2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/trace_events.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
72 * In order to export the rcu_state name to the tracing tools, it
73 * needs to be added in the __tracepoint_string section.
74 * This requires defining a separate variable tp_<sname>_varname
75 * that points to the string being used, and this will allow
76 * the tracing userspace tools to be able to decipher the string
77 * address to the matching string.
80 # define DEFINE_RCU_TPS(sname) \
81 static char sname##_varname[] = #sname; \
82 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
83 # define RCU_STATE_NAME(sname) sname##_varname
85 # define DEFINE_RCU_TPS(sname)
86 # define RCU_STATE_NAME(sname) __stringify(sname)
89 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
90 DEFINE_RCU_TPS(sname) \
91 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
92 struct rcu_state sname##_state = { \
93 .level = { &sname##_state.node[0] }, \
94 .rda = &sname##_data, \
96 .gp_state = RCU_GP_IDLE, \
97 .gpnum = 0UL - 300UL, \
98 .completed = 0UL - 300UL, \
99 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
100 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
101 .orphan_donetail = &sname##_state.orphan_donelist, \
102 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
103 .name = RCU_STATE_NAME(sname), \
105 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
106 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
109 RCU_STATE_INITIALIZER(rcu_sched
, 's', call_rcu_sched
);
110 RCU_STATE_INITIALIZER(rcu_bh
, 'b', call_rcu_bh
);
112 static struct rcu_state
*const rcu_state_p
;
113 LIST_HEAD(rcu_struct_flavors
);
115 /* Dump rcu_node combining tree at boot to verify correct setup. */
116 static bool dump_tree
;
117 module_param(dump_tree
, bool, 0444);
118 /* Control rcu_node-tree auto-balancing at boot time. */
119 static bool rcu_fanout_exact
;
120 module_param(rcu_fanout_exact
, bool, 0444);
121 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
122 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
123 module_param(rcu_fanout_leaf
, int, 0444);
124 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
125 /* Number of rcu_nodes at specified level. */
126 static int num_rcu_lvl
[] = NUM_RCU_LVL_INIT
;
127 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
128 /* panic() on RCU Stall sysctl. */
129 int sysctl_panic_on_rcu_stall __read_mostly
;
132 * The rcu_scheduler_active variable transitions from zero to one just
133 * before the first task is spawned. So when this variable is zero, RCU
134 * can assume that there is but one task, allowing RCU to (for example)
135 * optimize synchronize_rcu() to a simple barrier(). When this variable
136 * is one, RCU must actually do all the hard work required to detect real
137 * grace periods. This variable is also used to suppress boot-time false
138 * positives from lockdep-RCU error checking.
140 int rcu_scheduler_active __read_mostly
;
141 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
144 * The rcu_scheduler_fully_active variable transitions from zero to one
145 * during the early_initcall() processing, which is after the scheduler
146 * is capable of creating new tasks. So RCU processing (for example,
147 * creating tasks for RCU priority boosting) must be delayed until after
148 * rcu_scheduler_fully_active transitions from zero to one. We also
149 * currently delay invocation of any RCU callbacks until after this point.
151 * It might later prove better for people registering RCU callbacks during
152 * early boot to take responsibility for these callbacks, but one step at
155 static int rcu_scheduler_fully_active __read_mostly
;
157 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
158 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
159 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
162 static void rcu_report_exp_rdp(struct rcu_state
*rsp
,
163 struct rcu_data
*rdp
, bool wake
);
164 static void sync_sched_exp_online_cleanup(int cpu
);
166 /* rcuc/rcub kthread realtime priority */
167 #ifdef CONFIG_RCU_KTHREAD_PRIO
168 static int kthread_prio
= CONFIG_RCU_KTHREAD_PRIO
;
169 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
170 static int kthread_prio
= IS_ENABLED(CONFIG_RCU_BOOST
) ? 1 : 0;
171 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
172 module_param(kthread_prio
, int, 0644);
174 /* Delay in jiffies for grace-period initialization delays, debug only. */
176 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
177 static int gp_preinit_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY
;
178 module_param(gp_preinit_delay
, int, 0644);
179 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
180 static const int gp_preinit_delay
;
181 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
183 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
184 static int gp_init_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY
;
185 module_param(gp_init_delay
, int, 0644);
186 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
187 static const int gp_init_delay
;
188 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
190 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
191 static int gp_cleanup_delay
= CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY
;
192 module_param(gp_cleanup_delay
, int, 0644);
193 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
194 static const int gp_cleanup_delay
;
195 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
198 * Number of grace periods between delays, normalized by the duration of
199 * the delay. The longer the the delay, the more the grace periods between
200 * each delay. The reason for this normalization is that it means that,
201 * for non-zero delays, the overall slowdown of grace periods is constant
202 * regardless of the duration of the delay. This arrangement balances
203 * the need for long delays to increase some race probabilities with the
204 * need for fast grace periods to increase other race probabilities.
206 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
209 * Track the rcutorture test sequence number and the update version
210 * number within a given test. The rcutorture_testseq is incremented
211 * on every rcutorture module load and unload, so has an odd value
212 * when a test is running. The rcutorture_vernum is set to zero
213 * when rcutorture starts and is incremented on each rcutorture update.
214 * These variables enable correlating rcutorture output with the
215 * RCU tracing information.
217 unsigned long rcutorture_testseq
;
218 unsigned long rcutorture_vernum
;
221 * Compute the mask of online CPUs for the specified rcu_node structure.
222 * This will not be stable unless the rcu_node structure's ->lock is
223 * held, but the bit corresponding to the current CPU will be stable
226 unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
228 return READ_ONCE(rnp
->qsmaskinitnext
);
232 * Return true if an RCU grace period is in progress. The READ_ONCE()s
233 * permit this function to be invoked without holding the root rcu_node
234 * structure's ->lock, but of course results can be subject to change.
236 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
238 return READ_ONCE(rsp
->completed
) != READ_ONCE(rsp
->gpnum
);
242 * Note a quiescent state. Because we do not need to know
243 * how many quiescent states passed, just if there was at least
244 * one since the start of the grace period, this just sets a flag.
245 * The caller must have disabled preemption.
247 void rcu_sched_qs(void)
249 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.s
))
251 trace_rcu_grace_period(TPS("rcu_sched"),
252 __this_cpu_read(rcu_sched_data
.gpnum
),
254 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.norm
, false);
255 if (!__this_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))
257 __this_cpu_write(rcu_sched_data
.cpu_no_qs
.b
.exp
, false);
258 rcu_report_exp_rdp(&rcu_sched_state
,
259 this_cpu_ptr(&rcu_sched_data
), true);
264 if (__this_cpu_read(rcu_bh_data
.cpu_no_qs
.s
)) {
265 trace_rcu_grace_period(TPS("rcu_bh"),
266 __this_cpu_read(rcu_bh_data
.gpnum
),
268 __this_cpu_write(rcu_bh_data
.cpu_no_qs
.b
.norm
, false);
272 static DEFINE_PER_CPU(int, rcu_sched_qs_mask
);
274 static DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
275 .dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
,
276 .dynticks
= ATOMIC_INIT(1),
277 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
278 .dynticks_idle_nesting
= DYNTICK_TASK_NEST_VALUE
,
279 .dynticks_idle
= ATOMIC_INIT(1),
280 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
283 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr
);
284 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr
);
287 * Let the RCU core know that this CPU has gone through the scheduler,
288 * which is a quiescent state. This is called when the need for a
289 * quiescent state is urgent, so we burn an atomic operation and full
290 * memory barriers to let the RCU core know about it, regardless of what
291 * this CPU might (or might not) do in the near future.
293 * We inform the RCU core by emulating a zero-duration dyntick-idle
294 * period, which we in turn do by incrementing the ->dynticks counter
297 * The caller must have disabled interrupts.
299 static void rcu_momentary_dyntick_idle(void)
301 struct rcu_data
*rdp
;
302 struct rcu_dynticks
*rdtp
;
304 struct rcu_state
*rsp
;
307 * Yes, we can lose flag-setting operations. This is OK, because
308 * the flag will be set again after some delay.
310 resched_mask
= raw_cpu_read(rcu_sched_qs_mask
);
311 raw_cpu_write(rcu_sched_qs_mask
, 0);
313 /* Find the flavor that needs a quiescent state. */
314 for_each_rcu_flavor(rsp
) {
315 rdp
= raw_cpu_ptr(rsp
->rda
);
316 if (!(resched_mask
& rsp
->flavor_mask
))
318 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
319 if (READ_ONCE(rdp
->mynode
->completed
) !=
320 READ_ONCE(rdp
->cond_resched_completed
))
324 * Pretend to be momentarily idle for the quiescent state.
325 * This allows the grace-period kthread to record the
326 * quiescent state, with no need for this CPU to do anything
329 rdtp
= this_cpu_ptr(&rcu_dynticks
);
330 smp_mb__before_atomic(); /* Earlier stuff before QS. */
331 atomic_add(2, &rdtp
->dynticks
); /* QS. */
332 smp_mb__after_atomic(); /* Later stuff after QS. */
338 * Note a context switch. This is a quiescent state for RCU-sched,
339 * and requires special handling for preemptible RCU.
340 * The caller must have disabled interrupts.
342 void rcu_note_context_switch(void)
344 barrier(); /* Avoid RCU read-side critical sections leaking down. */
345 trace_rcu_utilization(TPS("Start context switch"));
347 rcu_preempt_note_context_switch();
348 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
)))
349 rcu_momentary_dyntick_idle();
350 trace_rcu_utilization(TPS("End context switch"));
351 barrier(); /* Avoid RCU read-side critical sections leaking up. */
353 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
356 * Register a quiescent state for all RCU flavors. If there is an
357 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
358 * dyntick-idle quiescent state visible to other CPUs (but only for those
359 * RCU flavors in desperate need of a quiescent state, which will normally
360 * be none of them). Either way, do a lightweight quiescent state for
363 * The barrier() calls are redundant in the common case when this is
364 * called externally, but just in case this is called from within this
368 void rcu_all_qs(void)
372 barrier(); /* Avoid RCU read-side critical sections leaking down. */
373 if (unlikely(raw_cpu_read(rcu_sched_qs_mask
))) {
374 local_irq_save(flags
);
375 rcu_momentary_dyntick_idle();
376 local_irq_restore(flags
);
378 if (unlikely(raw_cpu_read(rcu_sched_data
.cpu_no_qs
.b
.exp
))) {
380 * Yes, we just checked a per-CPU variable with preemption
381 * enabled, so we might be migrated to some other CPU at
382 * this point. That is OK because in that case, the
383 * migration will supply the needed quiescent state.
384 * We might end up needlessly disabling preemption and
385 * invoking rcu_sched_qs() on the destination CPU, but
386 * the probability and cost are both quite low, so this
387 * should not be a problem in practice.
393 this_cpu_inc(rcu_qs_ctr
);
394 barrier(); /* Avoid RCU read-side critical sections leaking up. */
396 EXPORT_SYMBOL_GPL(rcu_all_qs
);
398 static long blimit
= 10; /* Maximum callbacks per rcu_do_batch. */
399 static long qhimark
= 10000; /* If this many pending, ignore blimit. */
400 static long qlowmark
= 100; /* Once only this many pending, use blimit. */
402 module_param(blimit
, long, 0444);
403 module_param(qhimark
, long, 0444);
404 module_param(qlowmark
, long, 0444);
406 static ulong jiffies_till_first_fqs
= ULONG_MAX
;
407 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
408 static bool rcu_kick_kthreads
;
410 module_param(jiffies_till_first_fqs
, ulong
, 0644);
411 module_param(jiffies_till_next_fqs
, ulong
, 0644);
412 module_param(rcu_kick_kthreads
, bool, 0644);
415 * How long the grace period must be before we start recruiting
416 * quiescent-state help from rcu_note_context_switch().
418 static ulong jiffies_till_sched_qs
= HZ
/ 20;
419 module_param(jiffies_till_sched_qs
, ulong
, 0644);
421 static bool rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
422 struct rcu_data
*rdp
);
423 static void force_qs_rnp(struct rcu_state
*rsp
,
424 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
425 unsigned long *maxj
),
426 bool *isidle
, unsigned long *maxj
);
427 static void force_quiescent_state(struct rcu_state
*rsp
);
428 static int rcu_pending(void);
431 * Return the number of RCU batches started thus far for debug & stats.
433 unsigned long rcu_batches_started(void)
435 return rcu_state_p
->gpnum
;
437 EXPORT_SYMBOL_GPL(rcu_batches_started
);
440 * Return the number of RCU-sched batches started thus far for debug & stats.
442 unsigned long rcu_batches_started_sched(void)
444 return rcu_sched_state
.gpnum
;
446 EXPORT_SYMBOL_GPL(rcu_batches_started_sched
);
449 * Return the number of RCU BH batches started thus far for debug & stats.
451 unsigned long rcu_batches_started_bh(void)
453 return rcu_bh_state
.gpnum
;
455 EXPORT_SYMBOL_GPL(rcu_batches_started_bh
);
458 * Return the number of RCU batches completed thus far for debug & stats.
460 unsigned long rcu_batches_completed(void)
462 return rcu_state_p
->completed
;
464 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
467 * Return the number of RCU-sched batches completed thus far for debug & stats.
469 unsigned long rcu_batches_completed_sched(void)
471 return rcu_sched_state
.completed
;
473 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
476 * Return the number of RCU BH batches completed thus far for debug & stats.
478 unsigned long rcu_batches_completed_bh(void)
480 return rcu_bh_state
.completed
;
482 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
485 * Return the number of RCU expedited batches completed thus far for
486 * debug & stats. Odd numbers mean that a batch is in progress, even
487 * numbers mean idle. The value returned will thus be roughly double
488 * the cumulative batches since boot.
490 unsigned long rcu_exp_batches_completed(void)
492 return rcu_state_p
->expedited_sequence
;
494 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed
);
497 * Return the number of RCU-sched expedited batches completed thus far
498 * for debug & stats. Similar to rcu_exp_batches_completed().
500 unsigned long rcu_exp_batches_completed_sched(void)
502 return rcu_sched_state
.expedited_sequence
;
504 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched
);
507 * Force a quiescent state.
509 void rcu_force_quiescent_state(void)
511 force_quiescent_state(rcu_state_p
);
513 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
516 * Force a quiescent state for RCU BH.
518 void rcu_bh_force_quiescent_state(void)
520 force_quiescent_state(&rcu_bh_state
);
522 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
525 * Force a quiescent state for RCU-sched.
527 void rcu_sched_force_quiescent_state(void)
529 force_quiescent_state(&rcu_sched_state
);
531 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
534 * Show the state of the grace-period kthreads.
536 void show_rcu_gp_kthreads(void)
538 struct rcu_state
*rsp
;
540 for_each_rcu_flavor(rsp
) {
541 pr_info("%s: wait state: %d ->state: %#lx\n",
542 rsp
->name
, rsp
->gp_state
, rsp
->gp_kthread
->state
);
543 /* sched_show_task(rsp->gp_kthread); */
546 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads
);
549 * Record the number of times rcutorture tests have been initiated and
550 * terminated. This information allows the debugfs tracing stats to be
551 * correlated to the rcutorture messages, even when the rcutorture module
552 * is being repeatedly loaded and unloaded. In other words, we cannot
553 * store this state in rcutorture itself.
555 void rcutorture_record_test_transition(void)
557 rcutorture_testseq
++;
558 rcutorture_vernum
= 0;
560 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
563 * Send along grace-period-related data for rcutorture diagnostics.
565 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
566 unsigned long *gpnum
, unsigned long *completed
)
568 struct rcu_state
*rsp
= NULL
;
577 case RCU_SCHED_FLAVOR
:
578 rsp
= &rcu_sched_state
;
584 *flags
= READ_ONCE(rsp
->gp_flags
);
585 *gpnum
= READ_ONCE(rsp
->gpnum
);
586 *completed
= READ_ONCE(rsp
->completed
);
593 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
596 * Record the number of writer passes through the current rcutorture test.
597 * This is also used to correlate debugfs tracing stats with the rcutorture
600 void rcutorture_record_progress(unsigned long vernum
)
604 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
607 * Does the CPU have callbacks ready to be invoked?
610 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
612 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
] &&
613 rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
;
617 * Return the root node of the specified rcu_state structure.
619 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
621 return &rsp
->node
[0];
625 * Is there any need for future grace periods?
626 * Interrupts must be disabled. If the caller does not hold the root
627 * rnp_node structure's ->lock, the results are advisory only.
629 static int rcu_future_needs_gp(struct rcu_state
*rsp
)
631 struct rcu_node
*rnp
= rcu_get_root(rsp
);
632 int idx
= (READ_ONCE(rnp
->completed
) + 1) & 0x1;
633 int *fp
= &rnp
->need_future_gp
[idx
];
635 return READ_ONCE(*fp
);
639 * Does the current CPU require a not-yet-started grace period?
640 * The caller must have disabled interrupts to prevent races with
641 * normal callback registry.
644 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
648 if (rcu_gp_in_progress(rsp
))
649 return false; /* No, a grace period is already in progress. */
650 if (rcu_future_needs_gp(rsp
))
651 return true; /* Yes, a no-CBs CPU needs one. */
652 if (!rdp
->nxttail
[RCU_NEXT_TAIL
])
653 return false; /* No, this is a no-CBs (or offline) CPU. */
654 if (*rdp
->nxttail
[RCU_NEXT_READY_TAIL
])
655 return true; /* Yes, CPU has newly registered callbacks. */
656 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
657 if (rdp
->nxttail
[i
- 1] != rdp
->nxttail
[i
] &&
658 ULONG_CMP_LT(READ_ONCE(rsp
->completed
),
659 rdp
->nxtcompleted
[i
]))
660 return true; /* Yes, CBs for future grace period. */
661 return false; /* No grace period needed. */
665 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
667 * If the new value of the ->dynticks_nesting counter now is zero,
668 * we really have entered idle, and must do the appropriate accounting.
669 * The caller must have disabled interrupts.
671 static void rcu_eqs_enter_common(long long oldval
, bool user
)
673 struct rcu_state
*rsp
;
674 struct rcu_data
*rdp
;
675 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
677 trace_rcu_dyntick(TPS("Start"), oldval
, rdtp
->dynticks_nesting
);
678 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
679 !user
&& !is_idle_task(current
)) {
680 struct task_struct
*idle __maybe_unused
=
681 idle_task(smp_processor_id());
683 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval
, 0);
684 rcu_ftrace_dump(DUMP_ORIG
);
685 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
686 current
->pid
, current
->comm
,
687 idle
->pid
, idle
->comm
); /* must be idle task! */
689 for_each_rcu_flavor(rsp
) {
690 rdp
= this_cpu_ptr(rsp
->rda
);
691 do_nocb_deferred_wakeup(rdp
);
693 rcu_prepare_for_idle();
694 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
695 smp_mb__before_atomic(); /* See above. */
696 atomic_inc(&rdtp
->dynticks
);
697 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
698 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
699 atomic_read(&rdtp
->dynticks
) & 0x1);
700 rcu_dynticks_task_enter();
703 * It is illegal to enter an extended quiescent state while
704 * in an RCU read-side critical section.
706 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map
),
707 "Illegal idle entry in RCU read-side critical section.");
708 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
),
709 "Illegal idle entry in RCU-bh read-side critical section.");
710 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map
),
711 "Illegal idle entry in RCU-sched read-side critical section.");
715 * Enter an RCU extended quiescent state, which can be either the
716 * idle loop or adaptive-tickless usermode execution.
718 static void rcu_eqs_enter(bool user
)
721 struct rcu_dynticks
*rdtp
;
723 rdtp
= this_cpu_ptr(&rcu_dynticks
);
724 oldval
= rdtp
->dynticks_nesting
;
725 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
726 (oldval
& DYNTICK_TASK_NEST_MASK
) == 0);
727 if ((oldval
& DYNTICK_TASK_NEST_MASK
) == DYNTICK_TASK_NEST_VALUE
) {
728 rdtp
->dynticks_nesting
= 0;
729 rcu_eqs_enter_common(oldval
, user
);
731 rdtp
->dynticks_nesting
-= DYNTICK_TASK_NEST_VALUE
;
736 * rcu_idle_enter - inform RCU that current CPU is entering idle
738 * Enter idle mode, in other words, -leave- the mode in which RCU
739 * read-side critical sections can occur. (Though RCU read-side
740 * critical sections can occur in irq handlers in idle, a possibility
741 * handled by irq_enter() and irq_exit().)
743 * We crowbar the ->dynticks_nesting field to zero to allow for
744 * the possibility of usermode upcalls having messed up our count
745 * of interrupt nesting level during the prior busy period.
747 void rcu_idle_enter(void)
751 local_irq_save(flags
);
752 rcu_eqs_enter(false);
753 rcu_sysidle_enter(0);
754 local_irq_restore(flags
);
756 EXPORT_SYMBOL_GPL(rcu_idle_enter
);
758 #ifdef CONFIG_NO_HZ_FULL
760 * rcu_user_enter - inform RCU that we are resuming userspace.
762 * Enter RCU idle mode right before resuming userspace. No use of RCU
763 * is permitted between this call and rcu_user_exit(). This way the
764 * CPU doesn't need to maintain the tick for RCU maintenance purposes
765 * when the CPU runs in userspace.
767 void rcu_user_enter(void)
771 #endif /* CONFIG_NO_HZ_FULL */
774 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
776 * Exit from an interrupt handler, which might possibly result in entering
777 * idle mode, in other words, leaving the mode in which read-side critical
778 * sections can occur. The caller must have disabled interrupts.
780 * This code assumes that the idle loop never does anything that might
781 * result in unbalanced calls to irq_enter() and irq_exit(). If your
782 * architecture violates this assumption, RCU will give you what you
783 * deserve, good and hard. But very infrequently and irreproducibly.
785 * Use things like work queues to work around this limitation.
787 * You have been warned.
789 void rcu_irq_exit(void)
792 struct rcu_dynticks
*rdtp
;
794 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
795 rdtp
= this_cpu_ptr(&rcu_dynticks
);
796 oldval
= rdtp
->dynticks_nesting
;
797 rdtp
->dynticks_nesting
--;
798 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
799 rdtp
->dynticks_nesting
< 0);
800 if (rdtp
->dynticks_nesting
)
801 trace_rcu_dyntick(TPS("--="), oldval
, rdtp
->dynticks_nesting
);
803 rcu_eqs_enter_common(oldval
, true);
804 rcu_sysidle_enter(1);
808 * Wrapper for rcu_irq_exit() where interrupts are enabled.
810 void rcu_irq_exit_irqson(void)
814 local_irq_save(flags
);
816 local_irq_restore(flags
);
820 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
822 * If the new value of the ->dynticks_nesting counter was previously zero,
823 * we really have exited idle, and must do the appropriate accounting.
824 * The caller must have disabled interrupts.
826 static void rcu_eqs_exit_common(long long oldval
, int user
)
828 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
830 rcu_dynticks_task_exit();
831 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
832 atomic_inc(&rdtp
->dynticks
);
833 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
834 smp_mb__after_atomic(); /* See above. */
835 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
836 !(atomic_read(&rdtp
->dynticks
) & 0x1));
837 rcu_cleanup_after_idle();
838 trace_rcu_dyntick(TPS("End"), oldval
, rdtp
->dynticks_nesting
);
839 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
840 !user
&& !is_idle_task(current
)) {
841 struct task_struct
*idle __maybe_unused
=
842 idle_task(smp_processor_id());
844 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
845 oldval
, rdtp
->dynticks_nesting
);
846 rcu_ftrace_dump(DUMP_ORIG
);
847 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
848 current
->pid
, current
->comm
,
849 idle
->pid
, idle
->comm
); /* must be idle task! */
854 * Exit an RCU extended quiescent state, which can be either the
855 * idle loop or adaptive-tickless usermode execution.
857 static void rcu_eqs_exit(bool user
)
859 struct rcu_dynticks
*rdtp
;
862 rdtp
= this_cpu_ptr(&rcu_dynticks
);
863 oldval
= rdtp
->dynticks_nesting
;
864 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) && oldval
< 0);
865 if (oldval
& DYNTICK_TASK_NEST_MASK
) {
866 rdtp
->dynticks_nesting
+= DYNTICK_TASK_NEST_VALUE
;
868 rdtp
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
869 rcu_eqs_exit_common(oldval
, user
);
874 * rcu_idle_exit - inform RCU that current CPU is leaving idle
876 * Exit idle mode, in other words, -enter- the mode in which RCU
877 * read-side critical sections can occur.
879 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
880 * allow for the possibility of usermode upcalls messing up our count
881 * of interrupt nesting level during the busy period that is just
884 void rcu_idle_exit(void)
888 local_irq_save(flags
);
891 local_irq_restore(flags
);
893 EXPORT_SYMBOL_GPL(rcu_idle_exit
);
895 #ifdef CONFIG_NO_HZ_FULL
897 * rcu_user_exit - inform RCU that we are exiting userspace.
899 * Exit RCU idle mode while entering the kernel because it can
900 * run a RCU read side critical section anytime.
902 void rcu_user_exit(void)
906 #endif /* CONFIG_NO_HZ_FULL */
909 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
911 * Enter an interrupt handler, which might possibly result in exiting
912 * idle mode, in other words, entering the mode in which read-side critical
913 * sections can occur. The caller must have disabled interrupts.
915 * Note that the Linux kernel is fully capable of entering an interrupt
916 * handler that it never exits, for example when doing upcalls to
917 * user mode! This code assumes that the idle loop never does upcalls to
918 * user mode. If your architecture does do upcalls from the idle loop (or
919 * does anything else that results in unbalanced calls to the irq_enter()
920 * and irq_exit() functions), RCU will give you what you deserve, good
921 * and hard. But very infrequently and irreproducibly.
923 * Use things like work queues to work around this limitation.
925 * You have been warned.
927 void rcu_irq_enter(void)
929 struct rcu_dynticks
*rdtp
;
932 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
933 rdtp
= this_cpu_ptr(&rcu_dynticks
);
934 oldval
= rdtp
->dynticks_nesting
;
935 rdtp
->dynticks_nesting
++;
936 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG
) &&
937 rdtp
->dynticks_nesting
== 0);
939 trace_rcu_dyntick(TPS("++="), oldval
, rdtp
->dynticks_nesting
);
941 rcu_eqs_exit_common(oldval
, true);
946 * Wrapper for rcu_irq_enter() where interrupts are enabled.
948 void rcu_irq_enter_irqson(void)
952 local_irq_save(flags
);
954 local_irq_restore(flags
);
958 * rcu_nmi_enter - inform RCU of entry to NMI context
960 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
961 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
962 * that the CPU is active. This implementation permits nested NMIs, as
963 * long as the nesting level does not overflow an int. (You will probably
964 * run out of stack space first.)
966 void rcu_nmi_enter(void)
968 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
971 /* Complain about underflow. */
972 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
< 0);
975 * If idle from RCU viewpoint, atomically increment ->dynticks
976 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
977 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
978 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
979 * to be in the outermost NMI handler that interrupted an RCU-idle
980 * period (observation due to Andy Lutomirski).
982 if (!(atomic_read(&rdtp
->dynticks
) & 0x1)) {
983 smp_mb__before_atomic(); /* Force delay from prior write. */
984 atomic_inc(&rdtp
->dynticks
);
985 /* atomic_inc() before later RCU read-side crit sects */
986 smp_mb__after_atomic(); /* See above. */
987 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
990 rdtp
->dynticks_nmi_nesting
+= incby
;
995 * rcu_nmi_exit - inform RCU of exit from NMI context
997 * If we are returning from the outermost NMI handler that interrupted an
998 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
999 * to let the RCU grace-period handling know that the CPU is back to
1002 void rcu_nmi_exit(void)
1004 struct rcu_dynticks
*rdtp
= this_cpu_ptr(&rcu_dynticks
);
1007 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1008 * (We are exiting an NMI handler, so RCU better be paying attention
1011 WARN_ON_ONCE(rdtp
->dynticks_nmi_nesting
<= 0);
1012 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
1015 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1016 * leave it in non-RCU-idle state.
1018 if (rdtp
->dynticks_nmi_nesting
!= 1) {
1019 rdtp
->dynticks_nmi_nesting
-= 2;
1023 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1024 rdtp
->dynticks_nmi_nesting
= 0;
1025 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
1026 smp_mb__before_atomic(); /* See above. */
1027 atomic_inc(&rdtp
->dynticks
);
1028 smp_mb__after_atomic(); /* Force delay to next write. */
1029 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
1033 * __rcu_is_watching - are RCU read-side critical sections safe?
1035 * Return true if RCU is watching the running CPU, which means that
1036 * this CPU can safely enter RCU read-side critical sections. Unlike
1037 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1038 * least disabled preemption.
1040 bool notrace
__rcu_is_watching(void)
1042 return atomic_read(this_cpu_ptr(&rcu_dynticks
.dynticks
)) & 0x1;
1046 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1048 * If the current CPU is in its idle loop and is neither in an interrupt
1049 * or NMI handler, return true.
1051 bool notrace
rcu_is_watching(void)
1055 preempt_disable_notrace();
1056 ret
= __rcu_is_watching();
1057 preempt_enable_notrace();
1060 EXPORT_SYMBOL_GPL(rcu_is_watching
);
1062 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1065 * Is the current CPU online? Disable preemption to avoid false positives
1066 * that could otherwise happen due to the current CPU number being sampled,
1067 * this task being preempted, its old CPU being taken offline, resuming
1068 * on some other CPU, then determining that its old CPU is now offline.
1069 * It is OK to use RCU on an offline processor during initial boot, hence
1070 * the check for rcu_scheduler_fully_active. Note also that it is OK
1071 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1072 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1073 * offline to continue to use RCU for one jiffy after marking itself
1074 * offline in the cpu_online_mask. This leniency is necessary given the
1075 * non-atomic nature of the online and offline processing, for example,
1076 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1079 * This is also why RCU internally marks CPUs online during the
1080 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1082 * Disable checking if in an NMI handler because we cannot safely report
1083 * errors from NMI handlers anyway.
1085 bool rcu_lockdep_current_cpu_online(void)
1087 struct rcu_data
*rdp
;
1088 struct rcu_node
*rnp
;
1094 rdp
= this_cpu_ptr(&rcu_sched_data
);
1096 ret
= (rdp
->grpmask
& rcu_rnp_online_cpus(rnp
)) ||
1097 !rcu_scheduler_fully_active
;
1101 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
1103 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1106 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1108 * If the current CPU is idle or running at a first-level (not nested)
1109 * interrupt from idle, return true. The caller must have at least
1110 * disabled preemption.
1112 static int rcu_is_cpu_rrupt_from_idle(void)
1114 return __this_cpu_read(rcu_dynticks
.dynticks_nesting
) <= 1;
1118 * Snapshot the specified CPU's dynticks counter so that we can later
1119 * credit them with an implicit quiescent state. Return 1 if this CPU
1120 * is in dynticks idle mode, which is an extended quiescent state.
1122 static int dyntick_save_progress_counter(struct rcu_data
*rdp
,
1123 bool *isidle
, unsigned long *maxj
)
1125 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1126 rcu_sysidle_check_cpu(rdp
, isidle
, maxj
);
1127 if ((rdp
->dynticks_snap
& 0x1) == 0) {
1128 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1129 if (ULONG_CMP_LT(READ_ONCE(rdp
->gpnum
) + ULONG_MAX
/ 4,
1130 rdp
->mynode
->gpnum
))
1131 WRITE_ONCE(rdp
->gpwrap
, true);
1138 * Return true if the specified CPU has passed through a quiescent
1139 * state by virtue of being in or having passed through an dynticks
1140 * idle state since the last call to dyntick_save_progress_counter()
1141 * for this same CPU, or by virtue of having been offline.
1143 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
,
1144 bool *isidle
, unsigned long *maxj
)
1150 curr
= (unsigned int)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
1151 snap
= (unsigned int)rdp
->dynticks_snap
;
1154 * If the CPU passed through or entered a dynticks idle phase with
1155 * no active irq/NMI handlers, then we can safely pretend that the CPU
1156 * already acknowledged the request to pass through a quiescent
1157 * state. Either way, that CPU cannot possibly be in an RCU
1158 * read-side critical section that started before the beginning
1159 * of the current RCU grace period.
1161 if ((curr
& 0x1) == 0 || UINT_CMP_GE(curr
, snap
+ 2)) {
1162 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("dti"));
1163 rdp
->dynticks_fqs
++;
1168 * Check for the CPU being offline, but only if the grace period
1169 * is old enough. We don't need to worry about the CPU changing
1170 * state: If we see it offline even once, it has been through a
1173 * The reason for insisting that the grace period be at least
1174 * one jiffy old is that CPUs that are not quite online and that
1175 * have just gone offline can still execute RCU read-side critical
1178 if (ULONG_CMP_GE(rdp
->rsp
->gp_start
+ 2, jiffies
))
1179 return 0; /* Grace period is not old enough. */
1181 if (cpu_is_offline(rdp
->cpu
)) {
1182 trace_rcu_fqs(rdp
->rsp
->name
, rdp
->gpnum
, rdp
->cpu
, TPS("ofl"));
1188 * A CPU running for an extended time within the kernel can
1189 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1190 * even context-switching back and forth between a pair of
1191 * in-kernel CPU-bound tasks cannot advance grace periods.
1192 * So if the grace period is old enough, make the CPU pay attention.
1193 * Note that the unsynchronized assignments to the per-CPU
1194 * rcu_sched_qs_mask variable are safe. Yes, setting of
1195 * bits can be lost, but they will be set again on the next
1196 * force-quiescent-state pass. So lost bit sets do not result
1197 * in incorrect behavior, merely in a grace period lasting
1198 * a few jiffies longer than it might otherwise. Because
1199 * there are at most four threads involved, and because the
1200 * updates are only once every few jiffies, the probability of
1201 * lossage (and thus of slight grace-period extension) is
1204 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1205 * is set too high, we override with half of the RCU CPU stall
1208 rcrmp
= &per_cpu(rcu_sched_qs_mask
, rdp
->cpu
);
1209 if (ULONG_CMP_GE(jiffies
,
1210 rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
) ||
1211 ULONG_CMP_GE(jiffies
, rdp
->rsp
->jiffies_resched
)) {
1212 if (!(READ_ONCE(*rcrmp
) & rdp
->rsp
->flavor_mask
)) {
1213 WRITE_ONCE(rdp
->cond_resched_completed
,
1214 READ_ONCE(rdp
->mynode
->completed
));
1215 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1217 READ_ONCE(*rcrmp
) + rdp
->rsp
->flavor_mask
);
1219 rdp
->rsp
->jiffies_resched
+= 5; /* Re-enable beating. */
1222 /* And if it has been a really long time, kick the CPU as well. */
1223 if (ULONG_CMP_GE(jiffies
,
1224 rdp
->rsp
->gp_start
+ 2 * jiffies_till_sched_qs
) ||
1225 ULONG_CMP_GE(jiffies
, rdp
->rsp
->gp_start
+ jiffies_till_sched_qs
))
1226 resched_cpu(rdp
->cpu
); /* Force CPU into scheduler. */
1231 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
1233 unsigned long j
= jiffies
;
1237 smp_wmb(); /* Record start time before stall time. */
1238 j1
= rcu_jiffies_till_stall_check();
1239 WRITE_ONCE(rsp
->jiffies_stall
, j
+ j1
);
1240 rsp
->jiffies_resched
= j
+ j1
/ 2;
1241 rsp
->n_force_qs_gpstart
= READ_ONCE(rsp
->n_force_qs
);
1245 * Convert a ->gp_state value to a character string.
1247 static const char *gp_state_getname(short gs
)
1249 if (gs
< 0 || gs
>= ARRAY_SIZE(gp_state_names
))
1251 return gp_state_names
[gs
];
1255 * Complain about starvation of grace-period kthread.
1257 static void rcu_check_gp_kthread_starvation(struct rcu_state
*rsp
)
1263 gpa
= READ_ONCE(rsp
->gp_activity
);
1264 if (j
- gpa
> 2 * HZ
) {
1265 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1267 rsp
->gpnum
, rsp
->completed
,
1269 gp_state_getname(rsp
->gp_state
), rsp
->gp_state
,
1270 rsp
->gp_kthread
? rsp
->gp_kthread
->state
: ~0);
1271 if (rsp
->gp_kthread
) {
1272 sched_show_task(rsp
->gp_kthread
);
1273 wake_up_process(rsp
->gp_kthread
);
1279 * Dump stacks of all tasks running on stalled CPUs.
1281 static void rcu_dump_cpu_stacks(struct rcu_state
*rsp
)
1284 unsigned long flags
;
1285 struct rcu_node
*rnp
;
1287 rcu_for_each_leaf_node(rsp
, rnp
) {
1288 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1289 if (rnp
->qsmask
!= 0) {
1290 for_each_leaf_node_possible_cpu(rnp
, cpu
)
1291 if (rnp
->qsmask
& leaf_node_cpu_bit(rnp
, cpu
))
1294 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1299 * If too much time has passed in the current grace period, and if
1300 * so configured, go kick the relevant kthreads.
1302 static void rcu_stall_kick_kthreads(struct rcu_state
*rsp
)
1306 if (!rcu_kick_kthreads
)
1308 j
= READ_ONCE(rsp
->jiffies_kick_kthreads
);
1309 if (time_after(jiffies
, j
) && rsp
->gp_kthread
) {
1310 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp
->name
);
1311 rcu_ftrace_dump(DUMP_ALL
);
1312 wake_up_process(rsp
->gp_kthread
);
1313 WRITE_ONCE(rsp
->jiffies_kick_kthreads
, j
+ HZ
);
1317 static inline void panic_on_rcu_stall(void)
1319 if (sysctl_panic_on_rcu_stall
)
1320 panic("RCU Stall\n");
1323 static void print_other_cpu_stall(struct rcu_state
*rsp
, unsigned long gpnum
)
1327 unsigned long flags
;
1331 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1334 /* Kick and suppress, if so configured. */
1335 rcu_stall_kick_kthreads(rsp
);
1336 if (rcu_cpu_stall_suppress
)
1339 /* Only let one CPU complain about others per time interval. */
1341 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1342 delta
= jiffies
- READ_ONCE(rsp
->jiffies_stall
);
1343 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
1344 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1347 WRITE_ONCE(rsp
->jiffies_stall
,
1348 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1349 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1352 * OK, time to rat on our buddy...
1353 * See Documentation/RCU/stallwarn.txt for info on how to debug
1354 * RCU CPU stall warnings.
1356 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1358 print_cpu_stall_info_begin();
1359 rcu_for_each_leaf_node(rsp
, rnp
) {
1360 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1361 ndetected
+= rcu_print_task_stall(rnp
);
1362 if (rnp
->qsmask
!= 0) {
1363 for_each_leaf_node_possible_cpu(rnp
, cpu
)
1364 if (rnp
->qsmask
& leaf_node_cpu_bit(rnp
, cpu
)) {
1365 print_cpu_stall_info(rsp
, cpu
);
1369 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1372 print_cpu_stall_info_end();
1373 for_each_possible_cpu(cpu
)
1374 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1375 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1376 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
),
1377 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1379 rcu_dump_cpu_stacks(rsp
);
1381 if (READ_ONCE(rsp
->gpnum
) != gpnum
||
1382 READ_ONCE(rsp
->completed
) == gpnum
) {
1383 pr_err("INFO: Stall ended before state dump start\n");
1386 gpa
= READ_ONCE(rsp
->gp_activity
);
1387 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1388 rsp
->name
, j
- gpa
, j
, gpa
,
1389 jiffies_till_next_fqs
,
1390 rcu_get_root(rsp
)->qsmask
);
1391 /* In this case, the current CPU might be at fault. */
1392 sched_show_task(current
);
1396 /* Complain about tasks blocking the grace period. */
1397 rcu_print_detail_task_stall(rsp
);
1399 rcu_check_gp_kthread_starvation(rsp
);
1401 panic_on_rcu_stall();
1403 force_quiescent_state(rsp
); /* Kick them all. */
1406 static void print_cpu_stall(struct rcu_state
*rsp
)
1409 unsigned long flags
;
1410 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1413 /* Kick and suppress, if so configured. */
1414 rcu_stall_kick_kthreads(rsp
);
1415 if (rcu_cpu_stall_suppress
)
1419 * OK, time to rat on ourselves...
1420 * See Documentation/RCU/stallwarn.txt for info on how to debug
1421 * RCU CPU stall warnings.
1423 pr_err("INFO: %s self-detected stall on CPU", rsp
->name
);
1424 print_cpu_stall_info_begin();
1425 print_cpu_stall_info(rsp
, smp_processor_id());
1426 print_cpu_stall_info_end();
1427 for_each_possible_cpu(cpu
)
1428 totqlen
+= per_cpu_ptr(rsp
->rda
, cpu
)->qlen
;
1429 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1430 jiffies
- rsp
->gp_start
,
1431 (long)rsp
->gpnum
, (long)rsp
->completed
, totqlen
);
1433 rcu_check_gp_kthread_starvation(rsp
);
1435 rcu_dump_cpu_stacks(rsp
);
1437 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1438 if (ULONG_CMP_GE(jiffies
, READ_ONCE(rsp
->jiffies_stall
)))
1439 WRITE_ONCE(rsp
->jiffies_stall
,
1440 jiffies
+ 3 * rcu_jiffies_till_stall_check() + 3);
1441 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1443 panic_on_rcu_stall();
1446 * Attempt to revive the RCU machinery by forcing a context switch.
1448 * A context switch would normally allow the RCU state machine to make
1449 * progress and it could be we're stuck in kernel space without context
1450 * switches for an entirely unreasonable amount of time.
1452 resched_cpu(smp_processor_id());
1455 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1457 unsigned long completed
;
1458 unsigned long gpnum
;
1462 struct rcu_node
*rnp
;
1464 if ((rcu_cpu_stall_suppress
&& !rcu_kick_kthreads
) ||
1465 !rcu_gp_in_progress(rsp
))
1467 rcu_stall_kick_kthreads(rsp
);
1471 * Lots of memory barriers to reject false positives.
1473 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1474 * then rsp->gp_start, and finally rsp->completed. These values
1475 * are updated in the opposite order with memory barriers (or
1476 * equivalent) during grace-period initialization and cleanup.
1477 * Now, a false positive can occur if we get an new value of
1478 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1479 * the memory barriers, the only way that this can happen is if one
1480 * grace period ends and another starts between these two fetches.
1481 * Detect this by comparing rsp->completed with the previous fetch
1484 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1485 * and rsp->gp_start suffice to forestall false positives.
1487 gpnum
= READ_ONCE(rsp
->gpnum
);
1488 smp_rmb(); /* Pick up ->gpnum first... */
1489 js
= READ_ONCE(rsp
->jiffies_stall
);
1490 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1491 gps
= READ_ONCE(rsp
->gp_start
);
1492 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1493 completed
= READ_ONCE(rsp
->completed
);
1494 if (ULONG_CMP_GE(completed
, gpnum
) ||
1495 ULONG_CMP_LT(j
, js
) ||
1496 ULONG_CMP_GE(gps
, js
))
1497 return; /* No stall or GP completed since entering function. */
1499 if (rcu_gp_in_progress(rsp
) &&
1500 (READ_ONCE(rnp
->qsmask
) & rdp
->grpmask
)) {
1502 /* We haven't checked in, so go dump stack. */
1503 print_cpu_stall(rsp
);
1505 } else if (rcu_gp_in_progress(rsp
) &&
1506 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
1508 /* They had a few time units to dump stack, so complain. */
1509 print_other_cpu_stall(rsp
, gpnum
);
1514 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1516 * Set the stall-warning timeout way off into the future, thus preventing
1517 * any RCU CPU stall-warning messages from appearing in the current set of
1518 * RCU grace periods.
1520 * The caller must disable hard irqs.
1522 void rcu_cpu_stall_reset(void)
1524 struct rcu_state
*rsp
;
1526 for_each_rcu_flavor(rsp
)
1527 WRITE_ONCE(rsp
->jiffies_stall
, jiffies
+ ULONG_MAX
/ 2);
1531 * Initialize the specified rcu_data structure's default callback list
1532 * to empty. The default callback list is the one that is not used by
1533 * no-callbacks CPUs.
1535 static void init_default_callback_list(struct rcu_data
*rdp
)
1539 rdp
->nxtlist
= NULL
;
1540 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1541 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1545 * Initialize the specified rcu_data structure's callback list to empty.
1547 static void init_callback_list(struct rcu_data
*rdp
)
1549 if (init_nocb_callback_list(rdp
))
1551 init_default_callback_list(rdp
);
1555 * Determine the value that ->completed will have at the end of the
1556 * next subsequent grace period. This is used to tag callbacks so that
1557 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1558 * been dyntick-idle for an extended period with callbacks under the
1559 * influence of RCU_FAST_NO_HZ.
1561 * The caller must hold rnp->lock with interrupts disabled.
1563 static unsigned long rcu_cbs_completed(struct rcu_state
*rsp
,
1564 struct rcu_node
*rnp
)
1567 * If RCU is idle, we just wait for the next grace period.
1568 * But we can only be sure that RCU is idle if we are looking
1569 * at the root rcu_node structure -- otherwise, a new grace
1570 * period might have started, but just not yet gotten around
1571 * to initializing the current non-root rcu_node structure.
1573 if (rcu_get_root(rsp
) == rnp
&& rnp
->gpnum
== rnp
->completed
)
1574 return rnp
->completed
+ 1;
1577 * Otherwise, wait for a possible partial grace period and
1578 * then the subsequent full grace period.
1580 return rnp
->completed
+ 2;
1584 * Trace-event helper function for rcu_start_future_gp() and
1585 * rcu_nocb_wait_gp().
1587 static void trace_rcu_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1588 unsigned long c
, const char *s
)
1590 trace_rcu_future_grace_period(rdp
->rsp
->name
, rnp
->gpnum
,
1591 rnp
->completed
, c
, rnp
->level
,
1592 rnp
->grplo
, rnp
->grphi
, s
);
1596 * Start some future grace period, as needed to handle newly arrived
1597 * callbacks. The required future grace periods are recorded in each
1598 * rcu_node structure's ->need_future_gp field. Returns true if there
1599 * is reason to awaken the grace-period kthread.
1601 * The caller must hold the specified rcu_node structure's ->lock.
1603 static bool __maybe_unused
1604 rcu_start_future_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
1605 unsigned long *c_out
)
1610 struct rcu_node
*rnp_root
= rcu_get_root(rdp
->rsp
);
1613 * Pick up grace-period number for new callbacks. If this
1614 * grace period is already marked as needed, return to the caller.
1616 c
= rcu_cbs_completed(rdp
->rsp
, rnp
);
1617 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startleaf"));
1618 if (rnp
->need_future_gp
[c
& 0x1]) {
1619 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartleaf"));
1624 * If either this rcu_node structure or the root rcu_node structure
1625 * believe that a grace period is in progress, then we must wait
1626 * for the one following, which is in "c". Because our request
1627 * will be noticed at the end of the current grace period, we don't
1628 * need to explicitly start one. We only do the lockless check
1629 * of rnp_root's fields if the current rcu_node structure thinks
1630 * there is no grace period in flight, and because we hold rnp->lock,
1631 * the only possible change is when rnp_root's two fields are
1632 * equal, in which case rnp_root->gpnum might be concurrently
1633 * incremented. But that is OK, as it will just result in our
1634 * doing some extra useless work.
1636 if (rnp
->gpnum
!= rnp
->completed
||
1637 READ_ONCE(rnp_root
->gpnum
) != READ_ONCE(rnp_root
->completed
)) {
1638 rnp
->need_future_gp
[c
& 0x1]++;
1639 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleaf"));
1644 * There might be no grace period in progress. If we don't already
1645 * hold it, acquire the root rcu_node structure's lock in order to
1646 * start one (if needed).
1648 if (rnp
!= rnp_root
)
1649 raw_spin_lock_rcu_node(rnp_root
);
1652 * Get a new grace-period number. If there really is no grace
1653 * period in progress, it will be smaller than the one we obtained
1654 * earlier. Adjust callbacks as needed. Note that even no-CBs
1655 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1657 c
= rcu_cbs_completed(rdp
->rsp
, rnp_root
);
1658 for (i
= RCU_DONE_TAIL
; i
< RCU_NEXT_TAIL
; i
++)
1659 if (ULONG_CMP_LT(c
, rdp
->nxtcompleted
[i
]))
1660 rdp
->nxtcompleted
[i
] = c
;
1663 * If the needed for the required grace period is already
1664 * recorded, trace and leave.
1666 if (rnp_root
->need_future_gp
[c
& 0x1]) {
1667 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Prestartedroot"));
1671 /* Record the need for the future grace period. */
1672 rnp_root
->need_future_gp
[c
& 0x1]++;
1674 /* If a grace period is not already in progress, start one. */
1675 if (rnp_root
->gpnum
!= rnp_root
->completed
) {
1676 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedleafroot"));
1678 trace_rcu_future_gp(rnp
, rdp
, c
, TPS("Startedroot"));
1679 ret
= rcu_start_gp_advanced(rdp
->rsp
, rnp_root
, rdp
);
1682 if (rnp
!= rnp_root
)
1683 raw_spin_unlock_rcu_node(rnp_root
);
1691 * Clean up any old requests for the just-ended grace period. Also return
1692 * whether any additional grace periods have been requested. Also invoke
1693 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1694 * waiting for this grace period to complete.
1696 static int rcu_future_gp_cleanup(struct rcu_state
*rsp
, struct rcu_node
*rnp
)
1698 int c
= rnp
->completed
;
1700 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1702 rnp
->need_future_gp
[c
& 0x1] = 0;
1703 needmore
= rnp
->need_future_gp
[(c
+ 1) & 0x1];
1704 trace_rcu_future_gp(rnp
, rdp
, c
,
1705 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1710 * Awaken the grace-period kthread for the specified flavor of RCU.
1711 * Don't do a self-awaken, and don't bother awakening when there is
1712 * nothing for the grace-period kthread to do (as in several CPUs
1713 * raced to awaken, and we lost), and finally don't try to awaken
1714 * a kthread that has not yet been created.
1716 static void rcu_gp_kthread_wake(struct rcu_state
*rsp
)
1718 if (current
== rsp
->gp_kthread
||
1719 !READ_ONCE(rsp
->gp_flags
) ||
1722 swake_up(&rsp
->gp_wq
);
1726 * If there is room, assign a ->completed number to any callbacks on
1727 * this CPU that have not already been assigned. Also accelerate any
1728 * callbacks that were previously assigned a ->completed number that has
1729 * since proven to be too conservative, which can happen if callbacks get
1730 * assigned a ->completed number while RCU is idle, but with reference to
1731 * a non-root rcu_node structure. This function is idempotent, so it does
1732 * not hurt to call it repeatedly. Returns an flag saying that we should
1733 * awaken the RCU grace-period kthread.
1735 * The caller must hold rnp->lock with interrupts disabled.
1737 static bool rcu_accelerate_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1738 struct rcu_data
*rdp
)
1744 /* If the CPU has no callbacks, nothing to do. */
1745 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1749 * Starting from the sublist containing the callbacks most
1750 * recently assigned a ->completed number and working down, find the
1751 * first sublist that is not assignable to an upcoming grace period.
1752 * Such a sublist has something in it (first two tests) and has
1753 * a ->completed number assigned that will complete sooner than
1754 * the ->completed number for newly arrived callbacks (last test).
1756 * The key point is that any later sublist can be assigned the
1757 * same ->completed number as the newly arrived callbacks, which
1758 * means that the callbacks in any of these later sublist can be
1759 * grouped into a single sublist, whether or not they have already
1760 * been assigned a ->completed number.
1762 c
= rcu_cbs_completed(rsp
, rnp
);
1763 for (i
= RCU_NEXT_TAIL
- 1; i
> RCU_DONE_TAIL
; i
--)
1764 if (rdp
->nxttail
[i
] != rdp
->nxttail
[i
- 1] &&
1765 !ULONG_CMP_GE(rdp
->nxtcompleted
[i
], c
))
1769 * If there are no sublist for unassigned callbacks, leave.
1770 * At the same time, advance "i" one sublist, so that "i" will
1771 * index into the sublist where all the remaining callbacks should
1774 if (++i
>= RCU_NEXT_TAIL
)
1778 * Assign all subsequent callbacks' ->completed number to the next
1779 * full grace period and group them all in the sublist initially
1782 for (; i
<= RCU_NEXT_TAIL
; i
++) {
1783 rdp
->nxttail
[i
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1784 rdp
->nxtcompleted
[i
] = c
;
1786 /* Record any needed additional grace periods. */
1787 ret
= rcu_start_future_gp(rnp
, rdp
, NULL
);
1789 /* Trace depending on how much we were able to accelerate. */
1790 if (!*rdp
->nxttail
[RCU_WAIT_TAIL
])
1791 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccWaitCB"));
1793 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("AccReadyCB"));
1798 * Move any callbacks whose grace period has completed to the
1799 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1800 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1801 * sublist. This function is idempotent, so it does not hurt to
1802 * invoke it repeatedly. As long as it is not invoked -too- often...
1803 * Returns true if the RCU grace-period kthread needs to be awakened.
1805 * The caller must hold rnp->lock with interrupts disabled.
1807 static bool rcu_advance_cbs(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1808 struct rcu_data
*rdp
)
1812 /* If the CPU has no callbacks, nothing to do. */
1813 if (!rdp
->nxttail
[RCU_NEXT_TAIL
] || !*rdp
->nxttail
[RCU_DONE_TAIL
])
1817 * Find all callbacks whose ->completed numbers indicate that they
1818 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1820 for (i
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++) {
1821 if (ULONG_CMP_LT(rnp
->completed
, rdp
->nxtcompleted
[i
]))
1823 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[i
];
1825 /* Clean up any sublist tail pointers that were misordered above. */
1826 for (j
= RCU_WAIT_TAIL
; j
< i
; j
++)
1827 rdp
->nxttail
[j
] = rdp
->nxttail
[RCU_DONE_TAIL
];
1829 /* Copy down callbacks to fill in empty sublists. */
1830 for (j
= RCU_WAIT_TAIL
; i
< RCU_NEXT_TAIL
; i
++, j
++) {
1831 if (rdp
->nxttail
[j
] == rdp
->nxttail
[RCU_NEXT_TAIL
])
1833 rdp
->nxttail
[j
] = rdp
->nxttail
[i
];
1834 rdp
->nxtcompleted
[j
] = rdp
->nxtcompleted
[i
];
1837 /* Classify any remaining callbacks. */
1838 return rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1842 * Update CPU-local rcu_data state to record the beginnings and ends of
1843 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1844 * structure corresponding to the current CPU, and must have irqs disabled.
1845 * Returns true if the grace-period kthread needs to be awakened.
1847 static bool __note_gp_changes(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
1848 struct rcu_data
*rdp
)
1852 /* Handle the ends of any preceding grace periods first. */
1853 if (rdp
->completed
== rnp
->completed
&&
1854 !unlikely(READ_ONCE(rdp
->gpwrap
))) {
1856 /* No grace period end, so just accelerate recent callbacks. */
1857 ret
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
1861 /* Advance callbacks. */
1862 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
);
1864 /* Remember that we saw this grace-period completion. */
1865 rdp
->completed
= rnp
->completed
;
1866 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuend"));
1869 if (rdp
->gpnum
!= rnp
->gpnum
|| unlikely(READ_ONCE(rdp
->gpwrap
))) {
1871 * If the current grace period is waiting for this CPU,
1872 * set up to detect a quiescent state, otherwise don't
1873 * go looking for one.
1875 rdp
->gpnum
= rnp
->gpnum
;
1876 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpustart"));
1877 rdp
->cpu_no_qs
.b
.norm
= true;
1878 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
1879 rdp
->core_needs_qs
= !!(rnp
->qsmask
& rdp
->grpmask
);
1880 zero_cpu_stall_ticks(rdp
);
1881 WRITE_ONCE(rdp
->gpwrap
, false);
1886 static void note_gp_changes(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1888 unsigned long flags
;
1890 struct rcu_node
*rnp
;
1892 local_irq_save(flags
);
1894 if ((rdp
->gpnum
== READ_ONCE(rnp
->gpnum
) &&
1895 rdp
->completed
== READ_ONCE(rnp
->completed
) &&
1896 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1897 !raw_spin_trylock_rcu_node(rnp
)) { /* irqs already off, so later. */
1898 local_irq_restore(flags
);
1901 needwake
= __note_gp_changes(rsp
, rnp
, rdp
);
1902 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1904 rcu_gp_kthread_wake(rsp
);
1907 static void rcu_gp_slow(struct rcu_state
*rsp
, int delay
)
1910 !(rsp
->gpnum
% (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1911 schedule_timeout_uninterruptible(delay
);
1915 * Initialize a new grace period. Return false if no grace period required.
1917 static bool rcu_gp_init(struct rcu_state
*rsp
)
1919 unsigned long oldmask
;
1920 struct rcu_data
*rdp
;
1921 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1923 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
1924 raw_spin_lock_irq_rcu_node(rnp
);
1925 if (!READ_ONCE(rsp
->gp_flags
)) {
1926 /* Spurious wakeup, tell caller to go back to sleep. */
1927 raw_spin_unlock_irq_rcu_node(rnp
);
1930 WRITE_ONCE(rsp
->gp_flags
, 0); /* Clear all flags: New grace period. */
1932 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp
))) {
1934 * Grace period already in progress, don't start another.
1935 * Not supposed to be able to happen.
1937 raw_spin_unlock_irq_rcu_node(rnp
);
1941 /* Advance to a new grace period and initialize state. */
1942 record_gp_stall_check_time(rsp
);
1943 /* Record GP times before starting GP, hence smp_store_release(). */
1944 smp_store_release(&rsp
->gpnum
, rsp
->gpnum
+ 1);
1945 trace_rcu_grace_period(rsp
->name
, rsp
->gpnum
, TPS("start"));
1946 raw_spin_unlock_irq_rcu_node(rnp
);
1949 * Apply per-leaf buffered online and offline operations to the
1950 * rcu_node tree. Note that this new grace period need not wait
1951 * for subsequent online CPUs, and that quiescent-state forcing
1952 * will handle subsequent offline CPUs.
1954 rcu_for_each_leaf_node(rsp
, rnp
) {
1955 rcu_gp_slow(rsp
, gp_preinit_delay
);
1956 raw_spin_lock_irq_rcu_node(rnp
);
1957 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1958 !rnp
->wait_blkd_tasks
) {
1959 /* Nothing to do on this leaf rcu_node structure. */
1960 raw_spin_unlock_irq_rcu_node(rnp
);
1964 /* Record old state, apply changes to ->qsmaskinit field. */
1965 oldmask
= rnp
->qsmaskinit
;
1966 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1968 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1969 if (!oldmask
!= !rnp
->qsmaskinit
) {
1970 if (!oldmask
) /* First online CPU for this rcu_node. */
1971 rcu_init_new_rnp(rnp
);
1972 else if (rcu_preempt_has_tasks(rnp
)) /* blocked tasks */
1973 rnp
->wait_blkd_tasks
= true;
1974 else /* Last offline CPU and can propagate. */
1975 rcu_cleanup_dead_rnp(rnp
);
1979 * If all waited-on tasks from prior grace period are
1980 * done, and if all this rcu_node structure's CPUs are
1981 * still offline, propagate up the rcu_node tree and
1982 * clear ->wait_blkd_tasks. Otherwise, if one of this
1983 * rcu_node structure's CPUs has since come back online,
1984 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1985 * checks for this, so just call it unconditionally).
1987 if (rnp
->wait_blkd_tasks
&&
1988 (!rcu_preempt_has_tasks(rnp
) ||
1990 rnp
->wait_blkd_tasks
= false;
1991 rcu_cleanup_dead_rnp(rnp
);
1994 raw_spin_unlock_irq_rcu_node(rnp
);
1998 * Set the quiescent-state-needed bits in all the rcu_node
1999 * structures for all currently online CPUs in breadth-first order,
2000 * starting from the root rcu_node structure, relying on the layout
2001 * of the tree within the rsp->node[] array. Note that other CPUs
2002 * will access only the leaves of the hierarchy, thus seeing that no
2003 * grace period is in progress, at least until the corresponding
2004 * leaf node has been initialized.
2006 * The grace period cannot complete until the initialization
2007 * process finishes, because this kthread handles both.
2009 rcu_for_each_node_breadth_first(rsp
, rnp
) {
2010 rcu_gp_slow(rsp
, gp_init_delay
);
2011 raw_spin_lock_irq_rcu_node(rnp
);
2012 rdp
= this_cpu_ptr(rsp
->rda
);
2013 rcu_preempt_check_blocked_tasks(rnp
);
2014 rnp
->qsmask
= rnp
->qsmaskinit
;
2015 WRITE_ONCE(rnp
->gpnum
, rsp
->gpnum
);
2016 if (WARN_ON_ONCE(rnp
->completed
!= rsp
->completed
))
2017 WRITE_ONCE(rnp
->completed
, rsp
->completed
);
2018 if (rnp
== rdp
->mynode
)
2019 (void)__note_gp_changes(rsp
, rnp
, rdp
);
2020 rcu_preempt_boost_start_gp(rnp
);
2021 trace_rcu_grace_period_init(rsp
->name
, rnp
->gpnum
,
2022 rnp
->level
, rnp
->grplo
,
2023 rnp
->grphi
, rnp
->qsmask
);
2024 raw_spin_unlock_irq_rcu_node(rnp
);
2025 cond_resched_rcu_qs();
2026 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2033 * Helper function for wait_event_interruptible_timeout() wakeup
2034 * at force-quiescent-state time.
2036 static bool rcu_gp_fqs_check_wake(struct rcu_state
*rsp
, int *gfp
)
2038 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2040 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2041 *gfp
= READ_ONCE(rsp
->gp_flags
);
2042 if (*gfp
& RCU_GP_FLAG_FQS
)
2045 /* The current grace period has completed. */
2046 if (!READ_ONCE(rnp
->qsmask
) && !rcu_preempt_blocked_readers_cgp(rnp
))
2053 * Do one round of quiescent-state forcing.
2055 static void rcu_gp_fqs(struct rcu_state
*rsp
, bool first_time
)
2057 bool isidle
= false;
2059 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2061 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2064 /* Collect dyntick-idle snapshots. */
2065 if (is_sysidle_rcu_state(rsp
)) {
2067 maxj
= jiffies
- ULONG_MAX
/ 4;
2069 force_qs_rnp(rsp
, dyntick_save_progress_counter
,
2071 rcu_sysidle_report_gp(rsp
, isidle
, maxj
);
2073 /* Handle dyntick-idle and offline CPUs. */
2075 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
, &isidle
, &maxj
);
2077 /* Clear flag to prevent immediate re-entry. */
2078 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2079 raw_spin_lock_irq_rcu_node(rnp
);
2080 WRITE_ONCE(rsp
->gp_flags
,
2081 READ_ONCE(rsp
->gp_flags
) & ~RCU_GP_FLAG_FQS
);
2082 raw_spin_unlock_irq_rcu_node(rnp
);
2087 * Clean up after the old grace period.
2089 static void rcu_gp_cleanup(struct rcu_state
*rsp
)
2091 unsigned long gp_duration
;
2092 bool needgp
= false;
2094 struct rcu_data
*rdp
;
2095 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2096 struct swait_queue_head
*sq
;
2098 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2099 raw_spin_lock_irq_rcu_node(rnp
);
2100 gp_duration
= jiffies
- rsp
->gp_start
;
2101 if (gp_duration
> rsp
->gp_max
)
2102 rsp
->gp_max
= gp_duration
;
2105 * We know the grace period is complete, but to everyone else
2106 * it appears to still be ongoing. But it is also the case
2107 * that to everyone else it looks like there is nothing that
2108 * they can do to advance the grace period. It is therefore
2109 * safe for us to drop the lock in order to mark the grace
2110 * period as completed in all of the rcu_node structures.
2112 raw_spin_unlock_irq_rcu_node(rnp
);
2115 * Propagate new ->completed value to rcu_node structures so
2116 * that other CPUs don't have to wait until the start of the next
2117 * grace period to process their callbacks. This also avoids
2118 * some nasty RCU grace-period initialization races by forcing
2119 * the end of the current grace period to be completely recorded in
2120 * all of the rcu_node structures before the beginning of the next
2121 * grace period is recorded in any of the rcu_node structures.
2123 rcu_for_each_node_breadth_first(rsp
, rnp
) {
2124 raw_spin_lock_irq_rcu_node(rnp
);
2125 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
));
2126 WARN_ON_ONCE(rnp
->qsmask
);
2127 WRITE_ONCE(rnp
->completed
, rsp
->gpnum
);
2128 rdp
= this_cpu_ptr(rsp
->rda
);
2129 if (rnp
== rdp
->mynode
)
2130 needgp
= __note_gp_changes(rsp
, rnp
, rdp
) || needgp
;
2131 /* smp_mb() provided by prior unlock-lock pair. */
2132 nocb
+= rcu_future_gp_cleanup(rsp
, rnp
);
2133 sq
= rcu_nocb_gp_get(rnp
);
2134 raw_spin_unlock_irq_rcu_node(rnp
);
2135 rcu_nocb_gp_cleanup(sq
);
2136 cond_resched_rcu_qs();
2137 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2138 rcu_gp_slow(rsp
, gp_cleanup_delay
);
2140 rnp
= rcu_get_root(rsp
);
2141 raw_spin_lock_irq_rcu_node(rnp
); /* Order GP before ->completed update. */
2142 rcu_nocb_gp_set(rnp
, nocb
);
2144 /* Declare grace period done. */
2145 WRITE_ONCE(rsp
->completed
, rsp
->gpnum
);
2146 trace_rcu_grace_period(rsp
->name
, rsp
->completed
, TPS("end"));
2147 rsp
->gp_state
= RCU_GP_IDLE
;
2148 rdp
= this_cpu_ptr(rsp
->rda
);
2149 /* Advance CBs to reduce false positives below. */
2150 needgp
= rcu_advance_cbs(rsp
, rnp
, rdp
) || needgp
;
2151 if (needgp
|| cpu_needs_another_gp(rsp
, rdp
)) {
2152 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2153 trace_rcu_grace_period(rsp
->name
,
2154 READ_ONCE(rsp
->gpnum
),
2157 raw_spin_unlock_irq_rcu_node(rnp
);
2161 * Body of kthread that handles grace periods.
2163 static int __noreturn
rcu_gp_kthread(void *arg
)
2169 struct rcu_state
*rsp
= arg
;
2170 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2172 rcu_bind_gp_kthread();
2175 /* Handle grace-period start. */
2177 trace_rcu_grace_period(rsp
->name
,
2178 READ_ONCE(rsp
->gpnum
),
2180 rsp
->gp_state
= RCU_GP_WAIT_GPS
;
2181 swait_event_interruptible(rsp
->gp_wq
,
2182 READ_ONCE(rsp
->gp_flags
) &
2184 rsp
->gp_state
= RCU_GP_DONE_GPS
;
2185 /* Locking provides needed memory barrier. */
2186 if (rcu_gp_init(rsp
))
2188 cond_resched_rcu_qs();
2189 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2190 WARN_ON(signal_pending(current
));
2191 trace_rcu_grace_period(rsp
->name
,
2192 READ_ONCE(rsp
->gpnum
),
2196 /* Handle quiescent-state forcing. */
2197 first_gp_fqs
= true;
2198 j
= jiffies_till_first_fqs
;
2201 jiffies_till_first_fqs
= HZ
;
2206 rsp
->jiffies_force_qs
= jiffies
+ j
;
2207 WRITE_ONCE(rsp
->jiffies_kick_kthreads
,
2210 trace_rcu_grace_period(rsp
->name
,
2211 READ_ONCE(rsp
->gpnum
),
2213 rsp
->gp_state
= RCU_GP_WAIT_FQS
;
2214 ret
= swait_event_interruptible_timeout(rsp
->gp_wq
,
2215 rcu_gp_fqs_check_wake(rsp
, &gf
), j
);
2216 rsp
->gp_state
= RCU_GP_DOING_FQS
;
2217 /* Locking provides needed memory barriers. */
2218 /* If grace period done, leave loop. */
2219 if (!READ_ONCE(rnp
->qsmask
) &&
2220 !rcu_preempt_blocked_readers_cgp(rnp
))
2222 /* If time for quiescent-state forcing, do it. */
2223 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_force_qs
) ||
2224 (gf
& RCU_GP_FLAG_FQS
)) {
2225 trace_rcu_grace_period(rsp
->name
,
2226 READ_ONCE(rsp
->gpnum
),
2228 rcu_gp_fqs(rsp
, first_gp_fqs
);
2229 first_gp_fqs
= false;
2230 trace_rcu_grace_period(rsp
->name
,
2231 READ_ONCE(rsp
->gpnum
),
2233 cond_resched_rcu_qs();
2234 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2235 ret
= 0; /* Force full wait till next FQS. */
2236 j
= jiffies_till_next_fqs
;
2239 jiffies_till_next_fqs
= HZ
;
2242 jiffies_till_next_fqs
= 1;
2245 /* Deal with stray signal. */
2246 cond_resched_rcu_qs();
2247 WRITE_ONCE(rsp
->gp_activity
, jiffies
);
2248 WARN_ON(signal_pending(current
));
2249 trace_rcu_grace_period(rsp
->name
,
2250 READ_ONCE(rsp
->gpnum
),
2252 ret
= 1; /* Keep old FQS timing. */
2254 if (time_after(jiffies
, rsp
->jiffies_force_qs
))
2257 j
= rsp
->jiffies_force_qs
- j
;
2261 /* Handle grace-period end. */
2262 rsp
->gp_state
= RCU_GP_CLEANUP
;
2263 rcu_gp_cleanup(rsp
);
2264 rsp
->gp_state
= RCU_GP_CLEANED
;
2269 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2270 * in preparation for detecting the next grace period. The caller must hold
2271 * the root node's ->lock and hard irqs must be disabled.
2273 * Note that it is legal for a dying CPU (which is marked as offline) to
2274 * invoke this function. This can happen when the dying CPU reports its
2277 * Returns true if the grace-period kthread must be awakened.
2280 rcu_start_gp_advanced(struct rcu_state
*rsp
, struct rcu_node
*rnp
,
2281 struct rcu_data
*rdp
)
2283 if (!rsp
->gp_kthread
|| !cpu_needs_another_gp(rsp
, rdp
)) {
2285 * Either we have not yet spawned the grace-period
2286 * task, this CPU does not need another grace period,
2287 * or a grace period is already in progress.
2288 * Either way, don't start a new grace period.
2292 WRITE_ONCE(rsp
->gp_flags
, RCU_GP_FLAG_INIT
);
2293 trace_rcu_grace_period(rsp
->name
, READ_ONCE(rsp
->gpnum
),
2297 * We can't do wakeups while holding the rnp->lock, as that
2298 * could cause possible deadlocks with the rq->lock. Defer
2299 * the wakeup to our caller.
2305 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2306 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2307 * is invoked indirectly from rcu_advance_cbs(), which would result in
2308 * endless recursion -- or would do so if it wasn't for the self-deadlock
2309 * that is encountered beforehand.
2311 * Returns true if the grace-period kthread needs to be awakened.
2313 static bool rcu_start_gp(struct rcu_state
*rsp
)
2315 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
2316 struct rcu_node
*rnp
= rcu_get_root(rsp
);
2320 * If there is no grace period in progress right now, any
2321 * callbacks we have up to this point will be satisfied by the
2322 * next grace period. Also, advancing the callbacks reduces the
2323 * probability of false positives from cpu_needs_another_gp()
2324 * resulting in pointless grace periods. So, advance callbacks
2325 * then start the grace period!
2327 ret
= rcu_advance_cbs(rsp
, rnp
, rdp
) || ret
;
2328 ret
= rcu_start_gp_advanced(rsp
, rnp
, rdp
) || ret
;
2333 * Report a full set of quiescent states to the specified rcu_state data
2334 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2335 * kthread if another grace period is required. Whether we wake
2336 * the grace-period kthread or it awakens itself for the next round
2337 * of quiescent-state forcing, that kthread will clean up after the
2338 * just-completed grace period. Note that the caller must hold rnp->lock,
2339 * which is released before return.
2341 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
2342 __releases(rcu_get_root(rsp
)->lock
)
2344 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
2345 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2346 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
2347 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2351 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2352 * Allows quiescent states for a group of CPUs to be reported at one go
2353 * to the specified rcu_node structure, though all the CPUs in the group
2354 * must be represented by the same rcu_node structure (which need not be a
2355 * leaf rcu_node structure, though it often will be). The gps parameter
2356 * is the grace-period snapshot, which means that the quiescent states
2357 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2358 * must be held upon entry, and it is released before return.
2361 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
2362 struct rcu_node
*rnp
, unsigned long gps
, unsigned long flags
)
2363 __releases(rnp
->lock
)
2365 unsigned long oldmask
= 0;
2366 struct rcu_node
*rnp_c
;
2368 /* Walk up the rcu_node hierarchy. */
2370 if (!(rnp
->qsmask
& mask
) || rnp
->gpnum
!= gps
) {
2373 * Our bit has already been cleared, or the
2374 * relevant grace period is already over, so done.
2376 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2379 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
2380 rnp
->qsmask
&= ~mask
;
2381 trace_rcu_quiescent_state_report(rsp
->name
, rnp
->gpnum
,
2382 mask
, rnp
->qsmask
, rnp
->level
,
2383 rnp
->grplo
, rnp
->grphi
,
2385 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2387 /* Other bits still set at this level, so done. */
2388 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2391 mask
= rnp
->grpmask
;
2392 if (rnp
->parent
== NULL
) {
2394 /* No more levels. Exit loop holding root lock. */
2398 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2401 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2402 oldmask
= rnp_c
->qsmask
;
2406 * Get here if we are the last CPU to pass through a quiescent
2407 * state for this grace period. Invoke rcu_report_qs_rsp()
2408 * to clean up and start the next grace period if one is needed.
2410 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
2414 * Record a quiescent state for all tasks that were previously queued
2415 * on the specified rcu_node structure and that were blocking the current
2416 * RCU grace period. The caller must hold the specified rnp->lock with
2417 * irqs disabled, and this lock is released upon return, but irqs remain
2420 static void rcu_report_unblock_qs_rnp(struct rcu_state
*rsp
,
2421 struct rcu_node
*rnp
, unsigned long flags
)
2422 __releases(rnp
->lock
)
2426 struct rcu_node
*rnp_p
;
2428 if (rcu_state_p
== &rcu_sched_state
|| rsp
!= rcu_state_p
||
2429 rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
2430 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2431 return; /* Still need more quiescent states! */
2434 rnp_p
= rnp
->parent
;
2435 if (rnp_p
== NULL
) {
2437 * Only one rcu_node structure in the tree, so don't
2438 * try to report up to its nonexistent parent!
2440 rcu_report_qs_rsp(rsp
, flags
);
2444 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2446 mask
= rnp
->grpmask
;
2447 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2448 raw_spin_lock_rcu_node(rnp_p
); /* irqs already disabled. */
2449 rcu_report_qs_rnp(mask
, rsp
, rnp_p
, gps
, flags
);
2453 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2454 * structure. This must be called from the specified CPU.
2457 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2459 unsigned long flags
;
2462 struct rcu_node
*rnp
;
2465 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2466 if ((rdp
->cpu_no_qs
.b
.norm
&&
2467 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) ||
2468 rdp
->gpnum
!= rnp
->gpnum
|| rnp
->completed
== rnp
->gpnum
||
2472 * The grace period in which this quiescent state was
2473 * recorded has ended, so don't report it upwards.
2474 * We will instead need a new quiescent state that lies
2475 * within the current grace period.
2477 rdp
->cpu_no_qs
.b
.norm
= true; /* need qs for new gp. */
2478 rdp
->rcu_qs_ctr_snap
= __this_cpu_read(rcu_qs_ctr
);
2479 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2482 mask
= rdp
->grpmask
;
2483 if ((rnp
->qsmask
& mask
) == 0) {
2484 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2486 rdp
->core_needs_qs
= false;
2489 * This GP can't end until cpu checks in, so all of our
2490 * callbacks can be processed during the next GP.
2492 needwake
= rcu_accelerate_cbs(rsp
, rnp
, rdp
);
2494 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2495 /* ^^^ Released rnp->lock */
2497 rcu_gp_kthread_wake(rsp
);
2502 * Check to see if there is a new grace period of which this CPU
2503 * is not yet aware, and if so, set up local rcu_data state for it.
2504 * Otherwise, see if this CPU has just passed through its first
2505 * quiescent state for this grace period, and record that fact if so.
2508 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2510 /* Check for grace-period ends and beginnings. */
2511 note_gp_changes(rsp
, rdp
);
2514 * Does this CPU still need to do its part for current grace period?
2515 * If no, return and let the other CPUs do their part as well.
2517 if (!rdp
->core_needs_qs
)
2521 * Was there a quiescent state since the beginning of the grace
2522 * period? If no, then exit and wait for the next call.
2524 if (rdp
->cpu_no_qs
.b
.norm
&&
2525 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
))
2529 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2532 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
);
2536 * Send the specified CPU's RCU callbacks to the orphanage. The
2537 * specified CPU must be offline, and the caller must hold the
2541 rcu_send_cbs_to_orphanage(int cpu
, struct rcu_state
*rsp
,
2542 struct rcu_node
*rnp
, struct rcu_data
*rdp
)
2544 /* No-CBs CPUs do not have orphanable callbacks. */
2545 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) || rcu_is_nocb_cpu(rdp
->cpu
))
2549 * Orphan the callbacks. First adjust the counts. This is safe
2550 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2551 * cannot be running now. Thus no memory barrier is required.
2553 if (rdp
->nxtlist
!= NULL
) {
2554 rsp
->qlen_lazy
+= rdp
->qlen_lazy
;
2555 rsp
->qlen
+= rdp
->qlen
;
2556 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
2558 WRITE_ONCE(rdp
->qlen
, 0);
2562 * Next, move those callbacks still needing a grace period to
2563 * the orphanage, where some other CPU will pick them up.
2564 * Some of the callbacks might have gone partway through a grace
2565 * period, but that is too bad. They get to start over because we
2566 * cannot assume that grace periods are synchronized across CPUs.
2567 * We don't bother updating the ->nxttail[] array yet, instead
2568 * we just reset the whole thing later on.
2570 if (*rdp
->nxttail
[RCU_DONE_TAIL
] != NULL
) {
2571 *rsp
->orphan_nxttail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2572 rsp
->orphan_nxttail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
2573 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2577 * Then move the ready-to-invoke callbacks to the orphanage,
2578 * where some other CPU will pick them up. These will not be
2579 * required to pass though another grace period: They are done.
2581 if (rdp
->nxtlist
!= NULL
) {
2582 *rsp
->orphan_donetail
= rdp
->nxtlist
;
2583 rsp
->orphan_donetail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2587 * Finally, initialize the rcu_data structure's list to empty and
2588 * disallow further callbacks on this CPU.
2590 init_callback_list(rdp
);
2591 rdp
->nxttail
[RCU_NEXT_TAIL
] = NULL
;
2595 * Adopt the RCU callbacks from the specified rcu_state structure's
2596 * orphanage. The caller must hold the ->orphan_lock.
2598 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
, unsigned long flags
)
2601 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2603 /* No-CBs CPUs are handled specially. */
2604 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2605 rcu_nocb_adopt_orphan_cbs(rsp
, rdp
, flags
))
2608 /* Do the accounting first. */
2609 rdp
->qlen_lazy
+= rsp
->qlen_lazy
;
2610 rdp
->qlen
+= rsp
->qlen
;
2611 rdp
->n_cbs_adopted
+= rsp
->qlen
;
2612 if (rsp
->qlen_lazy
!= rsp
->qlen
)
2613 rcu_idle_count_callbacks_posted();
2618 * We do not need a memory barrier here because the only way we
2619 * can get here if there is an rcu_barrier() in flight is if
2620 * we are the task doing the rcu_barrier().
2623 /* First adopt the ready-to-invoke callbacks. */
2624 if (rsp
->orphan_donelist
!= NULL
) {
2625 *rsp
->orphan_donetail
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2626 *rdp
->nxttail
[RCU_DONE_TAIL
] = rsp
->orphan_donelist
;
2627 for (i
= RCU_NEXT_SIZE
- 1; i
>= RCU_DONE_TAIL
; i
--)
2628 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2629 rdp
->nxttail
[i
] = rsp
->orphan_donetail
;
2630 rsp
->orphan_donelist
= NULL
;
2631 rsp
->orphan_donetail
= &rsp
->orphan_donelist
;
2634 /* And then adopt the callbacks that still need a grace period. */
2635 if (rsp
->orphan_nxtlist
!= NULL
) {
2636 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxtlist
;
2637 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_nxttail
;
2638 rsp
->orphan_nxtlist
= NULL
;
2639 rsp
->orphan_nxttail
= &rsp
->orphan_nxtlist
;
2644 * Trace the fact that this CPU is going offline.
2646 static void rcu_cleanup_dying_cpu(struct rcu_state
*rsp
)
2648 RCU_TRACE(unsigned long mask
);
2649 RCU_TRACE(struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
));
2650 RCU_TRACE(struct rcu_node
*rnp
= rdp
->mynode
);
2652 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2655 RCU_TRACE(mask
= rdp
->grpmask
);
2656 trace_rcu_grace_period(rsp
->name
,
2657 rnp
->gpnum
+ 1 - !!(rnp
->qsmask
& mask
),
2662 * All CPUs for the specified rcu_node structure have gone offline,
2663 * and all tasks that were preempted within an RCU read-side critical
2664 * section while running on one of those CPUs have since exited their RCU
2665 * read-side critical section. Some other CPU is reporting this fact with
2666 * the specified rcu_node structure's ->lock held and interrupts disabled.
2667 * This function therefore goes up the tree of rcu_node structures,
2668 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2669 * the leaf rcu_node structure's ->qsmaskinit field has already been
2672 * This function does check that the specified rcu_node structure has
2673 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2674 * prematurely. That said, invoking it after the fact will cost you
2675 * a needless lock acquisition. So once it has done its work, don't
2678 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2681 struct rcu_node
*rnp
= rnp_leaf
;
2683 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2684 rnp
->qsmaskinit
|| rcu_preempt_has_tasks(rnp
))
2687 mask
= rnp
->grpmask
;
2691 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
2692 rnp
->qsmaskinit
&= ~mask
;
2693 rnp
->qsmask
&= ~mask
;
2694 if (rnp
->qsmaskinit
) {
2695 raw_spin_unlock_rcu_node(rnp
);
2696 /* irqs remain disabled. */
2699 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2704 * The CPU has been completely removed, and some other CPU is reporting
2705 * this fact from process context. Do the remainder of the cleanup,
2706 * including orphaning the outgoing CPU's RCU callbacks, and also
2707 * adopting them. There can only be one CPU hotplug operation at a time,
2708 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2710 static void rcu_cleanup_dead_cpu(int cpu
, struct rcu_state
*rsp
)
2712 unsigned long flags
;
2713 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
2714 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2716 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2719 /* Adjust any no-longer-needed kthreads. */
2720 rcu_boost_kthread_setaffinity(rnp
, -1);
2722 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2723 raw_spin_lock_irqsave(&rsp
->orphan_lock
, flags
);
2724 rcu_send_cbs_to_orphanage(cpu
, rsp
, rnp
, rdp
);
2725 rcu_adopt_orphan_cbs(rsp
, flags
);
2726 raw_spin_unlock_irqrestore(&rsp
->orphan_lock
, flags
);
2728 WARN_ONCE(rdp
->qlen
!= 0 || rdp
->nxtlist
!= NULL
,
2729 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2730 cpu
, rdp
->qlen
, rdp
->nxtlist
);
2734 * Invoke any RCU callbacks that have made it to the end of their grace
2735 * period. Thottle as specified by rdp->blimit.
2737 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
2739 unsigned long flags
;
2740 struct rcu_head
*next
, *list
, **tail
;
2741 long bl
, count
, count_lazy
;
2744 /* If no callbacks are ready, just return. */
2745 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
2746 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, 0);
2747 trace_rcu_batch_end(rsp
->name
, 0, !!READ_ONCE(rdp
->nxtlist
),
2748 need_resched(), is_idle_task(current
),
2749 rcu_is_callbacks_kthread());
2754 * Extract the list of ready callbacks, disabling to prevent
2755 * races with call_rcu() from interrupt handlers.
2757 local_irq_save(flags
);
2758 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2760 trace_rcu_batch_start(rsp
->name
, rdp
->qlen_lazy
, rdp
->qlen
, bl
);
2761 list
= rdp
->nxtlist
;
2762 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
2763 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
2764 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
2765 for (i
= RCU_NEXT_SIZE
- 1; i
>= 0; i
--)
2766 if (rdp
->nxttail
[i
] == rdp
->nxttail
[RCU_DONE_TAIL
])
2767 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
2768 local_irq_restore(flags
);
2770 /* Invoke callbacks. */
2771 count
= count_lazy
= 0;
2775 debug_rcu_head_unqueue(list
);
2776 if (__rcu_reclaim(rsp
->name
, list
))
2779 /* Stop only if limit reached and CPU has something to do. */
2780 if (++count
>= bl
&&
2782 (!is_idle_task(current
) && !rcu_is_callbacks_kthread())))
2786 local_irq_save(flags
);
2787 trace_rcu_batch_end(rsp
->name
, count
, !!list
, need_resched(),
2788 is_idle_task(current
),
2789 rcu_is_callbacks_kthread());
2791 /* Update count, and requeue any remaining callbacks. */
2793 *tail
= rdp
->nxtlist
;
2794 rdp
->nxtlist
= list
;
2795 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
2796 if (&rdp
->nxtlist
== rdp
->nxttail
[i
])
2797 rdp
->nxttail
[i
] = tail
;
2801 smp_mb(); /* List handling before counting for rcu_barrier(). */
2802 rdp
->qlen_lazy
-= count_lazy
;
2803 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
- count
);
2804 rdp
->n_cbs_invoked
+= count
;
2806 /* Reinstate batch limit if we have worked down the excess. */
2807 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
2808 rdp
->blimit
= blimit
;
2810 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2811 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2812 rdp
->qlen_last_fqs_check
= 0;
2813 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
2814 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
2815 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
2816 WARN_ON_ONCE((rdp
->nxtlist
== NULL
) != (rdp
->qlen
== 0));
2818 local_irq_restore(flags
);
2820 /* Re-invoke RCU core processing if there are callbacks remaining. */
2821 if (cpu_has_callbacks_ready_to_invoke(rdp
))
2826 * Check to see if this CPU is in a non-context-switch quiescent state
2827 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2828 * Also schedule RCU core processing.
2830 * This function must be called from hardirq context. It is normally
2831 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2832 * false, there is no point in invoking rcu_check_callbacks().
2834 void rcu_check_callbacks(int user
)
2836 trace_rcu_utilization(TPS("Start scheduler-tick"));
2837 increment_cpu_stall_ticks();
2838 if (user
|| rcu_is_cpu_rrupt_from_idle()) {
2841 * Get here if this CPU took its interrupt from user
2842 * mode or from the idle loop, and if this is not a
2843 * nested interrupt. In this case, the CPU is in
2844 * a quiescent state, so note it.
2846 * No memory barrier is required here because both
2847 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2848 * variables that other CPUs neither access nor modify,
2849 * at least not while the corresponding CPU is online.
2855 } else if (!in_softirq()) {
2858 * Get here if this CPU did not take its interrupt from
2859 * softirq, in other words, if it is not interrupting
2860 * a rcu_bh read-side critical section. This is an _bh
2861 * critical section, so note it.
2866 rcu_preempt_check_callbacks();
2870 rcu_note_voluntary_context_switch(current
);
2871 trace_rcu_utilization(TPS("End scheduler-tick"));
2875 * Scan the leaf rcu_node structures, processing dyntick state for any that
2876 * have not yet encountered a quiescent state, using the function specified.
2877 * Also initiate boosting for any threads blocked on the root rcu_node.
2879 * The caller must have suppressed start of new grace periods.
2881 static void force_qs_rnp(struct rcu_state
*rsp
,
2882 int (*f
)(struct rcu_data
*rsp
, bool *isidle
,
2883 unsigned long *maxj
),
2884 bool *isidle
, unsigned long *maxj
)
2887 unsigned long flags
;
2889 struct rcu_node
*rnp
;
2891 rcu_for_each_leaf_node(rsp
, rnp
) {
2892 cond_resched_rcu_qs();
2894 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2895 if (rnp
->qsmask
== 0) {
2896 if (rcu_state_p
== &rcu_sched_state
||
2897 rsp
!= rcu_state_p
||
2898 rcu_preempt_blocked_readers_cgp(rnp
)) {
2900 * No point in scanning bits because they
2901 * are all zero. But we might need to
2902 * priority-boost blocked readers.
2904 rcu_initiate_boost(rnp
, flags
);
2905 /* rcu_initiate_boost() releases rnp->lock */
2909 (rnp
->parent
->qsmask
& rnp
->grpmask
)) {
2911 * Race between grace-period
2912 * initialization and task exiting RCU
2913 * read-side critical section: Report.
2915 rcu_report_unblock_qs_rnp(rsp
, rnp
, flags
);
2916 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2920 for_each_leaf_node_possible_cpu(rnp
, cpu
) {
2921 unsigned long bit
= leaf_node_cpu_bit(rnp
, cpu
);
2922 if ((rnp
->qsmask
& bit
) != 0) {
2923 if (f(per_cpu_ptr(rsp
->rda
, cpu
), isidle
, maxj
))
2928 /* Idle/offline CPUs, report (releases rnp->lock. */
2929 rcu_report_qs_rnp(mask
, rsp
, rnp
, rnp
->gpnum
, flags
);
2931 /* Nothing to do here, so just drop the lock. */
2932 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2938 * Force quiescent states on reluctant CPUs, and also detect which
2939 * CPUs are in dyntick-idle mode.
2941 static void force_quiescent_state(struct rcu_state
*rsp
)
2943 unsigned long flags
;
2945 struct rcu_node
*rnp
;
2946 struct rcu_node
*rnp_old
= NULL
;
2948 /* Funnel through hierarchy to reduce memory contention. */
2949 rnp
= __this_cpu_read(rsp
->rda
->mynode
);
2950 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2951 ret
= (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) ||
2952 !raw_spin_trylock(&rnp
->fqslock
);
2953 if (rnp_old
!= NULL
)
2954 raw_spin_unlock(&rnp_old
->fqslock
);
2956 rsp
->n_force_qs_lh
++;
2961 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2963 /* Reached the root of the rcu_node tree, acquire lock. */
2964 raw_spin_lock_irqsave_rcu_node(rnp_old
, flags
);
2965 raw_spin_unlock(&rnp_old
->fqslock
);
2966 if (READ_ONCE(rsp
->gp_flags
) & RCU_GP_FLAG_FQS
) {
2967 rsp
->n_force_qs_lh
++;
2968 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2969 return; /* Someone beat us to it. */
2971 WRITE_ONCE(rsp
->gp_flags
, READ_ONCE(rsp
->gp_flags
) | RCU_GP_FLAG_FQS
);
2972 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2973 swake_up(&rsp
->gp_wq
); /* Memory barrier implied by swake_up() path. */
2977 * This does the RCU core processing work for the specified rcu_state
2978 * and rcu_data structures. This may be called only from the CPU to
2979 * whom the rdp belongs.
2982 __rcu_process_callbacks(struct rcu_state
*rsp
)
2984 unsigned long flags
;
2986 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
2988 WARN_ON_ONCE(rdp
->beenonline
== 0);
2990 /* Update RCU state based on any recent quiescent states. */
2991 rcu_check_quiescent_state(rsp
, rdp
);
2993 /* Does this CPU require a not-yet-started grace period? */
2994 local_irq_save(flags
);
2995 if (cpu_needs_another_gp(rsp
, rdp
)) {
2996 raw_spin_lock_rcu_node(rcu_get_root(rsp
)); /* irqs disabled. */
2997 needwake
= rcu_start_gp(rsp
);
2998 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp
), flags
);
3000 rcu_gp_kthread_wake(rsp
);
3002 local_irq_restore(flags
);
3005 /* If there are callbacks ready, invoke them. */
3006 if (cpu_has_callbacks_ready_to_invoke(rdp
))
3007 invoke_rcu_callbacks(rsp
, rdp
);
3009 /* Do any needed deferred wakeups of rcuo kthreads. */
3010 do_nocb_deferred_wakeup(rdp
);
3014 * Do RCU core processing for the current CPU.
3016 static void rcu_process_callbacks(struct softirq_action
*unused
)
3018 struct rcu_state
*rsp
;
3020 if (cpu_is_offline(smp_processor_id()))
3022 trace_rcu_utilization(TPS("Start RCU core"));
3023 for_each_rcu_flavor(rsp
)
3024 __rcu_process_callbacks(rsp
);
3025 trace_rcu_utilization(TPS("End RCU core"));
3029 * Schedule RCU callback invocation. If the specified type of RCU
3030 * does not support RCU priority boosting, just do a direct call,
3031 * otherwise wake up the per-CPU kernel kthread. Note that because we
3032 * are running on the current CPU with softirqs disabled, the
3033 * rcu_cpu_kthread_task cannot disappear out from under us.
3035 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3037 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active
)))
3039 if (likely(!rsp
->boost
)) {
3040 rcu_do_batch(rsp
, rdp
);
3043 invoke_rcu_callbacks_kthread();
3046 static void invoke_rcu_core(void)
3048 if (cpu_online(smp_processor_id()))
3049 raise_softirq(RCU_SOFTIRQ
);
3053 * Handle any core-RCU processing required by a call_rcu() invocation.
3055 static void __call_rcu_core(struct rcu_state
*rsp
, struct rcu_data
*rdp
,
3056 struct rcu_head
*head
, unsigned long flags
)
3061 * If called from an extended quiescent state, invoke the RCU
3062 * core in order to force a re-evaluation of RCU's idleness.
3064 if (!rcu_is_watching())
3067 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3068 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
3072 * Force the grace period if too many callbacks or too long waiting.
3073 * Enforce hysteresis, and don't invoke force_quiescent_state()
3074 * if some other CPU has recently done so. Also, don't bother
3075 * invoking force_quiescent_state() if the newly enqueued callback
3076 * is the only one waiting for a grace period to complete.
3078 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
3080 /* Are we ignoring a completed grace period? */
3081 note_gp_changes(rsp
, rdp
);
3083 /* Start a new grace period if one not already started. */
3084 if (!rcu_gp_in_progress(rsp
)) {
3085 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
3087 raw_spin_lock_rcu_node(rnp_root
);
3088 needwake
= rcu_start_gp(rsp
);
3089 raw_spin_unlock_rcu_node(rnp_root
);
3091 rcu_gp_kthread_wake(rsp
);
3093 /* Give the grace period a kick. */
3094 rdp
->blimit
= LONG_MAX
;
3095 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
3096 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
3097 force_quiescent_state(rsp
);
3098 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3099 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
3105 * RCU callback function to leak a callback.
3107 static void rcu_leak_callback(struct rcu_head
*rhp
)
3112 * Helper function for call_rcu() and friends. The cpu argument will
3113 * normally be -1, indicating "currently running CPU". It may specify
3114 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3115 * is expected to specify a CPU.
3118 __call_rcu(struct rcu_head
*head
, rcu_callback_t func
,
3119 struct rcu_state
*rsp
, int cpu
, bool lazy
)
3121 unsigned long flags
;
3122 struct rcu_data
*rdp
;
3124 WARN_ON_ONCE((unsigned long)head
& 0x1); /* Misaligned rcu_head! */
3125 if (debug_rcu_head_queue(head
)) {
3126 /* Probable double call_rcu(), so leak the callback. */
3127 WRITE_ONCE(head
->func
, rcu_leak_callback
);
3128 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3135 * Opportunistically note grace-period endings and beginnings.
3136 * Note that we might see a beginning right after we see an
3137 * end, but never vice versa, since this CPU has to pass through
3138 * a quiescent state betweentimes.
3140 local_irq_save(flags
);
3141 rdp
= this_cpu_ptr(rsp
->rda
);
3143 /* Add the callback to our list. */
3144 if (unlikely(rdp
->nxttail
[RCU_NEXT_TAIL
] == NULL
) || cpu
!= -1) {
3148 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3149 if (likely(rdp
->mynode
)) {
3150 /* Post-boot, so this should be for a no-CBs CPU. */
3151 offline
= !__call_rcu_nocb(rdp
, head
, lazy
, flags
);
3152 WARN_ON_ONCE(offline
);
3153 /* Offline CPU, _call_rcu() illegal, leak callback. */
3154 local_irq_restore(flags
);
3158 * Very early boot, before rcu_init(). Initialize if needed
3159 * and then drop through to queue the callback.
3162 WARN_ON_ONCE(!rcu_is_watching());
3163 if (!likely(rdp
->nxtlist
))
3164 init_default_callback_list(rdp
);
3166 WRITE_ONCE(rdp
->qlen
, rdp
->qlen
+ 1);
3170 rcu_idle_count_callbacks_posted();
3171 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3172 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
3173 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
3175 if (__is_kfree_rcu_offset((unsigned long)func
))
3176 trace_rcu_kfree_callback(rsp
->name
, head
, (unsigned long)func
,
3177 rdp
->qlen_lazy
, rdp
->qlen
);
3179 trace_rcu_callback(rsp
->name
, head
, rdp
->qlen_lazy
, rdp
->qlen
);
3181 /* Go handle any RCU core processing required. */
3182 __call_rcu_core(rsp
, rdp
, head
, flags
);
3183 local_irq_restore(flags
);
3187 * Queue an RCU-sched callback for invocation after a grace period.
3189 void call_rcu_sched(struct rcu_head
*head
, rcu_callback_t func
)
3191 __call_rcu(head
, func
, &rcu_sched_state
, -1, 0);
3193 EXPORT_SYMBOL_GPL(call_rcu_sched
);
3196 * Queue an RCU callback for invocation after a quicker grace period.
3198 void call_rcu_bh(struct rcu_head
*head
, rcu_callback_t func
)
3200 __call_rcu(head
, func
, &rcu_bh_state
, -1, 0);
3202 EXPORT_SYMBOL_GPL(call_rcu_bh
);
3205 * Queue an RCU callback for lazy invocation after a grace period.
3206 * This will likely be later named something like "call_rcu_lazy()",
3207 * but this change will require some way of tagging the lazy RCU
3208 * callbacks in the list of pending callbacks. Until then, this
3209 * function may only be called from __kfree_rcu().
3211 void kfree_call_rcu(struct rcu_head
*head
,
3212 rcu_callback_t func
)
3214 __call_rcu(head
, func
, rcu_state_p
, -1, 1);
3216 EXPORT_SYMBOL_GPL(kfree_call_rcu
);
3219 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3220 * any blocking grace-period wait automatically implies a grace period
3221 * if there is only one CPU online at any point time during execution
3222 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3223 * occasionally incorrectly indicate that there are multiple CPUs online
3224 * when there was in fact only one the whole time, as this just adds
3225 * some overhead: RCU still operates correctly.
3227 static inline int rcu_blocking_is_gp(void)
3231 might_sleep(); /* Check for RCU read-side critical section. */
3233 ret
= num_online_cpus() <= 1;
3239 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3241 * Control will return to the caller some time after a full rcu-sched
3242 * grace period has elapsed, in other words after all currently executing
3243 * rcu-sched read-side critical sections have completed. These read-side
3244 * critical sections are delimited by rcu_read_lock_sched() and
3245 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3246 * local_irq_disable(), and so on may be used in place of
3247 * rcu_read_lock_sched().
3249 * This means that all preempt_disable code sequences, including NMI and
3250 * non-threaded hardware-interrupt handlers, in progress on entry will
3251 * have completed before this primitive returns. However, this does not
3252 * guarantee that softirq handlers will have completed, since in some
3253 * kernels, these handlers can run in process context, and can block.
3255 * Note that this guarantee implies further memory-ordering guarantees.
3256 * On systems with more than one CPU, when synchronize_sched() returns,
3257 * each CPU is guaranteed to have executed a full memory barrier since the
3258 * end of its last RCU-sched read-side critical section whose beginning
3259 * preceded the call to synchronize_sched(). In addition, each CPU having
3260 * an RCU read-side critical section that extends beyond the return from
3261 * synchronize_sched() is guaranteed to have executed a full memory barrier
3262 * after the beginning of synchronize_sched() and before the beginning of
3263 * that RCU read-side critical section. Note that these guarantees include
3264 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3265 * that are executing in the kernel.
3267 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3268 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3269 * to have executed a full memory barrier during the execution of
3270 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3271 * again only if the system has more than one CPU).
3273 * This primitive provides the guarantees made by the (now removed)
3274 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3275 * guarantees that rcu_read_lock() sections will have completed.
3276 * In "classic RCU", these two guarantees happen to be one and
3277 * the same, but can differ in realtime RCU implementations.
3279 void synchronize_sched(void)
3281 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3282 lock_is_held(&rcu_lock_map
) ||
3283 lock_is_held(&rcu_sched_lock_map
),
3284 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3285 if (rcu_blocking_is_gp())
3287 if (rcu_gp_is_expedited())
3288 synchronize_sched_expedited();
3290 wait_rcu_gp(call_rcu_sched
);
3292 EXPORT_SYMBOL_GPL(synchronize_sched
);
3295 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3297 * Control will return to the caller some time after a full rcu_bh grace
3298 * period has elapsed, in other words after all currently executing rcu_bh
3299 * read-side critical sections have completed. RCU read-side critical
3300 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3301 * and may be nested.
3303 * See the description of synchronize_sched() for more detailed information
3304 * on memory ordering guarantees.
3306 void synchronize_rcu_bh(void)
3308 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3309 lock_is_held(&rcu_lock_map
) ||
3310 lock_is_held(&rcu_sched_lock_map
),
3311 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3312 if (rcu_blocking_is_gp())
3314 if (rcu_gp_is_expedited())
3315 synchronize_rcu_bh_expedited();
3317 wait_rcu_gp(call_rcu_bh
);
3319 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
3322 * get_state_synchronize_rcu - Snapshot current RCU state
3324 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3325 * to determine whether or not a full grace period has elapsed in the
3328 unsigned long get_state_synchronize_rcu(void)
3331 * Any prior manipulation of RCU-protected data must happen
3332 * before the load from ->gpnum.
3337 * Make sure this load happens before the purportedly
3338 * time-consuming work between get_state_synchronize_rcu()
3339 * and cond_synchronize_rcu().
3341 return smp_load_acquire(&rcu_state_p
->gpnum
);
3343 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3346 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3348 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3350 * If a full RCU grace period has elapsed since the earlier call to
3351 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3352 * synchronize_rcu() to wait for a full grace period.
3354 * Yes, this function does not take counter wrap into account. But
3355 * counter wrap is harmless. If the counter wraps, we have waited for
3356 * more than 2 billion grace periods (and way more on a 64-bit system!),
3357 * so waiting for one additional grace period should be just fine.
3359 void cond_synchronize_rcu(unsigned long oldstate
)
3361 unsigned long newstate
;
3364 * Ensure that this load happens before any RCU-destructive
3365 * actions the caller might carry out after we return.
3367 newstate
= smp_load_acquire(&rcu_state_p
->completed
);
3368 if (ULONG_CMP_GE(oldstate
, newstate
))
3371 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3374 * get_state_synchronize_sched - Snapshot current RCU-sched state
3376 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3377 * to determine whether or not a full grace period has elapsed in the
3380 unsigned long get_state_synchronize_sched(void)
3383 * Any prior manipulation of RCU-protected data must happen
3384 * before the load from ->gpnum.
3389 * Make sure this load happens before the purportedly
3390 * time-consuming work between get_state_synchronize_sched()
3391 * and cond_synchronize_sched().
3393 return smp_load_acquire(&rcu_sched_state
.gpnum
);
3395 EXPORT_SYMBOL_GPL(get_state_synchronize_sched
);
3398 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3400 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3402 * If a full RCU-sched grace period has elapsed since the earlier call to
3403 * get_state_synchronize_sched(), just return. Otherwise, invoke
3404 * synchronize_sched() to wait for a full grace period.
3406 * Yes, this function does not take counter wrap into account. But
3407 * counter wrap is harmless. If the counter wraps, we have waited for
3408 * more than 2 billion grace periods (and way more on a 64-bit system!),
3409 * so waiting for one additional grace period should be just fine.
3411 void cond_synchronize_sched(unsigned long oldstate
)
3413 unsigned long newstate
;
3416 * Ensure that this load happens before any RCU-destructive
3417 * actions the caller might carry out after we return.
3419 newstate
= smp_load_acquire(&rcu_sched_state
.completed
);
3420 if (ULONG_CMP_GE(oldstate
, newstate
))
3421 synchronize_sched();
3423 EXPORT_SYMBOL_GPL(cond_synchronize_sched
);
3425 /* Adjust sequence number for start of update-side operation. */
3426 static void rcu_seq_start(unsigned long *sp
)
3428 WRITE_ONCE(*sp
, *sp
+ 1);
3429 smp_mb(); /* Ensure update-side operation after counter increment. */
3430 WARN_ON_ONCE(!(*sp
& 0x1));
3433 /* Adjust sequence number for end of update-side operation. */
3434 static void rcu_seq_end(unsigned long *sp
)
3436 smp_mb(); /* Ensure update-side operation before counter increment. */
3437 WRITE_ONCE(*sp
, *sp
+ 1);
3438 WARN_ON_ONCE(*sp
& 0x1);
3441 /* Take a snapshot of the update side's sequence number. */
3442 static unsigned long rcu_seq_snap(unsigned long *sp
)
3446 s
= (READ_ONCE(*sp
) + 3) & ~0x1;
3447 smp_mb(); /* Above access must not bleed into critical section. */
3452 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3453 * full update-side operation has occurred.
3455 static bool rcu_seq_done(unsigned long *sp
, unsigned long s
)
3457 return ULONG_CMP_GE(READ_ONCE(*sp
), s
);
3461 * Check to see if there is any immediate RCU-related work to be done
3462 * by the current CPU, for the specified type of RCU, returning 1 if so.
3463 * The checks are in order of increasing expense: checks that can be
3464 * carried out against CPU-local state are performed first. However,
3465 * we must check for CPU stalls first, else we might not get a chance.
3467 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
3469 struct rcu_node
*rnp
= rdp
->mynode
;
3471 rdp
->n_rcu_pending
++;
3473 /* Check for CPU stalls, if enabled. */
3474 check_cpu_stall(rsp
, rdp
);
3476 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3477 if (rcu_nohz_full_cpu(rsp
))
3480 /* Is the RCU core waiting for a quiescent state from this CPU? */
3481 if (rcu_scheduler_fully_active
&&
3482 rdp
->core_needs_qs
&& rdp
->cpu_no_qs
.b
.norm
&&
3483 rdp
->rcu_qs_ctr_snap
== __this_cpu_read(rcu_qs_ctr
)) {
3484 rdp
->n_rp_core_needs_qs
++;
3485 } else if (rdp
->core_needs_qs
&&
3486 (!rdp
->cpu_no_qs
.b
.norm
||
3487 rdp
->rcu_qs_ctr_snap
!= __this_cpu_read(rcu_qs_ctr
))) {
3488 rdp
->n_rp_report_qs
++;
3492 /* Does this CPU have callbacks ready to invoke? */
3493 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
3494 rdp
->n_rp_cb_ready
++;
3498 /* Has RCU gone idle with this CPU needing another grace period? */
3499 if (cpu_needs_another_gp(rsp
, rdp
)) {
3500 rdp
->n_rp_cpu_needs_gp
++;
3504 /* Has another RCU grace period completed? */
3505 if (READ_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
3506 rdp
->n_rp_gp_completed
++;
3510 /* Has a new RCU grace period started? */
3511 if (READ_ONCE(rnp
->gpnum
) != rdp
->gpnum
||
3512 unlikely(READ_ONCE(rdp
->gpwrap
))) { /* outside lock */
3513 rdp
->n_rp_gp_started
++;
3517 /* Does this CPU need a deferred NOCB wakeup? */
3518 if (rcu_nocb_need_deferred_wakeup(rdp
)) {
3519 rdp
->n_rp_nocb_defer_wakeup
++;
3524 rdp
->n_rp_need_nothing
++;
3529 * Check to see if there is any immediate RCU-related work to be done
3530 * by the current CPU, returning 1 if so. This function is part of the
3531 * RCU implementation; it is -not- an exported member of the RCU API.
3533 static int rcu_pending(void)
3535 struct rcu_state
*rsp
;
3537 for_each_rcu_flavor(rsp
)
3538 if (__rcu_pending(rsp
, this_cpu_ptr(rsp
->rda
)))
3544 * Return true if the specified CPU has any callback. If all_lazy is
3545 * non-NULL, store an indication of whether all callbacks are lazy.
3546 * (If there are no callbacks, all of them are deemed to be lazy.)
3548 static bool __maybe_unused
rcu_cpu_has_callbacks(bool *all_lazy
)
3552 struct rcu_data
*rdp
;
3553 struct rcu_state
*rsp
;
3555 for_each_rcu_flavor(rsp
) {
3556 rdp
= this_cpu_ptr(rsp
->rda
);
3560 if (rdp
->qlen
!= rdp
->qlen_lazy
|| !all_lazy
) {
3571 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3572 * the compiler is expected to optimize this away.
3574 static void _rcu_barrier_trace(struct rcu_state
*rsp
, const char *s
,
3575 int cpu
, unsigned long done
)
3577 trace_rcu_barrier(rsp
->name
, s
, cpu
,
3578 atomic_read(&rsp
->barrier_cpu_count
), done
);
3582 * RCU callback function for _rcu_barrier(). If we are last, wake
3583 * up the task executing _rcu_barrier().
3585 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3587 struct rcu_data
*rdp
= container_of(rhp
, struct rcu_data
, barrier_head
);
3588 struct rcu_state
*rsp
= rdp
->rsp
;
3590 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
)) {
3591 _rcu_barrier_trace(rsp
, "LastCB", -1, rsp
->barrier_sequence
);
3592 complete(&rsp
->barrier_completion
);
3594 _rcu_barrier_trace(rsp
, "CB", -1, rsp
->barrier_sequence
);
3599 * Called with preemption disabled, and from cross-cpu IRQ context.
3601 static void rcu_barrier_func(void *type
)
3603 struct rcu_state
*rsp
= type
;
3604 struct rcu_data
*rdp
= raw_cpu_ptr(rsp
->rda
);
3606 _rcu_barrier_trace(rsp
, "IRQ", -1, rsp
->barrier_sequence
);
3607 atomic_inc(&rsp
->barrier_cpu_count
);
3608 rsp
->call(&rdp
->barrier_head
, rcu_barrier_callback
);
3612 * Orchestrate the specified type of RCU barrier, waiting for all
3613 * RCU callbacks of the specified type to complete.
3615 static void _rcu_barrier(struct rcu_state
*rsp
)
3618 struct rcu_data
*rdp
;
3619 unsigned long s
= rcu_seq_snap(&rsp
->barrier_sequence
);
3621 _rcu_barrier_trace(rsp
, "Begin", -1, s
);
3623 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3624 mutex_lock(&rsp
->barrier_mutex
);
3626 /* Did someone else do our work for us? */
3627 if (rcu_seq_done(&rsp
->barrier_sequence
, s
)) {
3628 _rcu_barrier_trace(rsp
, "EarlyExit", -1, rsp
->barrier_sequence
);
3629 smp_mb(); /* caller's subsequent code after above check. */
3630 mutex_unlock(&rsp
->barrier_mutex
);
3634 /* Mark the start of the barrier operation. */
3635 rcu_seq_start(&rsp
->barrier_sequence
);
3636 _rcu_barrier_trace(rsp
, "Inc1", -1, rsp
->barrier_sequence
);
3639 * Initialize the count to one rather than to zero in order to
3640 * avoid a too-soon return to zero in case of a short grace period
3641 * (or preemption of this task). Exclude CPU-hotplug operations
3642 * to ensure that no offline CPU has callbacks queued.
3644 init_completion(&rsp
->barrier_completion
);
3645 atomic_set(&rsp
->barrier_cpu_count
, 1);
3649 * Force each CPU with callbacks to register a new callback.
3650 * When that callback is invoked, we will know that all of the
3651 * corresponding CPU's preceding callbacks have been invoked.
3653 for_each_possible_cpu(cpu
) {
3654 if (!cpu_online(cpu
) && !rcu_is_nocb_cpu(cpu
))
3656 rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3657 if (rcu_is_nocb_cpu(cpu
)) {
3658 if (!rcu_nocb_cpu_needs_barrier(rsp
, cpu
)) {
3659 _rcu_barrier_trace(rsp
, "OfflineNoCB", cpu
,
3660 rsp
->barrier_sequence
);
3662 _rcu_barrier_trace(rsp
, "OnlineNoCB", cpu
,
3663 rsp
->barrier_sequence
);
3664 smp_mb__before_atomic();
3665 atomic_inc(&rsp
->barrier_cpu_count
);
3666 __call_rcu(&rdp
->barrier_head
,
3667 rcu_barrier_callback
, rsp
, cpu
, 0);
3669 } else if (READ_ONCE(rdp
->qlen
)) {
3670 _rcu_barrier_trace(rsp
, "OnlineQ", cpu
,
3671 rsp
->barrier_sequence
);
3672 smp_call_function_single(cpu
, rcu_barrier_func
, rsp
, 1);
3674 _rcu_barrier_trace(rsp
, "OnlineNQ", cpu
,
3675 rsp
->barrier_sequence
);
3681 * Now that we have an rcu_barrier_callback() callback on each
3682 * CPU, and thus each counted, remove the initial count.
3684 if (atomic_dec_and_test(&rsp
->barrier_cpu_count
))
3685 complete(&rsp
->barrier_completion
);
3687 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3688 wait_for_completion(&rsp
->barrier_completion
);
3690 /* Mark the end of the barrier operation. */
3691 _rcu_barrier_trace(rsp
, "Inc2", -1, rsp
->barrier_sequence
);
3692 rcu_seq_end(&rsp
->barrier_sequence
);
3694 /* Other rcu_barrier() invocations can now safely proceed. */
3695 mutex_unlock(&rsp
->barrier_mutex
);
3699 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3701 void rcu_barrier_bh(void)
3703 _rcu_barrier(&rcu_bh_state
);
3705 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
3708 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3710 void rcu_barrier_sched(void)
3712 _rcu_barrier(&rcu_sched_state
);
3714 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
3717 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3718 * first CPU in a given leaf rcu_node structure coming online. The caller
3719 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3722 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
3725 struct rcu_node
*rnp
= rnp_leaf
;
3728 mask
= rnp
->grpmask
;
3732 raw_spin_lock_rcu_node(rnp
); /* Interrupts already disabled. */
3733 rnp
->qsmaskinit
|= mask
;
3734 raw_spin_unlock_rcu_node(rnp
); /* Interrupts remain disabled. */
3739 * Do boot-time initialization of a CPU's per-CPU RCU data.
3742 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3744 unsigned long flags
;
3745 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3746 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3748 /* Set up local state, ensuring consistent view of global state. */
3749 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3750 rdp
->grpmask
= leaf_node_cpu_bit(rdp
->mynode
, cpu
);
3751 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
3752 WARN_ON_ONCE(rdp
->dynticks
->dynticks_nesting
!= DYNTICK_TASK_EXIT_IDLE
);
3753 WARN_ON_ONCE(atomic_read(&rdp
->dynticks
->dynticks
) != 1);
3756 rcu_boot_init_nocb_percpu_data(rdp
);
3757 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3761 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3762 * offline event can be happening at a given time. Note also that we
3763 * can accept some slop in the rsp->completed access due to the fact
3764 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3767 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
3769 unsigned long flags
;
3771 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3772 struct rcu_node
*rnp
= rcu_get_root(rsp
);
3774 /* Set up local state, ensuring consistent view of global state. */
3775 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3776 rdp
->qlen_last_fqs_check
= 0;
3777 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
3778 rdp
->blimit
= blimit
;
3780 init_callback_list(rdp
); /* Re-enable callbacks on this CPU. */
3781 rdp
->dynticks
->dynticks_nesting
= DYNTICK_TASK_EXIT_IDLE
;
3782 rcu_sysidle_init_percpu_data(rdp
->dynticks
);
3783 atomic_set(&rdp
->dynticks
->dynticks
,
3784 (atomic_read(&rdp
->dynticks
->dynticks
) & ~0x1) + 1);
3785 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
3788 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3789 * propagation up the rcu_node tree will happen at the beginning
3790 * of the next grace period.
3793 mask
= rdp
->grpmask
;
3794 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
3795 rnp
->qsmaskinitnext
|= mask
;
3796 rnp
->expmaskinitnext
|= mask
;
3797 if (!rdp
->beenonline
)
3798 WRITE_ONCE(rsp
->ncpus
, READ_ONCE(rsp
->ncpus
) + 1);
3799 rdp
->beenonline
= true; /* We have now been online. */
3800 rdp
->gpnum
= rnp
->completed
; /* Make CPU later note any new GP. */
3801 rdp
->completed
= rnp
->completed
;
3802 rdp
->cpu_no_qs
.b
.norm
= true;
3803 rdp
->rcu_qs_ctr_snap
= per_cpu(rcu_qs_ctr
, cpu
);
3804 rdp
->core_needs_qs
= false;
3805 trace_rcu_grace_period(rsp
->name
, rdp
->gpnum
, TPS("cpuonl"));
3806 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3809 static void rcu_prepare_cpu(int cpu
)
3811 struct rcu_state
*rsp
;
3813 for_each_rcu_flavor(rsp
)
3814 rcu_init_percpu_data(cpu
, rsp
);
3817 #ifdef CONFIG_HOTPLUG_CPU
3819 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3820 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3822 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3823 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3826 static void rcu_cleanup_dying_idle_cpu(int cpu
, struct rcu_state
*rsp
)
3828 unsigned long flags
;
3830 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
3831 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
3833 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3834 mask
= rdp
->grpmask
;
3835 raw_spin_lock_irqsave_rcu_node(rnp
, flags
); /* Enforce GP memory-order guarantee. */
3836 rnp
->qsmaskinitnext
&= ~mask
;
3837 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3840 void rcu_report_dead(unsigned int cpu
)
3842 struct rcu_state
*rsp
;
3844 /* QS for any half-done expedited RCU-sched GP. */
3846 rcu_report_exp_rdp(&rcu_sched_state
,
3847 this_cpu_ptr(rcu_sched_state
.rda
), true);
3849 for_each_rcu_flavor(rsp
)
3850 rcu_cleanup_dying_idle_cpu(cpu
, rsp
);
3855 * Handle CPU online/offline notification events.
3857 int rcu_cpu_notify(struct notifier_block
*self
,
3858 unsigned long action
, void *hcpu
)
3860 long cpu
= (long)hcpu
;
3861 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state_p
->rda
, cpu
);
3862 struct rcu_node
*rnp
= rdp
->mynode
;
3863 struct rcu_state
*rsp
;
3866 case CPU_UP_PREPARE
:
3867 case CPU_UP_PREPARE_FROZEN
:
3868 rcu_prepare_cpu(cpu
);
3869 rcu_prepare_kthreads(cpu
);
3870 rcu_spawn_all_nocb_kthreads(cpu
);
3873 case CPU_DOWN_FAILED
:
3874 sync_sched_exp_online_cleanup(cpu
);
3875 rcu_boost_kthread_setaffinity(rnp
, -1);
3877 case CPU_DOWN_PREPARE
:
3878 rcu_boost_kthread_setaffinity(rnp
, cpu
);
3881 case CPU_DYING_FROZEN
:
3882 for_each_rcu_flavor(rsp
)
3883 rcu_cleanup_dying_cpu(rsp
);
3886 case CPU_DEAD_FROZEN
:
3887 case CPU_UP_CANCELED
:
3888 case CPU_UP_CANCELED_FROZEN
:
3889 for_each_rcu_flavor(rsp
) {
3890 rcu_cleanup_dead_cpu(cpu
, rsp
);
3891 do_nocb_deferred_wakeup(per_cpu_ptr(rsp
->rda
, cpu
));
3900 static int rcu_pm_notify(struct notifier_block
*self
,
3901 unsigned long action
, void *hcpu
)
3904 case PM_HIBERNATION_PREPARE
:
3905 case PM_SUSPEND_PREPARE
:
3906 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3909 case PM_POST_HIBERNATION
:
3910 case PM_POST_SUSPEND
:
3911 if (nr_cpu_ids
<= 256) /* Expediting bad for large systems. */
3912 rcu_unexpedite_gp();
3921 * Spawn the kthreads that handle each RCU flavor's grace periods.
3923 static int __init
rcu_spawn_gp_kthread(void)
3925 unsigned long flags
;
3926 int kthread_prio_in
= kthread_prio
;
3927 struct rcu_node
*rnp
;
3928 struct rcu_state
*rsp
;
3929 struct sched_param sp
;
3930 struct task_struct
*t
;
3932 /* Force priority into range. */
3933 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
3935 else if (kthread_prio
< 0)
3937 else if (kthread_prio
> 99)
3939 if (kthread_prio
!= kthread_prio_in
)
3940 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3941 kthread_prio
, kthread_prio_in
);
3943 rcu_scheduler_fully_active
= 1;
3944 for_each_rcu_flavor(rsp
) {
3945 t
= kthread_create(rcu_gp_kthread
, rsp
, "%s", rsp
->name
);
3947 rnp
= rcu_get_root(rsp
);
3948 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
3949 rsp
->gp_kthread
= t
;
3951 sp
.sched_priority
= kthread_prio
;
3952 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
3954 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3957 rcu_spawn_nocb_kthreads();
3958 rcu_spawn_boost_kthreads();
3961 early_initcall(rcu_spawn_gp_kthread
);
3964 * This function is invoked towards the end of the scheduler's initialization
3965 * process. Before this is called, the idle task might contain
3966 * RCU read-side critical sections (during which time, this idle
3967 * task is booting the system). After this function is called, the
3968 * idle tasks are prohibited from containing RCU read-side critical
3969 * sections. This function also enables RCU lockdep checking.
3971 void rcu_scheduler_starting(void)
3973 WARN_ON(num_online_cpus() != 1);
3974 WARN_ON(nr_context_switches() > 0);
3975 rcu_scheduler_active
= 1;
3979 * Compute the per-level fanout, either using the exact fanout specified
3980 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
3982 static void __init
rcu_init_levelspread(int *levelspread
, const int *levelcnt
)
3986 if (rcu_fanout_exact
) {
3987 levelspread
[rcu_num_lvls
- 1] = rcu_fanout_leaf
;
3988 for (i
= rcu_num_lvls
- 2; i
>= 0; i
--)
3989 levelspread
[i
] = RCU_FANOUT
;
3995 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
3997 levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
4004 * Helper function for rcu_init() that initializes one rcu_state structure.
4006 static void __init
rcu_init_one(struct rcu_state
*rsp
)
4008 static const char * const buf
[] = RCU_NODE_NAME_INIT
;
4009 static const char * const fqs
[] = RCU_FQS_NAME_INIT
;
4010 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
4011 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
4012 static u8 fl_mask
= 0x1;
4014 int levelcnt
[RCU_NUM_LVLS
]; /* # nodes in each level. */
4015 int levelspread
[RCU_NUM_LVLS
]; /* kids/node in each level. */
4019 struct rcu_node
*rnp
;
4021 BUILD_BUG_ON(RCU_NUM_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4023 /* Silence gcc 4.8 false positive about array index out of range. */
4024 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4025 panic("rcu_init_one: rcu_num_lvls out of range");
4027 /* Initialize the level-tracking arrays. */
4029 for (i
= 0; i
< rcu_num_lvls
; i
++)
4030 levelcnt
[i
] = num_rcu_lvl
[i
];
4031 for (i
= 1; i
< rcu_num_lvls
; i
++)
4032 rsp
->level
[i
] = rsp
->level
[i
- 1] + levelcnt
[i
- 1];
4033 rcu_init_levelspread(levelspread
, levelcnt
);
4034 rsp
->flavor_mask
= fl_mask
;
4037 /* Initialize the elements themselves, starting from the leaves. */
4039 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4040 cpustride
*= levelspread
[i
];
4041 rnp
= rsp
->level
[i
];
4042 for (j
= 0; j
< levelcnt
[i
]; j
++, rnp
++) {
4043 raw_spin_lock_init(&ACCESS_PRIVATE(rnp
, lock
));
4044 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp
, lock
),
4045 &rcu_node_class
[i
], buf
[i
]);
4046 raw_spin_lock_init(&rnp
->fqslock
);
4047 lockdep_set_class_and_name(&rnp
->fqslock
,
4048 &rcu_fqs_class
[i
], fqs
[i
]);
4049 rnp
->gpnum
= rsp
->gpnum
;
4050 rnp
->completed
= rsp
->completed
;
4052 rnp
->qsmaskinit
= 0;
4053 rnp
->grplo
= j
* cpustride
;
4054 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4055 if (rnp
->grphi
>= nr_cpu_ids
)
4056 rnp
->grphi
= nr_cpu_ids
- 1;
4062 rnp
->grpnum
= j
% levelspread
[i
- 1];
4063 rnp
->grpmask
= 1UL << rnp
->grpnum
;
4064 rnp
->parent
= rsp
->level
[i
- 1] +
4065 j
/ levelspread
[i
- 1];
4068 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4069 rcu_init_one_nocb(rnp
);
4070 init_waitqueue_head(&rnp
->exp_wq
[0]);
4071 init_waitqueue_head(&rnp
->exp_wq
[1]);
4072 init_waitqueue_head(&rnp
->exp_wq
[2]);
4073 init_waitqueue_head(&rnp
->exp_wq
[3]);
4074 spin_lock_init(&rnp
->exp_lock
);
4078 init_swait_queue_head(&rsp
->gp_wq
);
4079 init_swait_queue_head(&rsp
->expedited_wq
);
4080 rnp
= rsp
->level
[rcu_num_lvls
- 1];
4081 for_each_possible_cpu(i
) {
4082 while (i
> rnp
->grphi
)
4084 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
4085 rcu_boot_init_percpu_data(i
, rsp
);
4087 list_add(&rsp
->flavors
, &rcu_struct_flavors
);
4091 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4092 * replace the definitions in tree.h because those are needed to size
4093 * the ->node array in the rcu_state structure.
4095 static void __init
rcu_init_geometry(void)
4099 int rcu_capacity
[RCU_NUM_LVLS
];
4102 * Initialize any unspecified boot parameters.
4103 * The default values of jiffies_till_first_fqs and
4104 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4105 * value, which is a function of HZ, then adding one for each
4106 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4108 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4109 if (jiffies_till_first_fqs
== ULONG_MAX
)
4110 jiffies_till_first_fqs
= d
;
4111 if (jiffies_till_next_fqs
== ULONG_MAX
)
4112 jiffies_till_next_fqs
= d
;
4114 /* If the compile-time values are accurate, just leave. */
4115 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4116 nr_cpu_ids
== NR_CPUS
)
4118 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4119 rcu_fanout_leaf
, nr_cpu_ids
);
4122 * The boot-time rcu_fanout_leaf parameter must be at least two
4123 * and cannot exceed the number of bits in the rcu_node masks.
4124 * Complain and fall back to the compile-time values if this
4125 * limit is exceeded.
4127 if (rcu_fanout_leaf
< 2 ||
4128 rcu_fanout_leaf
> sizeof(unsigned long) * 8) {
4129 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4135 * Compute number of nodes that can be handled an rcu_node tree
4136 * with the given number of levels.
4138 rcu_capacity
[0] = rcu_fanout_leaf
;
4139 for (i
= 1; i
< RCU_NUM_LVLS
; i
++)
4140 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4143 * The tree must be able to accommodate the configured number of CPUs.
4144 * If this limit is exceeded, fall back to the compile-time values.
4146 if (nr_cpu_ids
> rcu_capacity
[RCU_NUM_LVLS
- 1]) {
4147 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4152 /* Calculate the number of levels in the tree. */
4153 for (i
= 0; nr_cpu_ids
> rcu_capacity
[i
]; i
++) {
4155 rcu_num_lvls
= i
+ 1;
4157 /* Calculate the number of rcu_nodes at each level of the tree. */
4158 for (i
= 0; i
< rcu_num_lvls
; i
++) {
4159 int cap
= rcu_capacity
[(rcu_num_lvls
- 1) - i
];
4160 num_rcu_lvl
[i
] = DIV_ROUND_UP(nr_cpu_ids
, cap
);
4163 /* Calculate the total number of rcu_node structures. */
4165 for (i
= 0; i
< rcu_num_lvls
; i
++)
4166 rcu_num_nodes
+= num_rcu_lvl
[i
];
4170 * Dump out the structure of the rcu_node combining tree associated
4171 * with the rcu_state structure referenced by rsp.
4173 static void __init
rcu_dump_rcu_node_tree(struct rcu_state
*rsp
)
4176 struct rcu_node
*rnp
;
4178 pr_info("rcu_node tree layout dump\n");
4180 rcu_for_each_node_breadth_first(rsp
, rnp
) {
4181 if (rnp
->level
!= level
) {
4186 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4191 void __init
rcu_init(void)
4195 rcu_early_boot_tests();
4197 rcu_bootup_announce();
4198 rcu_init_geometry();
4199 rcu_init_one(&rcu_bh_state
);
4200 rcu_init_one(&rcu_sched_state
);
4202 rcu_dump_rcu_node_tree(&rcu_sched_state
);
4203 __rcu_init_preempt();
4204 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
4207 * We don't need protection against CPU-hotplug here because
4208 * this is called early in boot, before either interrupts
4209 * or the scheduler are operational.
4211 cpu_notifier(rcu_cpu_notify
, 0);
4212 pm_notifier(rcu_pm_notify
, 0);
4213 for_each_online_cpu(cpu
)
4214 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
4217 #include "tree_exp.h"
4218 #include "tree_plugin.h"