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), \
107 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
108 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
110 static struct rcu_state
*const rcu_state_p
;
111 LIST_HEAD(rcu_struct_flavors
);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree
;
115 module_param(dump_tree
, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact
;
118 module_param(rcu_fanout_exact
, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
121 module_param(rcu_fanout_leaf
, int, 0444);
122 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
123 /* Number of rcu_nodes at specified level. */
124 static int num_rcu_lvl
[] = NUM_RCU_LVL_INIT
;
125 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
128 * The rcu_scheduler_active variable transitions from zero to one just
129 * before the first task is spawned. So when this variable is zero, RCU
130 * can assume that there is but one task, allowing RCU to (for example)
131 * optimize synchronize_sched() to a simple barrier(). When this variable
132 * is one, RCU must actually do all the hard work required to detect real
133 * grace periods. This variable is also used to suppress boot-time false
134 * positives from lockdep-RCU error checking.
136 int rcu_scheduler_active __read_mostly
;
137 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
140 * The rcu_scheduler_fully_active variable transitions from zero to one
141 * during the early_initcall() processing, which is after the scheduler
142 * is capable of creating new tasks. So RCU processing (for example,
143 * creating tasks for RCU priority boosting) must be delayed until after
144 * rcu_scheduler_fully_active transitions from zero to one. We also
145 * currently delay invocation of any RCU callbacks until after this point.
147 * It might later prove better for people registering RCU callbacks during
148 * early boot to take responsibility for these callbacks, but one step at
151 static int rcu_scheduler_fully_active __read_mostly
;
153 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
154 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
155 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
156 static void invoke_rcu_core(void);
157 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
158 static void rcu_report_exp_rdp(struct rcu_state
*rsp
,
159 struct rcu_data
*rdp
, bool wake
);
161 /* rcuc/rcub kthread realtime priority */
162 #ifdef CONFIG_RCU_KTHREAD_PRIO
163 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
164 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
165 static int kthread_prio
= IS_ENABLED(CONFIG_RCU_BOOST
) ? 1 : 0;
166 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
167 module_param(kthread_prio
, int, 0644);
169 /* Delay in jiffies for grace-period initialization delays, debug only. */
171 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
172 static int gp_preinit_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY
;
173 module_param(gp_preinit_delay
, int, 0644);
174 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
175 static const int gp_preinit_delay
;
176 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
178 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
179 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
180 module_param(gp_init_delay
, int, 0644);
181 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
182 static const int gp_init_delay
;
183 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
185 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
186 static int gp_cleanup_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY
;
187 module_param(gp_cleanup_delay
, int, 0644);
188 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
189 static const int gp_cleanup_delay
;
190 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
193 * Number of grace periods between delays, normalized by the duration of
194 * the delay. The longer the the delay, the more the grace periods between
195 * each delay. The reason for this normalization is that it means that,
196 * for non-zero delays, the overall slowdown of grace periods is constant
197 * regardless of the duration of the delay. This arrangement balances
198 * the need for long delays to increase some race probabilities with the
199 * need for fast grace periods to increase other race probabilities.
201 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
204 * Track the rcutorture test sequence number and the update version
205 * number within a given test. The rcutorture_testseq is incremented
206 * on every rcutorture module load and unload, so has an odd value
207 * when a test is running. The rcutorture_vernum is set to zero
208 * when rcutorture starts and is incremented on each rcutorture update.
209 * These variables enable correlating rcutorture output with the
210 * RCU tracing information.
212 unsigned long rcutorture_testseq
;
213 unsigned long rcutorture_vernum
;
216 * Compute the mask of online CPUs for the specified rcu_node structure.
217 * This will not be stable unless the rcu_node structure's ->lock is
218 * held, but the bit corresponding to the current CPU will be stable
221 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
223 return READ_ONCE(rnp
->qsmaskinitnext
);
227 * Return true if an RCU grace period is in progress. The READ_ONCE()s
228 * permit this function to be invoked without holding the root rcu_node
229 * structure's ->lock, but of course results can be subject to change.
231 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
233 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
237 * Note a quiescent state. Because we do not need to know
238 * how many quiescent states passed, just if there was at least
239 * one since the start of the grace period, this just sets a flag.
240 * The caller must have disabled preemption.
242 void rcu_sched_qs(void)
244 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.s
))
246 trace_rcu_grace_period(TPS("rcu_sched"),
247 __this_cpu_read(rcu_sched_data
.gpnum
),
249 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.norm
, false);
250 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
252 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, false);
253 rcu_report_exp_rdp(&rcu_sched_state
,
254 this_cpu_ptr(&rcu_sched_data
), true);
259 if (__this_cpu_read(rcu_bh_data
.cpu_no_qs
.s
)) {
260 trace_rcu_grace_period(TPS("rcu_bh"),
261 __this_cpu_read(rcu_bh_data
.gpnum
),
263 __this_cpu_write(rcu_bh_data
.cpu_no_qs
.b
.norm
, false);
267 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
269 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
270 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
271 .dynticks
= ATOMIC_INIT(1),
272 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
273 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
274 .dynticks_idle
= ATOMIC_INIT(1),
275 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
278 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
279 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
282 * Let the RCU core know that this CPU has gone through the scheduler,
283 * which is a quiescent state. This is called when the need for a
284 * quiescent state is urgent, so we burn an atomic operation and full
285 * memory barriers to let the RCU core know about it, regardless of what
286 * this CPU might (or might not) do in the near future.
288 * We inform the RCU core by emulating a zero-duration dyntick-idle
289 * period, which we in turn do by incrementing the ->dynticks counter
292 * The caller must have disabled interrupts.
294 static void rcu_momentary_dyntick_idle(void)
296 struct rcu_data
*rdp
;
297 struct rcu_dynticks
*rdtp
;
299 struct rcu_state
*rsp
;
302 * Yes, we can lose flag-setting operations. This is OK, because
303 * the flag will be set again after some delay.
305 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
306 raw_cpu_write(rcu_sched_qs_mask
, 0);
308 /* Find the flavor that needs a quiescent state. */
309 for_each_rcu_flavor(rsp
) {
310 rdp
= raw_cpu_ptr(rsp
->rda
);
311 if (!(resched_mask
& rsp
->flavor_mask
))
313 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
314 if (READ_ONCE(rdp
->mynode
->completed
) !=
315 READ_ONCE(rdp
->cond_resched_completed
))
319 * Pretend to be momentarily idle for the quiescent state.
320 * This allows the grace-period kthread to record the
321 * quiescent state, with no need for this CPU to do anything
324 rdtp
= this_cpu_ptr(&rcu_dynticks
);
325 smp_mb__before_atomic(); /* Earlier stuff before QS. */
326 atomic_add(2, &rdtp
->dynticks
); /* QS. */
327 smp_mb__after_atomic(); /* Later stuff after QS. */
333 * Note a context switch. This is a quiescent state for RCU-sched,
334 * and requires special handling for preemptible RCU.
335 * The caller must have disabled interrupts.
337 void rcu_note_context_switch(void)
339 barrier(); /* Avoid RCU read-side critical sections leaking down. */
340 trace_rcu_utilization(TPS("Start context switch"));
342 rcu_preempt_note_context_switch();
343 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
344 rcu_momentary_dyntick_idle();
345 trace_rcu_utilization(TPS("End context switch"));
346 barrier(); /* Avoid RCU read-side critical sections leaking up. */
348 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
351 * Register a quiescent state for all RCU flavors. If there is an
352 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
353 * dyntick-idle quiescent state visible to other CPUs (but only for those
354 * RCU flavors in desperate need of a quiescent state, which will normally
355 * be none of them). Either way, do a lightweight quiescent state for
358 * The barrier() calls are redundant in the common case when this is
359 * called externally, but just in case this is called from within this
363 void rcu_all_qs(void)
367 barrier(); /* Avoid RCU read-side critical sections leaking down. */
368 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
))) {
369 local_irq_save(flags
);
370 rcu_momentary_dyntick_idle();
371 local_irq_restore(flags
);
373 this_cpu_inc(rcu_qs_ctr
);
374 barrier(); /* Avoid RCU read-side critical sections leaking up. */
376 EXPORT_SYMBOL_GPL(rcu_all_qs
);
378 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
379 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
380 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
382 module_param(blimit
, long, 0444);
383 module_param(qhimark
, long, 0444);
384 module_param(qlowmark
, long, 0444);
386 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
387 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
389 module_param(jiffies_till_first_fqs
, ulong
, 0644);
390 module_param(jiffies_till_next_fqs
, ulong
, 0644);
393 * How long the grace period must be before we start recruiting
394 * quiescent-state help from rcu_note_context_switch().
396 static ulong jiffies_till_sched_qs
= HZ
/ 20;
397 module_param(jiffies_till_sched_qs
, ulong
, 0644);
399 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
400 struct rcu_data
*rdp
);
401 static void force_qs_rnp(struct rcu_state
*rsp
,
402 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
403 unsigned long *maxj
),
404 bool *isidle
, unsigned long *maxj
);
405 static void force_quiescent_state(struct rcu_state
*rsp
);
406 static int rcu_pending(void);
409 * Return the number of RCU batches started thus far for debug & stats.
411 unsigned long rcu_batches_started(void)
413 return rcu_state_p
->gpnum
;
415 EXPORT_SYMBOL_GPL(rcu_batches_started
);
418 * Return the number of RCU-sched batches started thus far for debug & stats.
420 unsigned long rcu_batches_started_sched(void)
422 return rcu_sched_state
.gpnum
;
424 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
427 * Return the number of RCU BH batches started thus far for debug & stats.
429 unsigned long rcu_batches_started_bh(void)
431 return rcu_bh_state
.gpnum
;
433 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
436 * Return the number of RCU batches completed thus far for debug & stats.
438 unsigned long rcu_batches_completed(void)
440 return rcu_state_p
->completed
;
442 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
445 * Return the number of RCU-sched batches completed thus far for debug & stats.
447 unsigned long rcu_batches_completed_sched(void)
449 return rcu_sched_state
.completed
;
451 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
454 * Return the number of RCU BH batches completed thus far for debug & stats.
456 unsigned long rcu_batches_completed_bh(void)
458 return rcu_bh_state
.completed
;
460 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
463 * Force a quiescent state.
465 void rcu_force_quiescent_state(void)
467 force_quiescent_state(rcu_state_p
);
469 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
472 * Force a quiescent state for RCU BH.
474 void rcu_bh_force_quiescent_state(void)
476 force_quiescent_state(&rcu_bh_state
);
478 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
481 * Force a quiescent state for RCU-sched.
483 void rcu_sched_force_quiescent_state(void)
485 force_quiescent_state(&rcu_sched_state
);
487 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
490 * Show the state of the grace-period kthreads.
492 void show_rcu_gp_kthreads(void)
494 struct rcu_state
*rsp
;
496 for_each_rcu_flavor(rsp
) {
497 pr_info("%s: wait state: %d ->state: %#lx\n",
498 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
499 /* sched_show_task(rsp->gp_kthread); */
502 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
505 * Record the number of times rcutorture tests have been initiated and
506 * terminated. This information allows the debugfs tracing stats to be
507 * correlated to the rcutorture messages, even when the rcutorture module
508 * is being repeatedly loaded and unloaded. In other words, we cannot
509 * store this state in rcutorture itself.
511 void rcutorture_record_test_transition(void)
513 rcutorture_testseq
++;
514 rcutorture_vernum
= 0;
516 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
519 * Send along grace-period-related data for rcutorture diagnostics.
521 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
522 unsigned long *gpnum
, unsigned long *completed
)
524 struct rcu_state
*rsp
= NULL
;
533 case RCU_SCHED_FLAVOR
:
534 rsp
= &rcu_sched_state
;
540 *flags
= READ_ONCE(rsp
->gp_flags
);
541 *gpnum
= READ_ONCE(rsp
->gpnum
);
542 *completed
= READ_ONCE(rsp
->completed
);
549 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
552 * Record the number of writer passes through the current rcutorture test.
553 * This is also used to correlate debugfs tracing stats with the rcutorture
556 void rcutorture_record_progress(unsigned long vernum
)
560 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
563 * Does the CPU have callbacks ready to be invoked?
566 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
568 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
569 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
573 * Return the root node of the specified rcu_state structure.
575 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
577 return &rsp
->node
[0];
581 * Is there any need for future grace periods?
582 * Interrupts must be disabled. If the caller does not hold the root
583 * rnp_node structure's ->lock, the results are advisory only.
585 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
587 struct rcu_node
*rnp
= rcu_get_root(rsp
);
588 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
589 int *fp
= &rnp
->need_future_gp
[idx
];
591 return READ_ONCE(*fp
);
595 * Does the current CPU require a not-yet-started grace period?
596 * The caller must have disabled interrupts to prevent races with
597 * normal callback registry.
600 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
604 if (rcu_gp_in_progress(rsp
))
605 return false; /* No, a grace period is already in progress. */
606 if (rcu_future_needs_gp(rsp
))
607 return true; /* Yes, a no-CBs CPU needs one. */
608 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
609 return false; /* No, this is a no-CBs (or offline) CPU. */
610 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
611 return true; /* Yes, CPU has newly registered callbacks. */
612 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
613 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
614 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
615 rdp
->nxtcompleted
[i
]))
616 return true; /* Yes, CBs for future grace period. */
617 return false; /* No grace period needed. */
621 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
623 * If the new value of the ->dynticks_nesting counter now is zero,
624 * we really have entered idle, and must do the appropriate accounting.
625 * The caller must have disabled interrupts.
627 static void rcu_eqs_enter_common(long long oldval
, bool user
)
629 struct rcu_state
*rsp
;
630 struct rcu_data
*rdp
;
631 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
633 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
634 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
635 !user
&& !is_idle_task(current
)) {
636 struct task_struct
*idle __maybe_unused
=
637 idle_task(smp_processor_id());
639 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
640 rcu_ftrace_dump(DUMP_ORIG
);
641 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
642 current
->pid
, current
->comm
,
643 idle
->pid
, idle
->comm
); /* must be idle task! */
645 for_each_rcu_flavor(rsp
) {
646 rdp
= this_cpu_ptr(rsp
->rda
);
647 do_nocb_deferred_wakeup(rdp
);
649 rcu_prepare_for_idle();
650 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
651 smp_mb__before_atomic(); /* See above. */
652 atomic_inc(&rdtp
->dynticks
);
653 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
654 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
655 atomic_read(&rdtp
->dynticks
) & 0x1);
656 rcu_dynticks_task_enter();
659 * It is illegal to enter an extended quiescent state while
660 * in an RCU read-side critical section.
662 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map
),
663 "Illegal idle entry in RCU read-side critical section.");
664 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
),
665 "Illegal idle entry in RCU-bh read-side critical section.");
666 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map
),
667 "Illegal idle entry in RCU-sched read-side critical section.");
671 * Enter an RCU extended quiescent state, which can be either the
672 * idle loop or adaptive-tickless usermode execution.
674 static void rcu_eqs_enter(bool user
)
677 struct rcu_dynticks
*rdtp
;
679 rdtp
= this_cpu_ptr(&rcu_dynticks
);
680 oldval
= rdtp
->dynticks_nesting
;
681 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
682 (oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
683 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
684 rdtp
->dynticks_nesting
= 0;
685 rcu_eqs_enter_common(oldval
, user
);
687 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
692 * rcu_idle_enter - inform RCU that current CPU is entering idle
694 * Enter idle mode, in other words, -leave- the mode in which RCU
695 * read-side critical sections can occur. (Though RCU read-side
696 * critical sections can occur in irq handlers in idle, a possibility
697 * handled by irq_enter() and irq_exit().)
699 * We crowbar the ->dynticks_nesting field to zero to allow for
700 * the possibility of usermode upcalls having messed up our count
701 * of interrupt nesting level during the prior busy period.
703 void rcu_idle_enter(void)
707 local_irq_save(flags
);
708 rcu_eqs_enter(false);
709 rcu_sysidle_enter(0);
710 local_irq_restore(flags
);
712 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
714 #ifdef CONFIG_NO_HZ_FULL
716 * rcu_user_enter - inform RCU that we are resuming userspace.
718 * Enter RCU idle mode right before resuming userspace. No use of RCU
719 * is permitted between this call and rcu_user_exit(). This way the
720 * CPU doesn't need to maintain the tick for RCU maintenance purposes
721 * when the CPU runs in userspace.
723 void rcu_user_enter(void)
727 #endif /* CONFIG_NO_HZ_FULL */
730 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
732 * Exit from an interrupt handler, which might possibly result in entering
733 * idle mode, in other words, leaving the mode in which read-side critical
734 * sections can occur. The caller must have disabled interrupts.
736 * This code assumes that the idle loop never does anything that might
737 * result in unbalanced calls to irq_enter() and irq_exit(). If your
738 * architecture violates this assumption, RCU will give you what you
739 * deserve, good and hard. But very infrequently and irreproducibly.
741 * Use things like work queues to work around this limitation.
743 * You have been warned.
745 void rcu_irq_exit(void)
748 struct rcu_dynticks
*rdtp
;
750 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
751 rdtp
= this_cpu_ptr(&rcu_dynticks
);
752 oldval
= rdtp
->dynticks_nesting
;
753 rdtp
->dynticks_nesting
--;
754 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
755 rdtp
->dynticks_nesting
< 0);
756 if (rdtp
->dynticks_nesting
)
757 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
759 rcu_eqs_enter_common(oldval
, true);
760 rcu_sysidle_enter(1);
764 * Wrapper for rcu_irq_exit() where interrupts are enabled.
766 void rcu_irq_exit_irqson(void)
770 local_irq_save(flags
);
772 local_irq_restore(flags
);
776 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
778 * If the new value of the ->dynticks_nesting counter was previously zero,
779 * we really have exited idle, and must do the appropriate accounting.
780 * The caller must have disabled interrupts.
782 static void rcu_eqs_exit_common(long long oldval
, int user
)
784 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
786 rcu_dynticks_task_exit();
787 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
788 atomic_inc(&rdtp
->dynticks
);
789 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
790 smp_mb__after_atomic(); /* See above. */
791 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
792 !(atomic_read(&rdtp
->dynticks
) & 0x1));
793 rcu_cleanup_after_idle();
794 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
795 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
796 !user
&& !is_idle_task(current
)) {
797 struct task_struct
*idle __maybe_unused
=
798 idle_task(smp_processor_id());
800 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
801 oldval
, rdtp
->dynticks_nesting
);
802 rcu_ftrace_dump(DUMP_ORIG
);
803 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
804 current
->pid
, current
->comm
,
805 idle
->pid
, idle
->comm
); /* must be idle task! */
810 * Exit an RCU extended quiescent state, which can be either the
811 * idle loop or adaptive-tickless usermode execution.
813 static void rcu_eqs_exit(bool user
)
815 struct rcu_dynticks
*rdtp
;
818 rdtp
= this_cpu_ptr(&rcu_dynticks
);
819 oldval
= rdtp
->dynticks_nesting
;
820 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) && oldval
< 0);
821 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
822 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
824 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
825 rcu_eqs_exit_common(oldval
, user
);
830 * rcu_idle_exit - inform RCU that current CPU is leaving idle
832 * Exit idle mode, in other words, -enter- the mode in which RCU
833 * read-side critical sections can occur.
835 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
836 * allow for the possibility of usermode upcalls messing up our count
837 * of interrupt nesting level during the busy period that is just
840 void rcu_idle_exit(void)
844 local_irq_save(flags
);
847 local_irq_restore(flags
);
849 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
851 #ifdef CONFIG_NO_HZ_FULL
853 * rcu_user_exit - inform RCU that we are exiting userspace.
855 * Exit RCU idle mode while entering the kernel because it can
856 * run a RCU read side critical section anytime.
858 void rcu_user_exit(void)
862 #endif /* CONFIG_NO_HZ_FULL */
865 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
867 * Enter an interrupt handler, which might possibly result in exiting
868 * idle mode, in other words, entering the mode in which read-side critical
869 * sections can occur. The caller must have disabled interrupts.
871 * Note that the Linux kernel is fully capable of entering an interrupt
872 * handler that it never exits, for example when doing upcalls to
873 * user mode! This code assumes that the idle loop never does upcalls to
874 * user mode. If your architecture does do upcalls from the idle loop (or
875 * does anything else that results in unbalanced calls to the irq_enter()
876 * and irq_exit() functions), RCU will give you what you deserve, good
877 * and hard. But very infrequently and irreproducibly.
879 * Use things like work queues to work around this limitation.
881 * You have been warned.
883 void rcu_irq_enter(void)
885 struct rcu_dynticks
*rdtp
;
888 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
889 rdtp
= this_cpu_ptr(&rcu_dynticks
);
890 oldval
= rdtp
->dynticks_nesting
;
891 rdtp
->dynticks_nesting
++;
892 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
893 rdtp
->dynticks_nesting
== 0);
895 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
897 rcu_eqs_exit_common(oldval
, true);
902 * Wrapper for rcu_irq_enter() where interrupts are enabled.
904 void rcu_irq_enter_irqson(void)
908 local_irq_save(flags
);
910 local_irq_restore(flags
);
914 * rcu_nmi_enter - inform RCU of entry to NMI context
916 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
917 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
918 * that the CPU is active. This implementation permits nested NMIs, as
919 * long as the nesting level does not overflow an int. (You will probably
920 * run out of stack space first.)
922 void rcu_nmi_enter(void)
924 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
927 /* Complain about underflow. */
928 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
931 * If idle from RCU viewpoint, atomically increment ->dynticks
932 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
933 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
934 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
935 * to be in the outermost NMI handler that interrupted an RCU-idle
936 * period (observation due to Andy Lutomirski).
938 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
939 smp_mb__before_atomic(); /* Force delay from prior write. */
940 atomic_inc(&rdtp
->dynticks
);
941 /* atomic_inc() before later RCU read-side crit sects */
942 smp_mb__after_atomic(); /* See above. */
943 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
946 rdtp
->dynticks_nmi_nesting
+= incby
;
951 * rcu_nmi_exit - inform RCU of exit from NMI context
953 * If we are returning from the outermost NMI handler that interrupted an
954 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
955 * to let the RCU grace-period handling know that the CPU is back to
958 void rcu_nmi_exit(void)
960 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
963 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
964 * (We are exiting an NMI handler, so RCU better be paying attention
967 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
968 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
971 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
972 * leave it in non-RCU-idle state.
974 if (rdtp
->dynticks_nmi_nesting
!= 1) {
975 rdtp
->dynticks_nmi_nesting
-= 2;
979 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
980 rdtp
->dynticks_nmi_nesting
= 0;
981 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
982 smp_mb__before_atomic(); /* See above. */
983 atomic_inc(&rdtp
->dynticks
);
984 smp_mb__after_atomic(); /* Force delay to next write. */
985 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
989 * __rcu_is_watching - are RCU read-side critical sections safe?
991 * Return true if RCU is watching the running CPU, which means that
992 * this CPU can safely enter RCU read-side critical sections. Unlike
993 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
994 * least disabled preemption.
996 bool notrace
__rcu_is_watching(void)
998 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
1002 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1004 * If the current CPU is in its idle loop and is neither in an interrupt
1005 * or NMI handler, return true.
1007 bool notrace
rcu_is_watching(void)
1011 preempt_disable_notrace();
1012 ret
= __rcu_is_watching();
1013 preempt_enable_notrace();
1016 EXPORT_SYMBOL_GPL(rcu_is_watching
);
1018 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1021 * Is the current CPU online? Disable preemption to avoid false positives
1022 * that could otherwise happen due to the current CPU number being sampled,
1023 * this task being preempted, its old CPU being taken offline, resuming
1024 * on some other CPU, then determining that its old CPU is now offline.
1025 * It is OK to use RCU on an offline processor during initial boot, hence
1026 * the check for rcu_scheduler_fully_active. Note also that it is OK
1027 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1028 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1029 * offline to continue to use RCU for one jiffy after marking itself
1030 * offline in the cpu_online_mask. This leniency is necessary given the
1031 * non-atomic nature of the online and offline processing, for example,
1032 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1035 * This is also why RCU internally marks CPUs online during the
1036 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1038 * Disable checking if in an NMI handler because we cannot safely report
1039 * errors from NMI handlers anyway.
1041 bool rcu_lockdep_current_cpu_online(void)
1043 struct rcu_data
*rdp
;
1044 struct rcu_node
*rnp
;
1050 rdp
= this_cpu_ptr(&rcu_sched_data
);
1052 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
1053 !rcu_scheduler_fully_active
;
1057 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
1059 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1062 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1064 * If the current CPU is idle or running at a first-level (not nested)
1065 * interrupt from idle, return true. The caller must have at least
1066 * disabled preemption.
1068 static int rcu_is_cpu_rrupt_from_idle(void)
1070 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1074 * Snapshot the specified CPU's dynticks counter so that we can later
1075 * credit them with an implicit quiescent state. Return 1 if this CPU
1076 * is in dynticks idle mode, which is an extended quiescent state.
1078 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1079 bool *isidle
, unsigned long *maxj
)
1081 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1082 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1083 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1084 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1085 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1086 rdp
->mynode
->gpnum
))
1087 WRITE_ONCE(rdp
->gpwrap
, true);
1094 * Return true if the specified CPU has passed through a quiescent
1095 * state by virtue of being in or having passed through an dynticks
1096 * idle state since the last call to dyntick_save_progress_counter()
1097 * for this same CPU, or by virtue of having been offline.
1099 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1100 bool *isidle
, unsigned long *maxj
)
1106 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1107 snap
= (unsigned int)rdp
->dynticks_snap
;
1110 * If the CPU passed through or entered a dynticks idle phase with
1111 * no active irq/NMI handlers, then we can safely pretend that the CPU
1112 * already acknowledged the request to pass through a quiescent
1113 * state. Either way, that CPU cannot possibly be in an RCU
1114 * read-side critical section that started before the beginning
1115 * of the current RCU grace period.
1117 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1118 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1119 rdp
->dynticks_fqs
++;
1124 * Check for the CPU being offline, but only if the grace period
1125 * is old enough. We don't need to worry about the CPU changing
1126 * state: If we see it offline even once, it has been through a
1129 * The reason for insisting that the grace period be at least
1130 * one jiffy old is that CPUs that are not quite online and that
1131 * have just gone offline can still execute RCU read-side critical
1134 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1135 return 0; /* Grace period is not old enough. */
1137 if (cpu_is_offline(rdp
->cpu
)) {
1138 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1144 * A CPU running for an extended time within the kernel can
1145 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1146 * even context-switching back and forth between a pair of
1147 * in-kernel CPU-bound tasks cannot advance grace periods.
1148 * So if the grace period is old enough, make the CPU pay attention.
1149 * Note that the unsynchronized assignments to the per-CPU
1150 * rcu_sched_qs_mask variable are safe. Yes, setting of
1151 * bits can be lost, but they will be set again on the next
1152 * force-quiescent-state pass. So lost bit sets do not result
1153 * in incorrect behavior, merely in a grace period lasting
1154 * a few jiffies longer than it might otherwise. Because
1155 * there are at most four threads involved, and because the
1156 * updates are only once every few jiffies, the probability of
1157 * lossage (and thus of slight grace-period extension) is
1160 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1161 * is set too high, we override with half of the RCU CPU stall
1164 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1165 if (ULONG_CMP_GE(jiffies
,
1166 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1167 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1168 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1169 WRITE_ONCE(rdp
->cond_resched_completed
,
1170 READ_ONCE(rdp
->mynode
->completed
));
1171 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1173 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1175 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1178 /* And if it has been a really long time, kick the CPU as well. */
1179 if (ULONG_CMP_GE(jiffies
,
1180 rdp
->rsp
->gp_start
+ 2 * jiffies_till_sched_qs
) ||
1181 ULONG_CMP_GE(jiffies
, rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
))
1182 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1187 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1189 unsigned long j
= jiffies
;
1193 smp_wmb(); /* Record start time before stall time. */
1194 j1
= rcu_jiffies_till_stall_check();
1195 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1196 rsp
->jiffies_resched
= j
+ j1
/ 2;
1197 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1201 * Convert a ->gp_state value to a character string.
1203 static const char *gp_state_getname(short gs
)
1205 if (gs
< 0 || gs
>= ARRAY_SIZE(gp_state_names
))
1207 return gp_state_names
[gs
];
1211 * Complain about starvation of grace-period kthread.
1213 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1219 gpa
= READ_ONCE(rsp
->gp_activity
);
1220 if (j
- gpa
> 2 * HZ
) {
1221 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1223 rsp
->gpnum
, rsp
->completed
,
1225 gp_state_getname(rsp
->gp_state
), rsp
->gp_state
,
1226 rsp
->gp_kthread
? rsp
->gp_kthread
->state
: ~0);
1227 if (rsp
->gp_kthread
) {
1228 sched_show_task(rsp
->gp_kthread
);
1229 wake_up_process(rsp
->gp_kthread
);
1235 * Dump stacks of all tasks running on stalled CPUs.
1237 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1240 unsigned long flags
;
1241 struct rcu_node
*rnp
;
1243 rcu_for_each_leaf_node(rsp
, rnp
) {
1244 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1245 if (rnp
->qsmask
!= 0) {
1246 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1247 if (rnp
->qsmask
& (1UL << cpu
))
1248 dump_cpu_task(rnp
->grplo
+ cpu
);
1250 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1254 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1258 unsigned long flags
;
1262 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1265 /* Only let one CPU complain about others per time interval. */
1267 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1268 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1269 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1270 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1273 WRITE_ONCE(rsp
->jiffies_stall
,
1274 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1275 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1278 * OK, time to rat on our buddy...
1279 * See Documentation/RCU/stallwarn.txt for info on how to debug
1280 * RCU CPU stall warnings.
1282 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1284 print_cpu_stall_info_begin();
1285 rcu_for_each_leaf_node(rsp
, rnp
) {
1286 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1287 ndetected
+= rcu_print_task_stall(rnp
);
1288 if (rnp
->qsmask
!= 0) {
1289 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
1290 if (rnp
->qsmask
& (1UL << cpu
)) {
1291 print_cpu_stall_info(rsp
,
1296 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1299 print_cpu_stall_info_end();
1300 for_each_possible_cpu(cpu
)
1301 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1302 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1303 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1304 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1306 rcu_dump_cpu_stacks(rsp
);
1308 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1309 READ_ONCE(rsp
->completed
) == gpnum
) {
1310 pr_err("INFO: Stall ended before state dump start\n");
1313 gpa
= READ_ONCE(rsp
->gp_activity
);
1314 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1315 rsp
->name
, j
- gpa
, j
, gpa
,
1316 jiffies_till_next_fqs
,
1317 rcu_get_root(rsp
)->qsmask
);
1318 /* In this case, the current CPU might be at fault. */
1319 sched_show_task(current
);
1323 /* Complain about tasks blocking the grace period. */
1324 rcu_print_detail_task_stall(rsp
);
1326 rcu_check_gp_kthread_starvation(rsp
);
1328 force_quiescent_state(rsp
); /* Kick them all. */
1331 static void print_cpu_stall(struct rcu_state
*rsp
)
1334 unsigned long flags
;
1335 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1339 * OK, time to rat on ourselves...
1340 * See Documentation/RCU/stallwarn.txt for info on how to debug
1341 * RCU CPU stall warnings.
1343 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1344 print_cpu_stall_info_begin();
1345 print_cpu_stall_info(rsp
, smp_processor_id());
1346 print_cpu_stall_info_end();
1347 for_each_possible_cpu(cpu
)
1348 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1349 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1350 jiffies
- rsp
->gp_start
,
1351 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1353 rcu_check_gp_kthread_starvation(rsp
);
1355 rcu_dump_cpu_stacks(rsp
);
1357 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1358 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1359 WRITE_ONCE(rsp
->jiffies_stall
,
1360 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1361 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1364 * Attempt to revive the RCU machinery by forcing a context switch.
1366 * A context switch would normally allow the RCU state machine to make
1367 * progress and it could be we're stuck in kernel space without context
1368 * switches for an entirely unreasonable amount of time.
1370 resched_cpu(smp_processor_id());
1373 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1375 unsigned long completed
;
1376 unsigned long gpnum
;
1380 struct rcu_node
*rnp
;
1382 if (rcu_cpu_stall_suppress
|| !rcu_gp_in_progress(rsp
))
1387 * Lots of memory barriers to reject false positives.
1389 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1390 * then rsp->gp_start, and finally rsp->completed. These values
1391 * are updated in the opposite order with memory barriers (or
1392 * equivalent) during grace-period initialization and cleanup.
1393 * Now, a false positive can occur if we get an new value of
1394 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1395 * the memory barriers, the only way that this can happen is if one
1396 * grace period ends and another starts between these two fetches.
1397 * Detect this by comparing rsp->completed with the previous fetch
1400 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1401 * and rsp->gp_start suffice to forestall false positives.
1403 gpnum
= READ_ONCE(rsp
->gpnum
);
1404 smp_rmb(); /* Pick up ->gpnum first... */
1405 js
= READ_ONCE(rsp
->jiffies_stall
);
1406 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1407 gps
= READ_ONCE(rsp
->gp_start
);
1408 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1409 completed
= READ_ONCE(rsp
->completed
);
1410 if (ULONG_CMP_GE(completed
, gpnum
) ||
1411 ULONG_CMP_LT(j
, js
) ||
1412 ULONG_CMP_GE(gps
, js
))
1413 return; /* No stall or GP completed since entering function. */
1415 if (rcu_gp_in_progress(rsp
) &&
1416 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1418 /* We haven't checked in, so go dump stack. */
1419 print_cpu_stall(rsp
);
1421 } else if (rcu_gp_in_progress(rsp
) &&
1422 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1424 /* They had a few time units to dump stack, so complain. */
1425 print_other_cpu_stall(rsp
, gpnum
);
1430 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1432 * Set the stall-warning timeout way off into the future, thus preventing
1433 * any RCU CPU stall-warning messages from appearing in the current set of
1434 * RCU grace periods.
1436 * The caller must disable hard irqs.
1438 void rcu_cpu_stall_reset(void)
1440 struct rcu_state
*rsp
;
1442 for_each_rcu_flavor(rsp
)
1443 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1447 * Initialize the specified rcu_data structure's default callback list
1448 * to empty. The default callback list is the one that is not used by
1449 * no-callbacks CPUs.
1451 static void init_default_callback_list(struct rcu_data
*rdp
)
1455 rdp
->nxtlist
= NULL
;
1456 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1457 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1461 * Initialize the specified rcu_data structure's callback list to empty.
1463 static void init_callback_list(struct rcu_data
*rdp
)
1465 if (init_nocb_callback_list(rdp
))
1467 init_default_callback_list(rdp
);
1471 * Determine the value that ->completed will have at the end of the
1472 * next subsequent grace period. This is used to tag callbacks so that
1473 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1474 * been dyntick-idle for an extended period with callbacks under the
1475 * influence of RCU_FAST_NO_HZ.
1477 * The caller must hold rnp->lock with interrupts disabled.
1479 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1480 struct rcu_node
*rnp
)
1483 * If RCU is idle, we just wait for the next grace period.
1484 * But we can only be sure that RCU is idle if we are looking
1485 * at the root rcu_node structure -- otherwise, a new grace
1486 * period might have started, but just not yet gotten around
1487 * to initializing the current non-root rcu_node structure.
1489 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1490 return rnp
->completed
+ 1;
1493 * Otherwise, wait for a possible partial grace period and
1494 * then the subsequent full grace period.
1496 return rnp
->completed
+ 2;
1500 * Trace-event helper function for rcu_start_future_gp() and
1501 * rcu_nocb_wait_gp().
1503 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1504 unsigned long c
, const char *s
)
1506 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1507 rnp
->completed
, c
, rnp
->level
,
1508 rnp
->grplo
, rnp
->grphi
, s
);
1512 * Start some future grace period, as needed to handle newly arrived
1513 * callbacks. The required future grace periods are recorded in each
1514 * rcu_node structure's ->need_future_gp field. Returns true if there
1515 * is reason to awaken the grace-period kthread.
1517 * The caller must hold the specified rcu_node structure's ->lock.
1519 static bool __maybe_unused
1520 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1521 unsigned long *c_out
)
1526 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1529 * Pick up grace-period number for new callbacks. If this
1530 * grace period is already marked as needed, return to the caller.
1532 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1533 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1534 if (rnp
->need_future_gp
[c
& 0x1]) {
1535 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1540 * If either this rcu_node structure or the root rcu_node structure
1541 * believe that a grace period is in progress, then we must wait
1542 * for the one following, which is in "c". Because our request
1543 * will be noticed at the end of the current grace period, we don't
1544 * need to explicitly start one. We only do the lockless check
1545 * of rnp_root's fields if the current rcu_node structure thinks
1546 * there is no grace period in flight, and because we hold rnp->lock,
1547 * the only possible change is when rnp_root's two fields are
1548 * equal, in which case rnp_root->gpnum might be concurrently
1549 * incremented. But that is OK, as it will just result in our
1550 * doing some extra useless work.
1552 if (rnp
->gpnum
!= rnp
->completed
||
1553 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1554 rnp
->need_future_gp
[c
& 0x1]++;
1555 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1560 * There might be no grace period in progress. If we don't already
1561 * hold it, acquire the root rcu_node structure's lock in order to
1562 * start one (if needed).
1564 if (rnp
!= rnp_root
)
1565 raw_spin_lock_rcu_node(rnp_root
);
1568 * Get a new grace-period number. If there really is no grace
1569 * period in progress, it will be smaller than the one we obtained
1570 * earlier. Adjust callbacks as needed. Note that even no-CBs
1571 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1573 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1574 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1575 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1576 rdp
->nxtcompleted
[i
] = c
;
1579 * If the needed for the required grace period is already
1580 * recorded, trace and leave.
1582 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1583 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1587 /* Record the need for the future grace period. */
1588 rnp_root
->need_future_gp
[c
& 0x1]++;
1590 /* If a grace period is not already in progress, start one. */
1591 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1592 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1594 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1595 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1598 if (rnp
!= rnp_root
)
1599 raw_spin_unlock_rcu_node(rnp_root
);
1607 * Clean up any old requests for the just-ended grace period. Also return
1608 * whether any additional grace periods have been requested. Also invoke
1609 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1610 * waiting for this grace period to complete.
1612 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1614 int c
= rnp
->completed
;
1616 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1618 rnp
->need_future_gp
[c
& 0x1] = 0;
1619 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1620 trace_rcu_future_gp(rnp
, rdp
, c
,
1621 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1626 * Awaken the grace-period kthread for the specified flavor of RCU.
1627 * Don't do a self-awaken, and don't bother awakening when there is
1628 * nothing for the grace-period kthread to do (as in several CPUs
1629 * raced to awaken, and we lost), and finally don't try to awaken
1630 * a kthread that has not yet been created.
1632 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1634 if (current
== rsp
->gp_kthread
||
1635 !READ_ONCE(rsp
->gp_flags
) ||
1638 swake_up(&rsp
->gp_wq
);
1642 * If there is room, assign a ->completed number to any callbacks on
1643 * this CPU that have not already been assigned. Also accelerate any
1644 * callbacks that were previously assigned a ->completed number that has
1645 * since proven to be too conservative, which can happen if callbacks get
1646 * assigned a ->completed number while RCU is idle, but with reference to
1647 * a non-root rcu_node structure. This function is idempotent, so it does
1648 * not hurt to call it repeatedly. Returns an flag saying that we should
1649 * awaken the RCU grace-period kthread.
1651 * The caller must hold rnp->lock with interrupts disabled.
1653 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1654 struct rcu_data
*rdp
)
1660 /* If the CPU has no callbacks, nothing to do. */
1661 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1665 * Starting from the sublist containing the callbacks most
1666 * recently assigned a ->completed number and working down, find the
1667 * first sublist that is not assignable to an upcoming grace period.
1668 * Such a sublist has something in it (first two tests) and has
1669 * a ->completed number assigned that will complete sooner than
1670 * the ->completed number for newly arrived callbacks (last test).
1672 * The key point is that any later sublist can be assigned the
1673 * same ->completed number as the newly arrived callbacks, which
1674 * means that the callbacks in any of these later sublist can be
1675 * grouped into a single sublist, whether or not they have already
1676 * been assigned a ->completed number.
1678 c
= rcu_cbs_completed(rsp
, rnp
);
1679 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1680 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1681 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1685 * If there are no sublist for unassigned callbacks, leave.
1686 * At the same time, advance "i" one sublist, so that "i" will
1687 * index into the sublist where all the remaining callbacks should
1690 if (++i
>= RCU_NEXT_TAIL
)
1694 * Assign all subsequent callbacks' ->completed number to the next
1695 * full grace period and group them all in the sublist initially
1698 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1699 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1700 rdp
->nxtcompleted
[i
] = c
;
1702 /* Record any needed additional grace periods. */
1703 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1705 /* Trace depending on how much we were able to accelerate. */
1706 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1707 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1709 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1714 * Move any callbacks whose grace period has completed to the
1715 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1716 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1717 * sublist. This function is idempotent, so it does not hurt to
1718 * invoke it repeatedly. As long as it is not invoked -too- often...
1719 * Returns true if the RCU grace-period kthread needs to be awakened.
1721 * The caller must hold rnp->lock with interrupts disabled.
1723 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1724 struct rcu_data
*rdp
)
1728 /* If the CPU has no callbacks, nothing to do. */
1729 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1733 * Find all callbacks whose ->completed numbers indicate that they
1734 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1736 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1737 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1739 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1741 /* Clean up any sublist tail pointers that were misordered above. */
1742 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1743 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1745 /* Copy down callbacks to fill in empty sublists. */
1746 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1747 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1749 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1750 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1753 /* Classify any remaining callbacks. */
1754 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1758 * Update CPU-local rcu_data state to record the beginnings and ends of
1759 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1760 * structure corresponding to the current CPU, and must have irqs disabled.
1761 * Returns true if the grace-period kthread needs to be awakened.
1763 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1764 struct rcu_data
*rdp
)
1768 /* Handle the ends of any preceding grace periods first. */
1769 if (rdp
->completed
== rnp
->completed
&&
1770 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1772 /* No grace period end, so just accelerate recent callbacks. */
1773 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1777 /* Advance callbacks. */
1778 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1780 /* Remember that we saw this grace-period completion. */
1781 rdp
->completed
= rnp
->completed
;
1782 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1785 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1787 * If the current grace period is waiting for this CPU,
1788 * set up to detect a quiescent state, otherwise don't
1789 * go looking for one.
1791 rdp
->gpnum
= rnp
->gpnum
;
1792 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1793 rdp
->cpu_no_qs
.b
.norm
= true;
1794 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1795 rdp
->core_needs_qs
= !!(rnp
->qsmask
& rdp
->grpmask
);
1796 zero_cpu_stall_ticks(rdp
);
1797 WRITE_ONCE(rdp
->gpwrap
, false);
1802 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1804 unsigned long flags
;
1806 struct rcu_node
*rnp
;
1808 local_irq_save(flags
);
1810 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1811 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1812 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1813 !raw_spin_trylock_rcu_node(rnp
)) { /* irqs already off, so later. */
1814 local_irq_restore(flags
);
1817 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1818 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1820 rcu_gp_kthread_wake(rsp
);
1823 static void rcu_gp_slow(struct rcu_state
*rsp
, int delay
)
1826 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1827 schedule_timeout_uninterruptible(delay
);
1831 * Initialize a new grace period. Return false if no grace period required.
1833 static bool rcu_gp_init(struct rcu_state
*rsp
)
1835 unsigned long oldmask
;
1836 struct rcu_data
*rdp
;
1837 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1839 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1840 raw_spin_lock_irq_rcu_node(rnp
);
1841 if (!READ_ONCE(rsp
->gp_flags
)) {
1842 /* Spurious wakeup, tell caller to go back to sleep. */
1843 raw_spin_unlock_irq_rcu_node(rnp
);
1846 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1848 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1850 * Grace period already in progress, don't start another.
1851 * Not supposed to be able to happen.
1853 raw_spin_unlock_irq_rcu_node(rnp
);
1857 /* Advance to a new grace period and initialize state. */
1858 record_gp_stall_check_time(rsp
);
1859 /* Record GP times before starting GP, hence smp_store_release(). */
1860 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1861 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1862 raw_spin_unlock_irq_rcu_node(rnp
);
1865 * Apply per-leaf buffered online and offline operations to the
1866 * rcu_node tree. Note that this new grace period need not wait
1867 * for subsequent online CPUs, and that quiescent-state forcing
1868 * will handle subsequent offline CPUs.
1870 rcu_for_each_leaf_node(rsp
, rnp
) {
1871 rcu_gp_slow(rsp
, gp_preinit_delay
);
1872 raw_spin_lock_irq_rcu_node(rnp
);
1873 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1874 !rnp
->wait_blkd_tasks
) {
1875 /* Nothing to do on this leaf rcu_node structure. */
1876 raw_spin_unlock_irq_rcu_node(rnp
);
1880 /* Record old state, apply changes to ->qsmaskinit field. */
1881 oldmask
= rnp
->qsmaskinit
;
1882 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1884 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1885 if (!oldmask
!= !rnp
->qsmaskinit
) {
1886 if (!oldmask
) /* First online CPU for this rcu_node. */
1887 rcu_init_new_rnp(rnp
);
1888 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1889 rnp
->wait_blkd_tasks
= true;
1890 else /* Last offline CPU and can propagate. */
1891 rcu_cleanup_dead_rnp(rnp
);
1895 * If all waited-on tasks from prior grace period are
1896 * done, and if all this rcu_node structure's CPUs are
1897 * still offline, propagate up the rcu_node tree and
1898 * clear ->wait_blkd_tasks. Otherwise, if one of this
1899 * rcu_node structure's CPUs has since come back online,
1900 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1901 * checks for this, so just call it unconditionally).
1903 if (rnp
->wait_blkd_tasks
&&
1904 (!rcu_preempt_has_tasks(rnp
) ||
1906 rnp
->wait_blkd_tasks
= false;
1907 rcu_cleanup_dead_rnp(rnp
);
1910 raw_spin_unlock_irq_rcu_node(rnp
);
1914 * Set the quiescent-state-needed bits in all the rcu_node
1915 * structures for all currently online CPUs in breadth-first order,
1916 * starting from the root rcu_node structure, relying on the layout
1917 * of the tree within the rsp->node[] array. Note that other CPUs
1918 * will access only the leaves of the hierarchy, thus seeing that no
1919 * grace period is in progress, at least until the corresponding
1920 * leaf node has been initialized. In addition, we have excluded
1921 * CPU-hotplug operations.
1923 * The grace period cannot complete until the initialization
1924 * process finishes, because this kthread handles both.
1926 rcu_for_each_node_breadth_first(rsp
, rnp
) {
1927 rcu_gp_slow(rsp
, gp_init_delay
);
1928 raw_spin_lock_irq_rcu_node(rnp
);
1929 rdp
= this_cpu_ptr(rsp
->rda
);
1930 rcu_preempt_check_blocked_tasks(rnp
);
1931 rnp
->qsmask
= rnp
->qsmaskinit
;
1932 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
1933 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
1934 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
1935 if (rnp
== rdp
->mynode
)
1936 (void)__note_gp_changes(rsp
, rnp
, rdp
);
1937 rcu_preempt_boost_start_gp(rnp
);
1938 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
1939 rnp
->level
, rnp
->grplo
,
1940 rnp
->grphi
, rnp
->qsmask
);
1941 raw_spin_unlock_irq_rcu_node(rnp
);
1942 cond_resched_rcu_qs();
1943 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1950 * Helper function for wait_event_interruptible_timeout() wakeup
1951 * at force-quiescent-state time.
1953 static bool rcu_gp_fqs_check_wake(struct rcu_state
*rsp
, int *gfp
)
1955 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1957 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1958 *gfp
= READ_ONCE(rsp
->gp_flags
);
1959 if (*gfp
& RCU_GP_FLAG_FQS
)
1962 /* The current grace period has completed. */
1963 if (!READ_ONCE(rnp
->qsmask
) && !rcu_preempt_blocked_readers_cgp(rnp
))
1970 * Do one round of quiescent-state forcing.
1972 static void rcu_gp_fqs(struct rcu_state
*rsp
, bool first_time
)
1974 bool isidle
= false;
1976 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1978 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1981 /* Collect dyntick-idle snapshots. */
1982 if (is_sysidle_rcu_state(rsp
)) {
1984 maxj
= jiffies
- ULONG_MAX
/ 4;
1986 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
1988 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
1990 /* Handle dyntick-idle and offline CPUs. */
1992 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
1994 /* Clear flag to prevent immediate re-entry. */
1995 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
1996 raw_spin_lock_irq_rcu_node(rnp
);
1997 WRITE_ONCE(rsp
->gp_flags
,
1998 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
1999 raw_spin_unlock_irq_rcu_node(rnp
);
2004 * Clean up after the old grace period.
2006 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
2008 unsigned long gp_duration
;
2009 bool needgp
= false;
2011 struct rcu_data
*rdp
;
2012 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2013 struct swait_queue_head
*sq
;
2015 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2016 raw_spin_lock_irq_rcu_node(rnp
);
2017 gp_duration
= jiffies
- rsp
->gp_start
;
2018 if (gp_duration
> rsp
->gp_max
)
2019 rsp
->gp_max
= gp_duration
;
2022 * We know the grace period is complete, but to everyone else
2023 * it appears to still be ongoing. But it is also the case
2024 * that to everyone else it looks like there is nothing that
2025 * they can do to advance the grace period. It is therefore
2026 * safe for us to drop the lock in order to mark the grace
2027 * period as completed in all of the rcu_node structures.
2029 raw_spin_unlock_irq_rcu_node(rnp
);
2032 * Propagate new ->completed value to rcu_node structures so
2033 * that other CPUs don't have to wait until the start of the next
2034 * grace period to process their callbacks. This also avoids
2035 * some nasty RCU grace-period initialization races by forcing
2036 * the end of the current grace period to be completely recorded in
2037 * all of the rcu_node structures before the beginning of the next
2038 * grace period is recorded in any of the rcu_node structures.
2040 rcu_for_each_node_breadth_first(rsp
, rnp
) {
2041 raw_spin_lock_irq_rcu_node(rnp
);
2042 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
2043 WARN_ON_ONCE(rnp
->qsmask
);
2044 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
2045 rdp
= this_cpu_ptr(rsp
->rda
);
2046 if (rnp
== rdp
->mynode
)
2047 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
2048 /* smp_mb() provided by prior unlock-lock pair. */
2049 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
2050 sq
= rcu_nocb_gp_get(rnp
);
2051 raw_spin_unlock_irq_rcu_node(rnp
);
2052 rcu_nocb_gp_cleanup(sq
);
2053 cond_resched_rcu_qs();
2054 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2055 rcu_gp_slow(rsp
, gp_cleanup_delay
);
2057 rnp
= rcu_get_root(rsp
);
2058 raw_spin_lock_irq_rcu_node(rnp
); /* Order GP before ->completed update. */
2059 rcu_nocb_gp_set(rnp
, nocb
);
2061 /* Declare grace period done. */
2062 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
2063 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
2064 rsp
->gp_state
= RCU_GP_IDLE
;
2065 rdp
= this_cpu_ptr(rsp
->rda
);
2066 /* Advance CBs to reduce false positives below. */
2067 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
2068 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
2069 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2070 trace_rcu_grace_period(rsp
->name
,
2071 READ_ONCE(rsp
->gpnum
),
2074 raw_spin_unlock_irq_rcu_node(rnp
);
2078 * Body of kthread that handles grace periods.
2080 static int __noreturn
rcu_gp_kthread(void *arg
)
2086 struct rcu_state
*rsp
= arg
;
2087 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2089 rcu_bind_gp_kthread();
2092 /* Handle grace-period start. */
2094 trace_rcu_grace_period(rsp
->name
,
2095 READ_ONCE(rsp
->gpnum
),
2097 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2098 swait_event_interruptible(rsp
->gp_wq
,
2099 READ_ONCE(rsp
->gp_flags
) &
2101 rsp
->gp_state
= RCU_GP_DONE_GPS
;
2102 /* Locking provides needed memory barrier. */
2103 if (rcu_gp_init(rsp
))
2105 cond_resched_rcu_qs();
2106 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2107 WARN_ON(signal_pending(current
));
2108 trace_rcu_grace_period(rsp
->name
,
2109 READ_ONCE(rsp
->gpnum
),
2113 /* Handle quiescent-state forcing. */
2114 first_gp_fqs
= true;
2115 j
= jiffies_till_first_fqs
;
2118 jiffies_till_first_fqs
= HZ
;
2123 rsp
->jiffies_force_qs
= jiffies
+ j
;
2124 trace_rcu_grace_period(rsp
->name
,
2125 READ_ONCE(rsp
->gpnum
),
2127 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2128 ret
= swait_event_interruptible_timeout(rsp
->gp_wq
,
2129 rcu_gp_fqs_check_wake(rsp
, &gf
), j
);
2130 rsp
->gp_state
= RCU_GP_DOING_FQS
;
2131 /* Locking provides needed memory barriers. */
2132 /* If grace period done, leave loop. */
2133 if (!READ_ONCE(rnp
->qsmask
) &&
2134 !rcu_preempt_blocked_readers_cgp(rnp
))
2136 /* If time for quiescent-state forcing, do it. */
2137 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2138 (gf
& RCU_GP_FLAG_FQS
)) {
2139 trace_rcu_grace_period(rsp
->name
,
2140 READ_ONCE(rsp
->gpnum
),
2142 rcu_gp_fqs(rsp
, first_gp_fqs
);
2143 first_gp_fqs
= false;
2144 trace_rcu_grace_period(rsp
->name
,
2145 READ_ONCE(rsp
->gpnum
),
2147 cond_resched_rcu_qs();
2148 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2150 /* Deal with stray signal. */
2151 cond_resched_rcu_qs();
2152 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2153 WARN_ON(signal_pending(current
));
2154 trace_rcu_grace_period(rsp
->name
,
2155 READ_ONCE(rsp
->gpnum
),
2158 j
= jiffies_till_next_fqs
;
2161 jiffies_till_next_fqs
= HZ
;
2164 jiffies_till_next_fqs
= 1;
2168 /* Handle grace-period end. */
2169 rsp
->gp_state
= RCU_GP_CLEANUP
;
2170 rcu_gp_cleanup(rsp
);
2171 rsp
->gp_state
= RCU_GP_CLEANED
;
2176 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2177 * in preparation for detecting the next grace period. The caller must hold
2178 * the root node's ->lock and hard irqs must be disabled.
2180 * Note that it is legal for a dying CPU (which is marked as offline) to
2181 * invoke this function. This can happen when the dying CPU reports its
2184 * Returns true if the grace-period kthread must be awakened.
2187 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2188 struct rcu_data
*rdp
)
2190 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2192 * Either we have not yet spawned the grace-period
2193 * task, this CPU does not need another grace period,
2194 * or a grace period is already in progress.
2195 * Either way, don't start a new grace period.
2199 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2200 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2204 * We can't do wakeups while holding the rnp->lock, as that
2205 * could cause possible deadlocks with the rq->lock. Defer
2206 * the wakeup to our caller.
2212 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2213 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2214 * is invoked indirectly from rcu_advance_cbs(), which would result in
2215 * endless recursion -- or would do so if it wasn't for the self-deadlock
2216 * that is encountered beforehand.
2218 * Returns true if the grace-period kthread needs to be awakened.
2220 static bool rcu_start_gp(struct rcu_state
*rsp
)
2222 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2223 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2227 * If there is no grace period in progress right now, any
2228 * callbacks we have up to this point will be satisfied by the
2229 * next grace period. Also, advancing the callbacks reduces the
2230 * probability of false positives from cpu_needs_another_gp()
2231 * resulting in pointless grace periods. So, advance callbacks
2232 * then start the grace period!
2234 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2235 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2240 * Report a full set of quiescent states to the specified rcu_state data
2241 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2242 * kthread if another grace period is required. Whether we wake
2243 * the grace-period kthread or it awakens itself for the next round
2244 * of quiescent-state forcing, that kthread will clean up after the
2245 * just-completed grace period. Note that the caller must hold rnp->lock,
2246 * which is released before return.
2248 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2249 __releases(rcu_get_root(rsp
)->lock
)
2251 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2252 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2253 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2254 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2258 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2259 * Allows quiescent states for a group of CPUs to be reported at one go
2260 * to the specified rcu_node structure, though all the CPUs in the group
2261 * must be represented by the same rcu_node structure (which need not be a
2262 * leaf rcu_node structure, though it often will be). The gps parameter
2263 * is the grace-period snapshot, which means that the quiescent states
2264 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2265 * must be held upon entry, and it is released before return.
2268 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2269 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2270 __releases(rnp
->lock
)
2272 unsigned long oldmask
= 0;
2273 struct rcu_node
*rnp_c
;
2275 /* Walk up the rcu_node hierarchy. */
2277 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2280 * Our bit has already been cleared, or the
2281 * relevant grace period is already over, so done.
2283 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2286 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2287 rnp
->qsmask
&= ~mask
;
2288 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2289 mask
, rnp
->qsmask
, rnp
->level
,
2290 rnp
->grplo
, rnp
->grphi
,
2292 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2294 /* Other bits still set at this level, so done. */
2295 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2298 mask
= rnp
->grpmask
;
2299 if (rnp
->parent
== NULL
) {
2301 /* No more levels. Exit loop holding root lock. */
2305 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2308 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2309 oldmask
= rnp_c
->qsmask
;
2313 * Get here if we are the last CPU to pass through a quiescent
2314 * state for this grace period. Invoke rcu_report_qs_rsp()
2315 * to clean up and start the next grace period if one is needed.
2317 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2321 * Record a quiescent state for all tasks that were previously queued
2322 * on the specified rcu_node structure and that were blocking the current
2323 * RCU grace period. The caller must hold the specified rnp->lock with
2324 * irqs disabled, and this lock is released upon return, but irqs remain
2327 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2328 struct rcu_node
*rnp
, unsigned long flags
)
2329 __releases(rnp
->lock
)
2333 struct rcu_node
*rnp_p
;
2335 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2336 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2337 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2338 return; /* Still need more quiescent states! */
2341 rnp_p
= rnp
->parent
;
2342 if (rnp_p
== NULL
) {
2344 * Only one rcu_node structure in the tree, so don't
2345 * try to report up to its nonexistent parent!
2347 rcu_report_qs_rsp(rsp
, flags
);
2351 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2353 mask
= rnp
->grpmask
;
2354 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2355 raw_spin_lock_rcu_node(rnp_p
); /* irqs already disabled. */
2356 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2360 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2361 * structure. This must be called from the specified CPU.
2364 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2366 unsigned long flags
;
2369 struct rcu_node
*rnp
;
2372 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2373 if ((rdp
->cpu_no_qs
.b
.norm
&&
2374 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2375 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2379 * The grace period in which this quiescent state was
2380 * recorded has ended, so don't report it upwards.
2381 * We will instead need a new quiescent state that lies
2382 * within the current grace period.
2384 rdp
->cpu_no_qs
.b
.norm
= true; /* need qs for new gp. */
2385 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2386 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2389 mask
= rdp
->grpmask
;
2390 if ((rnp
->qsmask
& mask
) == 0) {
2391 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2393 rdp
->core_needs_qs
= false;
2396 * This GP can't end until cpu checks in, so all of our
2397 * callbacks can be processed during the next GP.
2399 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2401 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2402 /* ^^^ Released rnp->lock */
2404 rcu_gp_kthread_wake(rsp
);
2409 * Check to see if there is a new grace period of which this CPU
2410 * is not yet aware, and if so, set up local rcu_data state for it.
2411 * Otherwise, see if this CPU has just passed through its first
2412 * quiescent state for this grace period, and record that fact if so.
2415 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2417 /* Check for grace-period ends and beginnings. */
2418 note_gp_changes(rsp
, rdp
);
2421 * Does this CPU still need to do its part for current grace period?
2422 * If no, return and let the other CPUs do their part as well.
2424 if (!rdp
->core_needs_qs
)
2428 * Was there a quiescent state since the beginning of the grace
2429 * period? If no, then exit and wait for the next call.
2431 if (rdp
->cpu_no_qs
.b
.norm
&&
2432 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2436 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2439 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2443 * Send the specified CPU's RCU callbacks to the orphanage. The
2444 * specified CPU must be offline, and the caller must hold the
2448 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2449 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2451 /* No-CBs CPUs do not have orphanable callbacks. */
2452 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) || rcu_is_nocb_cpu(rdp
->cpu
))
2456 * Orphan the callbacks. First adjust the counts. This is safe
2457 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2458 * cannot be running now. Thus no memory barrier is required.
2460 if (rdp
->nxtlist
!= NULL
) {
2461 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2462 rsp
->qlen
+= rdp
->qlen
;
2463 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2465 WRITE_ONCE(rdp
->qlen
, 0);
2469 * Next, move those callbacks still needing a grace period to
2470 * the orphanage, where some other CPU will pick them up.
2471 * Some of the callbacks might have gone partway through a grace
2472 * period, but that is too bad. They get to start over because we
2473 * cannot assume that grace periods are synchronized across CPUs.
2474 * We don't bother updating the ->nxttail[] array yet, instead
2475 * we just reset the whole thing later on.
2477 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2478 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2479 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2480 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2484 * Then move the ready-to-invoke callbacks to the orphanage,
2485 * where some other CPU will pick them up. These will not be
2486 * required to pass though another grace period: They are done.
2488 if (rdp
->nxtlist
!= NULL
) {
2489 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2490 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2494 * Finally, initialize the rcu_data structure's list to empty and
2495 * disallow further callbacks on this CPU.
2497 init_callback_list(rdp
);
2498 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2502 * Adopt the RCU callbacks from the specified rcu_state structure's
2503 * orphanage. The caller must hold the ->orphan_lock.
2505 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2508 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2510 /* No-CBs CPUs are handled specially. */
2511 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2512 rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2515 /* Do the accounting first. */
2516 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2517 rdp
->qlen
+= rsp
->qlen
;
2518 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2519 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2520 rcu_idle_count_callbacks_posted();
2525 * We do not need a memory barrier here because the only way we
2526 * can get here if there is an rcu_barrier() in flight is if
2527 * we are the task doing the rcu_barrier().
2530 /* First adopt the ready-to-invoke callbacks. */
2531 if (rsp
->orphan_donelist
!= NULL
) {
2532 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2533 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2534 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2535 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2536 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2537 rsp
->orphan_donelist
= NULL
;
2538 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2541 /* And then adopt the callbacks that still need a grace period. */
2542 if (rsp
->orphan_nxtlist
!= NULL
) {
2543 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2544 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2545 rsp
->orphan_nxtlist
= NULL
;
2546 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2551 * Trace the fact that this CPU is going offline.
2553 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2555 RCU_TRACE(unsigned long mask
);
2556 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2557 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2559 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2562 RCU_TRACE(mask
= rdp
->grpmask
);
2563 trace_rcu_grace_period(rsp
->name
,
2564 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2569 * All CPUs for the specified rcu_node structure have gone offline,
2570 * and all tasks that were preempted within an RCU read-side critical
2571 * section while running on one of those CPUs have since exited their RCU
2572 * read-side critical section. Some other CPU is reporting this fact with
2573 * the specified rcu_node structure's ->lock held and interrupts disabled.
2574 * This function therefore goes up the tree of rcu_node structures,
2575 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2576 * the leaf rcu_node structure's ->qsmaskinit field has already been
2579 * This function does check that the specified rcu_node structure has
2580 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2581 * prematurely. That said, invoking it after the fact will cost you
2582 * a needless lock acquisition. So once it has done its work, don't
2585 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2588 struct rcu_node
*rnp
= rnp_leaf
;
2590 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2591 rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2594 mask
= rnp
->grpmask
;
2598 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
2599 rnp
->qsmaskinit
&= ~mask
;
2600 rnp
->qsmask
&= ~mask
;
2601 if (rnp
->qsmaskinit
) {
2602 raw_spin_unlock_rcu_node(rnp
);
2603 /* irqs remain disabled. */
2606 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2611 * The CPU has been completely removed, and some other CPU is reporting
2612 * this fact from process context. Do the remainder of the cleanup,
2613 * including orphaning the outgoing CPU's RCU callbacks, and also
2614 * adopting them. There can only be one CPU hotplug operation at a time,
2615 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2617 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2619 unsigned long flags
;
2620 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2621 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2623 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2626 /* Adjust any no-longer-needed kthreads. */
2627 rcu_boost_kthread_setaffinity(rnp
, -1);
2629 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2630 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2631 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2632 rcu_adopt_orphan_cbs(rsp
, flags
);
2633 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2635 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2636 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2637 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2641 * Invoke any RCU callbacks that have made it to the end of their grace
2642 * period. Thottle as specified by rdp->blimit.
2644 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2646 unsigned long flags
;
2647 struct rcu_head
*next
, *list
, **tail
;
2648 long bl
, count
, count_lazy
;
2651 /* If no callbacks are ready, just return. */
2652 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2653 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2654 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2655 need_resched(), is_idle_task(current
),
2656 rcu_is_callbacks_kthread());
2661 * Extract the list of ready callbacks, disabling to prevent
2662 * races with call_rcu() from interrupt handlers.
2664 local_irq_save(flags
);
2665 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2667 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2668 list
= rdp
->nxtlist
;
2669 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2670 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2671 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2672 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2673 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2674 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2675 local_irq_restore(flags
);
2677 /* Invoke callbacks. */
2678 count
= count_lazy
= 0;
2682 debug_rcu_head_unqueue(list
);
2683 if (__rcu_reclaim(rsp
->name
, list
))
2686 /* Stop only if limit reached and CPU has something to do. */
2687 if (++count
>= bl
&&
2689 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2693 local_irq_save(flags
);
2694 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2695 is_idle_task(current
),
2696 rcu_is_callbacks_kthread());
2698 /* Update count, and requeue any remaining callbacks. */
2700 *tail
= rdp
->nxtlist
;
2701 rdp
->nxtlist
= list
;
2702 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2703 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2704 rdp
->nxttail
[i
] = tail
;
2708 smp_mb(); /* List handling before counting for rcu_barrier(). */
2709 rdp
->qlen_lazy
-= count_lazy
;
2710 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2711 rdp
->n_cbs_invoked
+= count
;
2713 /* Reinstate batch limit if we have worked down the excess. */
2714 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2715 rdp
->blimit
= blimit
;
2717 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2718 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2719 rdp
->qlen_last_fqs_check
= 0;
2720 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2721 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2722 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2723 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2725 local_irq_restore(flags
);
2727 /* Re-invoke RCU core processing if there are callbacks remaining. */
2728 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2733 * Check to see if this CPU is in a non-context-switch quiescent state
2734 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2735 * Also schedule RCU core processing.
2737 * This function must be called from hardirq context. It is normally
2738 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2739 * false, there is no point in invoking rcu_check_callbacks().
2741 void rcu_check_callbacks(int user
)
2743 trace_rcu_utilization(TPS("Start scheduler-tick"));
2744 increment_cpu_stall_ticks();
2745 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2748 * Get here if this CPU took its interrupt from user
2749 * mode or from the idle loop, and if this is not a
2750 * nested interrupt. In this case, the CPU is in
2751 * a quiescent state, so note it.
2753 * No memory barrier is required here because both
2754 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2755 * variables that other CPUs neither access nor modify,
2756 * at least not while the corresponding CPU is online.
2762 } else if (!in_softirq()) {
2765 * Get here if this CPU did not take its interrupt from
2766 * softirq, in other words, if it is not interrupting
2767 * a rcu_bh read-side critical section. This is an _bh
2768 * critical section, so note it.
2773 rcu_preempt_check_callbacks();
2777 rcu_note_voluntary_context_switch(current
);
2778 trace_rcu_utilization(TPS("End scheduler-tick"));
2782 * Scan the leaf rcu_node structures, processing dyntick state for any that
2783 * have not yet encountered a quiescent state, using the function specified.
2784 * Also initiate boosting for any threads blocked on the root rcu_node.
2786 * The caller must have suppressed start of new grace periods.
2788 static void force_qs_rnp(struct rcu_state
*rsp
,
2789 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2790 unsigned long *maxj
),
2791 bool *isidle
, unsigned long *maxj
)
2795 unsigned long flags
;
2797 struct rcu_node
*rnp
;
2799 rcu_for_each_leaf_node(rsp
, rnp
) {
2800 cond_resched_rcu_qs();
2802 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2803 if (rnp
->qsmask
== 0) {
2804 if (rcu_state_p
== &rcu_sched_state
||
2805 rsp
!= rcu_state_p
||
2806 rcu_preempt_blocked_readers_cgp(rnp
)) {
2808 * No point in scanning bits because they
2809 * are all zero. But we might need to
2810 * priority-boost blocked readers.
2812 rcu_initiate_boost(rnp
, flags
);
2813 /* rcu_initiate_boost() releases rnp->lock */
2817 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2819 * Race between grace-period
2820 * initialization and task exiting RCU
2821 * read-side critical section: Report.
2823 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2824 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2830 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
2831 if ((rnp
->qsmask
& bit
) != 0) {
2832 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2837 /* Idle/offline CPUs, report (releases rnp->lock. */
2838 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2840 /* Nothing to do here, so just drop the lock. */
2841 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2847 * Force quiescent states on reluctant CPUs, and also detect which
2848 * CPUs are in dyntick-idle mode.
2850 static void force_quiescent_state(struct rcu_state
*rsp
)
2852 unsigned long flags
;
2854 struct rcu_node
*rnp
;
2855 struct rcu_node
*rnp_old
= NULL
;
2857 /* Funnel through hierarchy to reduce memory contention. */
2858 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2859 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2860 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2861 !raw_spin_trylock(&rnp
->fqslock
);
2862 if (rnp_old
!= NULL
)
2863 raw_spin_unlock(&rnp_old
->fqslock
);
2865 rsp
->n_force_qs_lh
++;
2870 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2872 /* Reached the root of the rcu_node tree, acquire lock. */
2873 raw_spin_lock_irqsave_rcu_node(rnp_old
, flags
);
2874 raw_spin_unlock(&rnp_old
->fqslock
);
2875 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2876 rsp
->n_force_qs_lh
++;
2877 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2878 return; /* Someone beat us to it. */
2880 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2881 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2882 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2886 * This does the RCU core processing work for the specified rcu_state
2887 * and rcu_data structures. This may be called only from the CPU to
2888 * whom the rdp belongs.
2891 __rcu_process_callbacks(struct rcu_state
*rsp
)
2893 unsigned long flags
;
2895 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2897 WARN_ON_ONCE(rdp
->beenonline
== 0);
2899 /* Update RCU state based on any recent quiescent states. */
2900 rcu_check_quiescent_state(rsp
, rdp
);
2902 /* Does this CPU require a not-yet-started grace period? */
2903 local_irq_save(flags
);
2904 if (cpu_needs_another_gp(rsp
, rdp
)) {
2905 raw_spin_lock_rcu_node(rcu_get_root(rsp
)); /* irqs disabled. */
2906 needwake
= rcu_start_gp(rsp
);
2907 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2909 rcu_gp_kthread_wake(rsp
);
2911 local_irq_restore(flags
);
2914 /* If there are callbacks ready, invoke them. */
2915 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2916 invoke_rcu_callbacks(rsp
, rdp
);
2918 /* Do any needed deferred wakeups of rcuo kthreads. */
2919 do_nocb_deferred_wakeup(rdp
);
2923 * Do RCU core processing for the current CPU.
2925 static void rcu_process_callbacks(struct softirq_action
*unused
)
2927 struct rcu_state
*rsp
;
2929 if (cpu_is_offline(smp_processor_id()))
2931 trace_rcu_utilization(TPS("Start RCU core"));
2932 for_each_rcu_flavor(rsp
)
2933 __rcu_process_callbacks(rsp
);
2934 trace_rcu_utilization(TPS("End RCU core"));
2938 * Schedule RCU callback invocation. If the specified type of RCU
2939 * does not support RCU priority boosting, just do a direct call,
2940 * otherwise wake up the per-CPU kernel kthread. Note that because we
2941 * are running on the current CPU with softirqs disabled, the
2942 * rcu_cpu_kthread_task cannot disappear out from under us.
2944 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2946 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
2948 if (likely(!rsp
->boost
)) {
2949 rcu_do_batch(rsp
, rdp
);
2952 invoke_rcu_callbacks_kthread();
2955 static void invoke_rcu_core(void)
2957 if (cpu_online(smp_processor_id()))
2958 raise_softirq(RCU_SOFTIRQ
);
2962 * Handle any core-RCU processing required by a call_rcu() invocation.
2964 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
2965 struct rcu_head
*head
, unsigned long flags
)
2970 * If called from an extended quiescent state, invoke the RCU
2971 * core in order to force a re-evaluation of RCU's idleness.
2973 if (!rcu_is_watching())
2976 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2977 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2981 * Force the grace period if too many callbacks or too long waiting.
2982 * Enforce hysteresis, and don't invoke force_quiescent_state()
2983 * if some other CPU has recently done so. Also, don't bother
2984 * invoking force_quiescent_state() if the newly enqueued callback
2985 * is the only one waiting for a grace period to complete.
2987 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
2989 /* Are we ignoring a completed grace period? */
2990 note_gp_changes(rsp
, rdp
);
2992 /* Start a new grace period if one not already started. */
2993 if (!rcu_gp_in_progress(rsp
)) {
2994 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
2996 raw_spin_lock_rcu_node(rnp_root
);
2997 needwake
= rcu_start_gp(rsp
);
2998 raw_spin_unlock_rcu_node(rnp_root
);
3000 rcu_gp_kthread_wake(rsp
);
3002 /* Give the grace period a kick. */
3003 rdp
->blimit
= LONG_MAX
;
3004 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
3005 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
3006 force_quiescent_state(rsp
);
3007 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3008 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
3014 * RCU callback function to leak a callback.
3016 static void rcu_leak_callback(struct rcu_head
*rhp
)
3021 * Helper function for call_rcu() and friends. The cpu argument will
3022 * normally be -1, indicating "currently running CPU". It may specify
3023 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3024 * is expected to specify a CPU.
3027 __call_rcu(struct rcu_head
*head
, rcu_callback_t func
,
3028 struct rcu_state
*rsp
, int cpu
, bool lazy
)
3030 unsigned long flags
;
3031 struct rcu_data
*rdp
;
3033 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
3034 if (debug_rcu_head_queue(head
)) {
3035 /* Probable double call_rcu(), so leak the callback. */
3036 WRITE_ONCE(head
->func
, rcu_leak_callback
);
3037 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3044 * Opportunistically note grace-period endings and beginnings.
3045 * Note that we might see a beginning right after we see an
3046 * end, but never vice versa, since this CPU has to pass through
3047 * a quiescent state betweentimes.
3049 local_irq_save(flags
);
3050 rdp
= this_cpu_ptr(rsp
->rda
);
3052 /* Add the callback to our list. */
3053 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3057 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3058 if (likely(rdp
->mynode
)) {
3059 /* Post-boot, so this should be for a no-CBs CPU. */
3060 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3061 WARN_ON_ONCE(offline
);
3062 /* Offline CPU, _call_rcu() illegal, leak callback. */
3063 local_irq_restore(flags
);
3067 * Very early boot, before rcu_init(). Initialize if needed
3068 * and then drop through to queue the callback.
3071 WARN_ON_ONCE(!rcu_is_watching());
3072 if (!likely(rdp
->nxtlist
))
3073 init_default_callback_list(rdp
);
3075 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3079 rcu_idle_count_callbacks_posted();
3080 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3081 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3082 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3084 if (__is_kfree_rcu_offset((unsigned long)func
))
3085 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3086 rdp
->qlen_lazy
, rdp
->qlen
);
3088 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3090 /* Go handle any RCU core processing required. */
3091 __call_rcu_core(rsp
, rdp
, head
, flags
);
3092 local_irq_restore(flags
);
3096 * Queue an RCU-sched callback for invocation after a grace period.
3098 void call_rcu_sched(struct rcu_head
*head
, rcu_callback_t func
)
3100 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3102 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3105 * Queue an RCU callback for invocation after a quicker grace period.
3107 void call_rcu_bh(struct rcu_head
*head
, rcu_callback_t func
)
3109 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3111 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3114 * Queue an RCU callback for lazy invocation after a grace period.
3115 * This will likely be later named something like "call_rcu_lazy()",
3116 * but this change will require some way of tagging the lazy RCU
3117 * callbacks in the list of pending callbacks. Until then, this
3118 * function may only be called from __kfree_rcu().
3120 void kfree_call_rcu(struct rcu_head
*head
,
3121 rcu_callback_t func
)
3123 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3125 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3128 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3129 * any blocking grace-period wait automatically implies a grace period
3130 * if there is only one CPU online at any point time during execution
3131 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3132 * occasionally incorrectly indicate that there are multiple CPUs online
3133 * when there was in fact only one the whole time, as this just adds
3134 * some overhead: RCU still operates correctly.
3136 static inline int rcu_blocking_is_gp(void)
3140 might_sleep(); /* Check for RCU read-side critical section. */
3142 ret
= num_online_cpus() <= 1;
3148 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3150 * Control will return to the caller some time after a full rcu-sched
3151 * grace period has elapsed, in other words after all currently executing
3152 * rcu-sched read-side critical sections have completed. These read-side
3153 * critical sections are delimited by rcu_read_lock_sched() and
3154 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3155 * local_irq_disable(), and so on may be used in place of
3156 * rcu_read_lock_sched().
3158 * This means that all preempt_disable code sequences, including NMI and
3159 * non-threaded hardware-interrupt handlers, in progress on entry will
3160 * have completed before this primitive returns. However, this does not
3161 * guarantee that softirq handlers will have completed, since in some
3162 * kernels, these handlers can run in process context, and can block.
3164 * Note that this guarantee implies further memory-ordering guarantees.
3165 * On systems with more than one CPU, when synchronize_sched() returns,
3166 * each CPU is guaranteed to have executed a full memory barrier since the
3167 * end of its last RCU-sched read-side critical section whose beginning
3168 * preceded the call to synchronize_sched(). In addition, each CPU having
3169 * an RCU read-side critical section that extends beyond the return from
3170 * synchronize_sched() is guaranteed to have executed a full memory barrier
3171 * after the beginning of synchronize_sched() and before the beginning of
3172 * that RCU read-side critical section. Note that these guarantees include
3173 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3174 * that are executing in the kernel.
3176 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3177 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3178 * to have executed a full memory barrier during the execution of
3179 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3180 * again only if the system has more than one CPU).
3182 * This primitive provides the guarantees made by the (now removed)
3183 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3184 * guarantees that rcu_read_lock() sections will have completed.
3185 * In "classic RCU", these two guarantees happen to be one and
3186 * the same, but can differ in realtime RCU implementations.
3188 void synchronize_sched(void)
3190 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3191 lock_is_held(&rcu_lock_map
) ||
3192 lock_is_held(&rcu_sched_lock_map
),
3193 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3194 if (rcu_blocking_is_gp())
3196 if (rcu_gp_is_expedited())
3197 synchronize_sched_expedited();
3199 wait_rcu_gp(call_rcu_sched
);
3201 EXPORT_SYMBOL_GPL(synchronize_sched
);
3204 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3206 * Control will return to the caller some time after a full rcu_bh grace
3207 * period has elapsed, in other words after all currently executing rcu_bh
3208 * read-side critical sections have completed. RCU read-side critical
3209 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3210 * and may be nested.
3212 * See the description of synchronize_sched() for more detailed information
3213 * on memory ordering guarantees.
3215 void synchronize_rcu_bh(void)
3217 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3218 lock_is_held(&rcu_lock_map
) ||
3219 lock_is_held(&rcu_sched_lock_map
),
3220 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3221 if (rcu_blocking_is_gp())
3223 if (rcu_gp_is_expedited())
3224 synchronize_rcu_bh_expedited();
3226 wait_rcu_gp(call_rcu_bh
);
3228 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3231 * get_state_synchronize_rcu - Snapshot current RCU state
3233 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3234 * to determine whether or not a full grace period has elapsed in the
3237 unsigned long get_state_synchronize_rcu(void)
3240 * Any prior manipulation of RCU-protected data must happen
3241 * before the load from ->gpnum.
3246 * Make sure this load happens before the purportedly
3247 * time-consuming work between get_state_synchronize_rcu()
3248 * and cond_synchronize_rcu().
3250 return smp_load_acquire(&rcu_state_p
->gpnum
);
3252 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3255 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3257 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3259 * If a full RCU grace period has elapsed since the earlier call to
3260 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3261 * synchronize_rcu() to wait for a full grace period.
3263 * Yes, this function does not take counter wrap into account. But
3264 * counter wrap is harmless. If the counter wraps, we have waited for
3265 * more than 2 billion grace periods (and way more on a 64-bit system!),
3266 * so waiting for one additional grace period should be just fine.
3268 void cond_synchronize_rcu(unsigned long oldstate
)
3270 unsigned long newstate
;
3273 * Ensure that this load happens before any RCU-destructive
3274 * actions the caller might carry out after we return.
3276 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3277 if (ULONG_CMP_GE(oldstate
, newstate
))
3280 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3283 * get_state_synchronize_sched - Snapshot current RCU-sched state
3285 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3286 * to determine whether or not a full grace period has elapsed in the
3289 unsigned long get_state_synchronize_sched(void)
3292 * Any prior manipulation of RCU-protected data must happen
3293 * before the load from ->gpnum.
3298 * Make sure this load happens before the purportedly
3299 * time-consuming work between get_state_synchronize_sched()
3300 * and cond_synchronize_sched().
3302 return smp_load_acquire(&rcu_sched_state
.gpnum
);
3304 EXPORT_SYMBOL_GPL(get_state_synchronize_sched
);
3307 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3309 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3311 * If a full RCU-sched grace period has elapsed since the earlier call to
3312 * get_state_synchronize_sched(), just return. Otherwise, invoke
3313 * synchronize_sched() to wait for a full grace period.
3315 * Yes, this function does not take counter wrap into account. But
3316 * counter wrap is harmless. If the counter wraps, we have waited for
3317 * more than 2 billion grace periods (and way more on a 64-bit system!),
3318 * so waiting for one additional grace period should be just fine.
3320 void cond_synchronize_sched(unsigned long oldstate
)
3322 unsigned long newstate
;
3325 * Ensure that this load happens before any RCU-destructive
3326 * actions the caller might carry out after we return.
3328 newstate
= smp_load_acquire(&rcu_sched_state
.completed
);
3329 if (ULONG_CMP_GE(oldstate
, newstate
))
3330 synchronize_sched();
3332 EXPORT_SYMBOL_GPL(cond_synchronize_sched
);
3334 /* Adjust sequence number for start of update-side operation. */
3335 static void rcu_seq_start(unsigned long *sp
)
3337 WRITE_ONCE(*sp
, *sp
+ 1);
3338 smp_mb(); /* Ensure update-side operation after counter increment. */
3339 WARN_ON_ONCE(!(*sp
& 0x1));
3342 /* Adjust sequence number for end of update-side operation. */
3343 static void rcu_seq_end(unsigned long *sp
)
3345 smp_mb(); /* Ensure update-side operation before counter increment. */
3346 WRITE_ONCE(*sp
, *sp
+ 1);
3347 WARN_ON_ONCE(*sp
& 0x1);
3350 /* Take a snapshot of the update side's sequence number. */
3351 static unsigned long rcu_seq_snap(unsigned long *sp
)
3355 s
= (READ_ONCE(*sp
) + 3) & ~0x1;
3356 smp_mb(); /* Above access must not bleed into critical section. */
3361 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3362 * full update-side operation has occurred.
3364 static bool rcu_seq_done(unsigned long *sp
, unsigned long s
)
3366 return ULONG_CMP_GE(READ_ONCE(*sp
), s
);
3369 /* Wrapper functions for expedited grace periods. */
3370 static void rcu_exp_gp_seq_start(struct rcu_state
*rsp
)
3372 rcu_seq_start(&rsp
->expedited_sequence
);
3374 static void rcu_exp_gp_seq_end(struct rcu_state
*rsp
)
3376 rcu_seq_end(&rsp
->expedited_sequence
);
3377 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3379 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state
*rsp
)
3381 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3382 return rcu_seq_snap(&rsp
->expedited_sequence
);
3384 static bool rcu_exp_gp_seq_done(struct rcu_state
*rsp
, unsigned long s
)
3386 return rcu_seq_done(&rsp
->expedited_sequence
, s
);
3390 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3391 * recent CPU-online activity. Note that these masks are not cleared
3392 * when CPUs go offline, so they reflect the union of all CPUs that have
3393 * ever been online. This means that this function normally takes its
3394 * no-work-to-do fastpath.
3396 static void sync_exp_reset_tree_hotplug(struct rcu_state
*rsp
)
3399 unsigned long flags
;
3401 unsigned long oldmask
;
3402 int ncpus
= READ_ONCE(rsp
->ncpus
);
3403 struct rcu_node
*rnp
;
3404 struct rcu_node
*rnp_up
;
3406 /* If no new CPUs onlined since last time, nothing to do. */
3407 if (likely(ncpus
== rsp
->ncpus_snap
))
3409 rsp
->ncpus_snap
= ncpus
;
3412 * Each pass through the following loop propagates newly onlined
3413 * CPUs for the current rcu_node structure up the rcu_node tree.
3415 rcu_for_each_leaf_node(rsp
, rnp
) {
3416 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3417 if (rnp
->expmaskinit
== rnp
->expmaskinitnext
) {
3418 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3419 continue; /* No new CPUs, nothing to do. */
3422 /* Update this node's mask, track old value for propagation. */
3423 oldmask
= rnp
->expmaskinit
;
3424 rnp
->expmaskinit
= rnp
->expmaskinitnext
;
3425 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3427 /* If was already nonzero, nothing to propagate. */
3431 /* Propagate the new CPU up the tree. */
3432 mask
= rnp
->grpmask
;
3433 rnp_up
= rnp
->parent
;
3436 raw_spin_lock_irqsave_rcu_node(rnp_up
, flags
);
3437 if (rnp_up
->expmaskinit
)
3439 rnp_up
->expmaskinit
|= mask
;
3440 raw_spin_unlock_irqrestore_rcu_node(rnp_up
, flags
);
3443 mask
= rnp_up
->grpmask
;
3444 rnp_up
= rnp_up
->parent
;
3450 * Reset the ->expmask values in the rcu_node tree in preparation for
3451 * a new expedited grace period.
3453 static void __maybe_unused
sync_exp_reset_tree(struct rcu_state
*rsp
)
3455 unsigned long flags
;
3456 struct rcu_node
*rnp
;
3458 sync_exp_reset_tree_hotplug(rsp
);
3459 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3460 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3461 WARN_ON_ONCE(rnp
->expmask
);
3462 rnp
->expmask
= rnp
->expmaskinit
;
3463 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3468 * Return non-zero if there is no RCU expedited grace period in progress
3469 * for the specified rcu_node structure, in other words, if all CPUs and
3470 * tasks covered by the specified rcu_node structure have done their bit
3471 * for the current expedited grace period. Works only for preemptible
3472 * RCU -- other RCU implementation use other means.
3474 * Caller must hold the root rcu_node's exp_funnel_mutex.
3476 static int sync_rcu_preempt_exp_done(struct rcu_node
*rnp
)
3478 return rnp
->exp_tasks
== NULL
&&
3479 READ_ONCE(rnp
->expmask
) == 0;
3483 * Report the exit from RCU read-side critical section for the last task
3484 * that queued itself during or before the current expedited preemptible-RCU
3485 * grace period. This event is reported either to the rcu_node structure on
3486 * which the task was queued or to one of that rcu_node structure's ancestors,
3487 * recursively up the tree. (Calm down, calm down, we do the recursion
3490 * Caller must hold the root rcu_node's exp_funnel_mutex and the
3491 * specified rcu_node structure's ->lock.
3493 static void __rcu_report_exp_rnp(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3494 bool wake
, unsigned long flags
)
3495 __releases(rnp
->lock
)
3500 if (!sync_rcu_preempt_exp_done(rnp
)) {
3502 rcu_initiate_boost(rnp
, flags
);
3504 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3507 if (rnp
->parent
== NULL
) {
3508 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3510 smp_mb(); /* EGP done before wake_up(). */
3511 swake_up(&rsp
->expedited_wq
);
3515 mask
= rnp
->grpmask
;
3516 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled */
3518 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled */
3519 WARN_ON_ONCE(!(rnp
->expmask
& mask
));
3520 rnp
->expmask
&= ~mask
;
3525 * Report expedited quiescent state for specified node. This is a
3526 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3528 * Caller must hold the root rcu_node's exp_funnel_mutex.
3530 static void __maybe_unused
rcu_report_exp_rnp(struct rcu_state
*rsp
,
3531 struct rcu_node
*rnp
, bool wake
)
3533 unsigned long flags
;
3535 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3536 __rcu_report_exp_rnp(rsp
, rnp
, wake
, flags
);
3540 * Report expedited quiescent state for multiple CPUs, all covered by the
3541 * specified leaf rcu_node structure. Caller must hold the root
3542 * rcu_node's exp_funnel_mutex.
3544 static void rcu_report_exp_cpu_mult(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3545 unsigned long mask
, bool wake
)
3547 unsigned long flags
;
3549 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3550 if (!(rnp
->expmask
& mask
)) {
3551 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3554 rnp
->expmask
&= ~mask
;
3555 __rcu_report_exp_rnp(rsp
, rnp
, wake
, flags
); /* Releases rnp->lock. */
3559 * Report expedited quiescent state for specified rcu_data (CPU).
3560 * Caller must hold the root rcu_node's exp_funnel_mutex.
3562 static void rcu_report_exp_rdp(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
3565 rcu_report_exp_cpu_mult(rsp
, rdp
->mynode
, rdp
->grpmask
, wake
);
3568 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3569 static bool sync_exp_work_done(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
3570 struct rcu_data
*rdp
,
3571 atomic_long_t
*stat
, unsigned long s
)
3573 if (rcu_exp_gp_seq_done(rsp
, s
)) {
3575 mutex_unlock(&rnp
->exp_funnel_mutex
);
3577 mutex_unlock(&rdp
->exp_funnel_mutex
);
3578 /* Ensure test happens before caller kfree(). */
3579 smp_mb__before_atomic(); /* ^^^ */
3580 atomic_long_inc(stat
);
3587 * Funnel-lock acquisition for expedited grace periods. Returns a
3588 * pointer to the root rcu_node structure, or NULL if some other
3589 * task did the expedited grace period for us.
3591 static struct rcu_node
*exp_funnel_lock(struct rcu_state
*rsp
, unsigned long s
)
3593 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, raw_smp_processor_id());
3594 struct rcu_node
*rnp0
;
3595 struct rcu_node
*rnp1
= NULL
;
3598 * First try directly acquiring the root lock in order to reduce
3599 * latency in the common case where expedited grace periods are
3600 * rare. We check mutex_is_locked() to avoid pathological levels of
3601 * memory contention on ->exp_funnel_mutex in the heavy-load case.
3603 rnp0
= rcu_get_root(rsp
);
3604 if (!mutex_is_locked(&rnp0
->exp_funnel_mutex
)) {
3605 if (mutex_trylock(&rnp0
->exp_funnel_mutex
)) {
3606 if (sync_exp_work_done(rsp
, rnp0
, NULL
,
3607 &rdp
->expedited_workdone0
, s
))
3614 * Each pass through the following loop works its way
3615 * up the rcu_node tree, returning if others have done the
3616 * work or otherwise falls through holding the root rnp's
3617 * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure
3618 * can be inexact, as it is just promoting locality and is not
3619 * strictly needed for correctness.
3621 if (sync_exp_work_done(rsp
, NULL
, NULL
, &rdp
->expedited_workdone1
, s
))
3623 mutex_lock(&rdp
->exp_funnel_mutex
);
3625 for (; rnp0
!= NULL
; rnp0
= rnp0
->parent
) {
3626 if (sync_exp_work_done(rsp
, rnp1
, rdp
,
3627 &rdp
->expedited_workdone2
, s
))
3629 mutex_lock(&rnp0
->exp_funnel_mutex
);
3631 mutex_unlock(&rnp1
->exp_funnel_mutex
);
3633 mutex_unlock(&rdp
->exp_funnel_mutex
);
3636 if (sync_exp_work_done(rsp
, rnp1
, rdp
,
3637 &rdp
->expedited_workdone3
, s
))
3642 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3643 static void sync_sched_exp_handler(void *data
)
3645 struct rcu_data
*rdp
;
3646 struct rcu_node
*rnp
;
3647 struct rcu_state
*rsp
= data
;
3649 rdp
= this_cpu_ptr(rsp
->rda
);
3651 if (!(READ_ONCE(rnp
->expmask
) & rdp
->grpmask
) ||
3652 __this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
3654 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, true);
3655 resched_cpu(smp_processor_id());
3658 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
3659 static void sync_sched_exp_online_cleanup(int cpu
)
3661 struct rcu_data
*rdp
;
3663 struct rcu_node
*rnp
;
3664 struct rcu_state
*rsp
= &rcu_sched_state
;
3666 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3668 if (!(READ_ONCE(rnp
->expmask
) & rdp
->grpmask
))
3670 ret
= smp_call_function_single(cpu
, sync_sched_exp_handler
, rsp
, 0);
3675 * Select the nodes that the upcoming expedited grace period needs
3678 static void sync_rcu_exp_select_cpus(struct rcu_state
*rsp
,
3679 smp_call_func_t func
)
3682 unsigned long flags
;
3684 unsigned long mask_ofl_test
;
3685 unsigned long mask_ofl_ipi
;
3687 struct rcu_node
*rnp
;
3689 sync_exp_reset_tree(rsp
);
3690 rcu_for_each_leaf_node(rsp
, rnp
) {
3691 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3693 /* Each pass checks a CPU for identity, offline, and idle. */
3695 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++) {
3696 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3697 struct rcu_dynticks
*rdtp
= &per_cpu(rcu_dynticks
, cpu
);
3699 if (raw_smp_processor_id() == cpu
||
3700 !(atomic_add_return(0, &rdtp
->dynticks
) & 0x1))
3701 mask_ofl_test
|= rdp
->grpmask
;
3703 mask_ofl_ipi
= rnp
->expmask
& ~mask_ofl_test
;
3706 * Need to wait for any blocked tasks as well. Note that
3707 * additional blocking tasks will also block the expedited
3708 * GP until such time as the ->expmask bits are cleared.
3710 if (rcu_preempt_has_tasks(rnp
))
3711 rnp
->exp_tasks
= rnp
->blkd_tasks
.next
;
3712 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3714 /* IPI the remaining CPUs for expedited quiescent state. */
3716 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3717 if (!(mask_ofl_ipi
& mask
))
3720 ret
= smp_call_function_single(cpu
, func
, rsp
, 0);
3722 mask_ofl_ipi
&= ~mask
;
3725 /* Failed, raced with offline. */
3726 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3727 if (cpu_online(cpu
) &&
3728 (rnp
->expmask
& mask
)) {
3729 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3730 schedule_timeout_uninterruptible(1);
3731 if (cpu_online(cpu
) &&
3732 (rnp
->expmask
& mask
))
3734 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3736 if (!(rnp
->expmask
& mask
))
3737 mask_ofl_ipi
&= ~mask
;
3738 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3740 /* Report quiescent states for those that went offline. */
3741 mask_ofl_test
|= mask_ofl_ipi
;
3743 rcu_report_exp_cpu_mult(rsp
, rnp
, mask_ofl_test
, false);
3747 static void synchronize_sched_expedited_wait(struct rcu_state
*rsp
)
3750 unsigned long jiffies_stall
;
3751 unsigned long jiffies_start
;
3754 struct rcu_node
*rnp
;
3755 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3758 jiffies_stall
= rcu_jiffies_till_stall_check();
3759 jiffies_start
= jiffies
;
3762 ret
= swait_event_timeout(
3764 sync_rcu_preempt_exp_done(rnp_root
),
3766 if (ret
> 0 || sync_rcu_preempt_exp_done(rnp_root
))
3769 /* Hit a signal, disable CPU stall warnings. */
3770 swait_event(rsp
->expedited_wq
,
3771 sync_rcu_preempt_exp_done(rnp_root
));
3774 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3777 rcu_for_each_leaf_node(rsp
, rnp
) {
3778 ndetected
= rcu_print_task_exp_stall(rnp
);
3780 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3781 struct rcu_data
*rdp
;
3783 if (!(rnp
->expmask
& mask
))
3786 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3787 pr_cont(" %d-%c%c%c", cpu
,
3788 "O."[cpu_online(cpu
)],
3789 "o."[!!(rdp
->grpmask
& rnp
->expmaskinit
)],
3790 "N."[!!(rdp
->grpmask
& rnp
->expmaskinitnext
)]);
3794 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
3795 jiffies
- jiffies_start
, rsp
->expedited_sequence
,
3796 rnp_root
->expmask
, ".T"[!!rnp_root
->exp_tasks
]);
3798 pr_err("blocking rcu_node structures:");
3799 rcu_for_each_node_breadth_first(rsp
, rnp
) {
3800 if (rnp
== rnp_root
)
3801 continue; /* printed unconditionally */
3802 if (sync_rcu_preempt_exp_done(rnp
))
3804 pr_cont(" l=%u:%d-%d:%#lx/%c",
3805 rnp
->level
, rnp
->grplo
, rnp
->grphi
,
3807 ".T"[!!rnp
->exp_tasks
]);
3811 rcu_for_each_leaf_node(rsp
, rnp
) {
3813 for (cpu
= rnp
->grplo
; cpu
<= rnp
->grphi
; cpu
++, mask
<<= 1) {
3814 if (!(rnp
->expmask
& mask
))
3819 jiffies_stall
= 3 * rcu_jiffies_till_stall_check() + 3;
3824 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3826 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3827 * approach to force the grace period to end quickly. This consumes
3828 * significant time on all CPUs and is unfriendly to real-time workloads,
3829 * so is thus not recommended for any sort of common-case code. In fact,
3830 * if you are using synchronize_sched_expedited() in a loop, please
3831 * restructure your code to batch your updates, and then use a single
3832 * synchronize_sched() instead.
3834 * This implementation can be thought of as an application of sequence
3835 * locking to expedited grace periods, but using the sequence counter to
3836 * determine when someone else has already done the work instead of for
3839 void synchronize_sched_expedited(void)
3842 struct rcu_node
*rnp
;
3843 struct rcu_state
*rsp
= &rcu_sched_state
;
3845 /* If only one CPU, this is automatically a grace period. */
3846 if (rcu_blocking_is_gp())
3849 /* If expedited grace periods are prohibited, fall back to normal. */
3850 if (rcu_gp_is_normal()) {
3851 wait_rcu_gp(call_rcu_sched
);
3855 /* Take a snapshot of the sequence number. */
3856 s
= rcu_exp_gp_seq_snap(rsp
);
3858 rnp
= exp_funnel_lock(rsp
, s
);
3860 return; /* Someone else did our work for us. */
3862 rcu_exp_gp_seq_start(rsp
);
3863 sync_rcu_exp_select_cpus(rsp
, sync_sched_exp_handler
);
3864 synchronize_sched_expedited_wait(rsp
);
3866 rcu_exp_gp_seq_end(rsp
);
3867 mutex_unlock(&rnp
->exp_funnel_mutex
);
3869 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
3872 * Check to see if there is any immediate RCU-related work to be done
3873 * by the current CPU, for the specified type of RCU, returning 1 if so.
3874 * The checks are in order of increasing expense: checks that can be
3875 * carried out against CPU-local state are performed first. However,
3876 * we must check for CPU stalls first, else we might not get a chance.
3878 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3880 struct rcu_node
*rnp
= rdp
->mynode
;
3882 rdp
->n_rcu_pending
++;
3884 /* Check for CPU stalls, if enabled. */
3885 check_cpu_stall(rsp
, rdp
);
3887 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3888 if (rcu_nohz_full_cpu(rsp
))
3891 /* Is the RCU core waiting for a quiescent state from this CPU? */
3892 if (rcu_scheduler_fully_active
&&
3893 rdp
->core_needs_qs
&& rdp
->cpu_no_qs
.b
.norm
&&
3894 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3895 rdp
->n_rp_core_needs_qs
++;
3896 } else if (rdp
->core_needs_qs
&&
3897 (!rdp
->cpu_no_qs
.b
.norm
||
3898 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3899 rdp
->n_rp_report_qs
++;
3903 /* Does this CPU have callbacks ready to invoke? */
3904 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3905 rdp
->n_rp_cb_ready
++;
3909 /* Has RCU gone idle with this CPU needing another grace period? */
3910 if (cpu_needs_another_gp(rsp
, rdp
)) {
3911 rdp
->n_rp_cpu_needs_gp
++;
3915 /* Has another RCU grace period completed? */
3916 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3917 rdp
->n_rp_gp_completed
++;
3921 /* Has a new RCU grace period started? */
3922 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3923 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3924 rdp
->n_rp_gp_started
++;
3928 /* Does this CPU need a deferred NOCB wakeup? */
3929 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3930 rdp
->n_rp_nocb_defer_wakeup
++;
3935 rdp
->n_rp_need_nothing
++;
3940 * Check to see if there is any immediate RCU-related work to be done
3941 * by the current CPU, returning 1 if so. This function is part of the
3942 * RCU implementation; it is -not- an exported member of the RCU API.
3944 static int rcu_pending(void)
3946 struct rcu_state
*rsp
;
3948 for_each_rcu_flavor(rsp
)
3949 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3955 * Return true if the specified CPU has any callback. If all_lazy is
3956 * non-NULL, store an indication of whether all callbacks are lazy.
3957 * (If there are no callbacks, all of them are deemed to be lazy.)
3959 static bool __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3963 struct rcu_data
*rdp
;
3964 struct rcu_state
*rsp
;
3966 for_each_rcu_flavor(rsp
) {
3967 rdp
= this_cpu_ptr(rsp
->rda
);
3971 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3982 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3983 * the compiler is expected to optimize this away.
3985 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3986 int cpu
, unsigned long done
)
3988 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3989 atomic_read(&rsp
->barrier_cpu_count
), done
);
3993 * RCU callback function for _rcu_barrier(). If we are last, wake
3994 * up the task executing _rcu_barrier().
3996 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3998 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3999 struct rcu_state
*rsp
= rdp
->rsp
;
4001 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
4002 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->barrier_sequence
);
4003 complete(&rsp
->barrier_completion
);
4005 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->barrier_sequence
);
4010 * Called with preemption disabled, and from cross-cpu IRQ context.
4012 static void rcu_barrier_func(void *type
)
4014 struct rcu_state
*rsp
= type
;
4015 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
4017 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->barrier_sequence
);
4018 atomic_inc(&rsp
->barrier_cpu_count
);
4019 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
4023 * Orchestrate the specified type of RCU barrier, waiting for all
4024 * RCU callbacks of the specified type to complete.
4026 static void _rcu_barrier(struct rcu_state
*rsp
)
4029 struct rcu_data
*rdp
;
4030 unsigned long s
= rcu_seq_snap(&rsp
->barrier_sequence
);
4032 _rcu_barrier_trace(rsp
, "Begin", -1, s
);
4034 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4035 mutex_lock(&rsp
->barrier_mutex
);
4037 /* Did someone else do our work for us? */
4038 if (rcu_seq_done(&rsp
->barrier_sequence
, s
)) {
4039 _rcu_barrier_trace(rsp
, "EarlyExit", -1, rsp
->barrier_sequence
);
4040 smp_mb(); /* caller's subsequent code after above check. */
4041 mutex_unlock(&rsp
->barrier_mutex
);
4045 /* Mark the start of the barrier operation. */
4046 rcu_seq_start(&rsp
->barrier_sequence
);
4047 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->barrier_sequence
);
4050 * Initialize the count to one rather than to zero in order to
4051 * avoid a too-soon return to zero in case of a short grace period
4052 * (or preemption of this task). Exclude CPU-hotplug operations
4053 * to ensure that no offline CPU has callbacks queued.
4055 init_completion(&rsp
->barrier_completion
);
4056 atomic_set(&rsp
->barrier_cpu_count
, 1);
4060 * Force each CPU with callbacks to register a new callback.
4061 * When that callback is invoked, we will know that all of the
4062 * corresponding CPU's preceding callbacks have been invoked.
4064 for_each_possible_cpu(cpu
) {
4065 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
4067 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4068 if (rcu_is_nocb_cpu(cpu
)) {
4069 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
4070 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
4071 rsp
->barrier_sequence
);
4073 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
4074 rsp
->barrier_sequence
);
4075 smp_mb__before_atomic();
4076 atomic_inc(&rsp
->barrier_cpu_count
);
4077 __call_rcu(&rdp
->barrier_head
,
4078 rcu_barrier_callback
, rsp
, cpu
, 0);
4080 } else if (READ_ONCE(rdp
->qlen
)) {
4081 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
4082 rsp
->barrier_sequence
);
4083 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
4085 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
4086 rsp
->barrier_sequence
);
4092 * Now that we have an rcu_barrier_callback() callback on each
4093 * CPU, and thus each counted, remove the initial count.
4095 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
4096 complete(&rsp
->barrier_completion
);
4098 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4099 wait_for_completion(&rsp
->barrier_completion
);
4101 /* Mark the end of the barrier operation. */
4102 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->barrier_sequence
);
4103 rcu_seq_end(&rsp
->barrier_sequence
);
4105 /* Other rcu_barrier() invocations can now safely proceed. */
4106 mutex_unlock(&rsp
->barrier_mutex
);
4110 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4112 void rcu_barrier_bh(void)
4114 _rcu_barrier(&rcu_bh_state
);
4116 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
4119 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4121 void rcu_barrier_sched(void)
4123 _rcu_barrier(&rcu_sched_state
);
4125 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
4128 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4129 * first CPU in a given leaf rcu_node structure coming online. The caller
4130 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4133 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
4136 struct rcu_node
*rnp
= rnp_leaf
;
4139 mask
= rnp
->grpmask
;
4143 raw_spin_lock_rcu_node(rnp
); /* Interrupts already disabled. */
4144 rnp
->qsmaskinit
|= mask
;
4145 raw_spin_unlock_rcu_node(rnp
); /* Interrupts remain disabled. */
4150 * Do boot-time initialization of a CPU's per-CPU RCU data.
4153 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
4155 unsigned long flags
;
4156 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4157 struct rcu_node
*rnp
= rcu_get_root(rsp
);
4159 /* Set up local state, ensuring consistent view of global state. */
4160 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4161 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
4162 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
4163 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
4164 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
4167 mutex_init(&rdp
->exp_funnel_mutex
);
4168 rcu_boot_init_nocb_percpu_data(rdp
);
4169 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4173 * Initialize a CPU's per-CPU RCU data. Note that only one online or
4174 * offline event can be happening at a given time. Note also that we
4175 * can accept some slop in the rsp->completed access due to the fact
4176 * that this CPU cannot possibly have any RCU callbacks in flight yet.
4179 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
4181 unsigned long flags
;
4183 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4184 struct rcu_node
*rnp
= rcu_get_root(rsp
);
4186 /* Set up local state, ensuring consistent view of global state. */
4187 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4188 rdp
->qlen_last_fqs_check
= 0;
4189 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
4190 rdp
->blimit
= blimit
;
4192 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
4193 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
4194 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
4195 atomic_set(&rdp
->dynticks
->dynticks
,
4196 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
4197 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
4200 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4201 * propagation up the rcu_node tree will happen at the beginning
4202 * of the next grace period.
4205 mask
= rdp
->grpmask
;
4206 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
4207 rnp
->qsmaskinitnext
|= mask
;
4208 rnp
->expmaskinitnext
|= mask
;
4209 if (!rdp
->beenonline
)
4210 WRITE_ONCE(rsp
->ncpus
, READ_ONCE(rsp
->ncpus
) + 1);
4211 rdp
->beenonline
= true; /* We have now been online. */
4212 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
4213 rdp
->completed
= rnp
->completed
;
4214 rdp
->cpu_no_qs
.b
.norm
= true;
4215 rdp
->rcu_qs_ctr_snap
= per_cpu(rcu_qs_ctr
, cpu
);
4216 rdp
->core_needs_qs
= false;
4217 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
4218 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4221 static void rcu_prepare_cpu(int cpu
)
4223 struct rcu_state
*rsp
;
4225 for_each_rcu_flavor(rsp
)
4226 rcu_init_percpu_data(cpu
, rsp
);
4229 #ifdef CONFIG_HOTPLUG_CPU
4231 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4232 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4234 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4235 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4238 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
4240 unsigned long flags
;
4242 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
4243 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
4245 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
4248 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4249 mask
= rdp
->grpmask
;
4250 raw_spin_lock_irqsave_rcu_node(rnp
, flags
); /* Enforce GP memory-order guarantee. */
4251 rnp
->qsmaskinitnext
&= ~mask
;
4252 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4255 void rcu_report_dead(unsigned int cpu
)
4257 struct rcu_state
*rsp
;
4259 /* QS for any half-done expedited RCU-sched GP. */
4261 rcu_report_exp_rdp(&rcu_sched_state
,
4262 this_cpu_ptr(rcu_sched_state
.rda
), true);
4264 for_each_rcu_flavor(rsp
)
4265 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
4270 * Handle CPU online/offline notification events.
4272 int rcu_cpu_notify(struct notifier_block
*self
,
4273 unsigned long action
, void *hcpu
)
4275 long cpu
= (long)hcpu
;
4276 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
4277 struct rcu_node
*rnp
= rdp
->mynode
;
4278 struct rcu_state
*rsp
;
4281 case CPU_UP_PREPARE
:
4282 case CPU_UP_PREPARE_FROZEN
:
4283 rcu_prepare_cpu(cpu
);
4284 rcu_prepare_kthreads(cpu
);
4285 rcu_spawn_all_nocb_kthreads(cpu
);
4288 case CPU_DOWN_FAILED
:
4289 sync_sched_exp_online_cleanup(cpu
);
4290 rcu_boost_kthread_setaffinity(rnp
, -1);
4292 case CPU_DOWN_PREPARE
:
4293 rcu_boost_kthread_setaffinity(rnp
, cpu
);
4296 case CPU_DYING_FROZEN
:
4297 for_each_rcu_flavor(rsp
)
4298 rcu_cleanup_dying_cpu(rsp
);
4301 case CPU_DEAD_FROZEN
:
4302 case CPU_UP_CANCELED
:
4303 case CPU_UP_CANCELED_FROZEN
:
4304 for_each_rcu_flavor(rsp
) {
4305 rcu_cleanup_dead_cpu(cpu
, rsp
);
4306 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
4315 static int rcu_pm_notify(struct notifier_block
*self
,
4316 unsigned long action
, void *hcpu
)
4319 case PM_HIBERNATION_PREPARE
:
4320 case PM_SUSPEND_PREPARE
:
4321 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
4324 case PM_POST_HIBERNATION
:
4325 case PM_POST_SUSPEND
:
4326 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
4327 rcu_unexpedite_gp();
4336 * Spawn the kthreads that handle each RCU flavor's grace periods.
4338 static int __init
rcu_spawn_gp_kthread(void)
4340 unsigned long flags
;
4341 int kthread_prio_in
= kthread_prio
;
4342 struct rcu_node
*rnp
;
4343 struct rcu_state
*rsp
;
4344 struct sched_param sp
;
4345 struct task_struct
*t
;
4347 /* Force priority into range. */
4348 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
4350 else if (kthread_prio
< 0)
4352 else if (kthread_prio
> 99)
4354 if (kthread_prio
!= kthread_prio_in
)
4355 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4356 kthread_prio
, kthread_prio_in
);
4358 rcu_scheduler_fully_active
= 1;
4359 for_each_rcu_flavor(rsp
) {
4360 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
4362 rnp
= rcu_get_root(rsp
);
4363 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4364 rsp
->gp_kthread
= t
;
4366 sp
.sched_priority
= kthread_prio
;
4367 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
4369 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4372 rcu_spawn_nocb_kthreads();
4373 rcu_spawn_boost_kthreads();
4376 early_initcall(rcu_spawn_gp_kthread
);
4379 * This function is invoked towards the end of the scheduler's initialization
4380 * process. Before this is called, the idle task might contain
4381 * RCU read-side critical sections (during which time, this idle
4382 * task is booting the system). After this function is called, the
4383 * idle tasks are prohibited from containing RCU read-side critical
4384 * sections. This function also enables RCU lockdep checking.
4386 void rcu_scheduler_starting(void)
4388 WARN_ON(num_online_cpus() != 1);
4389 WARN_ON(nr_context_switches() > 0);
4390 rcu_scheduler_active
= 1;
4394 * Compute the per-level fanout, either using the exact fanout specified
4395 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4397 static void __init
rcu_init_levelspread(int *levelspread
, const int *levelcnt
)
4401 if (rcu_fanout_exact
) {
4402 levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
4403 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
4404 levelspread
[i
] = RCU_FANOUT
;
4410 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4412 levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
4419 * Helper function for rcu_init() that initializes one rcu_state structure.
4421 static void __init
rcu_init_one(struct rcu_state
*rsp
)
4423 static const char * const buf
[] = RCU_NODE_NAME_INIT
;
4424 static const char * const fqs
[] = RCU_FQS_NAME_INIT
;
4425 static const char * const exp
[] = RCU_EXP_NAME_INIT
;
4426 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
4427 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
4428 static struct lock_class_key rcu_exp_class
[RCU_NUM_LVLS
];
4429 static u8 fl_mask
= 0x1;
4431 int levelcnt
[RCU_NUM_LVLS
]; /* # nodes in each level. */
4432 int levelspread
[RCU_NUM_LVLS
]; /* kids/node in each level. */
4436 struct rcu_node
*rnp
;
4438 BUILD_BUG_ON(RCU_NUM_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4440 /* Silence gcc 4.8 false positive about array index out of range. */
4441 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4442 panic("rcu_init_one: rcu_num_lvls out of range");
4444 /* Initialize the level-tracking arrays. */
4446 for (i
= 0; i
< rcu_num_lvls
; i
++)
4447 levelcnt
[i
] = num_rcu_lvl
[i
];
4448 for (i
= 1; i
< rcu_num_lvls
; i
++)
4449 rsp
->level
[i
] = rsp
->level
[i
- 1] + levelcnt
[i
- 1];
4450 rcu_init_levelspread(levelspread
, levelcnt
);
4451 rsp
->flavor_mask
= fl_mask
;
4454 /* Initialize the elements themselves, starting from the leaves. */
4456 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4457 cpustride
*= levelspread
[i
];
4458 rnp
= rsp
->level
[i
];
4459 for (j
= 0; j
< levelcnt
[i
]; j
++, rnp
++) {
4460 raw_spin_lock_init(&ACCESS_PRIVATE(rnp
, lock
));
4461 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp
, lock
),
4462 &rcu_node_class
[i
], buf
[i
]);
4463 raw_spin_lock_init(&rnp
->fqslock
);
4464 lockdep_set_class_and_name(&rnp
->fqslock
,
4465 &rcu_fqs_class
[i
], fqs
[i
]);
4466 rnp
->gpnum
= rsp
->gpnum
;
4467 rnp
->completed
= rsp
->completed
;
4469 rnp
->qsmaskinit
= 0;
4470 rnp
->grplo
= j
* cpustride
;
4471 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4472 if (rnp
->grphi
>= nr_cpu_ids
)
4473 rnp
->grphi
= nr_cpu_ids
- 1;
4479 rnp
->grpnum
= j
% levelspread
[i
- 1];
4480 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4481 rnp
->parent
= rsp
->level
[i
- 1] +
4482 j
/ levelspread
[i
- 1];
4485 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4486 rcu_init_one_nocb(rnp
);
4487 mutex_init(&rnp
->exp_funnel_mutex
);
4488 lockdep_set_class_and_name(&rnp
->exp_funnel_mutex
,
4489 &rcu_exp_class
[i
], exp
[i
]);
4493 init_swait_queue_head(&rsp
->gp_wq
);
4494 init_swait_queue_head(&rsp
->expedited_wq
);
4495 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4496 for_each_possible_cpu(i
) {
4497 while (i
> rnp
->grphi
)
4499 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4500 rcu_boot_init_percpu_data(i
, rsp
);
4502 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4506 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4507 * replace the definitions in tree.h because those are needed to size
4508 * the ->node array in the rcu_state structure.
4510 static void __init
rcu_init_geometry(void)
4514 int rcu_capacity
[RCU_NUM_LVLS
];
4517 * Initialize any unspecified boot parameters.
4518 * The default values of jiffies_till_first_fqs and
4519 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4520 * value, which is a function of HZ, then adding one for each
4521 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4523 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4524 if (jiffies_till_first_fqs
== ULONG_MAX
)
4525 jiffies_till_first_fqs
= d
;
4526 if (jiffies_till_next_fqs
== ULONG_MAX
)
4527 jiffies_till_next_fqs
= d
;
4529 /* If the compile-time values are accurate, just leave. */
4530 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4531 nr_cpu_ids
== NR_CPUS
)
4533 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4534 rcu_fanout_leaf
, nr_cpu_ids
);
4537 * The boot-time rcu_fanout_leaf parameter must be at least two
4538 * and cannot exceed the number of bits in the rcu_node masks.
4539 * Complain and fall back to the compile-time values if this
4540 * limit is exceeded.
4542 if (rcu_fanout_leaf
< 2 ||
4543 rcu_fanout_leaf
> sizeof(unsigned long) * 8) {
4544 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4550 * Compute number of nodes that can be handled an rcu_node tree
4551 * with the given number of levels.
4553 rcu_capacity
[0] = rcu_fanout_leaf
;
4554 for (i
= 1; i
< RCU_NUM_LVLS
; i
++)
4555 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4558 * The tree must be able to accommodate the configured number of CPUs.
4559 * If this limit is exceeded, fall back to the compile-time values.
4561 if (nr_cpu_ids
> rcu_capacity
[RCU_NUM_LVLS
- 1]) {
4562 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4567 /* Calculate the number of levels in the tree. */
4568 for (i
= 0; nr_cpu_ids
> rcu_capacity
[i
]; i
++) {
4570 rcu_num_lvls
= i
+ 1;
4572 /* Calculate the number of rcu_nodes at each level of the tree. */
4573 for (i
= 0; i
< rcu_num_lvls
; i
++) {
4574 int cap
= rcu_capacity
[(rcu_num_lvls
- 1) - i
];
4575 num_rcu_lvl
[i
] = DIV_ROUND_UP(nr_cpu_ids
, cap
);
4578 /* Calculate the total number of rcu_node structures. */
4580 for (i
= 0; i
< rcu_num_lvls
; i
++)
4581 rcu_num_nodes
+= num_rcu_lvl
[i
];
4585 * Dump out the structure of the rcu_node combining tree associated
4586 * with the rcu_state structure referenced by rsp.
4588 static void __init
rcu_dump_rcu_node_tree(struct rcu_state
*rsp
)
4591 struct rcu_node
*rnp
;
4593 pr_info("rcu_node tree layout dump\n");
4595 rcu_for_each_node_breadth_first(rsp
, rnp
) {
4596 if (rnp
->level
!= level
) {
4601 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4606 void __init
rcu_init(void)
4610 rcu_early_boot_tests();
4612 rcu_bootup_announce();
4613 rcu_init_geometry();
4614 rcu_init_one(&rcu_bh_state
);
4615 rcu_init_one(&rcu_sched_state
);
4617 rcu_dump_rcu_node_tree(&rcu_sched_state
);
4618 __rcu_init_preempt();
4619 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4622 * We don't need protection against CPU-hotplug here because
4623 * this is called early in boot, before either interrupts
4624 * or the scheduler are operational.
4626 cpu_notifier(rcu_cpu_notify
, 0);
4627 pm_notifier(rcu_pm_notify
, 0);
4628 for_each_online_cpu(cpu
)
4629 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4632 #include "tree_plugin.h"