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), \
106 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
109 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
110 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
112 static struct rcu_state
*const rcu_state_p
;
113 LIST_HEAD(rcu_struct_flavors
);
115 /* Dump rcu_node combining tree at boot to verify correct setup. */
116 static bool dump_tree
;
117 module_param(dump_tree
, bool, 0444);
118 /* Control rcu_node-tree auto-balancing at boot time. */
119 static bool rcu_fanout_exact
;
120 module_param(rcu_fanout_exact
, bool, 0444);
121 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
122 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
123 module_param(rcu_fanout_leaf
, int, 0444);
124 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
125 /* Number of rcu_nodes at specified level. */
126 static int num_rcu_lvl
[] = NUM_RCU_LVL_INIT
;
127 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
130 * The rcu_scheduler_active variable transitions from zero to one just
131 * before the first task is spawned. So when this variable is zero, RCU
132 * can assume that there is but one task, allowing RCU to (for example)
133 * optimize synchronize_sched() to a simple barrier(). When this variable
134 * is one, RCU must actually do all the hard work required to detect real
135 * grace periods. This variable is also used to suppress boot-time false
136 * positives from lockdep-RCU error checking.
138 int rcu_scheduler_active __read_mostly
;
139 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
142 * The rcu_scheduler_fully_active variable transitions from zero to one
143 * during the early_initcall() processing, which is after the scheduler
144 * is capable of creating new tasks. So RCU processing (for example,
145 * creating tasks for RCU priority boosting) must be delayed until after
146 * rcu_scheduler_fully_active transitions from zero to one. We also
147 * currently delay invocation of any RCU callbacks until after this point.
149 * It might later prove better for people registering RCU callbacks during
150 * early boot to take responsibility for these callbacks, but one step at
153 static int rcu_scheduler_fully_active __read_mostly
;
155 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
156 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
157 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
158 static void invoke_rcu_core(void);
159 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
160 static void rcu_report_exp_rdp(struct rcu_state
*rsp
,
161 struct rcu_data
*rdp
, bool wake
);
163 /* rcuc/rcub kthread realtime priority */
164 #ifdef CONFIG_RCU_KTHREAD_PRIO
165 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
166 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
167 static int kthread_prio
= IS_ENABLED(CONFIG_RCU_BOOST
) ? 1 : 0;
168 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
169 module_param(kthread_prio
, int, 0644);
171 /* Delay in jiffies for grace-period initialization delays, debug only. */
173 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
174 static int gp_preinit_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY
;
175 module_param(gp_preinit_delay
, int, 0644);
176 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
177 static const int gp_preinit_delay
;
178 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
180 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
181 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
182 module_param(gp_init_delay
, int, 0644);
183 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
184 static const int gp_init_delay
;
185 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
187 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
188 static int gp_cleanup_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY
;
189 module_param(gp_cleanup_delay
, int, 0644);
190 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
191 static const int gp_cleanup_delay
;
192 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
195 * Number of grace periods between delays, normalized by the duration of
196 * the delay. The longer the the delay, the more the grace periods between
197 * each delay. The reason for this normalization is that it means that,
198 * for non-zero delays, the overall slowdown of grace periods is constant
199 * regardless of the duration of the delay. This arrangement balances
200 * the need for long delays to increase some race probabilities with the
201 * need for fast grace periods to increase other race probabilities.
203 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
206 * Track the rcutorture test sequence number and the update version
207 * number within a given test. The rcutorture_testseq is incremented
208 * on every rcutorture module load and unload, so has an odd value
209 * when a test is running. The rcutorture_vernum is set to zero
210 * when rcutorture starts and is incremented on each rcutorture update.
211 * These variables enable correlating rcutorture output with the
212 * RCU tracing information.
214 unsigned long rcutorture_testseq
;
215 unsigned long rcutorture_vernum
;
218 * Compute the mask of online CPUs for the specified rcu_node structure.
219 * This will not be stable unless the rcu_node structure's ->lock is
220 * held, but the bit corresponding to the current CPU will be stable
223 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
225 return READ_ONCE(rnp
->qsmaskinitnext
);
229 * Return true if an RCU grace period is in progress. The READ_ONCE()s
230 * permit this function to be invoked without holding the root rcu_node
231 * structure's ->lock, but of course results can be subject to change.
233 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
235 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
239 * Note a quiescent state. Because we do not need to know
240 * how many quiescent states passed, just if there was at least
241 * one since the start of the grace period, this just sets a flag.
242 * The caller must have disabled preemption.
244 void rcu_sched_qs(void)
246 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.s
))
248 trace_rcu_grace_period(TPS("rcu_sched"),
249 __this_cpu_read(rcu_sched_data
.gpnum
),
251 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.norm
, false);
252 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
254 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, false);
255 rcu_report_exp_rdp(&rcu_sched_state
,
256 this_cpu_ptr(&rcu_sched_data
), true);
261 if (__this_cpu_read(rcu_bh_data
.cpu_no_qs
.s
)) {
262 trace_rcu_grace_period(TPS("rcu_bh"),
263 __this_cpu_read(rcu_bh_data
.gpnum
),
265 __this_cpu_write(rcu_bh_data
.cpu_no_qs
.b
.norm
, false);
269 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
271 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
272 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
273 .dynticks
= ATOMIC_INIT(1),
274 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
275 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
276 .dynticks_idle
= ATOMIC_INIT(1),
277 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
280 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
281 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
284 * Let the RCU core know that this CPU has gone through the scheduler,
285 * which is a quiescent state. This is called when the need for a
286 * quiescent state is urgent, so we burn an atomic operation and full
287 * memory barriers to let the RCU core know about it, regardless of what
288 * this CPU might (or might not) do in the near future.
290 * We inform the RCU core by emulating a zero-duration dyntick-idle
291 * period, which we in turn do by incrementing the ->dynticks counter
294 * The caller must have disabled interrupts.
296 static void rcu_momentary_dyntick_idle(void)
298 struct rcu_data
*rdp
;
299 struct rcu_dynticks
*rdtp
;
301 struct rcu_state
*rsp
;
304 * Yes, we can lose flag-setting operations. This is OK, because
305 * the flag will be set again after some delay.
307 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
308 raw_cpu_write(rcu_sched_qs_mask
, 0);
310 /* Find the flavor that needs a quiescent state. */
311 for_each_rcu_flavor(rsp
) {
312 rdp
= raw_cpu_ptr(rsp
->rda
);
313 if (!(resched_mask
& rsp
->flavor_mask
))
315 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
316 if (READ_ONCE(rdp
->mynode
->completed
) !=
317 READ_ONCE(rdp
->cond_resched_completed
))
321 * Pretend to be momentarily idle for the quiescent state.
322 * This allows the grace-period kthread to record the
323 * quiescent state, with no need for this CPU to do anything
326 rdtp
= this_cpu_ptr(&rcu_dynticks
);
327 smp_mb__before_atomic(); /* Earlier stuff before QS. */
328 atomic_add(2, &rdtp
->dynticks
); /* QS. */
329 smp_mb__after_atomic(); /* Later stuff after QS. */
335 * Note a context switch. This is a quiescent state for RCU-sched,
336 * and requires special handling for preemptible RCU.
337 * The caller must have disabled interrupts.
339 void rcu_note_context_switch(void)
341 barrier(); /* Avoid RCU read-side critical sections leaking down. */
342 trace_rcu_utilization(TPS("Start context switch"));
344 rcu_preempt_note_context_switch();
345 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
346 rcu_momentary_dyntick_idle();
347 trace_rcu_utilization(TPS("End context switch"));
348 barrier(); /* Avoid RCU read-side critical sections leaking up. */
350 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
353 * Register a quiescent state for all RCU flavors. If there is an
354 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
355 * dyntick-idle quiescent state visible to other CPUs (but only for those
356 * RCU flavors in desperate need of a quiescent state, which will normally
357 * be none of them). Either way, do a lightweight quiescent state for
360 * The barrier() calls are redundant in the common case when this is
361 * called externally, but just in case this is called from within this
365 void rcu_all_qs(void)
369 barrier(); /* Avoid RCU read-side critical sections leaking down. */
370 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
))) {
371 local_irq_save(flags
);
372 rcu_momentary_dyntick_idle();
373 local_irq_restore(flags
);
375 if (unlikely(raw_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))) {
377 * Yes, we just checked a per-CPU variable with preemption
378 * enabled, so we might be migrated to some other CPU at
379 * this point. That is OK because in that case, the
380 * migration will supply the needed quiescent state.
381 * We might end up needlessly disabling preemption and
382 * invoking rcu_sched_qs() on the destination CPU, but
383 * the probability and cost are both quite low, so this
384 * should not be a problem in practice.
390 this_cpu_inc(rcu_qs_ctr
);
391 barrier(); /* Avoid RCU read-side critical sections leaking up. */
393 EXPORT_SYMBOL_GPL(rcu_all_qs
);
395 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
396 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
397 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
399 module_param(blimit
, long, 0444);
400 module_param(qhimark
, long, 0444);
401 module_param(qlowmark
, long, 0444);
403 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
404 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
406 module_param(jiffies_till_first_fqs
, ulong
, 0644);
407 module_param(jiffies_till_next_fqs
, ulong
, 0644);
410 * How long the grace period must be before we start recruiting
411 * quiescent-state help from rcu_note_context_switch().
413 static ulong jiffies_till_sched_qs
= HZ
/ 20;
414 module_param(jiffies_till_sched_qs
, ulong
, 0644);
416 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
417 struct rcu_data
*rdp
);
418 static void force_qs_rnp(struct rcu_state
*rsp
,
419 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
420 unsigned long *maxj
),
421 bool *isidle
, unsigned long *maxj
);
422 static void force_quiescent_state(struct rcu_state
*rsp
);
423 static int rcu_pending(void);
426 * Return the number of RCU batches started thus far for debug & stats.
428 unsigned long rcu_batches_started(void)
430 return rcu_state_p
->gpnum
;
432 EXPORT_SYMBOL_GPL(rcu_batches_started
);
435 * Return the number of RCU-sched batches started thus far for debug & stats.
437 unsigned long rcu_batches_started_sched(void)
439 return rcu_sched_state
.gpnum
;
441 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
444 * Return the number of RCU BH batches started thus far for debug & stats.
446 unsigned long rcu_batches_started_bh(void)
448 return rcu_bh_state
.gpnum
;
450 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
453 * Return the number of RCU batches completed thus far for debug & stats.
455 unsigned long rcu_batches_completed(void)
457 return rcu_state_p
->completed
;
459 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
462 * Return the number of RCU-sched batches completed thus far for debug & stats.
464 unsigned long rcu_batches_completed_sched(void)
466 return rcu_sched_state
.completed
;
468 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
471 * Return the number of RCU BH batches completed thus far for debug & stats.
473 unsigned long rcu_batches_completed_bh(void)
475 return rcu_bh_state
.completed
;
477 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
480 * Force a quiescent state.
482 void rcu_force_quiescent_state(void)
484 force_quiescent_state(rcu_state_p
);
486 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
489 * Force a quiescent state for RCU BH.
491 void rcu_bh_force_quiescent_state(void)
493 force_quiescent_state(&rcu_bh_state
);
495 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
498 * Force a quiescent state for RCU-sched.
500 void rcu_sched_force_quiescent_state(void)
502 force_quiescent_state(&rcu_sched_state
);
504 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
507 * Show the state of the grace-period kthreads.
509 void show_rcu_gp_kthreads(void)
511 struct rcu_state
*rsp
;
513 for_each_rcu_flavor(rsp
) {
514 pr_info("%s: wait state: %d ->state: %#lx\n",
515 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
516 /* sched_show_task(rsp->gp_kthread); */
519 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
522 * Record the number of times rcutorture tests have been initiated and
523 * terminated. This information allows the debugfs tracing stats to be
524 * correlated to the rcutorture messages, even when the rcutorture module
525 * is being repeatedly loaded and unloaded. In other words, we cannot
526 * store this state in rcutorture itself.
528 void rcutorture_record_test_transition(void)
530 rcutorture_testseq
++;
531 rcutorture_vernum
= 0;
533 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
536 * Send along grace-period-related data for rcutorture diagnostics.
538 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
539 unsigned long *gpnum
, unsigned long *completed
)
541 struct rcu_state
*rsp
= NULL
;
550 case RCU_SCHED_FLAVOR
:
551 rsp
= &rcu_sched_state
;
557 *flags
= READ_ONCE(rsp
->gp_flags
);
558 *gpnum
= READ_ONCE(rsp
->gpnum
);
559 *completed
= READ_ONCE(rsp
->completed
);
566 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
569 * Record the number of writer passes through the current rcutorture test.
570 * This is also used to correlate debugfs tracing stats with the rcutorture
573 void rcutorture_record_progress(unsigned long vernum
)
577 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
580 * Does the CPU have callbacks ready to be invoked?
583 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
585 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
586 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
590 * Return the root node of the specified rcu_state structure.
592 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
594 return &rsp
->node
[0];
598 * Is there any need for future grace periods?
599 * Interrupts must be disabled. If the caller does not hold the root
600 * rnp_node structure's ->lock, the results are advisory only.
602 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
604 struct rcu_node
*rnp
= rcu_get_root(rsp
);
605 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
606 int *fp
= &rnp
->need_future_gp
[idx
];
608 return READ_ONCE(*fp
);
612 * Does the current CPU require a not-yet-started grace period?
613 * The caller must have disabled interrupts to prevent races with
614 * normal callback registry.
617 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
621 if (rcu_gp_in_progress(rsp
))
622 return false; /* No, a grace period is already in progress. */
623 if (rcu_future_needs_gp(rsp
))
624 return true; /* Yes, a no-CBs CPU needs one. */
625 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
626 return false; /* No, this is a no-CBs (or offline) CPU. */
627 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
628 return true; /* Yes, CPU has newly registered callbacks. */
629 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
630 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
631 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
632 rdp
->nxtcompleted
[i
]))
633 return true; /* Yes, CBs for future grace period. */
634 return false; /* No grace period needed. */
638 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
640 * If the new value of the ->dynticks_nesting counter now is zero,
641 * we really have entered idle, and must do the appropriate accounting.
642 * The caller must have disabled interrupts.
644 static void rcu_eqs_enter_common(long long oldval
, bool user
)
646 struct rcu_state
*rsp
;
647 struct rcu_data
*rdp
;
648 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
650 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
651 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
652 !user
&& !is_idle_task(current
)) {
653 struct task_struct
*idle __maybe_unused
=
654 idle_task(smp_processor_id());
656 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
657 rcu_ftrace_dump(DUMP_ORIG
);
658 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
659 current
->pid
, current
->comm
,
660 idle
->pid
, idle
->comm
); /* must be idle task! */
662 for_each_rcu_flavor(rsp
) {
663 rdp
= this_cpu_ptr(rsp
->rda
);
664 do_nocb_deferred_wakeup(rdp
);
666 rcu_prepare_for_idle();
667 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
668 smp_mb__before_atomic(); /* See above. */
669 atomic_inc(&rdtp
->dynticks
);
670 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
671 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
672 atomic_read(&rdtp
->dynticks
) & 0x1);
673 rcu_dynticks_task_enter();
676 * It is illegal to enter an extended quiescent state while
677 * in an RCU read-side critical section.
679 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map
),
680 "Illegal idle entry in RCU read-side critical section.");
681 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
),
682 "Illegal idle entry in RCU-bh read-side critical section.");
683 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map
),
684 "Illegal idle entry in RCU-sched read-side critical section.");
688 * Enter an RCU extended quiescent state, which can be either the
689 * idle loop or adaptive-tickless usermode execution.
691 static void rcu_eqs_enter(bool user
)
694 struct rcu_dynticks
*rdtp
;
696 rdtp
= this_cpu_ptr(&rcu_dynticks
);
697 oldval
= rdtp
->dynticks_nesting
;
698 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
699 (oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
700 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
701 rdtp
->dynticks_nesting
= 0;
702 rcu_eqs_enter_common(oldval
, user
);
704 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
709 * rcu_idle_enter - inform RCU that current CPU is entering idle
711 * Enter idle mode, in other words, -leave- the mode in which RCU
712 * read-side critical sections can occur. (Though RCU read-side
713 * critical sections can occur in irq handlers in idle, a possibility
714 * handled by irq_enter() and irq_exit().)
716 * We crowbar the ->dynticks_nesting field to zero to allow for
717 * the possibility of usermode upcalls having messed up our count
718 * of interrupt nesting level during the prior busy period.
720 void rcu_idle_enter(void)
724 local_irq_save(flags
);
725 rcu_eqs_enter(false);
726 rcu_sysidle_enter(0);
727 local_irq_restore(flags
);
729 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
731 #ifdef CONFIG_NO_HZ_FULL
733 * rcu_user_enter - inform RCU that we are resuming userspace.
735 * Enter RCU idle mode right before resuming userspace. No use of RCU
736 * is permitted between this call and rcu_user_exit(). This way the
737 * CPU doesn't need to maintain the tick for RCU maintenance purposes
738 * when the CPU runs in userspace.
740 void rcu_user_enter(void)
744 #endif /* CONFIG_NO_HZ_FULL */
747 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
749 * Exit from an interrupt handler, which might possibly result in entering
750 * idle mode, in other words, leaving the mode in which read-side critical
751 * sections can occur. The caller must have disabled interrupts.
753 * This code assumes that the idle loop never does anything that might
754 * result in unbalanced calls to irq_enter() and irq_exit(). If your
755 * architecture violates this assumption, RCU will give you what you
756 * deserve, good and hard. But very infrequently and irreproducibly.
758 * Use things like work queues to work around this limitation.
760 * You have been warned.
762 void rcu_irq_exit(void)
765 struct rcu_dynticks
*rdtp
;
767 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
768 rdtp
= this_cpu_ptr(&rcu_dynticks
);
769 oldval
= rdtp
->dynticks_nesting
;
770 rdtp
->dynticks_nesting
--;
771 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
772 rdtp
->dynticks_nesting
< 0);
773 if (rdtp
->dynticks_nesting
)
774 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
776 rcu_eqs_enter_common(oldval
, true);
777 rcu_sysidle_enter(1);
781 * Wrapper for rcu_irq_exit() where interrupts are enabled.
783 void rcu_irq_exit_irqson(void)
787 local_irq_save(flags
);
789 local_irq_restore(flags
);
793 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
795 * If the new value of the ->dynticks_nesting counter was previously zero,
796 * we really have exited idle, and must do the appropriate accounting.
797 * The caller must have disabled interrupts.
799 static void rcu_eqs_exit_common(long long oldval
, int user
)
801 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
803 rcu_dynticks_task_exit();
804 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
805 atomic_inc(&rdtp
->dynticks
);
806 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
807 smp_mb__after_atomic(); /* See above. */
808 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
809 !(atomic_read(&rdtp
->dynticks
) & 0x1));
810 rcu_cleanup_after_idle();
811 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
812 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
813 !user
&& !is_idle_task(current
)) {
814 struct task_struct
*idle __maybe_unused
=
815 idle_task(smp_processor_id());
817 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
818 oldval
, rdtp
->dynticks_nesting
);
819 rcu_ftrace_dump(DUMP_ORIG
);
820 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
821 current
->pid
, current
->comm
,
822 idle
->pid
, idle
->comm
); /* must be idle task! */
827 * Exit an RCU extended quiescent state, which can be either the
828 * idle loop or adaptive-tickless usermode execution.
830 static void rcu_eqs_exit(bool user
)
832 struct rcu_dynticks
*rdtp
;
835 rdtp
= this_cpu_ptr(&rcu_dynticks
);
836 oldval
= rdtp
->dynticks_nesting
;
837 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) && oldval
< 0);
838 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
839 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
841 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
842 rcu_eqs_exit_common(oldval
, user
);
847 * rcu_idle_exit - inform RCU that current CPU is leaving idle
849 * Exit idle mode, in other words, -enter- the mode in which RCU
850 * read-side critical sections can occur.
852 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
853 * allow for the possibility of usermode upcalls messing up our count
854 * of interrupt nesting level during the busy period that is just
857 void rcu_idle_exit(void)
861 local_irq_save(flags
);
864 local_irq_restore(flags
);
866 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
868 #ifdef CONFIG_NO_HZ_FULL
870 * rcu_user_exit - inform RCU that we are exiting userspace.
872 * Exit RCU idle mode while entering the kernel because it can
873 * run a RCU read side critical section anytime.
875 void rcu_user_exit(void)
879 #endif /* CONFIG_NO_HZ_FULL */
882 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
884 * Enter an interrupt handler, which might possibly result in exiting
885 * idle mode, in other words, entering the mode in which read-side critical
886 * sections can occur. The caller must have disabled interrupts.
888 * Note that the Linux kernel is fully capable of entering an interrupt
889 * handler that it never exits, for example when doing upcalls to
890 * user mode! This code assumes that the idle loop never does upcalls to
891 * user mode. If your architecture does do upcalls from the idle loop (or
892 * does anything else that results in unbalanced calls to the irq_enter()
893 * and irq_exit() functions), RCU will give you what you deserve, good
894 * and hard. But very infrequently and irreproducibly.
896 * Use things like work queues to work around this limitation.
898 * You have been warned.
900 void rcu_irq_enter(void)
902 struct rcu_dynticks
*rdtp
;
905 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
906 rdtp
= this_cpu_ptr(&rcu_dynticks
);
907 oldval
= rdtp
->dynticks_nesting
;
908 rdtp
->dynticks_nesting
++;
909 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
910 rdtp
->dynticks_nesting
== 0);
912 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
914 rcu_eqs_exit_common(oldval
, true);
919 * Wrapper for rcu_irq_enter() where interrupts are enabled.
921 void rcu_irq_enter_irqson(void)
925 local_irq_save(flags
);
927 local_irq_restore(flags
);
931 * rcu_nmi_enter - inform RCU of entry to NMI context
933 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
934 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
935 * that the CPU is active. This implementation permits nested NMIs, as
936 * long as the nesting level does not overflow an int. (You will probably
937 * run out of stack space first.)
939 void rcu_nmi_enter(void)
941 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
944 /* Complain about underflow. */
945 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
948 * If idle from RCU viewpoint, atomically increment ->dynticks
949 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
950 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
951 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
952 * to be in the outermost NMI handler that interrupted an RCU-idle
953 * period (observation due to Andy Lutomirski).
955 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
956 smp_mb__before_atomic(); /* Force delay from prior write. */
957 atomic_inc(&rdtp
->dynticks
);
958 /* atomic_inc() before later RCU read-side crit sects */
959 smp_mb__after_atomic(); /* See above. */
960 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
963 rdtp
->dynticks_nmi_nesting
+= incby
;
968 * rcu_nmi_exit - inform RCU of exit from NMI context
970 * If we are returning from the outermost NMI handler that interrupted an
971 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
972 * to let the RCU grace-period handling know that the CPU is back to
975 void rcu_nmi_exit(void)
977 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
980 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
981 * (We are exiting an NMI handler, so RCU better be paying attention
984 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
985 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
988 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
989 * leave it in non-RCU-idle state.
991 if (rdtp
->dynticks_nmi_nesting
!= 1) {
992 rdtp
->dynticks_nmi_nesting
-= 2;
996 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
997 rdtp
->dynticks_nmi_nesting
= 0;
998 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
999 smp_mb__before_atomic(); /* See above. */
1000 atomic_inc(&rdtp
->dynticks
);
1001 smp_mb__after_atomic(); /* Force delay to next write. */
1002 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
1006 * __rcu_is_watching - are RCU read-side critical sections safe?
1008 * Return true if RCU is watching the running CPU, which means that
1009 * this CPU can safely enter RCU read-side critical sections. Unlike
1010 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1011 * least disabled preemption.
1013 bool notrace
__rcu_is_watching(void)
1015 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
1019 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1021 * If the current CPU is in its idle loop and is neither in an interrupt
1022 * or NMI handler, return true.
1024 bool notrace
rcu_is_watching(void)
1028 preempt_disable_notrace();
1029 ret
= __rcu_is_watching();
1030 preempt_enable_notrace();
1033 EXPORT_SYMBOL_GPL(rcu_is_watching
);
1035 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1038 * Is the current CPU online? Disable preemption to avoid false positives
1039 * that could otherwise happen due to the current CPU number being sampled,
1040 * this task being preempted, its old CPU being taken offline, resuming
1041 * on some other CPU, then determining that its old CPU is now offline.
1042 * It is OK to use RCU on an offline processor during initial boot, hence
1043 * the check for rcu_scheduler_fully_active. Note also that it is OK
1044 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1045 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1046 * offline to continue to use RCU for one jiffy after marking itself
1047 * offline in the cpu_online_mask. This leniency is necessary given the
1048 * non-atomic nature of the online and offline processing, for example,
1049 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1052 * This is also why RCU internally marks CPUs online during the
1053 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1055 * Disable checking if in an NMI handler because we cannot safely report
1056 * errors from NMI handlers anyway.
1058 bool rcu_lockdep_current_cpu_online(void)
1060 struct rcu_data
*rdp
;
1061 struct rcu_node
*rnp
;
1067 rdp
= this_cpu_ptr(&rcu_sched_data
);
1069 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
1070 !rcu_scheduler_fully_active
;
1074 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
1076 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1079 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1081 * If the current CPU is idle or running at a first-level (not nested)
1082 * interrupt from idle, return true. The caller must have at least
1083 * disabled preemption.
1085 static int rcu_is_cpu_rrupt_from_idle(void)
1087 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1091 * Snapshot the specified CPU's dynticks counter so that we can later
1092 * credit them with an implicit quiescent state. Return 1 if this CPU
1093 * is in dynticks idle mode, which is an extended quiescent state.
1095 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1096 bool *isidle
, unsigned long *maxj
)
1098 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1099 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1100 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1101 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1102 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1103 rdp
->mynode
->gpnum
))
1104 WRITE_ONCE(rdp
->gpwrap
, true);
1111 * Return true if the specified CPU has passed through a quiescent
1112 * state by virtue of being in or having passed through an dynticks
1113 * idle state since the last call to dyntick_save_progress_counter()
1114 * for this same CPU, or by virtue of having been offline.
1116 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1117 bool *isidle
, unsigned long *maxj
)
1123 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1124 snap
= (unsigned int)rdp
->dynticks_snap
;
1127 * If the CPU passed through or entered a dynticks idle phase with
1128 * no active irq/NMI handlers, then we can safely pretend that the CPU
1129 * already acknowledged the request to pass through a quiescent
1130 * state. Either way, that CPU cannot possibly be in an RCU
1131 * read-side critical section that started before the beginning
1132 * of the current RCU grace period.
1134 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1135 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1136 rdp
->dynticks_fqs
++;
1141 * Check for the CPU being offline, but only if the grace period
1142 * is old enough. We don't need to worry about the CPU changing
1143 * state: If we see it offline even once, it has been through a
1146 * The reason for insisting that the grace period be at least
1147 * one jiffy old is that CPUs that are not quite online and that
1148 * have just gone offline can still execute RCU read-side critical
1151 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1152 return 0; /* Grace period is not old enough. */
1154 if (cpu_is_offline(rdp
->cpu
)) {
1155 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1161 * A CPU running for an extended time within the kernel can
1162 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1163 * even context-switching back and forth between a pair of
1164 * in-kernel CPU-bound tasks cannot advance grace periods.
1165 * So if the grace period is old enough, make the CPU pay attention.
1166 * Note that the unsynchronized assignments to the per-CPU
1167 * rcu_sched_qs_mask variable are safe. Yes, setting of
1168 * bits can be lost, but they will be set again on the next
1169 * force-quiescent-state pass. So lost bit sets do not result
1170 * in incorrect behavior, merely in a grace period lasting
1171 * a few jiffies longer than it might otherwise. Because
1172 * there are at most four threads involved, and because the
1173 * updates are only once every few jiffies, the probability of
1174 * lossage (and thus of slight grace-period extension) is
1177 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1178 * is set too high, we override with half of the RCU CPU stall
1181 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1182 if (ULONG_CMP_GE(jiffies
,
1183 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1184 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1185 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1186 WRITE_ONCE(rdp
->cond_resched_completed
,
1187 READ_ONCE(rdp
->mynode
->completed
));
1188 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1190 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1192 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1195 /* And if it has been a really long time, kick the CPU as well. */
1196 if (ULONG_CMP_GE(jiffies
,
1197 rdp
->rsp
->gp_start
+ 2 * jiffies_till_sched_qs
) ||
1198 ULONG_CMP_GE(jiffies
, rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
))
1199 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1204 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1206 unsigned long j
= jiffies
;
1210 smp_wmb(); /* Record start time before stall time. */
1211 j1
= rcu_jiffies_till_stall_check();
1212 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1213 rsp
->jiffies_resched
= j
+ j1
/ 2;
1214 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1218 * Convert a ->gp_state value to a character string.
1220 static const char *gp_state_getname(short gs
)
1222 if (gs
< 0 || gs
>= ARRAY_SIZE(gp_state_names
))
1224 return gp_state_names
[gs
];
1228 * Complain about starvation of grace-period kthread.
1230 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1236 gpa
= READ_ONCE(rsp
->gp_activity
);
1237 if (j
- gpa
> 2 * HZ
) {
1238 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1240 rsp
->gpnum
, rsp
->completed
,
1242 gp_state_getname(rsp
->gp_state
), rsp
->gp_state
,
1243 rsp
->gp_kthread
? rsp
->gp_kthread
->state
: ~0);
1244 if (rsp
->gp_kthread
)
1245 sched_show_task(rsp
->gp_kthread
);
1250 * Dump stacks of all tasks running on stalled CPUs.
1252 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1255 unsigned long flags
;
1256 struct rcu_node
*rnp
;
1258 rcu_for_each_leaf_node(rsp
, rnp
) {
1259 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1260 if (rnp
->qsmask
!= 0) {
1261 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1262 if (rnp
->qsmask
& (1UL << cpu
))
1263 dump_cpu_task(rnp
->grplo
+ cpu
);
1265 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1269 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1273 unsigned long flags
;
1277 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1280 /* Only let one CPU complain about others per time interval. */
1282 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1283 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1284 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1285 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1288 WRITE_ONCE(rsp
->jiffies_stall
,
1289 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1290 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1293 * OK, time to rat on our buddy...
1294 * See Documentation/RCU/stallwarn.txt for info on how to debug
1295 * RCU CPU stall warnings.
1297 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1299 print_cpu_stall_info_begin();
1300 rcu_for_each_leaf_node(rsp
, rnp
) {
1301 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1302 ndetected
+= rcu_print_task_stall(rnp
);
1303 if (rnp
->qsmask
!= 0) {
1304 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1305 if (rnp
->qsmask
& (1UL << cpu
)) {
1306 print_cpu_stall_info(rsp
,
1311 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1314 print_cpu_stall_info_end();
1315 for_each_possible_cpu(cpu
)
1316 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1317 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1318 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1319 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1321 rcu_dump_cpu_stacks(rsp
);
1323 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1324 READ_ONCE(rsp
->completed
) == gpnum
) {
1325 pr_err("INFO: Stall ended before state dump start\n");
1328 gpa
= READ_ONCE(rsp
->gp_activity
);
1329 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1330 rsp
->name
, j
- gpa
, j
, gpa
,
1331 jiffies_till_next_fqs
,
1332 rcu_get_root(rsp
)->qsmask
);
1333 /* In this case, the current CPU might be at fault. */
1334 sched_show_task(current
);
1338 /* Complain about tasks blocking the grace period. */
1339 rcu_print_detail_task_stall(rsp
);
1341 rcu_check_gp_kthread_starvation(rsp
);
1343 force_quiescent_state(rsp
); /* Kick them all. */
1346 static void print_cpu_stall(struct rcu_state
*rsp
)
1349 unsigned long flags
;
1350 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1354 * OK, time to rat on ourselves...
1355 * See Documentation/RCU/stallwarn.txt for info on how to debug
1356 * RCU CPU stall warnings.
1358 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1359 print_cpu_stall_info_begin();
1360 print_cpu_stall_info(rsp
, smp_processor_id());
1361 print_cpu_stall_info_end();
1362 for_each_possible_cpu(cpu
)
1363 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1364 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1365 jiffies
- rsp
->gp_start
,
1366 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1368 rcu_check_gp_kthread_starvation(rsp
);
1370 rcu_dump_cpu_stacks(rsp
);
1372 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1373 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1374 WRITE_ONCE(rsp
->jiffies_stall
,
1375 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1376 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1379 * Attempt to revive the RCU machinery by forcing a context switch.
1381 * A context switch would normally allow the RCU state machine to make
1382 * progress and it could be we're stuck in kernel space without context
1383 * switches for an entirely unreasonable amount of time.
1385 resched_cpu(smp_processor_id());
1388 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1390 unsigned long completed
;
1391 unsigned long gpnum
;
1395 struct rcu_node
*rnp
;
1397 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1402 * Lots of memory barriers to reject false positives.
1404 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1405 * then rsp->gp_start, and finally rsp->completed. These values
1406 * are updated in the opposite order with memory barriers (or
1407 * equivalent) during grace-period initialization and cleanup.
1408 * Now, a false positive can occur if we get an new value of
1409 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1410 * the memory barriers, the only way that this can happen is if one
1411 * grace period ends and another starts between these two fetches.
1412 * Detect this by comparing rsp->completed with the previous fetch
1415 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1416 * and rsp->gp_start suffice to forestall false positives.
1418 gpnum
= READ_ONCE(rsp
->gpnum
);
1419 smp_rmb(); /* Pick up ->gpnum first... */
1420 js
= READ_ONCE(rsp
->jiffies_stall
);
1421 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1422 gps
= READ_ONCE(rsp
->gp_start
);
1423 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1424 completed
= READ_ONCE(rsp
->completed
);
1425 if (ULONG_CMP_GE(completed
, gpnum
) ||
1426 ULONG_CMP_LT(j
, js
) ||
1427 ULONG_CMP_GE(gps
, js
))
1428 return; /* No stall or GP completed since entering function. */
1430 if (rcu_gp_in_progress(rsp
) &&
1431 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1433 /* We haven't checked in, so go dump stack. */
1434 print_cpu_stall(rsp
);
1436 } else if (rcu_gp_in_progress(rsp
) &&
1437 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1439 /* They had a few time units to dump stack, so complain. */
1440 print_other_cpu_stall(rsp
, gpnum
);
1445 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1447 * Set the stall-warning timeout way off into the future, thus preventing
1448 * any RCU CPU stall-warning messages from appearing in the current set of
1449 * RCU grace periods.
1451 * The caller must disable hard irqs.
1453 void rcu_cpu_stall_reset(void)
1455 struct rcu_state
*rsp
;
1457 for_each_rcu_flavor(rsp
)
1458 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1462 * Initialize the specified rcu_data structure's default callback list
1463 * to empty. The default callback list is the one that is not used by
1464 * no-callbacks CPUs.
1466 static void init_default_callback_list(struct rcu_data
*rdp
)
1470 rdp
->nxtlist
= NULL
;
1471 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1472 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1476 * Initialize the specified rcu_data structure's callback list to empty.
1478 static void init_callback_list(struct rcu_data
*rdp
)
1480 if (init_nocb_callback_list(rdp
))
1482 init_default_callback_list(rdp
);
1486 * Determine the value that ->completed will have at the end of the
1487 * next subsequent grace period. This is used to tag callbacks so that
1488 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1489 * been dyntick-idle for an extended period with callbacks under the
1490 * influence of RCU_FAST_NO_HZ.
1492 * The caller must hold rnp->lock with interrupts disabled.
1494 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1495 struct rcu_node
*rnp
)
1498 * If RCU is idle, we just wait for the next grace period.
1499 * But we can only be sure that RCU is idle if we are looking
1500 * at the root rcu_node structure -- otherwise, a new grace
1501 * period might have started, but just not yet gotten around
1502 * to initializing the current non-root rcu_node structure.
1504 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1505 return rnp
->completed
+ 1;
1508 * Otherwise, wait for a possible partial grace period and
1509 * then the subsequent full grace period.
1511 return rnp
->completed
+ 2;
1515 * Trace-event helper function for rcu_start_future_gp() and
1516 * rcu_nocb_wait_gp().
1518 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1519 unsigned long c
, const char *s
)
1521 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1522 rnp
->completed
, c
, rnp
->level
,
1523 rnp
->grplo
, rnp
->grphi
, s
);
1527 * Start some future grace period, as needed to handle newly arrived
1528 * callbacks. The required future grace periods are recorded in each
1529 * rcu_node structure's ->need_future_gp field. Returns true if there
1530 * is reason to awaken the grace-period kthread.
1532 * The caller must hold the specified rcu_node structure's ->lock.
1534 static bool __maybe_unused
1535 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1536 unsigned long *c_out
)
1541 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1544 * Pick up grace-period number for new callbacks. If this
1545 * grace period is already marked as needed, return to the caller.
1547 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1548 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1549 if (rnp
->need_future_gp
[c
& 0x1]) {
1550 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1555 * If either this rcu_node structure or the root rcu_node structure
1556 * believe that a grace period is in progress, then we must wait
1557 * for the one following, which is in "c". Because our request
1558 * will be noticed at the end of the current grace period, we don't
1559 * need to explicitly start one. We only do the lockless check
1560 * of rnp_root's fields if the current rcu_node structure thinks
1561 * there is no grace period in flight, and because we hold rnp->lock,
1562 * the only possible change is when rnp_root's two fields are
1563 * equal, in which case rnp_root->gpnum might be concurrently
1564 * incremented. But that is OK, as it will just result in our
1565 * doing some extra useless work.
1567 if (rnp
->gpnum
!= rnp
->completed
||
1568 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1569 rnp
->need_future_gp
[c
& 0x1]++;
1570 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1575 * There might be no grace period in progress. If we don't already
1576 * hold it, acquire the root rcu_node structure's lock in order to
1577 * start one (if needed).
1579 if (rnp
!= rnp_root
)
1580 raw_spin_lock_rcu_node(rnp_root
);
1583 * Get a new grace-period number. If there really is no grace
1584 * period in progress, it will be smaller than the one we obtained
1585 * earlier. Adjust callbacks as needed. Note that even no-CBs
1586 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1588 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1589 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1590 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1591 rdp
->nxtcompleted
[i
] = c
;
1594 * If the needed for the required grace period is already
1595 * recorded, trace and leave.
1597 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1598 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1602 /* Record the need for the future grace period. */
1603 rnp_root
->need_future_gp
[c
& 0x1]++;
1605 /* If a grace period is not already in progress, start one. */
1606 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1607 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1609 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1610 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1613 if (rnp
!= rnp_root
)
1614 raw_spin_unlock_rcu_node(rnp_root
);
1622 * Clean up any old requests for the just-ended grace period. Also return
1623 * whether any additional grace periods have been requested. Also invoke
1624 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1625 * waiting for this grace period to complete.
1627 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1629 int c
= rnp
->completed
;
1631 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1633 rnp
->need_future_gp
[c
& 0x1] = 0;
1634 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1635 trace_rcu_future_gp(rnp
, rdp
, c
,
1636 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1641 * Awaken the grace-period kthread for the specified flavor of RCU.
1642 * Don't do a self-awaken, and don't bother awakening when there is
1643 * nothing for the grace-period kthread to do (as in several CPUs
1644 * raced to awaken, and we lost), and finally don't try to awaken
1645 * a kthread that has not yet been created.
1647 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1649 if (current
== rsp
->gp_kthread
||
1650 !READ_ONCE(rsp
->gp_flags
) ||
1653 swake_up(&rsp
->gp_wq
);
1657 * If there is room, assign a ->completed number to any callbacks on
1658 * this CPU that have not already been assigned. Also accelerate any
1659 * callbacks that were previously assigned a ->completed number that has
1660 * since proven to be too conservative, which can happen if callbacks get
1661 * assigned a ->completed number while RCU is idle, but with reference to
1662 * a non-root rcu_node structure. This function is idempotent, so it does
1663 * not hurt to call it repeatedly. Returns an flag saying that we should
1664 * awaken the RCU grace-period kthread.
1666 * The caller must hold rnp->lock with interrupts disabled.
1668 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1669 struct rcu_data
*rdp
)
1675 /* If the CPU has no callbacks, nothing to do. */
1676 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1680 * Starting from the sublist containing the callbacks most
1681 * recently assigned a ->completed number and working down, find the
1682 * first sublist that is not assignable to an upcoming grace period.
1683 * Such a sublist has something in it (first two tests) and has
1684 * a ->completed number assigned that will complete sooner than
1685 * the ->completed number for newly arrived callbacks (last test).
1687 * The key point is that any later sublist can be assigned the
1688 * same ->completed number as the newly arrived callbacks, which
1689 * means that the callbacks in any of these later sublist can be
1690 * grouped into a single sublist, whether or not they have already
1691 * been assigned a ->completed number.
1693 c
= rcu_cbs_completed(rsp
, rnp
);
1694 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1695 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1696 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1700 * If there are no sublist for unassigned callbacks, leave.
1701 * At the same time, advance "i" one sublist, so that "i" will
1702 * index into the sublist where all the remaining callbacks should
1705 if (++i
>= RCU_NEXT_TAIL
)
1709 * Assign all subsequent callbacks' ->completed number to the next
1710 * full grace period and group them all in the sublist initially
1713 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1714 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1715 rdp
->nxtcompleted
[i
] = c
;
1717 /* Record any needed additional grace periods. */
1718 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1720 /* Trace depending on how much we were able to accelerate. */
1721 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1722 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1724 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1729 * Move any callbacks whose grace period has completed to the
1730 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1731 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1732 * sublist. This function is idempotent, so it does not hurt to
1733 * invoke it repeatedly. As long as it is not invoked -too- often...
1734 * Returns true if the RCU grace-period kthread needs to be awakened.
1736 * The caller must hold rnp->lock with interrupts disabled.
1738 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1739 struct rcu_data
*rdp
)
1743 /* If the CPU has no callbacks, nothing to do. */
1744 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1748 * Find all callbacks whose ->completed numbers indicate that they
1749 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1751 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1752 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1754 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1756 /* Clean up any sublist tail pointers that were misordered above. */
1757 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1758 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1760 /* Copy down callbacks to fill in empty sublists. */
1761 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1762 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1764 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1765 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1768 /* Classify any remaining callbacks. */
1769 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1773 * Update CPU-local rcu_data state to record the beginnings and ends of
1774 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1775 * structure corresponding to the current CPU, and must have irqs disabled.
1776 * Returns true if the grace-period kthread needs to be awakened.
1778 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1779 struct rcu_data
*rdp
)
1783 /* Handle the ends of any preceding grace periods first. */
1784 if (rdp
->completed
== rnp
->completed
&&
1785 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1787 /* No grace period end, so just accelerate recent callbacks. */
1788 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1792 /* Advance callbacks. */
1793 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1795 /* Remember that we saw this grace-period completion. */
1796 rdp
->completed
= rnp
->completed
;
1797 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1800 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1802 * If the current grace period is waiting for this CPU,
1803 * set up to detect a quiescent state, otherwise don't
1804 * go looking for one.
1806 rdp
->gpnum
= rnp
->gpnum
;
1807 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1808 rdp
->cpu_no_qs
.b
.norm
= true;
1809 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1810 rdp
->core_needs_qs
= !!(rnp
->qsmask
& rdp
->grpmask
);
1811 zero_cpu_stall_ticks(rdp
);
1812 WRITE_ONCE(rdp
->gpwrap
, false);
1817 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1819 unsigned long flags
;
1821 struct rcu_node
*rnp
;
1823 local_irq_save(flags
);
1825 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1826 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1827 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1828 !raw_spin_trylock_rcu_node(rnp
)) { /* irqs already off, so later. */
1829 local_irq_restore(flags
);
1832 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1833 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1835 rcu_gp_kthread_wake(rsp
);
1838 static void rcu_gp_slow(struct rcu_state
*rsp
, int delay
)
1841 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1842 schedule_timeout_uninterruptible(delay
);
1846 * Initialize a new grace period. Return false if no grace period required.
1848 static bool rcu_gp_init(struct rcu_state
*rsp
)
1850 unsigned long oldmask
;
1851 struct rcu_data
*rdp
;
1852 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1854 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1855 raw_spin_lock_irq_rcu_node(rnp
);
1856 if (!READ_ONCE(rsp
->gp_flags
)) {
1857 /* Spurious wakeup, tell caller to go back to sleep. */
1858 raw_spin_unlock_irq_rcu_node(rnp
);
1861 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1863 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1865 * Grace period already in progress, don't start another.
1866 * Not supposed to be able to happen.
1868 raw_spin_unlock_irq_rcu_node(rnp
);
1872 /* Advance to a new grace period and initialize state. */
1873 record_gp_stall_check_time(rsp
);
1874 /* Record GP times before starting GP, hence smp_store_release(). */
1875 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1876 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1877 raw_spin_unlock_irq_rcu_node(rnp
);
1880 * Apply per-leaf buffered online and offline operations to the
1881 * rcu_node tree. Note that this new grace period need not wait
1882 * for subsequent online CPUs, and that quiescent-state forcing
1883 * will handle subsequent offline CPUs.
1885 rcu_for_each_leaf_node(rsp
, rnp
) {
1886 rcu_gp_slow(rsp
, gp_preinit_delay
);
1887 raw_spin_lock_irq_rcu_node(rnp
);
1888 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1889 !rnp
->wait_blkd_tasks
) {
1890 /* Nothing to do on this leaf rcu_node structure. */
1891 raw_spin_unlock_irq_rcu_node(rnp
);
1895 /* Record old state, apply changes to ->qsmaskinit field. */
1896 oldmask
= rnp
->qsmaskinit
;
1897 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1899 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1900 if (!oldmask
!= !rnp
->qsmaskinit
) {
1901 if (!oldmask
) /* First online CPU for this rcu_node. */
1902 rcu_init_new_rnp(rnp
);
1903 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1904 rnp
->wait_blkd_tasks
= true;
1905 else /* Last offline CPU and can propagate. */
1906 rcu_cleanup_dead_rnp(rnp
);
1910 * If all waited-on tasks from prior grace period are
1911 * done, and if all this rcu_node structure's CPUs are
1912 * still offline, propagate up the rcu_node tree and
1913 * clear ->wait_blkd_tasks. Otherwise, if one of this
1914 * rcu_node structure's CPUs has since come back online,
1915 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1916 * checks for this, so just call it unconditionally).
1918 if (rnp
->wait_blkd_tasks
&&
1919 (!rcu_preempt_has_tasks(rnp
) ||
1921 rnp
->wait_blkd_tasks
= false;
1922 rcu_cleanup_dead_rnp(rnp
);
1925 raw_spin_unlock_irq_rcu_node(rnp
);
1929 * Set the quiescent-state-needed bits in all the rcu_node
1930 * structures for all currently online CPUs in breadth-first order,
1931 * starting from the root rcu_node structure, relying on the layout
1932 * of the tree within the rsp->node[] array. Note that other CPUs
1933 * will access only the leaves of the hierarchy, thus seeing that no
1934 * grace period is in progress, at least until the corresponding
1935 * leaf node has been initialized. In addition, we have excluded
1936 * CPU-hotplug operations.
1938 * The grace period cannot complete until the initialization
1939 * process finishes, because this kthread handles both.
1941 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1942 rcu_gp_slow(rsp
, gp_init_delay
);
1943 raw_spin_lock_irq_rcu_node(rnp
);
1944 rdp
= this_cpu_ptr(rsp
->rda
);
1945 rcu_preempt_check_blocked_tasks(rnp
);
1946 rnp
->qsmask
= rnp
->qsmaskinit
;
1947 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1948 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1949 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1950 if (rnp
== rdp
->mynode
)
1951 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1952 rcu_preempt_boost_start_gp(rnp
);
1953 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1954 rnp
->level
, rnp
->grplo
,
1955 rnp
->grphi
, rnp
->qsmask
);
1956 raw_spin_unlock_irq_rcu_node(rnp
);
1957 cond_resched_rcu_qs();
1958 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1965 * Helper function for wait_event_interruptible_timeout() wakeup
1966 * at force-quiescent-state time.
1968 static bool rcu_gp_fqs_check_wake(struct rcu_state
*rsp
, int *gfp
)
1970 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1972 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1973 *gfp
= READ_ONCE(rsp
->gp_flags
);
1974 if (*gfp
& RCU_GP_FLAG_FQS
)
1977 /* The current grace period has completed. */
1978 if (!READ_ONCE(rnp
->qsmask
) && !rcu_preempt_blocked_readers_cgp(rnp
))
1985 * Do one round of quiescent-state forcing.
1987 static void rcu_gp_fqs(struct rcu_state
*rsp
, bool first_time
)
1989 bool isidle
= false;
1991 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1993 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1996 /* Collect dyntick-idle snapshots. */
1997 if (is_sysidle_rcu_state(rsp
)) {
1999 maxj
= jiffies
- ULONG_MAX
/ 4;
2001 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
2003 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
2005 /* Handle dyntick-idle and offline CPUs. */
2007 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
2009 /* Clear flag to prevent immediate re-entry. */
2010 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2011 raw_spin_lock_irq_rcu_node(rnp
);
2012 WRITE_ONCE(rsp
->gp_flags
,
2013 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
2014 raw_spin_unlock_irq_rcu_node(rnp
);
2019 * Clean up after the old grace period.
2021 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
2023 unsigned long gp_duration
;
2024 bool needgp
= false;
2026 struct rcu_data
*rdp
;
2027 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2028 struct swait_queue_head
*sq
;
2030 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2031 raw_spin_lock_irq_rcu_node(rnp
);
2032 gp_duration
= jiffies
- rsp
->gp_start
;
2033 if (gp_duration
> rsp
->gp_max
)
2034 rsp
->gp_max
= gp_duration
;
2037 * We know the grace period is complete, but to everyone else
2038 * it appears to still be ongoing. But it is also the case
2039 * that to everyone else it looks like there is nothing that
2040 * they can do to advance the grace period. It is therefore
2041 * safe for us to drop the lock in order to mark the grace
2042 * period as completed in all of the rcu_node structures.
2044 raw_spin_unlock_irq_rcu_node(rnp
);
2047 * Propagate new ->completed value to rcu_node structures so
2048 * that other CPUs don't have to wait until the start of the next
2049 * grace period to process their callbacks. This also avoids
2050 * some nasty RCU grace-period initialization races by forcing
2051 * the end of the current grace period to be completely recorded in
2052 * all of the rcu_node structures before the beginning of the next
2053 * grace period is recorded in any of the rcu_node structures.
2055 rcu_for_each_node_breadth_first(rsp
, rnp
) {
2056 raw_spin_lock_irq_rcu_node(rnp
);
2057 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
2058 WARN_ON_ONCE(rnp
->qsmask
);
2059 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
2060 rdp
= this_cpu_ptr(rsp
->rda
);
2061 if (rnp
== rdp
->mynode
)
2062 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
2063 /* smp_mb() provided by prior unlock-lock pair. */
2064 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
2065 sq
= rcu_nocb_gp_get(rnp
);
2066 raw_spin_unlock_irq_rcu_node(rnp
);
2067 rcu_nocb_gp_cleanup(sq
);
2068 cond_resched_rcu_qs();
2069 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2070 rcu_gp_slow(rsp
, gp_cleanup_delay
);
2072 rnp
= rcu_get_root(rsp
);
2073 raw_spin_lock_irq_rcu_node(rnp
); /* Order GP before ->completed update. */
2074 rcu_nocb_gp_set(rnp
, nocb
);
2076 /* Declare grace period done. */
2077 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
2078 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
2079 rsp
->gp_state
= RCU_GP_IDLE
;
2080 rdp
= this_cpu_ptr(rsp
->rda
);
2081 /* Advance CBs to reduce false positives below. */
2082 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
2083 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
2084 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2085 trace_rcu_grace_period(rsp
->name
,
2086 READ_ONCE(rsp
->gpnum
),
2089 raw_spin_unlock_irq_rcu_node(rnp
);
2093 * Body of kthread that handles grace periods.
2095 static int __noreturn
rcu_gp_kthread(void *arg
)
2101 struct rcu_state
*rsp
= arg
;
2102 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2104 rcu_bind_gp_kthread();
2107 /* Handle grace-period start. */
2109 trace_rcu_grace_period(rsp
->name
,
2110 READ_ONCE(rsp
->gpnum
),
2112 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2113 swait_event_interruptible(rsp
->gp_wq
,
2114 READ_ONCE(rsp
->gp_flags
) &
2116 rsp
->gp_state
= RCU_GP_DONE_GPS
;
2117 /* Locking provides needed memory barrier. */
2118 if (rcu_gp_init(rsp
))
2120 cond_resched_rcu_qs();
2121 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2122 WARN_ON(signal_pending(current
));
2123 trace_rcu_grace_period(rsp
->name
,
2124 READ_ONCE(rsp
->gpnum
),
2128 /* Handle quiescent-state forcing. */
2129 first_gp_fqs
= true;
2130 j
= jiffies_till_first_fqs
;
2133 jiffies_till_first_fqs
= HZ
;
2138 rsp
->jiffies_force_qs
= jiffies
+ j
;
2139 trace_rcu_grace_period(rsp
->name
,
2140 READ_ONCE(rsp
->gpnum
),
2142 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2143 ret
= swait_event_interruptible_timeout(rsp
->gp_wq
,
2144 rcu_gp_fqs_check_wake(rsp
, &gf
), j
);
2145 rsp
->gp_state
= RCU_GP_DOING_FQS
;
2146 /* Locking provides needed memory barriers. */
2147 /* If grace period done, leave loop. */
2148 if (!READ_ONCE(rnp
->qsmask
) &&
2149 !rcu_preempt_blocked_readers_cgp(rnp
))
2151 /* If time for quiescent-state forcing, do it. */
2152 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2153 (gf
& RCU_GP_FLAG_FQS
)) {
2154 trace_rcu_grace_period(rsp
->name
,
2155 READ_ONCE(rsp
->gpnum
),
2157 rcu_gp_fqs(rsp
, first_gp_fqs
);
2158 first_gp_fqs
= false;
2159 trace_rcu_grace_period(rsp
->name
,
2160 READ_ONCE(rsp
->gpnum
),
2162 cond_resched_rcu_qs();
2163 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2165 /* Deal with stray signal. */
2166 cond_resched_rcu_qs();
2167 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2168 WARN_ON(signal_pending(current
));
2169 trace_rcu_grace_period(rsp
->name
,
2170 READ_ONCE(rsp
->gpnum
),
2173 j
= jiffies_till_next_fqs
;
2176 jiffies_till_next_fqs
= HZ
;
2179 jiffies_till_next_fqs
= 1;
2183 /* Handle grace-period end. */
2184 rsp
->gp_state
= RCU_GP_CLEANUP
;
2185 rcu_gp_cleanup(rsp
);
2186 rsp
->gp_state
= RCU_GP_CLEANED
;
2191 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2192 * in preparation for detecting the next grace period. The caller must hold
2193 * the root node's ->lock and hard irqs must be disabled.
2195 * Note that it is legal for a dying CPU (which is marked as offline) to
2196 * invoke this function. This can happen when the dying CPU reports its
2199 * Returns true if the grace-period kthread must be awakened.
2202 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2203 struct rcu_data
*rdp
)
2205 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2207 * Either we have not yet spawned the grace-period
2208 * task, this CPU does not need another grace period,
2209 * or a grace period is already in progress.
2210 * Either way, don't start a new grace period.
2214 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2215 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2219 * We can't do wakeups while holding the rnp->lock, as that
2220 * could cause possible deadlocks with the rq->lock. Defer
2221 * the wakeup to our caller.
2227 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2228 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2229 * is invoked indirectly from rcu_advance_cbs(), which would result in
2230 * endless recursion -- or would do so if it wasn't for the self-deadlock
2231 * that is encountered beforehand.
2233 * Returns true if the grace-period kthread needs to be awakened.
2235 static bool rcu_start_gp(struct rcu_state
*rsp
)
2237 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2238 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2242 * If there is no grace period in progress right now, any
2243 * callbacks we have up to this point will be satisfied by the
2244 * next grace period. Also, advancing the callbacks reduces the
2245 * probability of false positives from cpu_needs_another_gp()
2246 * resulting in pointless grace periods. So, advance callbacks
2247 * then start the grace period!
2249 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2250 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2255 * Report a full set of quiescent states to the specified rcu_state data
2256 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2257 * kthread if another grace period is required. Whether we wake
2258 * the grace-period kthread or it awakens itself for the next round
2259 * of quiescent-state forcing, that kthread will clean up after the
2260 * just-completed grace period. Note that the caller must hold rnp->lock,
2261 * which is released before return.
2263 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2264 __releases(rcu_get_root(rsp
)->lock
)
2266 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2267 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2268 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2269 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2273 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2274 * Allows quiescent states for a group of CPUs to be reported at one go
2275 * to the specified rcu_node structure, though all the CPUs in the group
2276 * must be represented by the same rcu_node structure (which need not be a
2277 * leaf rcu_node structure, though it often will be). The gps parameter
2278 * is the grace-period snapshot, which means that the quiescent states
2279 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2280 * must be held upon entry, and it is released before return.
2283 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2284 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2285 __releases(rnp
->lock
)
2287 unsigned long oldmask
= 0;
2288 struct rcu_node
*rnp_c
;
2290 /* Walk up the rcu_node hierarchy. */
2292 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2295 * Our bit has already been cleared, or the
2296 * relevant grace period is already over, so done.
2298 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2301 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2302 rnp
->qsmask
&= ~mask
;
2303 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2304 mask
, rnp
->qsmask
, rnp
->level
,
2305 rnp
->grplo
, rnp
->grphi
,
2307 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2309 /* Other bits still set at this level, so done. */
2310 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2313 mask
= rnp
->grpmask
;
2314 if (rnp
->parent
== NULL
) {
2316 /* No more levels. Exit loop holding root lock. */
2320 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2323 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2324 oldmask
= rnp_c
->qsmask
;
2328 * Get here if we are the last CPU to pass through a quiescent
2329 * state for this grace period. Invoke rcu_report_qs_rsp()
2330 * to clean up and start the next grace period if one is needed.
2332 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2336 * Record a quiescent state for all tasks that were previously queued
2337 * on the specified rcu_node structure and that were blocking the current
2338 * RCU grace period. The caller must hold the specified rnp->lock with
2339 * irqs disabled, and this lock is released upon return, but irqs remain
2342 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2343 struct rcu_node
*rnp
, unsigned long flags
)
2344 __releases(rnp
->lock
)
2348 struct rcu_node
*rnp_p
;
2350 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2351 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2352 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2353 return; /* Still need more quiescent states! */
2356 rnp_p
= rnp
->parent
;
2357 if (rnp_p
== NULL
) {
2359 * Only one rcu_node structure in the tree, so don't
2360 * try to report up to its nonexistent parent!
2362 rcu_report_qs_rsp(rsp
, flags
);
2366 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2368 mask
= rnp
->grpmask
;
2369 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2370 raw_spin_lock_rcu_node(rnp_p
); /* irqs already disabled. */
2371 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2375 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2376 * structure. This must be called from the specified CPU.
2379 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2381 unsigned long flags
;
2384 struct rcu_node
*rnp
;
2387 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2388 if ((rdp
->cpu_no_qs
.b
.norm
&&
2389 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2390 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2394 * The grace period in which this quiescent state was
2395 * recorded has ended, so don't report it upwards.
2396 * We will instead need a new quiescent state that lies
2397 * within the current grace period.
2399 rdp
->cpu_no_qs
.b
.norm
= true; /* need qs for new gp. */
2400 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2401 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2404 mask
= rdp
->grpmask
;
2405 if ((rnp
->qsmask
& mask
) == 0) {
2406 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2408 rdp
->core_needs_qs
= false;
2411 * This GP can't end until cpu checks in, so all of our
2412 * callbacks can be processed during the next GP.
2414 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2416 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2417 /* ^^^ Released rnp->lock */
2419 rcu_gp_kthread_wake(rsp
);
2424 * Check to see if there is a new grace period of which this CPU
2425 * is not yet aware, and if so, set up local rcu_data state for it.
2426 * Otherwise, see if this CPU has just passed through its first
2427 * quiescent state for this grace period, and record that fact if so.
2430 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2432 /* Check for grace-period ends and beginnings. */
2433 note_gp_changes(rsp
, rdp
);
2436 * Does this CPU still need to do its part for current grace period?
2437 * If no, return and let the other CPUs do their part as well.
2439 if (!rdp
->core_needs_qs
)
2443 * Was there a quiescent state since the beginning of the grace
2444 * period? If no, then exit and wait for the next call.
2446 if (rdp
->cpu_no_qs
.b
.norm
&&
2447 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2451 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2454 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2458 * Send the specified CPU's RCU callbacks to the orphanage. The
2459 * specified CPU must be offline, and the caller must hold the
2463 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2464 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2466 /* No-CBs CPUs do not have orphanable callbacks. */
2467 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) || rcu_is_nocb_cpu(rdp
->cpu
))
2471 * Orphan the callbacks. First adjust the counts. This is safe
2472 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2473 * cannot be running now. Thus no memory barrier is required.
2475 if (rdp
->nxtlist
!= NULL
) {
2476 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2477 rsp
->qlen
+= rdp
->qlen
;
2478 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2480 WRITE_ONCE(rdp
->qlen
, 0);
2484 * Next, move those callbacks still needing a grace period to
2485 * the orphanage, where some other CPU will pick them up.
2486 * Some of the callbacks might have gone partway through a grace
2487 * period, but that is too bad. They get to start over because we
2488 * cannot assume that grace periods are synchronized across CPUs.
2489 * We don't bother updating the ->nxttail[] array yet, instead
2490 * we just reset the whole thing later on.
2492 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2493 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2494 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2495 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2499 * Then move the ready-to-invoke callbacks to the orphanage,
2500 * where some other CPU will pick them up. These will not be
2501 * required to pass though another grace period: They are done.
2503 if (rdp
->nxtlist
!= NULL
) {
2504 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2505 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2509 * Finally, initialize the rcu_data structure's list to empty and
2510 * disallow further callbacks on this CPU.
2512 init_callback_list(rdp
);
2513 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2517 * Adopt the RCU callbacks from the specified rcu_state structure's
2518 * orphanage. The caller must hold the ->orphan_lock.
2520 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2523 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2525 /* No-CBs CPUs are handled specially. */
2526 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2527 rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2530 /* Do the accounting first. */
2531 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2532 rdp
->qlen
+= rsp
->qlen
;
2533 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2534 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2535 rcu_idle_count_callbacks_posted();
2540 * We do not need a memory barrier here because the only way we
2541 * can get here if there is an rcu_barrier() in flight is if
2542 * we are the task doing the rcu_barrier().
2545 /* First adopt the ready-to-invoke callbacks. */
2546 if (rsp
->orphan_donelist
!= NULL
) {
2547 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2548 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2549 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2550 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2551 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2552 rsp
->orphan_donelist
= NULL
;
2553 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2556 /* And then adopt the callbacks that still need a grace period. */
2557 if (rsp
->orphan_nxtlist
!= NULL
) {
2558 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2559 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2560 rsp
->orphan_nxtlist
= NULL
;
2561 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2566 * Trace the fact that this CPU is going offline.
2568 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2570 RCU_TRACE(unsigned long mask
);
2571 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2572 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2574 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2577 RCU_TRACE(mask
= rdp
->grpmask
);
2578 trace_rcu_grace_period(rsp
->name
,
2579 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2584 * All CPUs for the specified rcu_node structure have gone offline,
2585 * and all tasks that were preempted within an RCU read-side critical
2586 * section while running on one of those CPUs have since exited their RCU
2587 * read-side critical section. Some other CPU is reporting this fact with
2588 * the specified rcu_node structure's ->lock held and interrupts disabled.
2589 * This function therefore goes up the tree of rcu_node structures,
2590 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2591 * the leaf rcu_node structure's ->qsmaskinit field has already been
2594 * This function does check that the specified rcu_node structure has
2595 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2596 * prematurely. That said, invoking it after the fact will cost you
2597 * a needless lock acquisition. So once it has done its work, don't
2600 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2603 struct rcu_node
*rnp
= rnp_leaf
;
2605 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2606 rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2609 mask
= rnp
->grpmask
;
2613 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
2614 rnp
->qsmaskinit
&= ~mask
;
2615 rnp
->qsmask
&= ~mask
;
2616 if (rnp
->qsmaskinit
) {
2617 raw_spin_unlock_rcu_node(rnp
);
2618 /* irqs remain disabled. */
2621 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2626 * The CPU has been completely removed, and some other CPU is reporting
2627 * this fact from process context. Do the remainder of the cleanup,
2628 * including orphaning the outgoing CPU's RCU callbacks, and also
2629 * adopting them. There can only be one CPU hotplug operation at a time,
2630 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2632 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2634 unsigned long flags
;
2635 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2636 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2638 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2641 /* Adjust any no-longer-needed kthreads. */
2642 rcu_boost_kthread_setaffinity(rnp
, -1);
2644 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2645 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2646 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2647 rcu_adopt_orphan_cbs(rsp
, flags
);
2648 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2650 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2651 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2652 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2656 * Invoke any RCU callbacks that have made it to the end of their grace
2657 * period. Thottle as specified by rdp->blimit.
2659 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2661 unsigned long flags
;
2662 struct rcu_head
*next
, *list
, **tail
;
2663 long bl
, count
, count_lazy
;
2666 /* If no callbacks are ready, just return. */
2667 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2668 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2669 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2670 need_resched(), is_idle_task(current
),
2671 rcu_is_callbacks_kthread());
2676 * Extract the list of ready callbacks, disabling to prevent
2677 * races with call_rcu() from interrupt handlers.
2679 local_irq_save(flags
);
2680 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2682 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2683 list
= rdp
->nxtlist
;
2684 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2685 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2686 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2687 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2688 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2689 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2690 local_irq_restore(flags
);
2692 /* Invoke callbacks. */
2693 count
= count_lazy
= 0;
2697 debug_rcu_head_unqueue(list
);
2698 if (__rcu_reclaim(rsp
->name
, list
))
2701 /* Stop only if limit reached and CPU has something to do. */
2702 if (++count
>= bl
&&
2704 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2708 local_irq_save(flags
);
2709 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2710 is_idle_task(current
),
2711 rcu_is_callbacks_kthread());
2713 /* Update count, and requeue any remaining callbacks. */
2715 *tail
= rdp
->nxtlist
;
2716 rdp
->nxtlist
= list
;
2717 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2718 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2719 rdp
->nxttail
[i
] = tail
;
2723 smp_mb(); /* List handling before counting for rcu_barrier(). */
2724 rdp
->qlen_lazy
-= count_lazy
;
2725 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2726 rdp
->n_cbs_invoked
+= count
;
2728 /* Reinstate batch limit if we have worked down the excess. */
2729 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2730 rdp
->blimit
= blimit
;
2732 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2733 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2734 rdp
->qlen_last_fqs_check
= 0;
2735 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2736 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2737 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2738 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2740 local_irq_restore(flags
);
2742 /* Re-invoke RCU core processing if there are callbacks remaining. */
2743 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2748 * Check to see if this CPU is in a non-context-switch quiescent state
2749 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2750 * Also schedule RCU core processing.
2752 * This function must be called from hardirq context. It is normally
2753 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2754 * false, there is no point in invoking rcu_check_callbacks().
2756 void rcu_check_callbacks(int user
)
2758 trace_rcu_utilization(TPS("Start scheduler-tick"));
2759 increment_cpu_stall_ticks();
2760 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2763 * Get here if this CPU took its interrupt from user
2764 * mode or from the idle loop, and if this is not a
2765 * nested interrupt. In this case, the CPU is in
2766 * a quiescent state, so note it.
2768 * No memory barrier is required here because both
2769 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2770 * variables that other CPUs neither access nor modify,
2771 * at least not while the corresponding CPU is online.
2777 } else if (!in_softirq()) {
2780 * Get here if this CPU did not take its interrupt from
2781 * softirq, in other words, if it is not interrupting
2782 * a rcu_bh read-side critical section. This is an _bh
2783 * critical section, so note it.
2788 rcu_preempt_check_callbacks();
2792 rcu_note_voluntary_context_switch(current
);
2793 trace_rcu_utilization(TPS("End scheduler-tick"));
2797 * Scan the leaf rcu_node structures, processing dyntick state for any that
2798 * have not yet encountered a quiescent state, using the function specified.
2799 * Also initiate boosting for any threads blocked on the root rcu_node.
2801 * The caller must have suppressed start of new grace periods.
2803 static void force_qs_rnp(struct rcu_state
*rsp
,
2804 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2805 unsigned long *maxj
),
2806 bool *isidle
, unsigned long *maxj
)
2810 unsigned long flags
;
2812 struct rcu_node
*rnp
;
2814 rcu_for_each_leaf_node(rsp
, rnp
) {
2815 cond_resched_rcu_qs();
2817 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2818 if (rnp
->qsmask
== 0) {
2819 if (rcu_state_p
== &rcu_sched_state
||
2820 rsp
!= rcu_state_p
||
2821 rcu_preempt_blocked_readers_cgp(rnp
)) {
2823 * No point in scanning bits because they
2824 * are all zero. But we might need to
2825 * priority-boost blocked readers.
2827 rcu_initiate_boost(rnp
, flags
);
2828 /* rcu_initiate_boost() releases rnp->lock */
2832 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2834 * Race between grace-period
2835 * initialization and task exiting RCU
2836 * read-side critical section: Report.
2838 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2839 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2845 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2846 if ((rnp
->qsmask
& bit
) != 0) {
2847 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2852 /* Idle/offline CPUs, report (releases rnp->lock. */
2853 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2855 /* Nothing to do here, so just drop the lock. */
2856 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2862 * Force quiescent states on reluctant CPUs, and also detect which
2863 * CPUs are in dyntick-idle mode.
2865 static void force_quiescent_state(struct rcu_state
*rsp
)
2867 unsigned long flags
;
2869 struct rcu_node
*rnp
;
2870 struct rcu_node
*rnp_old
= NULL
;
2872 /* Funnel through hierarchy to reduce memory contention. */
2873 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2874 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2875 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2876 !raw_spin_trylock(&rnp
->fqslock
);
2877 if (rnp_old
!= NULL
)
2878 raw_spin_unlock(&rnp_old
->fqslock
);
2880 rsp
->n_force_qs_lh
++;
2885 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2887 /* Reached the root of the rcu_node tree, acquire lock. */
2888 raw_spin_lock_irqsave_rcu_node(rnp_old
, flags
);
2889 raw_spin_unlock(&rnp_old
->fqslock
);
2890 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2891 rsp
->n_force_qs_lh
++;
2892 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2893 return; /* Someone beat us to it. */
2895 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2896 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2897 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2901 * This does the RCU core processing work for the specified rcu_state
2902 * and rcu_data structures. This may be called only from the CPU to
2903 * whom the rdp belongs.
2906 __rcu_process_callbacks(struct rcu_state
*rsp
)
2908 unsigned long flags
;
2910 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2912 WARN_ON_ONCE(rdp
->beenonline
== 0);
2914 /* Update RCU state based on any recent quiescent states. */
2915 rcu_check_quiescent_state(rsp
, rdp
);
2917 /* Does this CPU require a not-yet-started grace period? */
2918 local_irq_save(flags
);
2919 if (cpu_needs_another_gp(rsp
, rdp
)) {
2920 raw_spin_lock_rcu_node(rcu_get_root(rsp
)); /* irqs disabled. */
2921 needwake
= rcu_start_gp(rsp
);
2922 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2924 rcu_gp_kthread_wake(rsp
);
2926 local_irq_restore(flags
);
2929 /* If there are callbacks ready, invoke them. */
2930 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2931 invoke_rcu_callbacks(rsp
, rdp
);
2933 /* Do any needed deferred wakeups of rcuo kthreads. */
2934 do_nocb_deferred_wakeup(rdp
);
2938 * Do RCU core processing for the current CPU.
2940 static void rcu_process_callbacks(struct softirq_action
*unused
)
2942 struct rcu_state
*rsp
;
2944 if (cpu_is_offline(smp_processor_id()))
2946 trace_rcu_utilization(TPS("Start RCU core"));
2947 for_each_rcu_flavor(rsp
)
2948 __rcu_process_callbacks(rsp
);
2949 trace_rcu_utilization(TPS("End RCU core"));
2953 * Schedule RCU callback invocation. If the specified type of RCU
2954 * does not support RCU priority boosting, just do a direct call,
2955 * otherwise wake up the per-CPU kernel kthread. Note that because we
2956 * are running on the current CPU with softirqs disabled, the
2957 * rcu_cpu_kthread_task cannot disappear out from under us.
2959 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2961 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2963 if (likely(!rsp
->boost
)) {
2964 rcu_do_batch(rsp
, rdp
);
2967 invoke_rcu_callbacks_kthread();
2970 static void invoke_rcu_core(void)
2972 if (cpu_online(smp_processor_id()))
2973 raise_softirq(RCU_SOFTIRQ
);
2977 * Handle any core-RCU processing required by a call_rcu() invocation.
2979 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2980 struct rcu_head
*head
, unsigned long flags
)
2985 * If called from an extended quiescent state, invoke the RCU
2986 * core in order to force a re-evaluation of RCU's idleness.
2988 if (!rcu_is_watching())
2991 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2992 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2996 * Force the grace period if too many callbacks or too long waiting.
2997 * Enforce hysteresis, and don't invoke force_quiescent_state()
2998 * if some other CPU has recently done so. Also, don't bother
2999 * invoking force_quiescent_state() if the newly enqueued callback
3000 * is the only one waiting for a grace period to complete.
3002 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
3004 /* Are we ignoring a completed grace period? */
3005 note_gp_changes(rsp
, rdp
);
3007 /* Start a new grace period if one not already started. */
3008 if (!rcu_gp_in_progress(rsp
)) {
3009 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3011 raw_spin_lock_rcu_node(rnp_root
);
3012 needwake
= rcu_start_gp(rsp
);
3013 raw_spin_unlock_rcu_node(rnp_root
);
3015 rcu_gp_kthread_wake(rsp
);
3017 /* Give the grace period a kick. */
3018 rdp
->blimit
= LONG_MAX
;
3019 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
3020 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
3021 force_quiescent_state(rsp
);
3022 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3023 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
3029 * RCU callback function to leak a callback.
3031 static void rcu_leak_callback(struct rcu_head
*rhp
)
3036 * Helper function for call_rcu() and friends. The cpu argument will
3037 * normally be -1, indicating "currently running CPU". It may specify
3038 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3039 * is expected to specify a CPU.
3042 __call_rcu(struct rcu_head
*head
, rcu_callback_t func
,
3043 struct rcu_state
*rsp
, int cpu
, bool lazy
)
3045 unsigned long flags
;
3046 struct rcu_data
*rdp
;
3048 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
3049 if (debug_rcu_head_queue(head
)) {
3050 /* Probable double call_rcu(), so leak the callback. */
3051 WRITE_ONCE(head
->func
, rcu_leak_callback
);
3052 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3059 * Opportunistically note grace-period endings and beginnings.
3060 * Note that we might see a beginning right after we see an
3061 * end, but never vice versa, since this CPU has to pass through
3062 * a quiescent state betweentimes.
3064 local_irq_save(flags
);
3065 rdp
= this_cpu_ptr(rsp
->rda
);
3067 /* Add the callback to our list. */
3068 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3072 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3073 if (likely(rdp
->mynode
)) {
3074 /* Post-boot, so this should be for a no-CBs CPU. */
3075 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3076 WARN_ON_ONCE(offline
);
3077 /* Offline CPU, _call_rcu() illegal, leak callback. */
3078 local_irq_restore(flags
);
3082 * Very early boot, before rcu_init(). Initialize if needed
3083 * and then drop through to queue the callback.
3086 WARN_ON_ONCE(!rcu_is_watching());
3087 if (!likely(rdp
->nxtlist
))
3088 init_default_callback_list(rdp
);
3090 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3094 rcu_idle_count_callbacks_posted();
3095 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3096 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3097 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3099 if (__is_kfree_rcu_offset((unsigned long)func
))
3100 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3101 rdp
->qlen_lazy
, rdp
->qlen
);
3103 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3105 /* Go handle any RCU core processing required. */
3106 __call_rcu_core(rsp
, rdp
, head
, flags
);
3107 local_irq_restore(flags
);
3111 * Queue an RCU-sched callback for invocation after a grace period.
3113 void call_rcu_sched(struct rcu_head
*head
, rcu_callback_t func
)
3115 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3117 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3120 * Queue an RCU callback for invocation after a quicker grace period.
3122 void call_rcu_bh(struct rcu_head
*head
, rcu_callback_t func
)
3124 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3126 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3129 * Queue an RCU callback for lazy invocation after a grace period.
3130 * This will likely be later named something like "call_rcu_lazy()",
3131 * but this change will require some way of tagging the lazy RCU
3132 * callbacks in the list of pending callbacks. Until then, this
3133 * function may only be called from __kfree_rcu().
3135 void kfree_call_rcu(struct rcu_head
*head
,
3136 rcu_callback_t func
)
3138 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3140 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3143 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3144 * any blocking grace-period wait automatically implies a grace period
3145 * if there is only one CPU online at any point time during execution
3146 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3147 * occasionally incorrectly indicate that there are multiple CPUs online
3148 * when there was in fact only one the whole time, as this just adds
3149 * some overhead: RCU still operates correctly.
3151 static inline int rcu_blocking_is_gp(void)
3155 might_sleep(); /* Check for RCU read-side critical section. */
3157 ret
= num_online_cpus() <= 1;
3163 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3165 * Control will return to the caller some time after a full rcu-sched
3166 * grace period has elapsed, in other words after all currently executing
3167 * rcu-sched read-side critical sections have completed. These read-side
3168 * critical sections are delimited by rcu_read_lock_sched() and
3169 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3170 * local_irq_disable(), and so on may be used in place of
3171 * rcu_read_lock_sched().
3173 * This means that all preempt_disable code sequences, including NMI and
3174 * non-threaded hardware-interrupt handlers, in progress on entry will
3175 * have completed before this primitive returns. However, this does not
3176 * guarantee that softirq handlers will have completed, since in some
3177 * kernels, these handlers can run in process context, and can block.
3179 * Note that this guarantee implies further memory-ordering guarantees.
3180 * On systems with more than one CPU, when synchronize_sched() returns,
3181 * each CPU is guaranteed to have executed a full memory barrier since the
3182 * end of its last RCU-sched read-side critical section whose beginning
3183 * preceded the call to synchronize_sched(). In addition, each CPU having
3184 * an RCU read-side critical section that extends beyond the return from
3185 * synchronize_sched() is guaranteed to have executed a full memory barrier
3186 * after the beginning of synchronize_sched() and before the beginning of
3187 * that RCU read-side critical section. Note that these guarantees include
3188 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3189 * that are executing in the kernel.
3191 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3192 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3193 * to have executed a full memory barrier during the execution of
3194 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3195 * again only if the system has more than one CPU).
3197 * This primitive provides the guarantees made by the (now removed)
3198 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3199 * guarantees that rcu_read_lock() sections will have completed.
3200 * In "classic RCU", these two guarantees happen to be one and
3201 * the same, but can differ in realtime RCU implementations.
3203 void synchronize_sched(void)
3205 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3206 lock_is_held(&rcu_lock_map
) ||
3207 lock_is_held(&rcu_sched_lock_map
),
3208 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3209 if (rcu_blocking_is_gp())
3211 if (rcu_gp_is_expedited())
3212 synchronize_sched_expedited();
3214 wait_rcu_gp(call_rcu_sched
);
3216 EXPORT_SYMBOL_GPL(synchronize_sched
);
3219 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3221 * Control will return to the caller some time after a full rcu_bh grace
3222 * period has elapsed, in other words after all currently executing rcu_bh
3223 * read-side critical sections have completed. RCU read-side critical
3224 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3225 * and may be nested.
3227 * See the description of synchronize_sched() for more detailed information
3228 * on memory ordering guarantees.
3230 void synchronize_rcu_bh(void)
3232 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3233 lock_is_held(&rcu_lock_map
) ||
3234 lock_is_held(&rcu_sched_lock_map
),
3235 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3236 if (rcu_blocking_is_gp())
3238 if (rcu_gp_is_expedited())
3239 synchronize_rcu_bh_expedited();
3241 wait_rcu_gp(call_rcu_bh
);
3243 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3246 * get_state_synchronize_rcu - Snapshot current RCU state
3248 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3249 * to determine whether or not a full grace period has elapsed in the
3252 unsigned long get_state_synchronize_rcu(void)
3255 * Any prior manipulation of RCU-protected data must happen
3256 * before the load from ->gpnum.
3261 * Make sure this load happens before the purportedly
3262 * time-consuming work between get_state_synchronize_rcu()
3263 * and cond_synchronize_rcu().
3265 return smp_load_acquire(&rcu_state_p
->gpnum
);
3267 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3270 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3272 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3274 * If a full RCU grace period has elapsed since the earlier call to
3275 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3276 * synchronize_rcu() to wait for a full grace period.
3278 * Yes, this function does not take counter wrap into account. But
3279 * counter wrap is harmless. If the counter wraps, we have waited for
3280 * more than 2 billion grace periods (and way more on a 64-bit system!),
3281 * so waiting for one additional grace period should be just fine.
3283 void cond_synchronize_rcu(unsigned long oldstate
)
3285 unsigned long newstate
;
3288 * Ensure that this load happens before any RCU-destructive
3289 * actions the caller might carry out after we return.
3291 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3292 if (ULONG_CMP_GE(oldstate
, newstate
))
3295 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3298 * get_state_synchronize_sched - Snapshot current RCU-sched state
3300 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3301 * to determine whether or not a full grace period has elapsed in the
3304 unsigned long get_state_synchronize_sched(void)
3307 * Any prior manipulation of RCU-protected data must happen
3308 * before the load from ->gpnum.
3313 * Make sure this load happens before the purportedly
3314 * time-consuming work between get_state_synchronize_sched()
3315 * and cond_synchronize_sched().
3317 return smp_load_acquire(&rcu_sched_state
.gpnum
);
3319 EXPORT_SYMBOL_GPL(get_state_synchronize_sched
);
3322 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3324 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3326 * If a full RCU-sched grace period has elapsed since the earlier call to
3327 * get_state_synchronize_sched(), just return. Otherwise, invoke
3328 * synchronize_sched() to wait for a full grace period.
3330 * Yes, this function does not take counter wrap into account. But
3331 * counter wrap is harmless. If the counter wraps, we have waited for
3332 * more than 2 billion grace periods (and way more on a 64-bit system!),
3333 * so waiting for one additional grace period should be just fine.
3335 void cond_synchronize_sched(unsigned long oldstate
)
3337 unsigned long newstate
;
3340 * Ensure that this load happens before any RCU-destructive
3341 * actions the caller might carry out after we return.
3343 newstate
= smp_load_acquire(&rcu_sched_state
.completed
);
3344 if (ULONG_CMP_GE(oldstate
, newstate
))
3345 synchronize_sched();
3347 EXPORT_SYMBOL_GPL(cond_synchronize_sched
);
3349 /* Adjust sequence number for start of update-side operation. */
3350 static void rcu_seq_start(unsigned long *sp
)
3352 WRITE_ONCE(*sp
, *sp
+ 1);
3353 smp_mb(); /* Ensure update-side operation after counter increment. */
3354 WARN_ON_ONCE(!(*sp
& 0x1));
3357 /* Adjust sequence number for end of update-side operation. */
3358 static void rcu_seq_end(unsigned long *sp
)
3360 smp_mb(); /* Ensure update-side operation before counter increment. */
3361 WRITE_ONCE(*sp
, *sp
+ 1);
3362 WARN_ON_ONCE(*sp
& 0x1);
3365 /* Take a snapshot of the update side's sequence number. */
3366 static unsigned long rcu_seq_snap(unsigned long *sp
)
3370 s
= (READ_ONCE(*sp
) + 3) & ~0x1;
3371 smp_mb(); /* Above access must not bleed into critical section. */
3376 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3377 * full update-side operation has occurred.
3379 static bool rcu_seq_done(unsigned long *sp
, unsigned long s
)
3381 return ULONG_CMP_GE(READ_ONCE(*sp
), s
);
3384 /* Wrapper functions for expedited grace periods. */
3385 static void rcu_exp_gp_seq_start(struct rcu_state
*rsp
)
3387 rcu_seq_start(&rsp
->expedited_sequence
);
3389 static void rcu_exp_gp_seq_end(struct rcu_state
*rsp
)
3391 rcu_seq_end(&rsp
->expedited_sequence
);
3392 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3394 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state
*rsp
)
3398 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3399 s
= rcu_seq_snap(&rsp
->expedited_sequence
);
3400 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("snap"));
3403 static bool rcu_exp_gp_seq_done(struct rcu_state
*rsp
, unsigned long s
)
3405 return rcu_seq_done(&rsp
->expedited_sequence
, s
);
3409 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3410 * recent CPU-online activity. Note that these masks are not cleared
3411 * when CPUs go offline, so they reflect the union of all CPUs that have
3412 * ever been online. This means that this function normally takes its
3413 * no-work-to-do fastpath.
3415 static void sync_exp_reset_tree_hotplug(struct rcu_state
*rsp
)
3418 unsigned long flags
;
3420 unsigned long oldmask
;
3421 int ncpus
= READ_ONCE(rsp
->ncpus
);
3422 struct rcu_node
*rnp
;
3423 struct rcu_node
*rnp_up
;
3425 /* If no new CPUs onlined since last time, nothing to do. */
3426 if (likely(ncpus
== rsp
->ncpus_snap
))
3428 rsp
->ncpus_snap
= ncpus
;
3431 * Each pass through the following loop propagates newly onlined
3432 * CPUs for the current rcu_node structure up the rcu_node tree.
3434 rcu_for_each_leaf_node(rsp
, rnp
) {
3435 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3436 if (rnp
->expmaskinit
== rnp
->expmaskinitnext
) {
3437 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3438 continue; /* No new CPUs, nothing to do. */
3441 /* Update this node's mask, track old value for propagation. */
3442 oldmask
= rnp
->expmaskinit
;
3443 rnp
->expmaskinit
= rnp
->expmaskinitnext
;
3444 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3446 /* If was already nonzero, nothing to propagate. */
3450 /* Propagate the new CPU up the tree. */
3451 mask
= rnp
->grpmask
;
3452 rnp_up
= rnp
->parent
;
3455 raw_spin_lock_irqsave_rcu_node(rnp_up
, flags
);
3456 if (rnp_up
->expmaskinit
)
3458 rnp_up
->expmaskinit
|= mask
;
3459 raw_spin_unlock_irqrestore_rcu_node(rnp_up
, flags
);
3462 mask
= rnp_up
->grpmask
;
3463 rnp_up
= rnp_up
->parent
;
3469 * Reset the ->expmask values in the rcu_node tree in preparation for
3470 * a new expedited grace period.
3472 static void __maybe_unused
sync_exp_reset_tree(struct rcu_state
*rsp
)
3474 unsigned long flags
;
3475 struct rcu_node
*rnp
;
3477 sync_exp_reset_tree_hotplug(rsp
);
3478 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3479 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3480 WARN_ON_ONCE(rnp
->expmask
);
3481 rnp
->expmask
= rnp
->expmaskinit
;
3482 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3487 * Return non-zero if there is no RCU expedited grace period in progress
3488 * for the specified rcu_node structure, in other words, if all CPUs and
3489 * tasks covered by the specified rcu_node structure have done their bit
3490 * for the current expedited grace period. Works only for preemptible
3491 * RCU -- other RCU implementation use other means.
3493 * Caller must hold the rcu_state's exp_mutex.
3495 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
3497 return rnp
->exp_tasks
== NULL
&&
3498 READ_ONCE(rnp
->expmask
) == 0;
3502 * Report the exit from RCU read-side critical section for the last task
3503 * that queued itself during or before the current expedited preemptible-RCU
3504 * grace period. This event is reported either to the rcu_node structure on
3505 * which the task was queued or to one of that rcu_node structure's ancestors,
3506 * recursively up the tree. (Calm down, calm down, we do the recursion
3509 * Caller must hold the rcu_state's exp_mutex and the specified rcu_node
3510 * structure's ->lock.
3512 static void __rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3513 bool wake
, unsigned long flags
)
3514 __releases(rnp
->lock
)
3519 if (!sync_rcu_preempt_exp_done(rnp
)) {
3521 rcu_initiate_boost(rnp
, flags
);
3523 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3526 if (rnp
->parent
== NULL
) {
3527 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3529 smp_mb(); /* EGP done before wake_up(). */
3530 swake_up(&rsp
->expedited_wq
);
3534 mask
= rnp
->grpmask
;
3535 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled */
3537 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled */
3538 WARN_ON_ONCE(!(rnp
->expmask
& mask
));
3539 rnp
->expmask
&= ~mask
;
3544 * Report expedited quiescent state for specified node. This is a
3545 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3547 * Caller must hold the rcu_state's exp_mutex.
3549 static void __maybe_unused
rcu_report_exp_rnp(struct rcu_state
*rsp
,
3550 struct rcu_node
*rnp
, bool wake
)
3552 unsigned long flags
;
3554 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3555 __rcu_report_exp_rnp(rsp
, rnp
, wake
, flags
);
3559 * Report expedited quiescent state for multiple CPUs, all covered by the
3560 * specified leaf rcu_node structure. Caller must hold the rcu_state's
3563 static void rcu_report_exp_cpu_mult(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3564 unsigned long mask
, bool wake
)
3566 unsigned long flags
;
3568 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3569 if (!(rnp
->expmask
& mask
)) {
3570 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3573 rnp
->expmask
&= ~mask
;
3574 __rcu_report_exp_rnp(rsp
, rnp
, wake
, flags
); /* Releases rnp->lock. */
3578 * Report expedited quiescent state for specified rcu_data (CPU).
3580 static void rcu_report_exp_rdp(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
3583 rcu_report_exp_cpu_mult(rsp
, rdp
->mynode
, rdp
->grpmask
, wake
);
3586 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3587 static bool sync_exp_work_done(struct rcu_state
*rsp
, atomic_long_t
*stat
,
3590 if (rcu_exp_gp_seq_done(rsp
, s
)) {
3591 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("done"));
3592 /* Ensure test happens before caller kfree(). */
3593 smp_mb__before_atomic(); /* ^^^ */
3594 atomic_long_inc(stat
);
3601 * Funnel-lock acquisition for expedited grace periods. Returns true
3602 * if some other task completed an expedited grace period that this task
3603 * can piggy-back on, and with no mutex held. Otherwise, returns false
3604 * with the mutex held, indicating that the caller must actually do the
3605 * expedited grace period.
3607 static bool exp_funnel_lock(struct rcu_state
*rsp
, unsigned long s
)
3609 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id());
3610 struct rcu_node
*rnp
= rdp
->mynode
;
3611 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3613 /* Low-contention fastpath. */
3614 if (ULONG_CMP_LT(READ_ONCE(rnp
->exp_seq_rq
), s
) &&
3616 ULONG_CMP_LT(READ_ONCE(rnp_root
->exp_seq_rq
), s
)) &&
3617 !mutex_is_locked(&rsp
->exp_mutex
) &&
3618 mutex_trylock(&rsp
->exp_mutex
))
3622 * Each pass through the following loop works its way up
3623 * the rcu_node tree, returning if others have done the work or
3624 * otherwise falls through to acquire rsp->exp_mutex. The mapping
3625 * from CPU to rcu_node structure can be inexact, as it is just
3626 * promoting locality and is not strictly needed for correctness.
3628 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
3629 if (sync_exp_work_done(rsp
, &rdp
->exp_workdone1
, s
))
3632 /* Work not done, either wait here or go up. */
3633 spin_lock(&rnp
->exp_lock
);
3634 if (ULONG_CMP_GE(rnp
->exp_seq_rq
, s
)) {
3636 /* Someone else doing GP, so wait for them. */
3637 spin_unlock(&rnp
->exp_lock
);
3638 trace_rcu_exp_funnel_lock(rsp
->name
, rnp
->level
,
3639 rnp
->grplo
, rnp
->grphi
,
3641 wait_event(rnp
->exp_wq
[(s
>> 1) & 0x3],
3642 sync_exp_work_done(rsp
,
3643 &rdp
->exp_workdone2
, s
));
3646 rnp
->exp_seq_rq
= s
; /* Followers can wait on us. */
3647 spin_unlock(&rnp
->exp_lock
);
3648 trace_rcu_exp_funnel_lock(rsp
->name
, rnp
->level
, rnp
->grplo
,
3649 rnp
->grphi
, TPS("nxtlvl"));
3651 mutex_lock(&rsp
->exp_mutex
);
3653 if (sync_exp_work_done(rsp
, &rdp
->exp_workdone3
, s
)) {
3654 mutex_unlock(&rsp
->exp_mutex
);
3657 rcu_exp_gp_seq_start(rsp
);
3658 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("start"));
3662 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3663 static void sync_sched_exp_handler(void *data
)
3665 struct rcu_data
*rdp
;
3666 struct rcu_node
*rnp
;
3667 struct rcu_state
*rsp
= data
;
3669 rdp
= this_cpu_ptr(rsp
->rda
);
3671 if (!(READ_ONCE(rnp
->expmask
) & rdp
->grpmask
) ||
3672 __this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
3674 if (rcu_is_cpu_rrupt_from_idle()) {
3675 rcu_report_exp_rdp(&rcu_sched_state
,
3676 this_cpu_ptr(&rcu_sched_data
), true);
3679 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, true);
3680 resched_cpu(smp_processor_id());
3683 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
3684 static void sync_sched_exp_online_cleanup(int cpu
)
3686 struct rcu_data
*rdp
;
3688 struct rcu_node
*rnp
;
3689 struct rcu_state
*rsp
= &rcu_sched_state
;
3691 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3693 if (!(READ_ONCE(rnp
->expmask
) & rdp
->grpmask
))
3695 ret
= smp_call_function_single(cpu
, sync_sched_exp_handler
, rsp
, 0);
3700 * Select the nodes that the upcoming expedited grace period needs
3703 static void sync_rcu_exp_select_cpus(struct rcu_state
*rsp
,
3704 smp_call_func_t func
)
3707 unsigned long flags
;
3709 unsigned long mask_ofl_test
;
3710 unsigned long mask_ofl_ipi
;
3712 struct rcu_node
*rnp
;
3714 sync_exp_reset_tree(rsp
);
3715 rcu_for_each_leaf_node(rsp
, rnp
) {
3716 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3718 /* Each pass checks a CPU for identity, offline, and idle. */
3720 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++) {
3721 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3722 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3724 if (raw_smp_processor_id() == cpu
||
3725 !(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3726 mask_ofl_test
|= rdp
->grpmask
;
3728 mask_ofl_ipi
= rnp
->expmask
& ~mask_ofl_test
;
3731 * Need to wait for any blocked tasks as well. Note that
3732 * additional blocking tasks will also block the expedited
3733 * GP until such time as the ->expmask bits are cleared.
3735 if (rcu_preempt_has_tasks(rnp
))
3736 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
3737 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3739 /* IPI the remaining CPUs for expedited quiescent state. */
3741 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3742 if (!(mask_ofl_ipi
& mask
))
3745 ret
= smp_call_function_single(cpu
, func
, rsp
, 0);
3747 mask_ofl_ipi
&= ~mask
;
3750 /* Failed, raced with offline. */
3751 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3752 if (cpu_online(cpu
) &&
3753 (rnp
->expmask
& mask
)) {
3754 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3755 schedule_timeout_uninterruptible(1);
3756 if (cpu_online(cpu
) &&
3757 (rnp
->expmask
& mask
))
3759 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3761 if (!(rnp
->expmask
& mask
))
3762 mask_ofl_ipi
&= ~mask
;
3763 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3765 /* Report quiescent states for those that went offline. */
3766 mask_ofl_test
|= mask_ofl_ipi
;
3768 rcu_report_exp_cpu_mult(rsp
, rnp
, mask_ofl_test
, false);
3772 static void synchronize_sched_expedited_wait(struct rcu_state
*rsp
)
3775 unsigned long jiffies_stall
;
3776 unsigned long jiffies_start
;
3779 struct rcu_node
*rnp
;
3780 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3783 jiffies_stall
= rcu_jiffies_till_stall_check();
3784 jiffies_start
= jiffies
;
3787 ret
= swait_event_timeout(
3789 sync_rcu_preempt_exp_done(rnp_root
),
3791 if (ret
> 0 || sync_rcu_preempt_exp_done(rnp_root
))
3794 /* Hit a signal, disable CPU stall warnings. */
3795 swait_event(rsp
->expedited_wq
,
3796 sync_rcu_preempt_exp_done(rnp_root
));
3799 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3802 rcu_for_each_leaf_node(rsp
, rnp
) {
3803 ndetected
+= rcu_print_task_exp_stall(rnp
);
3805 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3806 struct rcu_data
*rdp
;
3808 if (!(rnp
->expmask
& mask
))
3811 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3812 pr_cont(" %d-%c%c%c", cpu
,
3813 "O."[!!cpu_online(cpu
)],
3814 "o."[!!(rdp
->grpmask
& rnp
->expmaskinit
)],
3815 "N."[!!(rdp
->grpmask
& rnp
->expmaskinitnext
)]);
3819 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
3820 jiffies
- jiffies_start
, rsp
->expedited_sequence
,
3821 rnp_root
->expmask
, ".T"[!!rnp_root
->exp_tasks
]);
3823 pr_err("blocking rcu_node structures:");
3824 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3825 if (rnp
== rnp_root
)
3826 continue; /* printed unconditionally */
3827 if (sync_rcu_preempt_exp_done(rnp
))
3829 pr_cont(" l=%u:%d-%d:%#lx/%c",
3830 rnp
->level
, rnp
->grplo
, rnp
->grphi
,
3832 ".T"[!!rnp
->exp_tasks
]);
3836 rcu_for_each_leaf_node(rsp
, rnp
) {
3838 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3839 if (!(rnp
->expmask
& mask
))
3844 jiffies_stall
= 3 * rcu_jiffies_till_stall_check() + 3;
3849 * Wait for the current expedited grace period to complete, and then
3850 * wake up everyone who piggybacked on the just-completed expedited
3851 * grace period. Also update all the ->exp_seq_rq counters as needed
3852 * in order to avoid counter-wrap problems.
3854 static void rcu_exp_wait_wake(struct rcu_state
*rsp
, unsigned long s
)
3856 struct rcu_node
*rnp
;
3858 synchronize_sched_expedited_wait(rsp
);
3859 rcu_exp_gp_seq_end(rsp
);
3860 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("end"));
3863 * Switch over to wakeup mode, allowing the next GP, but -only- the
3864 * next GP, to proceed.
3866 mutex_lock(&rsp
->exp_wake_mutex
);
3867 mutex_unlock(&rsp
->exp_mutex
);
3869 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3870 if (ULONG_CMP_LT(READ_ONCE(rnp
->exp_seq_rq
), s
)) {
3871 spin_lock(&rnp
->exp_lock
);
3872 /* Recheck, avoid hang in case someone just arrived. */
3873 if (ULONG_CMP_LT(rnp
->exp_seq_rq
, s
))
3874 rnp
->exp_seq_rq
= s
;
3875 spin_unlock(&rnp
->exp_lock
);
3877 wake_up_all(&rnp
->exp_wq
[(rsp
->expedited_sequence
>> 1) & 0x3]);
3879 trace_rcu_exp_grace_period(rsp
->name
, s
, TPS("endwake"));
3880 mutex_unlock(&rsp
->exp_wake_mutex
);
3884 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3886 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3887 * approach to force the grace period to end quickly. This consumes
3888 * significant time on all CPUs and is unfriendly to real-time workloads,
3889 * so is thus not recommended for any sort of common-case code. In fact,
3890 * if you are using synchronize_sched_expedited() in a loop, please
3891 * restructure your code to batch your updates, and then use a single
3892 * synchronize_sched() instead.
3894 * This implementation can be thought of as an application of sequence
3895 * locking to expedited grace periods, but using the sequence counter to
3896 * determine when someone else has already done the work instead of for
3899 void synchronize_sched_expedited(void)
3902 struct rcu_state
*rsp
= &rcu_sched_state
;
3904 /* If only one CPU, this is automatically a grace period. */
3905 if (rcu_blocking_is_gp())
3908 /* If expedited grace periods are prohibited, fall back to normal. */
3909 if (rcu_gp_is_normal()) {
3910 wait_rcu_gp(call_rcu_sched
);
3914 /* Take a snapshot of the sequence number. */
3915 s
= rcu_exp_gp_seq_snap(rsp
);
3916 if (exp_funnel_lock(rsp
, s
))
3917 return; /* Someone else did our work for us. */
3919 /* Initialize the rcu_node tree in preparation for the wait. */
3920 sync_rcu_exp_select_cpus(rsp
, sync_sched_exp_handler
);
3922 /* Wait and clean up, including waking everyone. */
3923 rcu_exp_wait_wake(rsp
, s
);
3925 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3928 * Check to see if there is any immediate RCU-related work to be done
3929 * by the current CPU, for the specified type of RCU, returning 1 if so.
3930 * The checks are in order of increasing expense: checks that can be
3931 * carried out against CPU-local state are performed first. However,
3932 * we must check for CPU stalls first, else we might not get a chance.
3934 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3936 struct rcu_node
*rnp
= rdp
->mynode
;
3938 rdp
->n_rcu_pending
++;
3940 /* Check for CPU stalls, if enabled. */
3941 check_cpu_stall(rsp
, rdp
);
3943 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3944 if (rcu_nohz_full_cpu(rsp
))
3947 /* Is the RCU core waiting for a quiescent state from this CPU? */
3948 if (rcu_scheduler_fully_active
&&
3949 rdp
->core_needs_qs
&& rdp
->cpu_no_qs
.b
.norm
&&
3950 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3951 rdp
->n_rp_core_needs_qs
++;
3952 } else if (rdp
->core_needs_qs
&&
3953 (!rdp
->cpu_no_qs
.b
.norm
||
3954 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3955 rdp
->n_rp_report_qs
++;
3959 /* Does this CPU have callbacks ready to invoke? */
3960 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3961 rdp
->n_rp_cb_ready
++;
3965 /* Has RCU gone idle with this CPU needing another grace period? */
3966 if (cpu_needs_another_gp(rsp
, rdp
)) {
3967 rdp
->n_rp_cpu_needs_gp
++;
3971 /* Has another RCU grace period completed? */
3972 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3973 rdp
->n_rp_gp_completed
++;
3977 /* Has a new RCU grace period started? */
3978 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3979 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3980 rdp
->n_rp_gp_started
++;
3984 /* Does this CPU need a deferred NOCB wakeup? */
3985 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3986 rdp
->n_rp_nocb_defer_wakeup
++;
3991 rdp
->n_rp_need_nothing
++;
3996 * Check to see if there is any immediate RCU-related work to be done
3997 * by the current CPU, returning 1 if so. This function is part of the
3998 * RCU implementation; it is -not- an exported member of the RCU API.
4000 static int rcu_pending(void)
4002 struct rcu_state
*rsp
;
4004 for_each_rcu_flavor(rsp
)
4005 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
4011 * Return true if the specified CPU has any callback. If all_lazy is
4012 * non-NULL, store an indication of whether all callbacks are lazy.
4013 * (If there are no callbacks, all of them are deemed to be lazy.)
4015 static bool __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
4019 struct rcu_data
*rdp
;
4020 struct rcu_state
*rsp
;
4022 for_each_rcu_flavor(rsp
) {
4023 rdp
= this_cpu_ptr(rsp
->rda
);
4027 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
4038 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
4039 * the compiler is expected to optimize this away.
4041 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
4042 int cpu
, unsigned long done
)
4044 trace_rcu_barrier(rsp
->name
, s
, cpu
,
4045 atomic_read(&rsp
->barrier_cpu_count
), done
);
4049 * RCU callback function for _rcu_barrier(). If we are last, wake
4050 * up the task executing _rcu_barrier().
4052 static void rcu_barrier_callback(struct rcu_head
*rhp
)
4054 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
4055 struct rcu_state
*rsp
= rdp
->rsp
;
4057 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
4058 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->barrier_sequence
);
4059 complete(&rsp
->barrier_completion
);
4061 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->barrier_sequence
);
4066 * Called with preemption disabled, and from cross-cpu IRQ context.
4068 static void rcu_barrier_func(void *type
)
4070 struct rcu_state
*rsp
= type
;
4071 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
4073 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->barrier_sequence
);
4074 atomic_inc(&rsp
->barrier_cpu_count
);
4075 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
4079 * Orchestrate the specified type of RCU barrier, waiting for all
4080 * RCU callbacks of the specified type to complete.
4082 static void _rcu_barrier(struct rcu_state
*rsp
)
4085 struct rcu_data
*rdp
;
4086 unsigned long s
= rcu_seq_snap(&rsp
->barrier_sequence
);
4088 _rcu_barrier_trace(rsp
, "Begin", -1, s
);
4090 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4091 mutex_lock(&rsp
->barrier_mutex
);
4093 /* Did someone else do our work for us? */
4094 if (rcu_seq_done(&rsp
->barrier_sequence
, s
)) {
4095 _rcu_barrier_trace(rsp
, "EarlyExit", -1, rsp
->barrier_sequence
);
4096 smp_mb(); /* caller's subsequent code after above check. */
4097 mutex_unlock(&rsp
->barrier_mutex
);
4101 /* Mark the start of the barrier operation. */
4102 rcu_seq_start(&rsp
->barrier_sequence
);
4103 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->barrier_sequence
);
4106 * Initialize the count to one rather than to zero in order to
4107 * avoid a too-soon return to zero in case of a short grace period
4108 * (or preemption of this task). Exclude CPU-hotplug operations
4109 * to ensure that no offline CPU has callbacks queued.
4111 init_completion(&rsp
->barrier_completion
);
4112 atomic_set(&rsp
->barrier_cpu_count
, 1);
4116 * Force each CPU with callbacks to register a new callback.
4117 * When that callback is invoked, we will know that all of the
4118 * corresponding CPU's preceding callbacks have been invoked.
4120 for_each_possible_cpu(cpu
) {
4121 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
4123 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4124 if (rcu_is_nocb_cpu(cpu
)) {
4125 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
4126 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
4127 rsp
->barrier_sequence
);
4129 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
4130 rsp
->barrier_sequence
);
4131 smp_mb__before_atomic();
4132 atomic_inc(&rsp
->barrier_cpu_count
);
4133 __call_rcu(&rdp
->barrier_head
,
4134 rcu_barrier_callback
, rsp
, cpu
, 0);
4136 } else if (READ_ONCE(rdp
->qlen
)) {
4137 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
4138 rsp
->barrier_sequence
);
4139 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
4141 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
4142 rsp
->barrier_sequence
);
4148 * Now that we have an rcu_barrier_callback() callback on each
4149 * CPU, and thus each counted, remove the initial count.
4151 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
4152 complete(&rsp
->barrier_completion
);
4154 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4155 wait_for_completion(&rsp
->barrier_completion
);
4157 /* Mark the end of the barrier operation. */
4158 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->barrier_sequence
);
4159 rcu_seq_end(&rsp
->barrier_sequence
);
4161 /* Other rcu_barrier() invocations can now safely proceed. */
4162 mutex_unlock(&rsp
->barrier_mutex
);
4166 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4168 void rcu_barrier_bh(void)
4170 _rcu_barrier(&rcu_bh_state
);
4172 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
4175 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4177 void rcu_barrier_sched(void)
4179 _rcu_barrier(&rcu_sched_state
);
4181 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
4184 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4185 * first CPU in a given leaf rcu_node structure coming online. The caller
4186 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4189 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
4192 struct rcu_node
*rnp
= rnp_leaf
;
4195 mask
= rnp
->grpmask
;
4199 raw_spin_lock_rcu_node(rnp
); /* Interrupts already disabled. */
4200 rnp
->qsmaskinit
|= mask
;
4201 raw_spin_unlock_rcu_node(rnp
); /* Interrupts remain disabled. */
4206 * Do boot-time initialization of a CPU's per-CPU RCU data.
4209 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
4211 unsigned long flags
;
4212 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4213 struct rcu_node
*rnp
= rcu_get_root(rsp
);
4215 /* Set up local state, ensuring consistent view of global state. */
4216 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4217 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
4218 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
4219 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
4220 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
4223 rcu_boot_init_nocb_percpu_data(rdp
);
4224 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4228 * Initialize a CPU's per-CPU RCU data. Note that only one online or
4229 * offline event can be happening at a given time. Note also that we
4230 * can accept some slop in the rsp->completed access due to the fact
4231 * that this CPU cannot possibly have any RCU callbacks in flight yet.
4234 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
4236 unsigned long flags
;
4238 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4239 struct rcu_node
*rnp
= rcu_get_root(rsp
);
4241 /* Set up local state, ensuring consistent view of global state. */
4242 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4243 rdp
->qlen_last_fqs_check
= 0;
4244 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
4245 rdp
->blimit
= blimit
;
4247 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
4248 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
4249 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
4250 atomic_set(&rdp
->dynticks
->dynticks
,
4251 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
4252 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
4255 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4256 * propagation up the rcu_node tree will happen at the beginning
4257 * of the next grace period.
4260 mask
= rdp
->grpmask
;
4261 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
4262 rnp
->qsmaskinitnext
|= mask
;
4263 rnp
->expmaskinitnext
|= mask
;
4264 if (!rdp
->beenonline
)
4265 WRITE_ONCE(rsp
->ncpus
, READ_ONCE(rsp
->ncpus
) + 1);
4266 rdp
->beenonline
= true; /* We have now been online. */
4267 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
4268 rdp
->completed
= rnp
->completed
;
4269 rdp
->cpu_no_qs
.b
.norm
= true;
4270 rdp
->rcu_qs_ctr_snap
= per_cpu(rcu_qs_ctr
, cpu
);
4271 rdp
->core_needs_qs
= false;
4272 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
4273 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4276 static void rcu_prepare_cpu(int cpu
)
4278 struct rcu_state
*rsp
;
4280 for_each_rcu_flavor(rsp
)
4281 rcu_init_percpu_data(cpu
, rsp
);
4284 #ifdef CONFIG_HOTPLUG_CPU
4286 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4287 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4289 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4290 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4293 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
4295 unsigned long flags
;
4297 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4298 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
4300 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
4303 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4304 mask
= rdp
->grpmask
;
4305 raw_spin_lock_irqsave_rcu_node(rnp
, flags
); /* Enforce GP memory-order guarantee. */
4306 rnp
->qsmaskinitnext
&= ~mask
;
4307 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4310 void rcu_report_dead(unsigned int cpu
)
4312 struct rcu_state
*rsp
;
4314 /* QS for any half-done expedited RCU-sched GP. */
4316 rcu_report_exp_rdp(&rcu_sched_state
,
4317 this_cpu_ptr(rcu_sched_state
.rda
), true);
4319 for_each_rcu_flavor(rsp
)
4320 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
4325 * Handle CPU online/offline notification events.
4327 int rcu_cpu_notify(struct notifier_block
*self
,
4328 unsigned long action
, void *hcpu
)
4330 long cpu
= (long)hcpu
;
4331 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
4332 struct rcu_node
*rnp
= rdp
->mynode
;
4333 struct rcu_state
*rsp
;
4336 case CPU_UP_PREPARE
:
4337 case CPU_UP_PREPARE_FROZEN
:
4338 rcu_prepare_cpu(cpu
);
4339 rcu_prepare_kthreads(cpu
);
4340 rcu_spawn_all_nocb_kthreads(cpu
);
4343 case CPU_DOWN_FAILED
:
4344 sync_sched_exp_online_cleanup(cpu
);
4345 rcu_boost_kthread_setaffinity(rnp
, -1);
4347 case CPU_DOWN_PREPARE
:
4348 rcu_boost_kthread_setaffinity(rnp
, cpu
);
4351 case CPU_DYING_FROZEN
:
4352 for_each_rcu_flavor(rsp
)
4353 rcu_cleanup_dying_cpu(rsp
);
4356 case CPU_DEAD_FROZEN
:
4357 case CPU_UP_CANCELED
:
4358 case CPU_UP_CANCELED_FROZEN
:
4359 for_each_rcu_flavor(rsp
) {
4360 rcu_cleanup_dead_cpu(cpu
, rsp
);
4361 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
4370 static int rcu_pm_notify(struct notifier_block
*self
,
4371 unsigned long action
, void *hcpu
)
4374 case PM_HIBERNATION_PREPARE
:
4375 case PM_SUSPEND_PREPARE
:
4376 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
4379 case PM_POST_HIBERNATION
:
4380 case PM_POST_SUSPEND
:
4381 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
4382 rcu_unexpedite_gp();
4391 * Spawn the kthreads that handle each RCU flavor's grace periods.
4393 static int __init
rcu_spawn_gp_kthread(void)
4395 unsigned long flags
;
4396 int kthread_prio_in
= kthread_prio
;
4397 struct rcu_node
*rnp
;
4398 struct rcu_state
*rsp
;
4399 struct sched_param sp
;
4400 struct task_struct
*t
;
4402 /* Force priority into range. */
4403 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
4405 else if (kthread_prio
< 0)
4407 else if (kthread_prio
> 99)
4409 if (kthread_prio
!= kthread_prio_in
)
4410 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4411 kthread_prio
, kthread_prio_in
);
4413 rcu_scheduler_fully_active
= 1;
4414 for_each_rcu_flavor(rsp
) {
4415 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
4417 rnp
= rcu_get_root(rsp
);
4418 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4419 rsp
->gp_kthread
= t
;
4421 sp
.sched_priority
= kthread_prio
;
4422 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
4424 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4427 rcu_spawn_nocb_kthreads();
4428 rcu_spawn_boost_kthreads();
4431 early_initcall(rcu_spawn_gp_kthread
);
4434 * This function is invoked towards the end of the scheduler's initialization
4435 * process. Before this is called, the idle task might contain
4436 * RCU read-side critical sections (during which time, this idle
4437 * task is booting the system). After this function is called, the
4438 * idle tasks are prohibited from containing RCU read-side critical
4439 * sections. This function also enables RCU lockdep checking.
4441 void rcu_scheduler_starting(void)
4443 WARN_ON(num_online_cpus() != 1);
4444 WARN_ON(nr_context_switches() > 0);
4445 rcu_scheduler_active
= 1;
4449 * Compute the per-level fanout, either using the exact fanout specified
4450 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4452 static void __init
rcu_init_levelspread(int *levelspread
, const int *levelcnt
)
4456 if (rcu_fanout_exact
) {
4457 levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
4458 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
4459 levelspread
[i
] = RCU_FANOUT
;
4465 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4467 levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
4474 * Helper function for rcu_init() that initializes one rcu_state structure.
4476 static void __init
rcu_init_one(struct rcu_state
*rsp
)
4478 static const char * const buf
[] = RCU_NODE_NAME_INIT
;
4479 static const char * const fqs
[] = RCU_FQS_NAME_INIT
;
4480 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
4481 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
4482 static u8 fl_mask
= 0x1;
4484 int levelcnt
[RCU_NUM_LVLS
]; /* # nodes in each level. */
4485 int levelspread
[RCU_NUM_LVLS
]; /* kids/node in each level. */
4489 struct rcu_node
*rnp
;
4491 BUILD_BUG_ON(RCU_NUM_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4493 /* Silence gcc 4.8 false positive about array index out of range. */
4494 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4495 panic("rcu_init_one: rcu_num_lvls out of range");
4497 /* Initialize the level-tracking arrays. */
4499 for (i
= 0; i
< rcu_num_lvls
; i
++)
4500 levelcnt
[i
] = num_rcu_lvl
[i
];
4501 for (i
= 1; i
< rcu_num_lvls
; i
++)
4502 rsp
->level
[i
] = rsp
->level
[i
- 1] + levelcnt
[i
- 1];
4503 rcu_init_levelspread(levelspread
, levelcnt
);
4504 rsp
->flavor_mask
= fl_mask
;
4507 /* Initialize the elements themselves, starting from the leaves. */
4509 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4510 cpustride
*= levelspread
[i
];
4511 rnp
= rsp
->level
[i
];
4512 for (j
= 0; j
< levelcnt
[i
]; j
++, rnp
++) {
4513 raw_spin_lock_init(&ACCESS_PRIVATE(rnp
, lock
));
4514 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp
, lock
),
4515 &rcu_node_class
[i
], buf
[i
]);
4516 raw_spin_lock_init(&rnp
->fqslock
);
4517 lockdep_set_class_and_name(&rnp
->fqslock
,
4518 &rcu_fqs_class
[i
], fqs
[i
]);
4519 rnp
->gpnum
= rsp
->gpnum
;
4520 rnp
->completed
= rsp
->completed
;
4522 rnp
->qsmaskinit
= 0;
4523 rnp
->grplo
= j
* cpustride
;
4524 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4525 if (rnp
->grphi
>= nr_cpu_ids
)
4526 rnp
->grphi
= nr_cpu_ids
- 1;
4532 rnp
->grpnum
= j
% levelspread
[i
- 1];
4533 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4534 rnp
->parent
= rsp
->level
[i
- 1] +
4535 j
/ levelspread
[i
- 1];
4538 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4539 rcu_init_one_nocb(rnp
);
4540 init_waitqueue_head(&rnp
->exp_wq
[0]);
4541 init_waitqueue_head(&rnp
->exp_wq
[1]);
4542 init_waitqueue_head(&rnp
->exp_wq
[2]);
4543 init_waitqueue_head(&rnp
->exp_wq
[3]);
4544 spin_lock_init(&rnp
->exp_lock
);
4548 init_swait_queue_head(&rsp
->gp_wq
);
4549 init_swait_queue_head(&rsp
->expedited_wq
);
4550 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4551 for_each_possible_cpu(i
) {
4552 while (i
> rnp
->grphi
)
4554 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4555 rcu_boot_init_percpu_data(i
, rsp
);
4557 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4561 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4562 * replace the definitions in tree.h because those are needed to size
4563 * the ->node array in the rcu_state structure.
4565 static void __init
rcu_init_geometry(void)
4569 int rcu_capacity
[RCU_NUM_LVLS
];
4572 * Initialize any unspecified boot parameters.
4573 * The default values of jiffies_till_first_fqs and
4574 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4575 * value, which is a function of HZ, then adding one for each
4576 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4578 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4579 if (jiffies_till_first_fqs
== ULONG_MAX
)
4580 jiffies_till_first_fqs
= d
;
4581 if (jiffies_till_next_fqs
== ULONG_MAX
)
4582 jiffies_till_next_fqs
= d
;
4584 /* If the compile-time values are accurate, just leave. */
4585 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4586 nr_cpu_ids
== NR_CPUS
)
4588 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4589 rcu_fanout_leaf
, nr_cpu_ids
);
4592 * The boot-time rcu_fanout_leaf parameter must be at least two
4593 * and cannot exceed the number of bits in the rcu_node masks.
4594 * Complain and fall back to the compile-time values if this
4595 * limit is exceeded.
4597 if (rcu_fanout_leaf
< 2 ||
4598 rcu_fanout_leaf
> sizeof(unsigned long) * 8) {
4599 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4605 * Compute number of nodes that can be handled an rcu_node tree
4606 * with the given number of levels.
4608 rcu_capacity
[0] = rcu_fanout_leaf
;
4609 for (i
= 1; i
< RCU_NUM_LVLS
; i
++)
4610 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4613 * The tree must be able to accommodate the configured number of CPUs.
4614 * If this limit is exceeded, fall back to the compile-time values.
4616 if (nr_cpu_ids
> rcu_capacity
[RCU_NUM_LVLS
- 1]) {
4617 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4622 /* Calculate the number of levels in the tree. */
4623 for (i
= 0; nr_cpu_ids
> rcu_capacity
[i
]; i
++) {
4625 rcu_num_lvls
= i
+ 1;
4627 /* Calculate the number of rcu_nodes at each level of the tree. */
4628 for (i
= 0; i
< rcu_num_lvls
; i
++) {
4629 int cap
= rcu_capacity
[(rcu_num_lvls
- 1) - i
];
4630 num_rcu_lvl
[i
] = DIV_ROUND_UP(nr_cpu_ids
, cap
);
4633 /* Calculate the total number of rcu_node structures. */
4635 for (i
= 0; i
< rcu_num_lvls
; i
++)
4636 rcu_num_nodes
+= num_rcu_lvl
[i
];
4640 * Dump out the structure of the rcu_node combining tree associated
4641 * with the rcu_state structure referenced by rsp.
4643 static void __init
rcu_dump_rcu_node_tree(struct rcu_state
*rsp
)
4646 struct rcu_node
*rnp
;
4648 pr_info("rcu_node tree layout dump\n");
4650 rcu_for_each_node_breadth_first(rsp
, rnp
) {
4651 if (rnp
->level
!= level
) {
4656 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4661 void __init
rcu_init(void)
4665 rcu_early_boot_tests();
4667 rcu_bootup_announce();
4668 rcu_init_geometry();
4669 rcu_init_one(&rcu_bh_state
);
4670 rcu_init_one(&rcu_sched_state
);
4672 rcu_dump_rcu_node_tree(&rcu_sched_state
);
4673 __rcu_init_preempt();
4674 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4677 * We don't need protection against CPU-hotplug here because
4678 * this is called early in boot, before either interrupts
4679 * or the scheduler are operational.
4681 cpu_notifier(rcu_cpu_notify
, 0);
4682 pm_notifier(rcu_pm_notify
, 0);
4683 for_each_online_cpu(cpu
)
4684 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4687 #include "tree_plugin.h"