * index into the sublist where all the remaining callbacks should
1704 if (++i
>= RCU_NEXT_TAIL
)
1708 * Assign all subsequent callbacks' ->completed number to the next
1709 * full grace period and group them all in the sublist initially
1712 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1713 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1714 rdp
->nxtcompleted
[i
] = c
;
1716 /* Record any needed additional grace periods. */
1717 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1719 /* Trace depending on how much we were able to accelerate. */
1720 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1721 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1723 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1728 * Move any callbacks whose grace period has completed to the
1729 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1730 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1731 * sublist. This function is idempotent, so it does not hurt to
1732 * invoke it repeatedly. As long as it is not invoked -too- often...
1733 * Returns true if the RCU grace-period kthread needs to be awakened.
1735 * The caller must hold rnp->lock with interrupts disabled.
1737 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1738 struct rcu_data
*rdp
)
1742 /* If the CPU has no callbacks, nothing to do. */
1743 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1747 * Find all callbacks whose ->completed numbers indicate that they
1748 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1750 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1751 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1753 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1755 /* Clean up any sublist tail pointers that were misordered above. */
1756 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1757 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1759 /* Copy down callbacks to fill in empty sublists. */
1760 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1761 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1763 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1764 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1767 /* Classify any remaining callbacks. */
1768 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1772 * Update CPU-local rcu_data state to record the beginnings and ends of
1773 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1774 * structure corresponding to the current CPU, and must have irqs disabled.
1775 * Returns true if the grace-period kthread needs to be awakened.
1777 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1778 struct rcu_data
*rdp
)
1782 /* Handle the ends of any preceding grace periods first. */
1783 if (rdp
->completed
== rnp
->completed
&&
1784 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1786 /* No grace period end, so just accelerate recent callbacks. */
1787 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1791 /* Advance callbacks. */
1792 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1794 /* Remember that we saw this grace-period completion. */
1795 rdp
->completed
= rnp
->completed
;
1796 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1799 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1801 * If the current grace period is waiting for this CPU,
1802 * set up to detect a quiescent state, otherwise don't
1803 * go looking for one.
1805 rdp
->gpnum
= rnp
->gpnum
;
1806 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1807 rdp
->cpu_no_qs
.b
.norm
= true;
1808 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1809 rdp
->core_needs_qs
= !!(rnp
->qsmask
& rdp
->grpmask
);
1810 zero_cpu_stall_ticks(rdp
);
1811 WRITE_ONCE(rdp
->gpwrap
, false);
1816 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1818 unsigned long flags
;
1820 struct rcu_node
*rnp
;
1822 local_irq_save(flags
);
1824 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1825 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1826 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1827 !raw_spin_trylock_rcu_node(rnp
)) { /* irqs already off, so later. */
1828 local_irq_restore(flags
);
1831 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1832 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1834 rcu_gp_kthread_wake(rsp
);
1837 static void rcu_gp_slow(struct rcu_state
*rsp
, int delay
)
1840 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1841 schedule_timeout_uninterruptible(delay
);
1845 * Initialize a new grace period. Return false if no grace period required.
1847 static bool rcu_gp_init(struct rcu_state
*rsp
)
1849 unsigned long oldmask
;
1850 struct rcu_data
*rdp
;
1851 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1853 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1854 raw_spin_lock_irq_rcu_node(rnp
);
1855 if (!READ_ONCE(rsp
->gp_flags
)) {
1856 /* Spurious wakeup, tell caller to go back to sleep. */
1857 raw_spin_unlock_irq_rcu_node(rnp
);
1860 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1862 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1864 * Grace period already in progress, don't start another.
1865 * Not supposed to be able to happen.
1867 raw_spin_unlock_irq_rcu_node(rnp
);
1871 /* Advance to a new grace period and initialize state. */
1872 record_gp_stall_check_time(rsp
);
1873 /* Record GP times before starting GP, hence smp_store_release(). */
1874 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1875 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1876 raw_spin_unlock_irq_rcu_node(rnp
);
1879 * Apply per-leaf buffered online and offline operations to the
1880 * rcu_node tree. Note that this new grace period need not wait
1881 * for subsequent online CPUs, and that quiescent-state forcing
1882 * will handle subsequent offline CPUs.
1884 rcu_for_each_leaf_node(rsp
, rnp
) {
1885 rcu_gp_slow(rsp
, gp_preinit_delay
);
1886 raw_spin_lock_irq_rcu_node(rnp
);
1887 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1888 !rnp
->wait_blkd_tasks
) {
1889 /* Nothing to do on this leaf rcu_node structure. */
1890 raw_spin_unlock_irq_rcu_node(rnp
);
1894 /* Record old state, apply changes to ->qsmaskinit field. */
1895 oldmask
= rnp
->qsmaskinit
;
1896 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1898 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1899 if (!oldmask
!= !rnp
->qsmaskinit
) {
1900 if (!oldmask
) /* First online CPU for this rcu_node. */
1901 rcu_init_new_rnp(rnp
);
1902 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1903 rnp
->wait_blkd_tasks
= true;
1904 else /* Last offline CPU and can propagate. */
1905 rcu_cleanup_dead_rnp(rnp
);
1909 * If all waited-on tasks from prior grace period are
1910 * done, and if all this rcu_node structure's CPUs are
1911 * still offline, propagate up the rcu_node tree and
1912 * clear ->wait_blkd_tasks. Otherwise, if one of this
1913 * rcu_node structure's CPUs has since come back online,
1914 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1915 * checks for this, so just call it unconditionally).
1917 if (rnp
->wait_blkd_tasks
&&
1918 (!rcu_preempt_has_tasks(rnp
) ||
1920 rnp
->wait_blkd_tasks
= false;
1921 rcu_cleanup_dead_rnp(rnp
);
1924 raw_spin_unlock_irq_rcu_node(rnp
);
1928 * Set the quiescent-state-needed bits in all the rcu_node
1929 * structures for all currently online CPUs in breadth-first order,
1930 * starting from the root rcu_node structure, relying on the layout
1931 * of the tree within the rsp->node[] array. Note that other CPUs
1932 * will access only the leaves of the hierarchy, thus seeing that no
1933 * grace period is in progress, at least until the corresponding
1934 * leaf node has been initialized. In addition, we have excluded
1935 * CPU-hotplug operations.
1937 * The grace period cannot complete until the initialization
1938 * process finishes, because this kthread handles both.
1940 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1941 rcu_gp_slow(rsp
, gp_init_delay
);
1942 raw_spin_lock_irq_rcu_node(rnp
);
1943 rdp
= this_cpu_ptr(rsp
->rda
);
1944 rcu_preempt_check_blocked_tasks(rnp
);
1945 rnp
->qsmask
= rnp
->qsmaskinit
;
1946 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1947 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1948 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1949 if (rnp
== rdp
->mynode
)
1950 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1951 rcu_preempt_boost_start_gp(rnp
);
1952 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1953 rnp
->level
, rnp
->grplo
,
1954 rnp
->grphi
, rnp
->qsmask
);
1955 raw_spin_unlock_irq_rcu_node(rnp
);
1956 cond_resched_rcu_qs();
1957 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1964 * Helper function for wait_event_interruptible_timeout() wakeup
1965 * at force-quiescent-state time.
1967 static bool rcu_gp_fqs_check_wake(struct rcu_state
*rsp
, int *gfp
)
1969 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1971 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1972 *gfp
= READ_ONCE(rsp
->gp_flags
);
1973 if (*gfp
& RCU_GP_FLAG_FQS
)
1976 /* The current grace period has completed. */
1977 if (!READ_ONCE(rnp
->qsmask
) && !rcu_preempt_blocked_readers_cgp(rnp
))
1984 * Do one round of quiescent-state forcing.
1986 static void rcu_gp_fqs(struct rcu_state
*rsp
, bool first_time
)
1988 bool isidle
= false;
1990 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1992 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1995 /* Collect dyntick-idle snapshots. */
1996 if (is_sysidle_rcu_state(rsp
)) {
1998 maxj
= jiffies
- ULONG_MAX
/ 4;
2000 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
2002 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
2004 /* Handle dyntick-idle and offline CPUs. */
2006 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
2008 /* Clear flag to prevent immediate re-entry. */
2009 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2010 raw_spin_lock_irq_rcu_node(rnp
);
2011 WRITE_ONCE(rsp
->gp_flags
,
2012 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
2013 raw_spin_unlock_irq_rcu_node(rnp
);
2018 * Clean up after the old grace period.
2020 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
2022 unsigned long gp_duration
;
2023 bool needgp
= false;
2025 struct rcu_data
*rdp
;
2026 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2027 struct swait_queue_head
*sq
;
2029 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2030 raw_spin_lock_irq_rcu_node(rnp
);
2031 gp_duration
= jiffies
- rsp
->gp_start
;
2032 if (gp_duration
> rsp
->gp_max
)
2033 rsp
->gp_max
= gp_duration
;
2036 * We know the grace period is complete, but to everyone else
2037 * it appears to still be ongoing. But it is also the case
2038 * that to everyone else it looks like there is nothing that
2039 * they can do to advance the grace period. It is therefore
2040 * safe for us to drop the lock in order to mark the grace
2041 * period as completed in all of the rcu_node structures.
2043 raw_spin_unlock_irq_rcu_node(rnp
);
2046 * Propagate new ->completed value to rcu_node structures so
2047 * that other CPUs don't have to wait until the start of the next
2048 * grace period to process their callbacks. This also avoids
2049 * some nasty RCU grace-period initialization races by forcing
2050 * the end of the current grace period to be completely recorded in
2051 * all of the rcu_node structures before the beginning of the next
2052 * grace period is recorded in any of the rcu_node structures.
2054 rcu_for_each_node_breadth_first(rsp
, rnp
) {
2055 raw_spin_lock_irq_rcu_node(rnp
);
2056 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
2057 WARN_ON_ONCE(rnp
->qsmask
);
2058 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
2059 rdp
= this_cpu_ptr(rsp
->rda
);
2060 if (rnp
== rdp
->mynode
)
2061 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
2062 /* smp_mb() provided by prior unlock-lock pair. */
2063 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
2064 sq
= rcu_nocb_gp_get(rnp
);
2065 raw_spin_unlock_irq_rcu_node(rnp
);
2066 rcu_nocb_gp_cleanup(sq
);
2067 cond_resched_rcu_qs();
2068 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2069 rcu_gp_slow(rsp
, gp_cleanup_delay
);
2071 rnp
= rcu_get_root(rsp
);
2072 raw_spin_lock_irq_rcu_node(rnp
); /* Order GP before ->completed update. */
2073 rcu_nocb_gp_set(rnp
, nocb
);
2075 /* Declare grace period done. */
2076 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
2077 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
2078 rsp
->gp_state
= RCU_GP_IDLE
;
2079 rdp
= this_cpu_ptr(rsp
->rda
);
2080 /* Advance CBs to reduce false positives below. */
2081 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
2082 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
2083 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2084 trace_rcu_grace_period(rsp
->name
,
2085 READ_ONCE(rsp
->gpnum
),
2088 raw_spin_unlock_irq_rcu_node(rnp
);
2092 * Body of kthread that handles grace periods.
2094 static int __noreturn
rcu_gp_kthread(void *arg
)
2100 struct rcu_state
*rsp
= arg
;
2101 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2103 rcu_bind_gp_kthread();
2106 /* Handle grace-period start. */
2108 trace_rcu_grace_period(rsp
->name
,
2109 READ_ONCE(rsp
->gpnum
),
2111 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2112 swait_event_interruptible(rsp
->gp_wq
,
2113 READ_ONCE(rsp
->gp_flags
) &
2115 rsp
->gp_state
= RCU_GP_DONE_GPS
;
2116 /* Locking provides needed memory barrier. */
2117 if (rcu_gp_init(rsp
))
2119 cond_resched_rcu_qs();
2120 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2121 WARN_ON(signal_pending(current
));
2122 trace_rcu_grace_period(rsp
->name
,
2123 READ_ONCE(rsp
->gpnum
),
2127 /* Handle quiescent-state forcing. */
2128 first_gp_fqs
= true;
2129 j
= jiffies_till_first_fqs
;
2132 jiffies_till_first_fqs
= HZ
;
2137 rsp
->jiffies_force_qs
= jiffies
+ j
;
2138 trace_rcu_grace_period(rsp
->name
,
2139 READ_ONCE(rsp
->gpnum
),
2141 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2142 ret
= swait_event_interruptible_timeout(rsp
->gp_wq
,
2143 rcu_gp_fqs_check_wake(rsp
, &gf
), j
);
2144 rsp
->gp_state
= RCU_GP_DOING_FQS
;
2145 /* Locking provides needed memory barriers. */
2146 /* If grace period done, leave loop. */
2147 if (!READ_ONCE(rnp
->qsmask
) &&
2148 !rcu_preempt_blocked_readers_cgp(rnp
))
2150 /* If time for quiescent-state forcing, do it. */
2151 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2152 (gf
& RCU_GP_FLAG_FQS
)) {
2153 trace_rcu_grace_period(rsp
->name
,
2154 READ_ONCE(rsp
->gpnum
),
2156 rcu_gp_fqs(rsp
, first_gp_fqs
);
2157 first_gp_fqs
= false;
2158 trace_rcu_grace_period(rsp
->name
,
2159 READ_ONCE(rsp
->gpnum
),
2161 cond_resched_rcu_qs();
2162 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2164 /* Deal with stray signal. */
2165 cond_resched_rcu_qs();
2166 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2167 WARN_ON(signal_pending(current
));
2168 trace_rcu_grace_period(rsp
->name
,
2169 READ_ONCE(rsp
->gpnum
),
2172 j
= jiffies_till_next_fqs
;
2175 jiffies_till_next_fqs
= HZ
;
2178 jiffies_till_next_fqs
= 1;
2182 /* Handle grace-period end. */
2183 rsp
->gp_state
= RCU_GP_CLEANUP
;
2184 rcu_gp_cleanup(rsp
);
2185 rsp
->gp_state
= RCU_GP_CLEANED
;
2190 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2191 * in preparation for detecting the next grace period. The caller must hold
2192 * the root node's ->lock and hard irqs must be disabled.
2194 * Note that it is legal for a dying CPU (which is marked as offline) to
2195 * invoke this function. This can happen when the dying CPU reports its
2198 * Returns true if the grace-period kthread must be awakened.
2201 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2202 struct rcu_data
*rdp
)
2204 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2206 * Either we have not yet spawned the grace-period
2207 * task, this CPU does not need another grace period,
2208 * or a grace period is already in progress.
2209 * Either way, don't start a new grace period.
2213 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2214 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2218 * We can't do wakeups while holding the rnp->lock, as that
2219 * could cause possible deadlocks with the rq->lock. Defer
2220 * the wakeup to our caller.
2226 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2227 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2228 * is invoked indirectly from rcu_advance_cbs(), which would result in
2229 * endless recursion -- or would do so if it wasn't for the self-deadlock
2230 * that is encountered beforehand.
2232 * Returns true if the grace-period kthread needs to be awakened.
2234 static bool rcu_start_gp(struct rcu_state
*rsp
)
2236 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2237 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2241 * If there is no grace period in progress right now, any
2242 * callbacks we have up to this point will be satisfied by the
2243 * next grace period. Also, advancing the callbacks reduces the
2244 * probability of false positives from cpu_needs_another_gp()
2245 * resulting in pointless grace periods. So, advance callbacks
2246 * then start the grace period!
2248 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2249 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2254 * Report a full set of quiescent states to the specified rcu_state data
2255 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2256 * kthread if another grace period is required. Whether we wake
2257 * the grace-period kthread or it awakens itself for the next round
2258 * of quiescent-state forcing, that kthread will clean up after the
2259 * just-completed grace period. Note that the caller must hold rnp->lock,
2260 * which is released before return.
2262 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2263 __releases(rcu_get_root(rsp
)->lock
)
2265 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2266 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2267 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2268 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2272 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2273 * Allows quiescent states for a group of CPUs to be reported at one go
2274 * to the specified rcu_node structure, though all the CPUs in the group
2275 * must be represented by the same rcu_node structure (which need not be a
2276 * leaf rcu_node structure, though it often will be). The gps parameter
2277 * is the grace-period snapshot, which means that the quiescent states
2278 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2279 * must be held upon entry, and it is released before return.
2282 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2283 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2284 __releases(rnp
->lock
)
2286 unsigned long oldmask
= 0;
2287 struct rcu_node
*rnp_c
;
2289 /* Walk up the rcu_node hierarchy. */
2291 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2294 * Our bit has already been cleared, or the
2295 * relevant grace period is already over, so done.
2297 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2300 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2301 rnp
->qsmask
&= ~mask
;
2302 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2303 mask
, rnp
->qsmask
, rnp
->level
,
2304 rnp
->grplo
, rnp
->grphi
,
2306 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2308 /* Other bits still set at this level, so done. */
2309 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2312 mask
= rnp
->grpmask
;
2313 if (rnp
->parent
== NULL
) {
2315 /* No more levels. Exit loop holding root lock. */
2319 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2322 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2323 oldmask
= rnp_c
->qsmask
;
2327 * Get here if we are the last CPU to pass through a quiescent
2328 * state for this grace period. Invoke rcu_report_qs_rsp()
2329 * to clean up and start the next grace period if one is needed.
2331 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2335 * Record a quiescent state for all tasks that were previously queued
2336 * on the specified rcu_node structure and that were blocking the current
2337 * RCU grace period. The caller must hold the specified rnp->lock with
2338 * irqs disabled, and this lock is released upon return, but irqs remain
2341 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2342 struct rcu_node
*rnp
, unsigned long flags
)
2343 __releases(rnp
->lock
)
2347 struct rcu_node
*rnp_p
;
2349 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2350 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2351 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2352 return; /* Still need more quiescent states! */
2355 rnp_p
= rnp
->parent
;
2356 if (rnp_p
== NULL
) {
2358 * Only one rcu_node structure in the tree, so don't
2359 * try to report up to its nonexistent parent!
2361 rcu_report_qs_rsp(rsp
, flags
);
2365 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2367 mask
= rnp
->grpmask
;
2368 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2369 raw_spin_lock_rcu_node(rnp_p
); /* irqs already disabled. */
2370 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2374 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2375 * structure. This must be called from the specified CPU.
2378 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2380 unsigned long flags
;
2383 struct rcu_node
*rnp
;
2386 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2387 if ((rdp
->cpu_no_qs
.b
.norm
&&
2388 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2389 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2393 * The grace period in which this quiescent state was
2394 * recorded has ended, so don't report it upwards.
2395 * We will instead need a new quiescent state that lies
2396 * within the current grace period.
2398 rdp
->cpu_no_qs
.b
.norm
= true; /* need qs for new gp. */
2399 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2400 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2403 mask
= rdp
->grpmask
;
2404 if ((rnp
->qsmask
& mask
) == 0) {
2405 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2407 rdp
->core_needs_qs
= false;
2410 * This GP can't end until cpu checks in, so all of our
2411 * callbacks can be processed during the next GP.
2413 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2415 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2416 /* ^^^ Released rnp->lock */
2418 rcu_gp_kthread_wake(rsp
);
2423 * Check to see if there is a new grace period of which this CPU
2424 * is not yet aware, and if so, set up local rcu_data state for it.
2425 * Otherwise, see if this CPU has just passed through its first
2426 * quiescent state for this grace period, and record that fact if so.
2429 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2431 /* Check for grace-period ends and beginnings. */
2432 note_gp_changes(rsp
, rdp
);
2435 * Does this CPU still need to do its part for current grace period?
2436 * If no, return and let the other CPUs do their part as well.
2438 if (!rdp
->core_needs_qs
)
2442 * Was there a quiescent state since the beginning of the grace
2443 * period? If no, then exit and wait for the next call.
2445 if (rdp
->cpu_no_qs
.b
.norm
&&
2446 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2450 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2453 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2457 * Send the specified CPU's RCU callbacks to the orphanage. The
2458 * specified CPU must be offline, and the caller must hold the
2462 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2463 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2465 /* No-CBs CPUs do not have orphanable callbacks. */
2466 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) || rcu_is_nocb_cpu(rdp
->cpu
))
2470 * Orphan the callbacks. First adjust the counts. This is safe
2471 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2472 * cannot be running now. Thus no memory barrier is required.
2474 if (rdp
->nxtlist
!= NULL
) {
2475 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2476 rsp
->qlen
+= rdp
->qlen
;
2477 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2479 WRITE_ONCE(rdp
->qlen
, 0);
2483 * Next, move those callbacks still needing a grace period to
2484 * the orphanage, where some other CPU will pick them up.
2485 * Some of the callbacks might have gone partway through a grace
2486 * period, but that is too bad. They get to start over because we
2487 * cannot assume that grace periods are synchronized across CPUs.
2488 * We don't bother updating the ->nxttail[] array yet, instead
2489 * we just reset the whole thing later on.
2491 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2492 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2493 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2494 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2498 * Then move the ready-to-invoke callbacks to the orphanage,
2499 * where some other CPU will pick them up. These will not be
2500 * required to pass though another grace period: They are done.
2502 if (rdp
->nxtlist
!= NULL
) {
2503 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2504 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2508 * Finally, initialize the rcu_data structure's list to empty and
2509 * disallow further callbacks on this CPU.
2511 init_callback_list(rdp
);
2512 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2516 * Adopt the RCU callbacks from the specified rcu_state structure's
2517 * orphanage. The caller must hold the ->orphan_lock.
2519 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2522 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2524 /* No-CBs CPUs are handled specially. */
2525 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2526 rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2529 /* Do the accounting first. */
2530 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2531 rdp
->qlen
+= rsp
->qlen
;
2532 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2533 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2534 rcu_idle_count_callbacks_posted();
2539 * We do not need a memory barrier here because the only way we
2540 * can get here if there is an rcu_barrier() in flight is if
2541 * we are the task doing the rcu_barrier().
2544 /* First adopt the ready-to-invoke callbacks. */
2545 if (rsp
->orphan_donelist
!= NULL
) {
2546 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2547 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2548 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2549 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2550 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2551 rsp
->orphan_donelist
= NULL
;
2552 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2555 /* And then adopt the callbacks that still need a grace period. */
2556 if (rsp
->orphan_nxtlist
!= NULL
) {
2557 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2558 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2559 rsp
->orphan_nxtlist
= NULL
;
2560 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2565 * Trace the fact that this CPU is going offline.
2567 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2569 RCU_TRACE(unsigned long mask
);
2570 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2571 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2573 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2576 RCU_TRACE(mask
= rdp
->grpmask
);
2577 trace_rcu_grace_period(rsp
->name
,
2578 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2583 * All CPUs for the specified rcu_node structure have gone offline,
2584 * and all tasks that were preempted within an RCU read-side critical
2585 * section while running on one of those CPUs have since exited their RCU
2586 * read-side critical section. Some other CPU is reporting this fact with
2587 * the specified rcu_node structure's ->lock held and interrupts disabled.
2588 * This function therefore goes up the tree of rcu_node structures,
2589 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2590 * the leaf rcu_node structure's ->qsmaskinit field has already been
2593 * This function does check that the specified rcu_node structure has
2594 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2595 * prematurely. That said, invoking it after the fact will cost you
2596 * a needless lock acquisition. So once it has done its work, don't
2599 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2602 struct rcu_node
*rnp
= rnp_leaf
;
2604 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2605 rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2608 mask
= rnp
->grpmask
;
2612 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
2613 rnp
->qsmaskinit
&= ~mask
;
2614 rnp
->qsmask
&= ~mask
;
2615 if (rnp
->qsmaskinit
) {
2616 raw_spin_unlock_rcu_node(rnp
);
2617 /* irqs remain disabled. */
2620 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2625 * The CPU has been completely removed, and some other CPU is reporting
2626 * this fact from process context. Do the remainder of the cleanup,
2627 * including orphaning the outgoing CPU's RCU callbacks, and also
2628 * adopting them. There can only be one CPU hotplug operation at a time,
2629 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2631 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2633 unsigned long flags
;
2634 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2635 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2637 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2640 /* Adjust any no-longer-needed kthreads. */
2641 rcu_boost_kthread_setaffinity(rnp
, -1);
2643 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2644 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2645 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2646 rcu_adopt_orphan_cbs(rsp
, flags
);
2647 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2649 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2650 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2651 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2655 * Invoke any RCU callbacks that have made it to the end of their grace
2656 * period. Thottle as specified by rdp->blimit.
2658 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2660 unsigned long flags
;
2661 struct rcu_head
*next
, *list
, **tail
;
2662 long bl
, count
, count_lazy
;
2665 /* If no callbacks are ready, just return. */
2666 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2667 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2668 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2669 need_resched(), is_idle_task(current
),
2670 rcu_is_callbacks_kthread());
2675 * Extract the list of ready callbacks, disabling to prevent
2676 * races with call_rcu() from interrupt handlers.
2678 local_irq_save(flags
);
2679 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2681 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2682 list
= rdp
->nxtlist
;
2683 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2684 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2685 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2686 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2687 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2688 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2689 local_irq_restore(flags
);
2691 /* Invoke callbacks. */
2692 count
= count_lazy
= 0;
2696 debug_rcu_head_unqueue(list
);
2697 if (__rcu_reclaim(rsp
->name
, list
))
2700 /* Stop only if limit reached and CPU has something to do. */
2701 if (++count
>= bl
&&
2703 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2707 local_irq_save(flags
);
2708 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2709 is_idle_task(current
),
2710 rcu_is_callbacks_kthread());
2712 /* Update count, and requeue any remaining callbacks. */
2714 *tail
= rdp
->nxtlist
;
2715 rdp
->nxtlist
= list
;
2716 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2717 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2718 rdp
->nxttail
[i
] = tail
;
2722 smp_mb(); /* List handling before counting for rcu_barrier(). */
2723 rdp
->qlen_lazy
-= count_lazy
;
2724 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2725 rdp
->n_cbs_invoked
+= count
;
2727 /* Reinstate batch limit if we have worked down the excess. */
2728 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2729 rdp
->blimit
= blimit
;
2731 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2732 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2733 rdp
->qlen_last_fqs_check
= 0;
2734 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2735 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2736 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2737 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2739 local_irq_restore(flags
);
2741 /* Re-invoke RCU core processing if there are callbacks remaining. */
2742 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2747 * Check to see if this CPU is in a non-context-switch quiescent state
2748 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2749 * Also schedule RCU core processing.
2751 * This function must be called from hardirq context. It is normally
2752 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2753 * false, there is no point in invoking rcu_check_callbacks().
2755 void rcu_check_callbacks(int user
)
2757 trace_rcu_utilization(TPS("Start scheduler-tick"));
2758 increment_cpu_stall_ticks();
2759 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2762 * Get here if this CPU took its interrupt from user
2763 * mode or from the idle loop, and if this is not a
2764 * nested interrupt. In this case, the CPU is in
2765 * a quiescent state, so note it.
2767 * No memory barrier is required here because both
2768 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2769 * variables that other CPUs neither access nor modify,
2770 * at least not while the corresponding CPU is online.
2776 } else if (!in_softirq()) {
2779 * Get here if this CPU did not take its interrupt from
2780 * softirq, in other words, if it is not interrupting
2781 * a rcu_bh read-side critical section. This is an _bh
2782 * critical section, so note it.
2787 rcu_preempt_check_callbacks();
2791 rcu_note_voluntary_context_switch(current
);
2792 trace_rcu_utilization(TPS("End scheduler-tick"));
2796 * Scan the leaf rcu_node structures, processing dyntick state for any that
2797 * have not yet encountered a quiescent state, using the function specified.
2798 * Also initiate boosting for any threads blocked on the root rcu_node.
2800 * The caller must have suppressed start of new grace periods.
2802 static void force_qs_rnp(struct rcu_state
*rsp
,
2803 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2804 unsigned long *maxj
),
2805 bool *isidle
, unsigned long *maxj
)
2809 unsigned long flags
;
2811 struct rcu_node
*rnp
;
2813 rcu_for_each_leaf_node(rsp
, rnp
) {
2814 cond_resched_rcu_qs();
2816 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2817 if (rnp
->qsmask
== 0) {
2818 if (rcu_state_p
== &rcu_sched_state
||
2819 rsp
!= rcu_state_p
||
2820 rcu_preempt_blocked_readers_cgp(rnp
)) {
2822 * No point in scanning bits because they
2823 * are all zero. But we might need to
2824 * priority-boost blocked readers.
2826 rcu_initiate_boost(rnp
, flags
);
2827 /* rcu_initiate_boost() releases rnp->lock */
2831 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2833 * Race between grace-period
2834 * initialization and task exiting RCU
2835 * read-side critical section: Report.
2837 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2838 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2844 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2845 if ((rnp
->qsmask
& bit
) != 0) {
2846 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2851 /* Idle/offline CPUs, report (releases rnp->lock. */
2852 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2854 /* Nothing to do here, so just drop the lock. */
2855 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2861 * Force quiescent states on reluctant CPUs, and also detect which
2862 * CPUs are in dyntick-idle mode.
2864 static void force_quiescent_state(struct rcu_state
*rsp
)
2866 unsigned long flags
;
2868 struct rcu_node
*rnp
;
2869 struct rcu_node
*rnp_old
= NULL
;
2871 /* Funnel through hierarchy to reduce memory contention. */
2872 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2873 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2874 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2875 !raw_spin_trylock(&rnp
->fqslock
);
2876 if (rnp_old
!= NULL
)
2877 raw_spin_unlock(&rnp_old
->fqslock
);
2879 rsp
->n_force_qs_lh
++;
2884 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2886 /* Reached the root of the rcu_node tree, acquire lock. */
2887 raw_spin_lock_irqsave_rcu_node(rnp_old
, flags
);
2888 raw_spin_unlock(&rnp_old
->fqslock
);
2889 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2890 rsp
->n_force_qs_lh
++;
2891 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2892 return; /* Someone beat us to it. */
2894 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2895 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2896 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2900 * This does the RCU core processing work for the specified rcu_state
2901 * and rcu_data structures. This may be called only from the CPU to
2902 * whom the rdp belongs.
2905 __rcu_process_callbacks(struct rcu_state
*rsp
)
2907 unsigned long flags
;
2909 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2911 WARN_ON_ONCE(rdp
->beenonline
== 0);
2913 /* Update RCU state based on any recent quiescent states. */
2914 rcu_check_quiescent_state(rsp
, rdp
);
2916 /* Does this CPU require a not-yet-started grace period? */
2917 local_irq_save(flags
);
2918 if (cpu_needs_another_gp(rsp
, rdp
)) {
2919 raw_spin_lock_rcu_node(rcu_get_root(rsp
)); /* irqs disabled. */
2920 needwake
= rcu_start_gp(rsp
);
2921 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2923 rcu_gp_kthread_wake(rsp
);
2925 local_irq_restore(flags
);
2928 /* If there are callbacks ready, invoke them. */
2929 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2930 invoke_rcu_callbacks(rsp
, rdp
);
2932 /* Do any needed deferred wakeups of rcuo kthreads. */
2933 do_nocb_deferred_wakeup(rdp
);
2937 * Do RCU core processing for the current CPU.
2939 static void rcu_process_callbacks(struct softirq_action
*unused
)
2941 struct rcu_state
*rsp
;
2943 if (cpu_is_offline(smp_processor_id()))
2945 trace_rcu_utilization(TPS("Start RCU core"));
2946 for_each_rcu_flavor(rsp
)
2947 __rcu_process_callbacks(rsp
);
2948 trace_rcu_utilization(TPS("End RCU core"));
2952 * Schedule RCU callback invocation. If the specified type of RCU
2953 * does not support RCU priority boosting, just do a direct call,
2954 * otherwise wake up the per-CPU kernel kthread. Note that because we
2955 * are running on the current CPU with softirqs disabled, the
2956 * rcu_cpu_kthread_task cannot disappear out from under us.
2958 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2960 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2962 if (likely(!rsp
->boost
)) {
2963 rcu_do_batch(rsp
, rdp
);
2966 invoke_rcu_callbacks_kthread();
2969 static void invoke_rcu_core(void)
2971 if (cpu_online(smp_processor_id()))
2972 raise_softirq(RCU_SOFTIRQ
);
2976 * Handle any core-RCU processing required by a call_rcu() invocation.
2978 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2979 struct rcu_head
*head
, unsigned long flags
)
2984 * If called from an extended quiescent state, invoke the RCU
2985 * core in order to force a re-evaluation of RCU's idleness.
2987 if (!rcu_is_watching())
2990 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2991 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2995 * Force the grace period if too many callbacks or too long waiting.
2996 * Enforce hysteresis, and don't invoke force_quiescent_state()
2997 * if some other CPU has recently done so. Also, don't bother
2998 * invoking force_quiescent_state() if the newly enqueued callback
2999 * is the only one waiting for a grace period to complete.
3001 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
3003 /* Are we ignoring a completed grace period? */
3004 note_gp_changes(rsp
, rdp
);
3006 /* Start a new grace period if one not already started. */
3007 if (!rcu_gp_in_progress(rsp
)) {
3008 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3010 raw_spin_lock_rcu_node(rnp_root
);
3011 needwake
= rcu_start_gp(rsp
);
3012 raw_spin_unlock_rcu_node(rnp_root
);
3014 rcu_gp_kthread_wake(rsp
);
3016 /* Give the grace period a kick. */
3017 rdp
->blimit
= LONG_MAX
;
3018 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
3019 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
3020 force_quiescent_state(rsp
);
3021 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3022 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
3028 * RCU callback function to leak a callback.
3030 static void rcu_leak_callback(struct rcu_head
*rhp
)
3035 * Helper function for call_rcu() and friends. The cpu argument will
3036 * normally be -1, indicating "currently running CPU". It may specify
3037 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3038 * is expected to specify a CPU.
3041 __call_rcu(struct rcu_head
*head
, rcu_callback_t func
,
3042 struct rcu_state
*rsp
, int cpu
, bool lazy
)
3044 unsigned long flags
;
3045 struct rcu_data
*rdp
;
3047 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
3048 if (debug_rcu_head_queue(head
)) {
3049 /* Probable double call_rcu(), so leak the callback. */
3050 WRITE_ONCE(head
->func
, rcu_leak_callback
);
3051 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3058 * Opportunistically note grace-period endings and beginnings.
3059 * Note that we might see a beginning right after we see an
3060 * end, but never vice versa, since this CPU has to pass through
3061 * a quiescent state betweentimes.
3063 local_irq_save(flags
);
3064 rdp
= this_cpu_ptr(rsp
->rda
);
3066 /* Add the callback to our list. */
3067 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3071 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3072 if (likely(rdp
->mynode
)) {
3073 /* Post-boot, so this should be for a no-CBs CPU. */
3074 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3075 WARN_ON_ONCE(offline
);
3076 /* Offline CPU, _call_rcu() illegal, leak callback. */
3077 local_irq_restore(flags
);
3081 * Very early boot, before rcu_init(). Initialize if needed
3082 * and then drop through to queue the callback.
3085 WARN_ON_ONCE(!rcu_is_watching());
3086 if (!likely(rdp
->nxtlist
))
3087 init_default_callback_list(rdp
);
3089 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3093 rcu_idle_count_callbacks_posted();
3094 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3095 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3096 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3098 if (__is_kfree_rcu_offset((unsigned long)func
))
3099 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3100 rdp
->qlen_lazy
, rdp
->qlen
);
3102 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3104 /* Go handle any RCU core processing required. */
3105 __call_rcu_core(rsp
, rdp
, head
, flags
);
3106 local_irq_restore(flags
);
3110 * Queue an RCU-sched callback for invocation after a grace period.
3112 void call_rcu_sched(struct rcu_head
*head
, rcu_callback_t func
)
3114 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3116 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3119 * Queue an RCU callback for invocation after a quicker grace period.
3121 void call_rcu_bh(struct rcu_head
*head
, rcu_callback_t func
)
3123 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3125 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3128 * Queue an RCU callback for lazy invocation after a grace period.
3129 * This will likely be later named something like "call_rcu_lazy()",
3130 * but this change will require some way of tagging the lazy RCU
3131 * callbacks in the list of pending callbacks. Until then, this
3132 * function may only be called from __kfree_rcu().
3134 void kfree_call_rcu(struct rcu_head
*head
,
3135 rcu_callback_t func
)
3137 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3139 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3142 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3143 * any blocking grace-period wait automatically implies a grace period
3144 * if there is only one CPU online at any point time during execution
3145 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3146 * occasionally incorrectly indicate that there are multiple CPUs online
3147 * when there was in fact only one the whole time, as this just adds
3148 * some overhead: RCU still operates correctly.
3150 static inline int rcu_blocking_is_gp(void)
3154 might_sleep(); /* Check for RCU read-side critical section. */
3156 ret
= num_online_cpus() <= 1;
3162 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3164 * Control will return to the caller some time after a full rcu-sched
3165 * grace period has elapsed, in other words after all currently executing
3166 * rcu-sched read-side critical sections have completed. These read-side
3167 * critical sections are delimited by rcu_read_lock_sched() and
3168 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3169 * local_irq_disable(), and so on may be used in place of
3170 * rcu_read_lock_sched().
3172 * This means that all preempt_disable code sequences, including NMI and
3173 * non-threaded hardware-interrupt handlers, in progress on entry will
3174 * have completed before this primitive returns. However, this does not
3175 * guarantee that softirq handlers will have completed, since in some
3176 * kernels, these handlers can run in process context, and can block.
3178 * Note that this guarantee implies further memory-ordering guarantees.
3179 * On systems with more than one CPU, when synchronize_sched() returns,
3180 * each CPU is guaranteed to have executed a full memory barrier since the
3181 * end of its last RCU-sched read-side critical section whose beginning
3182 * preceded the call to synchronize_sched(). In addition, each CPU having
3183 * an RCU read-side critical section that extends beyond the return from
3184 * synchronize_sched() is guaranteed to have executed a full memory barrier
3185 * after the beginning of synchronize_sched() and before the beginning of
3186 * that RCU read-side critical section. Note that these guarantees include
3187 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3188 * that are executing in the kernel.
3190 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3191 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3192 * to have executed a full memory barrier during the execution of
3193 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3194 * again only if the system has more than one CPU).
3196 * This primitive provides the guarantees made by the (now removed)
3197 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3198 * guarantees that rcu_read_lock() sections will have completed.
3199 * In "classic RCU", these two guarantees happen to be one and
3200 * the same, but can differ in realtime RCU implementations.
3202 void synchronize_sched(void)
3204 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3205 lock_is_held(&rcu_lock_map
) ||
3206 lock_is_held(&rcu_sched_lock_map
),
3207 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3208 if (rcu_blocking_is_gp())
3210 if (rcu_gp_is_expedited())
3211 synchronize_sched_expedited();
3213 wait_rcu_gp(call_rcu_sched
);
3215 EXPORT_SYMBOL_GPL(synchronize_sched
);
3218 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3220 * Control will return to the caller some time after a full rcu_bh grace
3221 * period has elapsed, in other words after all currently executing rcu_bh
3222 * read-side critical sections have completed. RCU read-side critical
3223 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3224 * and may be nested.
3226 * See the description of synchronize_sched() for more detailed information
3227 * on memory ordering guarantees.
3229 void synchronize_rcu_bh(void)
3231 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3232 lock_is_held(&rcu_lock_map
) ||
3233 lock_is_held(&rcu_sched_lock_map
),
3234 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3235 if (rcu_blocking_is_gp())
3237 if (rcu_gp_is_expedited())
3238 synchronize_rcu_bh_expedited();
3240 wait_rcu_gp(call_rcu_bh
);
3242 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3245 * get_state_synchronize_rcu - Snapshot current RCU state
3247 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3248 * to determine whether or not a full grace period has elapsed in the
3251 unsigned long get_state_synchronize_rcu(void)
3254 * Any prior manipulation of RCU-protected data must happen
3255 * before the load from ->gpnum.
3260 * Make sure this load happens before the purportedly
3261 * time-consuming work between get_state_synchronize_rcu()
3262 * and cond_synchronize_rcu().
3264 return smp_load_acquire(&rcu_state_p
->gpnum
);
3266 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3269 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3271 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3273 * If a full RCU grace period has elapsed since the earlier call to
3274 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3275 * synchronize_rcu() to wait for a full grace period.
3277 * Yes, this function does not take counter wrap into account. But
3278 * counter wrap is harmless. If the counter wraps, we have waited for
3279 * more than 2 billion grace periods (and way more on a 64-bit system!),
3280 * so waiting for one additional grace period should be just fine.
3282 void cond_synchronize_rcu(unsigned long oldstate
)
3284 unsigned long newstate
;
3287 * Ensure that this load happens before any RCU-destructive
3288 * actions the caller might carry out after we return.
3290 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3291 if (ULONG_CMP_GE(oldstate
, newstate
))
3294 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3297 * get_state_synchronize_sched - Snapshot current RCU-sched state
3299 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3300 * to determine whether or not a full grace period has elapsed in the
3303 unsigned long get_state_synchronize_sched(void)
3306 * Any prior manipulation of RCU-protected data must happen
3307 * before the load from ->gpnum.
3312 * Make sure this load happens before the purportedly
3313 * time-consuming work between get_state_synchronize_sched()
3314 * and cond_synchronize_sched().
3316 return smp_load_acquire(&rcu_sched_state
.gpnum
);
3318 EXPORT_SYMBOL_GPL(get_state_synchronize_sched
);
3321 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3323 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3325 * If a full RCU-sched grace period has elapsed since the earlier call to
3326 * get_state_synchronize_sched(), just return. Otherwise, invoke
3327 * synchronize_sched() to wait for a full grace period.
3329 * Yes, this function does not take counter wrap into account. But
3330 * counter wrap is harmless. If the counter wraps, we have waited for
3331 * more than 2 billion grace periods (and way more on a 64-bit system!),
3332 * so waiting for one additional grace period should be just fine.
3334 void cond_synchronize_sched(unsigned long oldstate
)
3336 unsigned long newstate
;
3339 * Ensure that this load happens before any RCU-destructive
3340 * actions the caller might carry out after we return.
3342 newstate
= smp_load_acquire(&rcu_sched_state
.completed
);
3343 if (ULONG_CMP_GE(oldstate
, newstate
))
3344 synchronize_sched();
3346 EXPORT_SYMBOL_GPL(cond_synchronize_sched
);
3348 /* Adjust sequence number for start of update-side operation. */
3349 static void rcu_seq_start(unsigned long *sp
)
3351 WRITE_ONCE(*sp
, *sp
+ 1);
3352 smp_mb(); /* Ensure update-side operation after counter increment. */
3353 WARN_ON_ONCE(!(*sp
& 0x1));
3356 /* Adjust sequence number for end of update-side operation. */
3357 static void rcu_seq_end(unsigned long *sp
)
3359 smp_mb(); /* Ensure update-side operation before counter increment. */
3360 WRITE_ONCE(*sp
, *sp
+ 1);
3361 WARN_ON_ONCE(*sp
& 0x1);
3364 /* Take a snapshot of the update side's sequence number. */
3365 static unsigned long rcu_seq_snap(unsigned long *sp
)
3369 s
= (READ_ONCE(*sp
) + 3) & ~0x1;
3370 smp_mb(); /* Above access must not bleed into critical section. */
3375 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3376 * full update-side operation has occurred.
3378 static bool rcu_seq_done(unsigned long *sp
, unsigned long s
)
3380 return ULONG_CMP_GE(READ_ONCE(*sp
), s
);
3383 /* Wrapper functions for expedited grace periods. */
3384 static void rcu_exp_gp_seq_start(struct rcu_state
*rsp
)
3386 rcu_seq_start(&rsp
->expedited_sequence
);
3388 static void rcu_exp_gp_seq_end(struct rcu_state
*rsp
)
3390 rcu_seq_end(&rsp
->expedited_sequence
);
3391 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3393 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state
*rsp
)
3397 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3398 s
= rcu_seq_snap(&rsp
->expedited_sequence
);
3399 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("snap"));
3402 static bool rcu_exp_gp_seq_done(struct rcu_state
*rsp
, unsigned long s
)
3404 return rcu_seq_done(&rsp
->expedited_sequence
, s
);
3408 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3409 * recent CPU-online activity. Note that these masks are not cleared
3410 * when CPUs go offline, so they reflect the union of all CPUs that have
3411 * ever been online. This means that this function normally takes its
3412 * no-work-to-do fastpath.
3414 static void sync_exp_reset_tree_hotplug(struct rcu_state
*rsp
)
3417 unsigned long flags
;
3419 unsigned long oldmask
;
3420 int ncpus
= READ_ONCE(rsp
->ncpus
);
3421 struct rcu_node
*rnp
;
3422 struct rcu_node
*rnp_up
;
3424 /* If no new CPUs onlined since last time, nothing to do. */
3425 if (likely(ncpus
== rsp
->ncpus_snap
))
3427 rsp
->ncpus_snap
= ncpus
;
3430 * Each pass through the following loop propagates newly onlined
3431 * CPUs for the current rcu_node structure up the rcu_node tree.
3433 rcu_for_each_leaf_node(rsp
, rnp
) {
3434 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3435 if (rnp
->expmaskinit
== rnp
->expmaskinitnext
) {
3436 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3437 continue; /* No new CPUs, nothing to do. */
3440 /* Update this node's mask, track old value for propagation. */
3441 oldmask
= rnp
->expmaskinit
;
3442 rnp
->expmaskinit
= rnp
->expmaskinitnext
;
3443 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3445 /* If was already nonzero, nothing to propagate. */
3449 /* Propagate the new CPU up the tree. */
3450 mask
= rnp
->grpmask
;
3451 rnp_up
= rnp
->parent
;
3454 raw_spin_lock_irqsave_rcu_node(rnp_up
, flags
);
3455 if (rnp_up
->expmaskinit
)
3457 rnp_up
->expmaskinit
|= mask
;
3458 raw_spin_unlock_irqrestore_rcu_node(rnp_up
, flags
);
3461 mask
= rnp_up
->grpmask
;
3462 rnp_up
= rnp_up
->parent
;
3468 * Reset the ->expmask values in the rcu_node tree in preparation for
3469 * a new expedited grace period.
3471 static void __maybe_unused
sync_exp_reset_tree(struct rcu_state
*rsp
)
3473 unsigned long flags
;
3474 struct rcu_node
*rnp
;
3476 sync_exp_reset_tree_hotplug(rsp
);
3477 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3478 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3479 WARN_ON_ONCE(rnp
->expmask
);
3480 rnp
->expmask
= rnp
->expmaskinit
;
3481 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3486 * Return non-zero if there is no RCU expedited grace period in progress
3487 * for the specified rcu_node structure, in other words, if all CPUs and
3488 * tasks covered by the specified rcu_node structure have done their bit
3489 * for the current expedited grace period. Works only for preemptible
3490 * RCU -- other RCU implementation use other means.
3492 * Caller must hold the rcu_state's exp_mutex.
3494 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
3496 return rnp
->exp_tasks
== NULL
&&
3497 READ_ONCE(rnp
->expmask
) == 0;
3501 * Report the exit from RCU read-side critical section for the last task
3502 * that queued itself during or before the current expedited preemptible-RCU
3503 * grace period. This event is reported either to the rcu_node structure on
3504 * which the task was queued or to one of that rcu_node structure's ancestors,
3505 * recursively up the tree. (Calm down, calm down, we do the recursion
3508 * Caller must hold the rcu_state's exp_mutex and the specified rcu_node
3509 * structure's ->lock.
3511 static void __rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3512 bool wake
, unsigned long flags
)
3513 __releases(rnp
->lock
)
3518 if (!sync_rcu_preempt_exp_done(rnp
)) {
3520 rcu_initiate_boost(rnp
, flags
);
3522 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3525 if (rnp
->parent
== NULL
) {
3526 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3528 smp_mb(); /* EGP done before wake_up(). */
3529 swake_up(&rsp
->expedited_wq
);
3533 mask
= rnp
->grpmask
;
3534 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled */
3536 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled */
3537 WARN_ON_ONCE(!(rnp
->expmask
& mask
));
3538 rnp
->expmask
&= ~mask
;
3543 * Report expedited quiescent state for specified node. This is a
3544 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3546 * Caller must hold the rcu_state's exp_mutex.
3548 static void __maybe_unused
rcu_report_exp_rnp(struct rcu_state
*rsp
,
3549 struct rcu_node
*rnp
, bool wake
)
3551 unsigned long flags
;
3553 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3554 __rcu_report_exp_rnp(rsp
, rnp
, wake
, flags
);
3558 * Report expedited quiescent state for multiple CPUs, all covered by the
3559 * specified leaf rcu_node structure. Caller must hold the rcu_state's
3562 static void rcu_report_exp_cpu_mult(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3563 unsigned long mask
, bool wake
)
3565 unsigned long flags
;
3567 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3568 if (!(rnp
->expmask
& mask
)) {
3569 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3572 rnp
->expmask
&= ~mask
;
3573 __rcu_report_exp_rnp(rsp
, rnp
, wake
, flags
); /* Releases rnp->lock. */
3577 * Report expedited quiescent state for specified rcu_data (CPU).
3579 static void rcu_report_exp_rdp(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
3582 rcu_report_exp_cpu_mult(rsp
, rdp
->mynode
, rdp
->grpmask
, wake
);
3585 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3586 static bool sync_exp_work_done(struct rcu_state
*rsp
, atomic_long_t
*stat
,
3589 if (rcu_exp_gp_seq_done(rsp
, s
)) {
3590 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("done"));
3591 /* Ensure test happens before caller kfree(). */
3592 smp_mb__before_atomic(); /* ^^^ */
3593 atomic_long_inc(stat
);
3600 * Funnel-lock acquisition for expedited grace periods. Returns true
3601 * if some other task completed an expedited grace period that this task
3602 * can piggy-back on, and with no mutex held. Otherwise, returns false
3603 * with the mutex held, indicating that the caller must actually do the
3604 * expedited grace period.
3606 static bool exp_funnel_lock(struct rcu_state
*rsp
, unsigned long s
)
3608 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id());
3609 struct rcu_node
*rnp
= rdp
->mynode
;
3610 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3612 /* Low-contention fastpath. */
3613 if (ULONG_CMP_LT(READ_ONCE(rnp
->exp_seq_rq
), s
) &&
3615 ULONG_CMP_LT(READ_ONCE(rnp_root
->exp_seq_rq
), s
)) &&
3616 !mutex_is_locked(&rsp
->exp_mutex
) &&
3617 mutex_trylock(&rsp
->exp_mutex
))
3621 * Each pass through the following loop works its way up
3622 * the rcu_node tree, returning if others have done the work or
3623 * otherwise falls through to acquire rsp->exp_mutex. The mapping
3624 * from CPU to rcu_node structure can be inexact, as it is just
3625 * promoting locality and is not strictly needed for correctness.
3627 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
3628 if (sync_exp_work_done(rsp
, &rdp
->exp_workdone1
, s
))
3631 /* Work not done, either wait here or go up. */
3632 spin_lock(&rnp
->exp_lock
);
3633 if (ULONG_CMP_GE(rnp
->exp_seq_rq
, s
)) {
3635 /* Someone else doing GP, so wait for them. */
3636 spin_unlock(&rnp
->exp_lock
);
3637 trace_rcu_exp_funnel_lock(rsp
->name
, rnp
->level
,
3638 rnp
->grplo
, rnp
->grphi
,
3640 wait_event(rnp
->exp_wq
[(s
>> 1) & 0x1],
3641 sync_exp_work_done(rsp
,
3642 &rdp
->exp_workdone2
, s
));
3645 rnp
->exp_seq_rq
= s
; /* Followers can wait on us. */
3646 spin_unlock(&rnp
->exp_lock
);
3647 trace_rcu_exp_funnel_lock(rsp
->name
, rnp
->level
, rnp
->grplo
,
3648 rnp
->grphi
, TPS("nxtlvl"));
3650 mutex_lock(&rsp
->exp_mutex
);
3652 if (sync_exp_work_done(rsp
, &rdp
->exp_workdone3
, s
)) {
3653 mutex_unlock(&rsp
->exp_mutex
);
3656 rcu_exp_gp_seq_start(rsp
);
3657 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("start"));
3661 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3662 static void sync_sched_exp_handler(void *data
)
3664 struct rcu_data
*rdp
;
3665 struct rcu_node
*rnp
;
3666 struct rcu_state
*rsp
= data
;
3668 rdp
= this_cpu_ptr(rsp
->rda
);
3670 if (!(READ_ONCE(rnp
->expmask
) & rdp
->grpmask
) ||
3671 __this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
3673 if (rcu_is_cpu_rrupt_from_idle()) {
3674 rcu_report_exp_rdp(&rcu_sched_state
,
3675 this_cpu_ptr(&rcu_sched_data
), true);
3678 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, true);
3679 resched_cpu(smp_processor_id());
3682 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
3683 static void sync_sched_exp_online_cleanup(int cpu
)
3685 struct rcu_data
*rdp
;
3687 struct rcu_node
*rnp
;
3688 struct rcu_state
*rsp
= &rcu_sched_state
;
3690 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3692 if (!(READ_ONCE(rnp
->expmask
) & rdp
->grpmask
))
3694 ret
= smp_call_function_single(cpu
, sync_sched_exp_handler
, rsp
, 0);
3699 * Select the nodes that the upcoming expedited grace period needs
3702 static void sync_rcu_exp_select_cpus(struct rcu_state
*rsp
,
3703 smp_call_func_t func
)
3706 unsigned long flags
;
3708 unsigned long mask_ofl_test
;
3709 unsigned long mask_ofl_ipi
;
3711 struct rcu_node
*rnp
;
3713 sync_exp_reset_tree(rsp
);
3714 rcu_for_each_leaf_node(rsp
, rnp
) {
3715 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3717 /* Each pass checks a CPU for identity, offline, and idle. */
3719 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++) {
3720 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3721 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3723 if (raw_smp_processor_id() == cpu
||
3724 !(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3725 mask_ofl_test
|= rdp
->grpmask
;
3727 mask_ofl_ipi
= rnp
->expmask
& ~mask_ofl_test
;
3730 * Need to wait for any blocked tasks as well. Note that
3731 * additional blocking tasks will also block the expedited
3732 * GP until such time as the ->expmask bits are cleared.
3734 if (rcu_preempt_has_tasks(rnp
))
3735 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
3736 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3738 /* IPI the remaining CPUs for expedited quiescent state. */
3740 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3741 if (!(mask_ofl_ipi
& mask
))
3744 ret
= smp_call_function_single(cpu
, func
, rsp
, 0);
3746 mask_ofl_ipi
&= ~mask
;
3749 /* Failed, raced with offline. */
3750 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3751 if (cpu_online(cpu
) &&
3752 (rnp
->expmask
& mask
)) {
3753 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3754 schedule_timeout_uninterruptible(1);
3755 if (cpu_online(cpu
) &&
3756 (rnp
->expmask
& mask
))
3758 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3760 if (!(rnp
->expmask
& mask
))
3761 mask_ofl_ipi
&= ~mask
;
3762 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3764 /* Report quiescent states for those that went offline. */
3765 mask_ofl_test
|= mask_ofl_ipi
;
3767 rcu_report_exp_cpu_mult(rsp
, rnp
, mask_ofl_test
, false);
3771 static void synchronize_sched_expedited_wait(struct rcu_state
*rsp
)
3774 unsigned long jiffies_stall
;
3775 unsigned long jiffies_start
;
3778 struct rcu_node
*rnp
;
3779 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3782 jiffies_stall
= rcu_jiffies_till_stall_check();
3783 jiffies_start
= jiffies
;
3786 ret
= swait_event_timeout(
3788 sync_rcu_preempt_exp_done(rnp_root
),
3790 if (ret
> 0 || sync_rcu_preempt_exp_done(rnp_root
))
3793 /* Hit a signal, disable CPU stall warnings. */
3794 swait_event(rsp
->expedited_wq
,
3795 sync_rcu_preempt_exp_done(rnp_root
));
3798 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3801 rcu_for_each_leaf_node(rsp
, rnp
) {
3802 ndetected
+= rcu_print_task_exp_stall(rnp
);
3804 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3805 struct rcu_data
*rdp
;
3807 if (!(rnp
->expmask
& mask
))
3810 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3811 pr_cont(" %d-%c%c%c", cpu
,
3812 "O."[!!cpu_online(cpu
)],
3813 "o."[!!(rdp
->grpmask
& rnp
->expmaskinit
)],
3814 "N."[!!(rdp
->grpmask
& rnp
->expmaskinitnext
)]);
3818 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
3819 jiffies
- jiffies_start
, rsp
->expedited_sequence
,
3820 rnp_root
->expmask
, ".T"[!!rnp_root
->exp_tasks
]);
3822 pr_err("blocking rcu_node structures:");
3823 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3824 if (rnp
== rnp_root
)
3825 continue; /* printed unconditionally */
3826 if (sync_rcu_preempt_exp_done(rnp
))
3828 pr_cont(" l=%u:%d-%d:%#lx/%c",
3829 rnp
->level
, rnp
->grplo
, rnp
->grphi
,
3831 ".T"[!!rnp
->exp_tasks
]);
3835 rcu_for_each_leaf_node(rsp
, rnp
) {
3837 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3838 if (!(rnp
->expmask
& mask
))
3843 jiffies_stall
= 3 * rcu_jiffies_till_stall_check() + 3;
3848 * Wait for the current expedited grace period to complete, and then
3849 * wake up everyone who piggybacked on the just-completed expedited
3850 * grace period. Also update all the ->exp_seq_rq counters as needed
3851 * in order to avoid counter-wrap problems.
3853 static void rcu_exp_wait_wake(struct rcu_state
*rsp
, unsigned long s
)
3855 struct rcu_node
*rnp
;
3857 synchronize_sched_expedited_wait(rsp
);
3858 rcu_exp_gp_seq_end(rsp
);
3859 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("end"));
3860 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3861 if (ULONG_CMP_LT(READ_ONCE(rnp
->exp_seq_rq
), s
)) {
3862 spin_lock(&rnp
->exp_lock
);
3863 /* Recheck, avoid hang in case someone just arrived. */
3864 if (ULONG_CMP_LT(rnp
->exp_seq_rq
, s
))
3865 rnp
->exp_seq_rq
= s
;
3866 spin_unlock(&rnp
->exp_lock
);
3868 wake_up_all(&rnp
->exp_wq
[(rsp
->expedited_sequence
>> 1) & 0x1]);
3870 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("endwake"));
3871 mutex_unlock(&rsp
->exp_mutex
);
3875 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3877 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3878 * approach to force the grace period to end quickly. This consumes
3879 * significant time on all CPUs and is unfriendly to real-time workloads,
3880 * so is thus not recommended for any sort of common-case code. In fact,
3881 * if you are using synchronize_sched_expedited() in a loop, please
3882 * restructure your code to batch your updates, and then use a single
3883 * synchronize_sched() instead.
3885 * This implementation can be thought of as an application of sequence
3886 * locking to expedited grace periods, but using the sequence counter to
3887 * determine when someone else has already done the work instead of for
3890 void synchronize_sched_expedited(void)
3893 struct rcu_state
*rsp
= &rcu_sched_state
;
3895 /* If only one CPU, this is automatically a grace period. */
3896 if (rcu_blocking_is_gp())
3899 /* If expedited grace periods are prohibited, fall back to normal. */
3900 if (rcu_gp_is_normal()) {
3901 wait_rcu_gp(call_rcu_sched
);
3905 /* Take a snapshot of the sequence number. */
3906 s
= rcu_exp_gp_seq_snap(rsp
);
3907 if (exp_funnel_lock(rsp
, s
))
3908 return; /* Someone else did our work for us. */
3910 /* Initialize the rcu_node tree in preparation for the wait. */
3911 sync_rcu_exp_select_cpus(rsp
, sync_sched_exp_handler
);
3913 /* Wait and clean up, including waking everyone. */
3914 rcu_exp_wait_wake(rsp
, s
);
3916 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3919 * Check to see if there is any immediate RCU-related work to be done
3920 * by the current CPU, for the specified type of RCU, returning 1 if so.
3921 * The checks are in order of increasing expense: checks that can be
3922 * carried out against CPU-local state are performed first. However,
3923 * we must check for CPU stalls first, else we might not get a chance.
3925 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3927 struct rcu_node
*rnp
= rdp
->mynode
;
3929 rdp
->n_rcu_pending
++;
3931 /* Check for CPU stalls, if enabled. */
3932 check_cpu_stall(rsp
, rdp
);
3934 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3935 if (rcu_nohz_full_cpu(rsp
))
3938 /* Is the RCU core waiting for a quiescent state from this CPU? */
3939 if (rcu_scheduler_fully_active
&&
3940 rdp
->core_needs_qs
&& rdp
->cpu_no_qs
.b
.norm
&&
3941 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3942 rdp
->n_rp_core_needs_qs
++;
3943 } else if (rdp
->core_needs_qs
&&
3944 (!rdp
->cpu_no_qs
.b
.norm
||
3945 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3946 rdp
->n_rp_report_qs
++;
3950 /* Does this CPU have callbacks ready to invoke? */
3951 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3952 rdp
->n_rp_cb_ready
++;
3956 /* Has RCU gone idle with this CPU needing another grace period? */
3957 if (cpu_needs_another_gp(rsp
, rdp
)) {
3958 rdp
->n_rp_cpu_needs_gp
++;
3962 /* Has another RCU grace period completed? */
3963 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3964 rdp
->n_rp_gp_completed
++;
3968 /* Has a new RCU grace period started? */
3969 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3970 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3971 rdp
->n_rp_gp_started
++;
3975 /* Does this CPU need a deferred NOCB wakeup? */
3976 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3977 rdp
->n_rp_nocb_defer_wakeup
++;
3982 rdp
->n_rp_need_nothing
++;
3987 * Check to see if there is any immediate RCU-related work to be done
3988 * by the current CPU, returning 1 if so. This function is part of the
3989 * RCU implementation; it is -not- an exported member of the RCU API.
3991 static int rcu_pending(void)
3993 struct rcu_state
*rsp
;
3995 for_each_rcu_flavor(rsp
)
3996 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
4002 * Return true if the specified CPU has any callback. If all_lazy is
4003 * non-NULL, store an indication of whether all callbacks are lazy.
4004 * (If there are no callbacks, all of them are deemed to be lazy.)
4006 static bool __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
4010 struct rcu_data
*rdp
;
4011 struct rcu_state
*rsp
;
4013 for_each_rcu_flavor(rsp
) {
4014 rdp
= this_cpu_ptr(rsp
->rda
);
4018 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
4029 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
4030 * the compiler is expected to optimize this away.
4032 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
4033 int cpu
, unsigned long done
)
4035 trace_rcu_barrier(rsp
->name
, s
, cpu
,
4036 atomic_read(&rsp
->barrier_cpu_count
), done
);
4040 * RCU callback function for _rcu_barrier(). If we are last, wake
4041 * up the task executing _rcu_barrier().
4043 static void rcu_barrier_callback(struct rcu_head
*rhp
)
4045 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
4046 struct rcu_state
*rsp
= rdp
->rsp
;
4048 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
4049 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->barrier_sequence
);
4050 complete(&rsp
->barrier_completion
);
4052 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->barrier_sequence
);
4057 * Called with preemption disabled, and from cross-cpu IRQ context.
4059 static void rcu_barrier_func(void *type
)
4061 struct rcu_state
*rsp
= type
;
4062 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
4064 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->barrier_sequence
);
4065 atomic_inc(&rsp
->barrier_cpu_count
);
4066 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
4070 * Orchestrate the specified type of RCU barrier, waiting for all
4071 * RCU callbacks of the specified type to complete.
4073 static void _rcu_barrier(struct rcu_state
*rsp
)
4076 struct rcu_data
*rdp
;
4077 unsigned long s
= rcu_seq_snap(&rsp
->barrier_sequence
);
4079 _rcu_barrier_trace(rsp
, "Begin", -1, s
);
4081 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4082 mutex_lock(&rsp
->barrier_mutex
);
4084 /* Did someone else do our work for us? */
4085 if (rcu_seq_done(&rsp
->barrier_sequence
, s
)) {
4086 _rcu_barrier_trace(rsp
, "EarlyExit", -1, rsp
->barrier_sequence
);
4087 smp_mb(); /* caller's subsequent code after above check. */
4088 mutex_unlock(&rsp
->barrier_mutex
);
4092 /* Mark the start of the barrier operation. */
4093 rcu_seq_start(&rsp
->barrier_sequence
);
4094 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->barrier_sequence
);
4097 * Initialize the count to one rather than to zero in order to
4098 * avoid a too-soon return to zero in case of a short grace period
4099 * (or preemption of this task). Exclude CPU-hotplug operations
4100 * to ensure that no offline CPU has callbacks queued.
4102 init_completion(&rsp
->barrier_completion
);
4103 atomic_set(&rsp
->barrier_cpu_count
, 1);
4107 * Force each CPU with callbacks to register a new callback.
4108 * When that callback is invoked, we will know that all of the
4109 * corresponding CPU's preceding callbacks have been invoked.
4111 for_each_possible_cpu(cpu
) {
4112 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
4114 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4115 if (rcu_is_nocb_cpu(cpu
)) {
4116 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
4117 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
4118 rsp
->barrier_sequence
);
4120 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
4121 rsp
->barrier_sequence
);
4122 smp_mb__before_atomic();
4123 atomic_inc(&rsp
->barrier_cpu_count
);
4124 __call_rcu(&rdp
->barrier_head
,
4125 rcu_barrier_callback
, rsp
, cpu
, 0);
4127 } else if (READ_ONCE(rdp
->qlen
)) {
4128 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
4129 rsp
->barrier_sequence
);
4130 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
4132 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
4133 rsp
->barrier_sequence
);
4139 * Now that we have an rcu_barrier_callback() callback on each
4140 * CPU, and thus each counted, remove the initial count.
4142 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
4143 complete(&rsp
->barrier_completion
);
4145 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4146 wait_for_completion(&rsp
->barrier_completion
);
4148 /* Mark the end of the barrier operation. */
4149 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->barrier_sequence
);
4150 rcu_seq_end(&rsp
->barrier_sequence
);
4152 /* Other rcu_barrier() invocations can now safely proceed. */
4153 mutex_unlock(&rsp
->barrier_mutex
);
4157 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4159 void rcu_barrier_bh(void)
4161 _rcu_barrier(&rcu_bh_state
);
4163 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
4166 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4168 void rcu_barrier_sched(void)
4170 _rcu_barrier(&rcu_sched_state
);
4172 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
4175 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4176 * first CPU in a given leaf rcu_node structure coming online. The caller
4177 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4180 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
4183 struct rcu_node
*rnp
= rnp_leaf
;
4186 mask
= rnp
->grpmask
;
4190 raw_spin_lock_rcu_node(rnp
); /* Interrupts already disabled. */
4191 rnp
->qsmaskinit
|= mask
;
4192 raw_spin_unlock_rcu_node(rnp
); /* Interrupts remain disabled. */
4197 * Do boot-time initialization of a CPU's per-CPU RCU data.
4200 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
4202 unsigned long flags
;
4203 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4204 struct rcu_node
*rnp
= rcu_get_root(rsp
);
4206 /* Set up local state, ensuring consistent view of global state. */
4207 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4208 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
4209 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
4210 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
4211 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
4214 rcu_boot_init_nocb_percpu_data(rdp
);
4215 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4219 * Initialize a CPU's per-CPU RCU data. Note that only one online or
4220 * offline event can be happening at a given time. Note also that we
4221 * can accept some slop in the rsp->completed access due to the fact
4222 * that this CPU cannot possibly have any RCU callbacks in flight yet.
4225 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
4227 unsigned long flags
;
4229 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4230 struct rcu_node
*rnp
= rcu_get_root(rsp
);
4232 /* Set up local state, ensuring consistent view of global state. */
4233 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4234 rdp
->qlen_last_fqs_check
= 0;
4235 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
4236 rdp
->blimit
= blimit
;
4238 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
4239 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
4240 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
4241 atomic_set(&rdp
->dynticks
->dynticks
,
4242 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
4243 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
4246 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4247 * propagation up the rcu_node tree will happen at the beginning
4248 * of the next grace period.
4251 mask
= rdp
->grpmask
;
4252 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
4253 rnp
->qsmaskinitnext
|= mask
;
4254 rnp
->expmaskinitnext
|= mask
;
4255 if (!rdp
->beenonline
)
4256 WRITE_ONCE(rsp
->ncpus
, READ_ONCE(rsp
->ncpus
) + 1);
4257 rdp
->beenonline
= true; /* We have now been online. */
4258 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
4259 rdp
->completed
= rnp
->completed
;
4260 rdp
->cpu_no_qs
.b
.norm
= true;
4261 rdp
->rcu_qs_ctr_snap
= per_cpu(rcu_qs_ctr
, cpu
);
4262 rdp
->core_needs_qs
= false;
4263 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
4264 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4267 static void rcu_prepare_cpu(int cpu
)
4269 struct rcu_state
*rsp
;
4271 for_each_rcu_flavor(rsp
)
4272 rcu_init_percpu_data(cpu
, rsp
);
4275 #ifdef CONFIG_HOTPLUG_CPU
4277 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4278 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4280 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4281 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4284 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
4286 unsigned long flags
;
4288 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4289 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
4291 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
4294 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4295 mask
= rdp
->grpmask
;
4296 raw_spin_lock_irqsave_rcu_node(rnp
, flags
); /* Enforce GP memory-order guarantee. */
4297 rnp
->qsmaskinitnext
&= ~mask
;
4298 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4301 void rcu_report_dead(unsigned int cpu
)
4303 struct rcu_state
*rsp
;
4305 /* QS for any half-done expedited RCU-sched GP. */
4307 rcu_report_exp_rdp(&rcu_sched_state
,
4308 this_cpu_ptr(rcu_sched_state
.rda
), true);
4310 for_each_rcu_flavor(rsp
)
4311 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
4316 * Handle CPU online/offline notification events.
4318 int rcu_cpu_notify(struct notifier_block
*self
,
4319 unsigned long action
, void *hcpu
)
4321 long cpu
= (long)hcpu
;
4322 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
4323 struct rcu_node
*rnp
= rdp
->mynode
;
4324 struct rcu_state
*rsp
;
4327 case CPU_UP_PREPARE
:
4328 case CPU_UP_PREPARE_FROZEN
:
4329 rcu_prepare_cpu(cpu
);
4330 rcu_prepare_kthreads(cpu
);
4331 rcu_spawn_all_nocb_kthreads(cpu
);
4334 case CPU_DOWN_FAILED
:
4335 sync_sched_exp_online_cleanup(cpu
);
4336 rcu_boost_kthread_setaffinity(rnp
, -1);
4338 case CPU_DOWN_PREPARE
:
4339 rcu_boost_kthread_setaffinity(rnp
, cpu
);
4342 case CPU_DYING_FROZEN
:
4343 for_each_rcu_flavor(rsp
)
4344 rcu_cleanup_dying_cpu(rsp
);
4347 case CPU_DEAD_FROZEN
:
4348 case CPU_UP_CANCELED
:
4349 case CPU_UP_CANCELED_FROZEN
:
4350 for_each_rcu_flavor(rsp
) {
4351 rcu_cleanup_dead_cpu(cpu
, rsp
);
4352 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
4361 static int rcu_pm_notify(struct notifier_block
*self
,
4362 unsigned long action
, void *hcpu
)
4365 case PM_HIBERNATION_PREPARE
:
4366 case PM_SUSPEND_PREPARE
:
4367 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
4370 case PM_POST_HIBERNATION
:
4371 case PM_POST_SUSPEND
:
4372 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
4373 rcu_unexpedite_gp();
4382 * Spawn the kthreads that handle each RCU flavor's grace periods.
4384 static int __init
rcu_spawn_gp_kthread(void)
4386 unsigned long flags
;
4387 int kthread_prio_in
= kthread_prio
;
4388 struct rcu_node
*rnp
;
4389 struct rcu_state
*rsp
;
4390 struct sched_param sp
;
4391 struct task_struct
*t
;
4393 /* Force priority into range. */
4394 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
4396 else if (kthread_prio
< 0)
4398 else if (kthread_prio
> 99)
4400 if (kthread_prio
!= kthread_prio_in
)
4401 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4402 kthread_prio
, kthread_prio_in
);
4404 rcu_scheduler_fully_active
= 1;
4405 for_each_rcu_flavor(rsp
) {
4406 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
4408 rnp
= rcu_get_root(rsp
);
4409 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4410 rsp
->gp_kthread
= t
;
4412 sp
.sched_priority
= kthread_prio
;
4413 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
4415 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4418 rcu_spawn_nocb_kthreads();
4419 rcu_spawn_boost_kthreads();
4422 early_initcall(rcu_spawn_gp_kthread
);
4425 * This function is invoked towards the end of the scheduler's initialization
4426 * process. Before this is called, the idle task might contain
4427 * RCU read-side critical sections (during which time, this idle
4428 * task is booting the system). After this function is called, the
4429 * idle tasks are prohibited from containing RCU read-side critical
4430 * sections. This function also enables RCU lockdep checking.
4432 void rcu_scheduler_starting(void)
4434 WARN_ON(num_online_cpus() != 1);
4435 WARN_ON(nr_context_switches() > 0);
4436 rcu_scheduler_active
= 1;
4440 * Compute the per-level fanout, either using the exact fanout specified
4441 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4443 static void __init
rcu_init_levelspread(int *levelspread
, const int *levelcnt
)
4447 if (rcu_fanout_exact
) {
4448 levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
4449 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
4450 levelspread
[i
] = RCU_FANOUT
;
4456 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4458 levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
4465 * Helper function for rcu_init() that initializes one rcu_state structure.
4467 static void __init
rcu_init_one(struct rcu_state
*rsp
)
4469 static const char * const buf
[] = RCU_NODE_NAME_INIT
;
4470 static const char * const fqs
[] = RCU_FQS_NAME_INIT
;
4471 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
4472 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
4473 static u8 fl_mask
= 0x1;
4475 int levelcnt
[RCU_NUM_LVLS
]; /* # nodes in each level. */
4476 int levelspread
[RCU_NUM_LVLS
]; /* kids/node in each level. */
4480 struct rcu_node
*rnp
;
4482 BUILD_BUG_ON(RCU_NUM_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4484 /* Silence gcc 4.8 false positive about array index out of range. */
4485 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4486 panic("rcu_init_one: rcu_num_lvls out of range");
4488 /* Initialize the level-tracking arrays. */
4490 for (i
= 0; i
< rcu_num_lvls
; i
++)
4491 levelcnt
[i
] = num_rcu_lvl
[i
];
4492 for (i
= 1; i
< rcu_num_lvls
; i
++)
4493 rsp
->level
[i
] = rsp
->level
[i
- 1] + levelcnt
[i
- 1];
4494 rcu_init_levelspread(levelspread
, levelcnt
);
4495 rsp
->flavor_mask
= fl_mask
;
4498 /* Initialize the elements themselves, starting from the leaves. */
4500 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4501 cpustride
*= levelspread
[i
];
4502 rnp
= rsp
->level
[i
];
4503 for (j
= 0; j
< levelcnt
[i
]; j
++, rnp
++) {
4504 raw_spin_lock_init(&ACCESS_PRIVATE(rnp
, lock
));
4505 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp
, lock
),
4506 &rcu_node_class
[i
], buf
[i
]);
4507 raw_spin_lock_init(&rnp
->fqslock
);
4508 lockdep_set_class_and_name(&rnp
->fqslock
,
4509 &rcu_fqs_class
[i
], fqs
[i
]);
4510 rnp
->gpnum
= rsp
->gpnum
;
4511 rnp
->completed
= rsp
->completed
;
4513 rnp
->qsmaskinit
= 0;
4514 rnp
->grplo
= j
* cpustride
;
4515 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4516 if (rnp
->grphi
>= nr_cpu_ids
)
4517 rnp
->grphi
= nr_cpu_ids
- 1;
4523 rnp
->grpnum
= j
% levelspread
[i
- 1];
4524 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4525 rnp
->parent
= rsp
->level
[i
- 1] +
4526 j
/ levelspread
[i
- 1];
4529 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4530 rcu_init_one_nocb(rnp
);
4531 init_waitqueue_head(&rnp
->exp_wq
[0]);
4532 init_waitqueue_head(&rnp
->exp_wq
[1]);
4533 spin_lock_init(&rnp
->exp_lock
);
4537 init_swait_queue_head(&rsp
->gp_wq
);
4538 init_swait_queue_head(&rsp
->expedited_wq
);
4539 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4540 for_each_possible_cpu(i
) {
4541 while (i
> rnp
->grphi
)
4543 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4544 rcu_boot_init_percpu_data(i
, rsp
);
4546 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4550 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4551 * replace the definitions in tree.h because those are needed to size
4552 * the ->node array in the rcu_state structure.
4554 static void __init
rcu_init_geometry(void)
4558 int rcu_capacity
[RCU_NUM_LVLS
];
4561 * Initialize any unspecified boot parameters.
4562 * The default values of jiffies_till_first_fqs and
4563 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4564 * value, which is a function of HZ, then adding one for each
4565 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4567 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4568 if (jiffies_till_first_fqs
== ULONG_MAX
)
4569 jiffies_till_first_fqs
= d
;
4570 if (jiffies_till_next_fqs
== ULONG_MAX
)
4571 jiffies_till_next_fqs
= d
;
4573 /* If the compile-time values are accurate, just leave. */
4574 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4575 nr_cpu_ids
== NR_CPUS
)
4577 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4578 rcu_fanout_leaf
, nr_cpu_ids
);
4581 * The boot-time rcu_fanout_leaf parameter must be at least two
4582 * and cannot exceed the number of bits in the rcu_node masks.
4583 * Complain and fall back to the compile-time values if this
4584 * limit is exceeded.
4586 if (rcu_fanout_leaf
< 2 ||
4587 rcu_fanout_leaf
> sizeof(unsigned long) * 8) {
4588 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4594 * Compute number of nodes that can be handled an rcu_node tree
4595 * with the given number of levels.
4597 rcu_capacity
[0] = rcu_fanout_leaf
;
4598 for (i
= 1; i
< RCU_NUM_LVLS
; i
++)
4599 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4602 * The tree must be able to accommodate the configured number of CPUs.
4603 * If this limit is exceeded, fall back to the compile-time values.
4605 if (nr_cpu_ids
> rcu_capacity
[RCU_NUM_LVLS
- 1]) {
4606 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4611 /* Calculate the number of levels in the tree. */
4612 for (i
= 0; nr_cpu_ids
> rcu_capacity
[i
]; i
++) {
4614 rcu_num_lvls
= i
+ 1;
4616 /* Calculate the number of rcu_nodes at each level of the tree. */
4617 for (i
= 0; i
< rcu_num_lvls
; i
++) {
4618 int cap
= rcu_capacity
[(rcu_num_lvls
- 1) - i
];
4619 num_rcu_lvl
[i
] = DIV_ROUND_UP(nr_cpu_ids
, cap
);
4622 /* Calculate the total number of rcu_node structures. */
4624 for (i
= 0; i
< rcu_num_lvls
; i
++)
4625 rcu_num_nodes
+= num_rcu_lvl
[i
];
4629 * Dump out the structure of the rcu_node combining tree associated
4630 * with the rcu_state structure referenced by rsp.
4632 static void __init
rcu_dump_rcu_node_tree(struct rcu_state
*rsp
)
4635 struct rcu_node
*rnp
;
4637 pr_info("rcu_node tree layout dump\n");
4639 rcu_for_each_node_breadth_first(rsp
, rnp
) {
4640 if (rnp
->level
!= level
) {
4645 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4650 void __init
rcu_init(void)
4654 rcu_early_boot_tests();
4656 rcu_bootup_announce();
4657 rcu_init_geometry();
4658 rcu_init_one(&rcu_bh_state
);
4659 rcu_init_one(&rcu_sched_state
);
4661 rcu_dump_rcu_node_tree(&rcu_sched_state
);
4662 __rcu_init_preempt();
4663 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4666 * We don't need protection against CPU-hotplug here because
4667 * this is called early in boot, before either interrupts
4668 * or the scheduler are operational.
4670 cpu_notifier(rcu_cpu_notify
, 0);
4671 pm_notifier(rcu_pm_notify
, 0);
4672 for_each_online_cpu(cpu
)
4673 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4676 #include "tree_plugin.h"