2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/trace_events.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
73 static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS];
76 * In order to export the rcu_state name to the tracing tools, it
77 * needs to be added in the __tracepoint_string section.
78 * This requires defining a separate variable tp_<sname>_varname
79 * that points to the string being used, and this will allow
80 * the tracing userspace tools to be able to decipher the string
81 * address to the matching string.
84 # define DEFINE_RCU_TPS(sname) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
87 # define RCU_STATE_NAME(sname) sname##_varname
89 # define DEFINE_RCU_TPS(sname)
90 # define RCU_STATE_NAME(sname) __stringify(sname)
93 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
94 DEFINE_RCU_TPS(sname) \
95 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
96 struct rcu_state sname##_state = { \
97 .level = { &sname##_state.node[0] }, \
98 .rda = &sname##_data, \
100 .gp_state = RCU_GP_IDLE, \
101 .gpnum = 0UL - 300UL, \
102 .completed = 0UL - 300UL, \
103 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
104 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
105 .orphan_donetail = &sname##_state.orphan_donelist, \
106 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
107 .name = RCU_STATE_NAME(sname), \
111 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
112 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
114 static struct rcu_state *const rcu_state_p;
115 static struct rcu_data __percpu *const rcu_data_p;
116 LIST_HEAD(rcu_struct_flavors);
118 /* Dump rcu_node combining tree at boot to verify correct setup. */
119 static bool dump_tree;
120 module_param(dump_tree, bool, 0444);
121 /* Control rcu_node-tree auto-balancing at boot time. */
122 static bool rcu_fanout_exact;
123 module_param(rcu_fanout_exact, bool, 0444);
124 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
125 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
126 module_param(rcu_fanout_leaf, int, 0444);
127 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
128 /* Number of rcu_nodes at specified level. */
129 static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
130 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
133 * The rcu_scheduler_active variable transitions from zero to one just
134 * before the first task is spawned. So when this variable is zero, RCU
135 * can assume that there is but one task, allowing RCU to (for example)
136 * optimize synchronize_sched() to a simple barrier(). When this variable
137 * is one, RCU must actually do all the hard work required to detect real
138 * grace periods. This variable is also used to suppress boot-time false
139 * positives from lockdep-RCU error checking.
141 int rcu_scheduler_active __read_mostly;
142 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
156 static int rcu_scheduler_fully_active __read_mostly;
158 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
159 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163 static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
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)
251 if (__this_cpu_read(rcu_sched_data.cpu_no_qs.s)) {
252 trace_rcu_grace_period(TPS("rcu_sched"),
253 __this_cpu_read(rcu_sched_data.gpnum),
255 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
256 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
258 local_irq_save(flags);
259 if (__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) {
260 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
261 rcu_report_exp_rdp(&rcu_sched_state,
262 this_cpu_ptr(&rcu_sched_data),
265 local_irq_restore(flags);
271 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
272 trace_rcu_grace_period(TPS("rcu_bh"),
273 __this_cpu_read(rcu_bh_data.gpnum),
275 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
279 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
281 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
282 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
283 .dynticks = ATOMIC_INIT(1),
284 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
285 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
286 .dynticks_idle = ATOMIC_INIT(1),
287 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
290 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
291 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
294 * Let the RCU core know that this CPU has gone through the scheduler,
295 * which is a quiescent state. This is called when the need for a
296 * quiescent state is urgent, so we burn an atomic operation and full
297 * memory barriers to let the RCU core know about it, regardless of what
298 * this CPU might (or might not) do in the near future.
300 * We inform the RCU core by emulating a zero-duration dyntick-idle
301 * period, which we in turn do by incrementing the ->dynticks counter
304 static void rcu_momentary_dyntick_idle(void)
307 struct rcu_data *rdp;
308 struct rcu_dynticks *rdtp;
310 struct rcu_state *rsp;
312 local_irq_save(flags);
315 * Yes, we can lose flag-setting operations. This is OK, because
316 * the flag will be set again after some delay.
318 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
319 raw_cpu_write(rcu_sched_qs_mask, 0);
321 /* Find the flavor that needs a quiescent state. */
322 for_each_rcu_flavor(rsp) {
323 rdp = raw_cpu_ptr(rsp->rda);
324 if (!(resched_mask & rsp->flavor_mask))
326 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
327 if (READ_ONCE(rdp->mynode->completed) !=
328 READ_ONCE(rdp->cond_resched_completed))
332 * Pretend to be momentarily idle for the quiescent state.
333 * This allows the grace-period kthread to record the
334 * quiescent state, with no need for this CPU to do anything
337 rdtp = this_cpu_ptr(&rcu_dynticks);
338 smp_mb__before_atomic(); /* Earlier stuff before QS. */
339 atomic_add(2, &rdtp->dynticks); /* QS. */
340 smp_mb__after_atomic(); /* Later stuff after QS. */
343 local_irq_restore(flags);
347 * Note a context switch. This is a quiescent state for RCU-sched,
348 * and requires special handling for preemptible RCU.
349 * The caller must have disabled preemption.
351 void rcu_note_context_switch(void)
353 barrier(); /* Avoid RCU read-side critical sections leaking down. */
354 trace_rcu_utilization(TPS("Start context switch"));
356 rcu_preempt_note_context_switch();
357 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
358 rcu_momentary_dyntick_idle();
359 trace_rcu_utilization(TPS("End context switch"));
360 barrier(); /* Avoid RCU read-side critical sections leaking up. */
362 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
365 * Register a quiescent state for all RCU flavors. If there is an
366 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
367 * dyntick-idle quiescent state visible to other CPUs (but only for those
368 * RCU flavors in desperate need of a quiescent state, which will normally
369 * be none of them). Either way, do a lightweight quiescent state for
372 * The barrier() calls are redundant in the common case when this is
373 * called externally, but just in case this is called from within this
377 void rcu_all_qs(void)
379 barrier(); /* Avoid RCU read-side critical sections leaking down. */
380 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
381 rcu_momentary_dyntick_idle();
382 this_cpu_inc(rcu_qs_ctr);
383 barrier(); /* Avoid RCU read-side critical sections leaking up. */
385 EXPORT_SYMBOL_GPL(rcu_all_qs);
387 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
388 static long qhimark = 10000; /* If this many pending, ignore blimit. */
389 static long qlowmark = 100; /* Once only this many pending, use blimit. */
391 module_param(blimit, long, 0444);
392 module_param(qhimark, long, 0444);
393 module_param(qlowmark, long, 0444);
395 static ulong jiffies_till_first_fqs = ULONG_MAX;
396 static ulong jiffies_till_next_fqs = ULONG_MAX;
398 module_param(jiffies_till_first_fqs, ulong, 0644);
399 module_param(jiffies_till_next_fqs, ulong, 0644);
402 * How long the grace period must be before we start recruiting
403 * quiescent-state help from rcu_note_context_switch().
405 static ulong jiffies_till_sched_qs = HZ / 20;
406 module_param(jiffies_till_sched_qs, ulong, 0644);
408 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
409 struct rcu_data *rdp);
410 static void force_qs_rnp(struct rcu_state *rsp,
411 int (*f)(struct rcu_data *rsp, bool *isidle,
412 unsigned long *maxj),
413 bool *isidle, unsigned long *maxj);
414 static void force_quiescent_state(struct rcu_state *rsp);
415 static int rcu_pending(void);
418 * Return the number of RCU batches started thus far for debug & stats.
420 unsigned long rcu_batches_started(void)
422 return rcu_state_p->gpnum;
424 EXPORT_SYMBOL_GPL(rcu_batches_started);
427 * Return the number of RCU-sched batches started thus far for debug & stats.
429 unsigned long rcu_batches_started_sched(void)
431 return rcu_sched_state.gpnum;
433 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
436 * Return the number of RCU BH batches started thus far for debug & stats.
438 unsigned long rcu_batches_started_bh(void)
440 return rcu_bh_state.gpnum;
442 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
445 * Return the number of RCU batches completed thus far for debug & stats.
447 unsigned long rcu_batches_completed(void)
449 return rcu_state_p->completed;
451 EXPORT_SYMBOL_GPL(rcu_batches_completed);
454 * Return the number of RCU-sched batches completed thus far for debug & stats.
456 unsigned long rcu_batches_completed_sched(void)
458 return rcu_sched_state.completed;
460 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
463 * Return the number of RCU BH batches completed thus far for debug & stats.
465 unsigned long rcu_batches_completed_bh(void)
467 return rcu_bh_state.completed;
469 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
472 * Force a quiescent state.
474 void rcu_force_quiescent_state(void)
476 force_quiescent_state(rcu_state_p);
478 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
481 * Force a quiescent state for RCU BH.
483 void rcu_bh_force_quiescent_state(void)
485 force_quiescent_state(&rcu_bh_state);
487 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
490 * Force a quiescent state for RCU-sched.
492 void rcu_sched_force_quiescent_state(void)
494 force_quiescent_state(&rcu_sched_state);
496 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
499 * Show the state of the grace-period kthreads.
501 void show_rcu_gp_kthreads(void)
503 struct rcu_state *rsp;
505 for_each_rcu_flavor(rsp) {
506 pr_info("%s: wait state: %d ->state: %#lx\n",
507 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
508 /* sched_show_task(rsp->gp_kthread); */
511 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
514 * Record the number of times rcutorture tests have been initiated and
515 * terminated. This information allows the debugfs tracing stats to be
516 * correlated to the rcutorture messages, even when the rcutorture module
517 * is being repeatedly loaded and unloaded. In other words, we cannot
518 * store this state in rcutorture itself.
520 void rcutorture_record_test_transition(void)
522 rcutorture_testseq++;
523 rcutorture_vernum = 0;
525 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
528 * Send along grace-period-related data for rcutorture diagnostics.
530 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
531 unsigned long *gpnum, unsigned long *completed)
533 struct rcu_state *rsp = NULL;
542 case RCU_SCHED_FLAVOR:
543 rsp = &rcu_sched_state;
549 *flags = READ_ONCE(rsp->gp_flags);
550 *gpnum = READ_ONCE(rsp->gpnum);
551 *completed = READ_ONCE(rsp->completed);
558 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
561 * Record the number of writer passes through the current rcutorture test.
562 * This is also used to correlate debugfs tracing stats with the rcutorture
565 void rcutorture_record_progress(unsigned long vernum)
569 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
572 * Does the CPU have callbacks ready to be invoked?
575 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
577 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
578 rdp->nxttail[RCU_DONE_TAIL] != NULL;
582 * Return the root node of the specified rcu_state structure.
584 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
586 return &rsp->node[0];
590 * Is there any need for future grace periods?
591 * Interrupts must be disabled. If the caller does not hold the root
592 * rnp_node structure's ->lock, the results are advisory only.
594 static int rcu_future_needs_gp(struct rcu_state *rsp)
596 struct rcu_node *rnp = rcu_get_root(rsp);
597 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
598 int *fp = &rnp->need_future_gp[idx];
600 return READ_ONCE(*fp);
604 * Does the current CPU require a not-yet-started grace period?
605 * The caller must have disabled interrupts to prevent races with
606 * normal callback registry.
609 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
613 if (rcu_gp_in_progress(rsp))
614 return 0; /* No, a grace period is already in progress. */
615 if (rcu_future_needs_gp(rsp))
616 return 1; /* Yes, a no-CBs CPU needs one. */
617 if (!rdp->nxttail[RCU_NEXT_TAIL])
618 return 0; /* No, this is a no-CBs (or offline) CPU. */
619 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
620 return 1; /* Yes, this CPU has newly registered callbacks. */
621 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
622 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
623 ULONG_CMP_LT(READ_ONCE(rsp->completed),
624 rdp->nxtcompleted[i]))
625 return 1; /* Yes, CBs for future grace period. */
626 return 0; /* No grace period needed. */
630 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
632 * If the new value of the ->dynticks_nesting counter now is zero,
633 * we really have entered idle, and must do the appropriate accounting.
634 * The caller must have disabled interrupts.
636 static void rcu_eqs_enter_common(long long oldval, bool user)
638 struct rcu_state *rsp;
639 struct rcu_data *rdp;
640 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
642 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
643 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
644 !user && !is_idle_task(current)) {
645 struct task_struct *idle __maybe_unused =
646 idle_task(smp_processor_id());
648 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
649 ftrace_dump(DUMP_ORIG);
650 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
651 current->pid, current->comm,
652 idle->pid, idle->comm); /* must be idle task! */
654 for_each_rcu_flavor(rsp) {
655 rdp = this_cpu_ptr(rsp->rda);
656 do_nocb_deferred_wakeup(rdp);
658 rcu_prepare_for_idle();
659 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
660 smp_mb__before_atomic(); /* See above. */
661 atomic_inc(&rdtp->dynticks);
662 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
663 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
664 atomic_read(&rdtp->dynticks) & 0x1);
665 rcu_dynticks_task_enter();
668 * It is illegal to enter an extended quiescent state while
669 * in an RCU read-side critical section.
671 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
672 "Illegal idle entry in RCU read-side critical section.");
673 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
674 "Illegal idle entry in RCU-bh read-side critical section.");
675 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
676 "Illegal idle entry in RCU-sched read-side critical section.");
680 * Enter an RCU extended quiescent state, which can be either the
681 * idle loop or adaptive-tickless usermode execution.
683 static void rcu_eqs_enter(bool user)
686 struct rcu_dynticks *rdtp;
688 rdtp = this_cpu_ptr(&rcu_dynticks);
689 oldval = rdtp->dynticks_nesting;
690 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
691 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
692 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
693 rdtp->dynticks_nesting = 0;
694 rcu_eqs_enter_common(oldval, user);
696 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
701 * rcu_idle_enter - inform RCU that current CPU is entering idle
703 * Enter idle mode, in other words, -leave- the mode in which RCU
704 * read-side critical sections can occur. (Though RCU read-side
705 * critical sections can occur in irq handlers in idle, a possibility
706 * handled by irq_enter() and irq_exit().)
708 * We crowbar the ->dynticks_nesting field to zero to allow for
709 * the possibility of usermode upcalls having messed up our count
710 * of interrupt nesting level during the prior busy period.
712 void rcu_idle_enter(void)
716 local_irq_save(flags);
717 rcu_eqs_enter(false);
718 rcu_sysidle_enter(0);
719 local_irq_restore(flags);
721 EXPORT_SYMBOL_GPL(rcu_idle_enter);
723 #ifdef CONFIG_NO_HZ_FULL
725 * rcu_user_enter - inform RCU that we are resuming userspace.
727 * Enter RCU idle mode right before resuming userspace. No use of RCU
728 * is permitted between this call and rcu_user_exit(). This way the
729 * CPU doesn't need to maintain the tick for RCU maintenance purposes
730 * when the CPU runs in userspace.
732 void rcu_user_enter(void)
736 #endif /* CONFIG_NO_HZ_FULL */
739 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
741 * Exit from an interrupt handler, which might possibly result in entering
742 * idle mode, in other words, leaving the mode in which read-side critical
743 * sections can occur.
745 * This code assumes that the idle loop never does anything that might
746 * result in unbalanced calls to irq_enter() and irq_exit(). If your
747 * architecture violates this assumption, RCU will give you what you
748 * deserve, good and hard. But very infrequently and irreproducibly.
750 * Use things like work queues to work around this limitation.
752 * You have been warned.
754 void rcu_irq_exit(void)
758 struct rcu_dynticks *rdtp;
760 local_irq_save(flags);
761 rdtp = this_cpu_ptr(&rcu_dynticks);
763 /* Page faults can happen in NMI handlers, so check... */
764 if (READ_ONCE(rdtp->dynticks_nmi_nesting))
767 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
768 oldval = rdtp->dynticks_nesting;
769 rdtp->dynticks_nesting--;
770 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
771 rdtp->dynticks_nesting < 0);
772 if (rdtp->dynticks_nesting)
773 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
775 rcu_eqs_enter_common(oldval, true);
776 rcu_sysidle_enter(1);
777 local_irq_restore(flags);
781 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
783 * If the new value of the ->dynticks_nesting counter was previously zero,
784 * we really have exited idle, and must do the appropriate accounting.
785 * The caller must have disabled interrupts.
787 static void rcu_eqs_exit_common(long long oldval, int user)
789 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
791 rcu_dynticks_task_exit();
792 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
793 atomic_inc(&rdtp->dynticks);
794 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
795 smp_mb__after_atomic(); /* See above. */
796 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
797 !(atomic_read(&rdtp->dynticks) & 0x1));
798 rcu_cleanup_after_idle();
799 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
800 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
801 !user && !is_idle_task(current)) {
802 struct task_struct *idle __maybe_unused =
803 idle_task(smp_processor_id());
805 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
806 oldval, rdtp->dynticks_nesting);
807 ftrace_dump(DUMP_ORIG);
808 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
809 current->pid, current->comm,
810 idle->pid, idle->comm); /* must be idle task! */
815 * Exit an RCU extended quiescent state, which can be either the
816 * idle loop or adaptive-tickless usermode execution.
818 static void rcu_eqs_exit(bool user)
820 struct rcu_dynticks *rdtp;
823 rdtp = this_cpu_ptr(&rcu_dynticks);
824 oldval = rdtp->dynticks_nesting;
825 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
826 if (oldval & DYNTICK_TASK_NEST_MASK) {
827 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
829 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
830 rcu_eqs_exit_common(oldval, user);
835 * rcu_idle_exit - inform RCU that current CPU is leaving idle
837 * Exit idle mode, in other words, -enter- the mode in which RCU
838 * read-side critical sections can occur.
840 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
841 * allow for the possibility of usermode upcalls messing up our count
842 * of interrupt nesting level during the busy period that is just
845 void rcu_idle_exit(void)
849 local_irq_save(flags);
852 local_irq_restore(flags);
854 EXPORT_SYMBOL_GPL(rcu_idle_exit);
856 #ifdef CONFIG_NO_HZ_FULL
858 * rcu_user_exit - inform RCU that we are exiting userspace.
860 * Exit RCU idle mode while entering the kernel because it can
861 * run a RCU read side critical section anytime.
863 void rcu_user_exit(void)
867 #endif /* CONFIG_NO_HZ_FULL */
870 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
872 * Enter an interrupt handler, which might possibly result in exiting
873 * idle mode, in other words, entering the mode in which read-side critical
874 * sections can occur.
876 * Note that the Linux kernel is fully capable of entering an interrupt
877 * handler that it never exits, for example when doing upcalls to
878 * user mode! This code assumes that the idle loop never does upcalls to
879 * user mode. If your architecture does do upcalls from the idle loop (or
880 * does anything else that results in unbalanced calls to the irq_enter()
881 * and irq_exit() functions), RCU will give you what you deserve, good
882 * and hard. But very infrequently and irreproducibly.
884 * Use things like work queues to work around this limitation.
886 * You have been warned.
888 void rcu_irq_enter(void)
891 struct rcu_dynticks *rdtp;
894 local_irq_save(flags);
895 rdtp = this_cpu_ptr(&rcu_dynticks);
897 /* Page faults can happen in NMI handlers, so check... */
898 if (READ_ONCE(rdtp->dynticks_nmi_nesting))
901 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
902 oldval = rdtp->dynticks_nesting;
903 rdtp->dynticks_nesting++;
904 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
905 rdtp->dynticks_nesting == 0);
907 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
909 rcu_eqs_exit_common(oldval, true);
911 local_irq_restore(flags);
915 * rcu_nmi_enter - inform RCU of entry to NMI context
917 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
918 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
919 * that the CPU is active. This implementation permits nested NMIs, as
920 * long as the nesting level does not overflow an int. (You will probably
921 * run out of stack space first.)
923 void rcu_nmi_enter(void)
925 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
928 /* Complain about underflow. */
929 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
932 * If idle from RCU viewpoint, atomically increment ->dynticks
933 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
934 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
935 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
936 * to be in the outermost NMI handler that interrupted an RCU-idle
937 * period (observation due to Andy Lutomirski).
939 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
940 smp_mb__before_atomic(); /* Force delay from prior write. */
941 atomic_inc(&rdtp->dynticks);
942 /* atomic_inc() before later RCU read-side crit sects */
943 smp_mb__after_atomic(); /* See above. */
944 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
947 rdtp->dynticks_nmi_nesting += incby;
952 * rcu_nmi_exit - inform RCU of exit from NMI context
954 * If we are returning from the outermost NMI handler that interrupted an
955 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
956 * to let the RCU grace-period handling know that the CPU is back to
959 void rcu_nmi_exit(void)
961 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
964 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
965 * (We are exiting an NMI handler, so RCU better be paying attention
968 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
969 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
972 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
973 * leave it in non-RCU-idle state.
975 if (rdtp->dynticks_nmi_nesting != 1) {
976 rdtp->dynticks_nmi_nesting -= 2;
980 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
981 rdtp->dynticks_nmi_nesting = 0;
982 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
983 smp_mb__before_atomic(); /* See above. */
984 atomic_inc(&rdtp->dynticks);
985 smp_mb__after_atomic(); /* Force delay to next write. */
986 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
990 * __rcu_is_watching - are RCU read-side critical sections safe?
992 * Return true if RCU is watching the running CPU, which means that
993 * this CPU can safely enter RCU read-side critical sections. Unlike
994 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
995 * least disabled preemption.
997 bool notrace __rcu_is_watching(void)
999 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
1003 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1005 * If the current CPU is in its idle loop and is neither in an interrupt
1006 * or NMI handler, return true.
1008 bool notrace rcu_is_watching(void)
1012 preempt_disable_notrace();
1013 ret = __rcu_is_watching();
1014 preempt_enable_notrace();
1017 EXPORT_SYMBOL_GPL(rcu_is_watching);
1019 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1022 * Is the current CPU online? Disable preemption to avoid false positives
1023 * that could otherwise happen due to the current CPU number being sampled,
1024 * this task being preempted, its old CPU being taken offline, resuming
1025 * on some other CPU, then determining that its old CPU is now offline.
1026 * It is OK to use RCU on an offline processor during initial boot, hence
1027 * the check for rcu_scheduler_fully_active. Note also that it is OK
1028 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1029 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1030 * offline to continue to use RCU for one jiffy after marking itself
1031 * offline in the cpu_online_mask. This leniency is necessary given the
1032 * non-atomic nature of the online and offline processing, for example,
1033 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1036 * This is also why RCU internally marks CPUs online during the
1037 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1039 * Disable checking if in an NMI handler because we cannot safely report
1040 * errors from NMI handlers anyway.
1042 bool rcu_lockdep_current_cpu_online(void)
1044 struct rcu_data *rdp;
1045 struct rcu_node *rnp;
1051 rdp = this_cpu_ptr(&rcu_sched_data);
1053 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1054 !rcu_scheduler_fully_active;
1058 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1060 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1063 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1065 * If the current CPU is idle or running at a first-level (not nested)
1066 * interrupt from idle, return true. The caller must have at least
1067 * disabled preemption.
1069 static int rcu_is_cpu_rrupt_from_idle(void)
1071 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1075 * Snapshot the specified CPU's dynticks counter so that we can later
1076 * credit them with an implicit quiescent state. Return 1 if this CPU
1077 * is in dynticks idle mode, which is an extended quiescent state.
1079 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1080 bool *isidle, unsigned long *maxj)
1082 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1083 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1084 if ((rdp->dynticks_snap & 0x1) == 0) {
1085 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1088 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1089 rdp->mynode->gpnum))
1090 WRITE_ONCE(rdp->gpwrap, true);
1096 * Return true if the specified CPU has passed through a quiescent
1097 * state by virtue of being in or having passed through an dynticks
1098 * idle state since the last call to dyntick_save_progress_counter()
1099 * for this same CPU, or by virtue of having been offline.
1101 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1102 bool *isidle, unsigned long *maxj)
1108 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1109 snap = (unsigned int)rdp->dynticks_snap;
1112 * If the CPU passed through or entered a dynticks idle phase with
1113 * no active irq/NMI handlers, then we can safely pretend that the CPU
1114 * already acknowledged the request to pass through a quiescent
1115 * state. Either way, that CPU cannot possibly be in an RCU
1116 * read-side critical section that started before the beginning
1117 * of the current RCU grace period.
1119 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1120 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1121 rdp->dynticks_fqs++;
1126 * Check for the CPU being offline, but only if the grace period
1127 * is old enough. We don't need to worry about the CPU changing
1128 * state: If we see it offline even once, it has been through a
1131 * The reason for insisting that the grace period be at least
1132 * one jiffy old is that CPUs that are not quite online and that
1133 * have just gone offline can still execute RCU read-side critical
1136 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1137 return 0; /* Grace period is not old enough. */
1139 if (cpu_is_offline(rdp->cpu)) {
1140 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1146 * A CPU running for an extended time within the kernel can
1147 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1148 * even context-switching back and forth between a pair of
1149 * in-kernel CPU-bound tasks cannot advance grace periods.
1150 * So if the grace period is old enough, make the CPU pay attention.
1151 * Note that the unsynchronized assignments to the per-CPU
1152 * rcu_sched_qs_mask variable are safe. Yes, setting of
1153 * bits can be lost, but they will be set again on the next
1154 * force-quiescent-state pass. So lost bit sets do not result
1155 * in incorrect behavior, merely in a grace period lasting
1156 * a few jiffies longer than it might otherwise. Because
1157 * there are at most four threads involved, and because the
1158 * updates are only once every few jiffies, the probability of
1159 * lossage (and thus of slight grace-period extension) is
1162 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1163 * is set too high, we override with half of the RCU CPU stall
1166 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1167 if (ULONG_CMP_GE(jiffies,
1168 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1169 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1170 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1171 WRITE_ONCE(rdp->cond_resched_completed,
1172 READ_ONCE(rdp->mynode->completed));
1173 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1175 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1176 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1177 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
1178 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1179 /* Time to beat on that CPU again! */
1180 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1181 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1188 static void record_gp_stall_check_time(struct rcu_state *rsp)
1190 unsigned long j = jiffies;
1194 smp_wmb(); /* Record start time before stall time. */
1195 j1 = rcu_jiffies_till_stall_check();
1196 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1197 rsp->jiffies_resched = j + j1 / 2;
1198 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1202 * Complain about starvation of grace-period kthread.
1204 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1210 gpa = READ_ONCE(rsp->gp_activity);
1211 if (j - gpa > 2 * HZ)
1212 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n",
1214 rsp->gpnum, rsp->completed,
1215 rsp->gp_flags, rsp->gp_state,
1216 rsp->gp_kthread ? rsp->gp_kthread->state : 0);
1220 * Dump stacks of all tasks running on stalled CPUs.
1222 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1225 unsigned long flags;
1226 struct rcu_node *rnp;
1228 rcu_for_each_leaf_node(rsp, rnp) {
1229 raw_spin_lock_irqsave(&rnp->lock, flags);
1230 if (rnp->qsmask != 0) {
1231 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1232 if (rnp->qsmask & (1UL << cpu))
1233 dump_cpu_task(rnp->grplo + cpu);
1235 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1239 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1243 unsigned long flags;
1247 struct rcu_node *rnp = rcu_get_root(rsp);
1250 /* Only let one CPU complain about others per time interval. */
1252 raw_spin_lock_irqsave(&rnp->lock, flags);
1253 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1254 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1255 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1258 WRITE_ONCE(rsp->jiffies_stall,
1259 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1260 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1263 * OK, time to rat on our buddy...
1264 * See Documentation/RCU/stallwarn.txt for info on how to debug
1265 * RCU CPU stall warnings.
1267 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1269 print_cpu_stall_info_begin();
1270 rcu_for_each_leaf_node(rsp, rnp) {
1271 raw_spin_lock_irqsave(&rnp->lock, flags);
1272 ndetected += rcu_print_task_stall(rnp);
1273 if (rnp->qsmask != 0) {
1274 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1275 if (rnp->qsmask & (1UL << cpu)) {
1276 print_cpu_stall_info(rsp,
1281 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1284 print_cpu_stall_info_end();
1285 for_each_possible_cpu(cpu)
1286 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1287 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1288 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1289 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1291 rcu_dump_cpu_stacks(rsp);
1293 if (READ_ONCE(rsp->gpnum) != gpnum ||
1294 READ_ONCE(rsp->completed) == gpnum) {
1295 pr_err("INFO: Stall ended before state dump start\n");
1298 gpa = READ_ONCE(rsp->gp_activity);
1299 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1300 rsp->name, j - gpa, j, gpa,
1301 jiffies_till_next_fqs,
1302 rcu_get_root(rsp)->qsmask);
1303 /* In this case, the current CPU might be at fault. */
1304 sched_show_task(current);
1308 /* Complain about tasks blocking the grace period. */
1309 rcu_print_detail_task_stall(rsp);
1311 rcu_check_gp_kthread_starvation(rsp);
1313 force_quiescent_state(rsp); /* Kick them all. */
1316 static void print_cpu_stall(struct rcu_state *rsp)
1319 unsigned long flags;
1320 struct rcu_node *rnp = rcu_get_root(rsp);
1324 * OK, time to rat on ourselves...
1325 * See Documentation/RCU/stallwarn.txt for info on how to debug
1326 * RCU CPU stall warnings.
1328 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1329 print_cpu_stall_info_begin();
1330 print_cpu_stall_info(rsp, smp_processor_id());
1331 print_cpu_stall_info_end();
1332 for_each_possible_cpu(cpu)
1333 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1334 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1335 jiffies - rsp->gp_start,
1336 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1338 rcu_check_gp_kthread_starvation(rsp);
1340 rcu_dump_cpu_stacks(rsp);
1342 raw_spin_lock_irqsave(&rnp->lock, flags);
1343 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1344 WRITE_ONCE(rsp->jiffies_stall,
1345 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1346 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1349 * Attempt to revive the RCU machinery by forcing a context switch.
1351 * A context switch would normally allow the RCU state machine to make
1352 * progress and it could be we're stuck in kernel space without context
1353 * switches for an entirely unreasonable amount of time.
1355 resched_cpu(smp_processor_id());
1358 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1360 unsigned long completed;
1361 unsigned long gpnum;
1365 struct rcu_node *rnp;
1367 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1372 * Lots of memory barriers to reject false positives.
1374 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1375 * then rsp->gp_start, and finally rsp->completed. These values
1376 * are updated in the opposite order with memory barriers (or
1377 * equivalent) during grace-period initialization and cleanup.
1378 * Now, a false positive can occur if we get an new value of
1379 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1380 * the memory barriers, the only way that this can happen is if one
1381 * grace period ends and another starts between these two fetches.
1382 * Detect this by comparing rsp->completed with the previous fetch
1385 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1386 * and rsp->gp_start suffice to forestall false positives.
1388 gpnum = READ_ONCE(rsp->gpnum);
1389 smp_rmb(); /* Pick up ->gpnum first... */
1390 js = READ_ONCE(rsp->jiffies_stall);
1391 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1392 gps = READ_ONCE(rsp->gp_start);
1393 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1394 completed = READ_ONCE(rsp->completed);
1395 if (ULONG_CMP_GE(completed, gpnum) ||
1396 ULONG_CMP_LT(j, js) ||
1397 ULONG_CMP_GE(gps, js))
1398 return; /* No stall or GP completed since entering function. */
1400 if (rcu_gp_in_progress(rsp) &&
1401 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1403 /* We haven't checked in, so go dump stack. */
1404 print_cpu_stall(rsp);
1406 } else if (rcu_gp_in_progress(rsp) &&
1407 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1409 /* They had a few time units to dump stack, so complain. */
1410 print_other_cpu_stall(rsp, gpnum);
1415 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1417 * Set the stall-warning timeout way off into the future, thus preventing
1418 * any RCU CPU stall-warning messages from appearing in the current set of
1419 * RCU grace periods.
1421 * The caller must disable hard irqs.
1423 void rcu_cpu_stall_reset(void)
1425 struct rcu_state *rsp;
1427 for_each_rcu_flavor(rsp)
1428 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1432 * Initialize the specified rcu_data structure's default callback list
1433 * to empty. The default callback list is the one that is not used by
1434 * no-callbacks CPUs.
1436 static void init_default_callback_list(struct rcu_data *rdp)
1440 rdp->nxtlist = NULL;
1441 for (i = 0; i < RCU_NEXT_SIZE; i++)
1442 rdp->nxttail[i] = &rdp->nxtlist;
1446 * Initialize the specified rcu_data structure's callback list to empty.
1448 static void init_callback_list(struct rcu_data *rdp)
1450 if (init_nocb_callback_list(rdp))
1452 init_default_callback_list(rdp);
1456 * Determine the value that ->completed will have at the end of the
1457 * next subsequent grace period. This is used to tag callbacks so that
1458 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1459 * been dyntick-idle for an extended period with callbacks under the
1460 * influence of RCU_FAST_NO_HZ.
1462 * The caller must hold rnp->lock with interrupts disabled.
1464 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1465 struct rcu_node *rnp)
1468 * If RCU is idle, we just wait for the next grace period.
1469 * But we can only be sure that RCU is idle if we are looking
1470 * at the root rcu_node structure -- otherwise, a new grace
1471 * period might have started, but just not yet gotten around
1472 * to initializing the current non-root rcu_node structure.
1474 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1475 return rnp->completed + 1;
1478 * Otherwise, wait for a possible partial grace period and
1479 * then the subsequent full grace period.
1481 return rnp->completed + 2;
1485 * Trace-event helper function for rcu_start_future_gp() and
1486 * rcu_nocb_wait_gp().
1488 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1489 unsigned long c, const char *s)
1491 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1492 rnp->completed, c, rnp->level,
1493 rnp->grplo, rnp->grphi, s);
1497 * Start some future grace period, as needed to handle newly arrived
1498 * callbacks. The required future grace periods are recorded in each
1499 * rcu_node structure's ->need_future_gp field. Returns true if there
1500 * is reason to awaken the grace-period kthread.
1502 * The caller must hold the specified rcu_node structure's ->lock.
1504 static bool __maybe_unused
1505 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1506 unsigned long *c_out)
1511 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1514 * Pick up grace-period number for new callbacks. If this
1515 * grace period is already marked as needed, return to the caller.
1517 c = rcu_cbs_completed(rdp->rsp, rnp);
1518 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1519 if (rnp->need_future_gp[c & 0x1]) {
1520 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1525 * If either this rcu_node structure or the root rcu_node structure
1526 * believe that a grace period is in progress, then we must wait
1527 * for the one following, which is in "c". Because our request
1528 * will be noticed at the end of the current grace period, we don't
1529 * need to explicitly start one. We only do the lockless check
1530 * of rnp_root's fields if the current rcu_node structure thinks
1531 * there is no grace period in flight, and because we hold rnp->lock,
1532 * the only possible change is when rnp_root's two fields are
1533 * equal, in which case rnp_root->gpnum might be concurrently
1534 * incremented. But that is OK, as it will just result in our
1535 * doing some extra useless work.
1537 if (rnp->gpnum != rnp->completed ||
1538 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1539 rnp->need_future_gp[c & 0x1]++;
1540 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1545 * There might be no grace period in progress. If we don't already
1546 * hold it, acquire the root rcu_node structure's lock in order to
1547 * start one (if needed).
1549 if (rnp != rnp_root) {
1550 raw_spin_lock(&rnp_root->lock);
1551 smp_mb__after_unlock_lock();
1555 * Get a new grace-period number. If there really is no grace
1556 * period in progress, it will be smaller than the one we obtained
1557 * earlier. Adjust callbacks as needed. Note that even no-CBs
1558 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1560 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1561 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1562 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1563 rdp->nxtcompleted[i] = c;
1566 * If the needed for the required grace period is already
1567 * recorded, trace and leave.
1569 if (rnp_root->need_future_gp[c & 0x1]) {
1570 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1574 /* Record the need for the future grace period. */
1575 rnp_root->need_future_gp[c & 0x1]++;
1577 /* If a grace period is not already in progress, start one. */
1578 if (rnp_root->gpnum != rnp_root->completed) {
1579 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1581 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1582 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1585 if (rnp != rnp_root)
1586 raw_spin_unlock(&rnp_root->lock);
1594 * Clean up any old requests for the just-ended grace period. Also return
1595 * whether any additional grace periods have been requested. Also invoke
1596 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1597 * waiting for this grace period to complete.
1599 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1601 int c = rnp->completed;
1603 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1605 rcu_nocb_gp_cleanup(rsp, rnp);
1606 rnp->need_future_gp[c & 0x1] = 0;
1607 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1608 trace_rcu_future_gp(rnp, rdp, c,
1609 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1614 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1615 * an interrupt or softirq handler), and don't bother awakening when there
1616 * is nothing for the grace-period kthread to do (as in several CPUs raced
1617 * to awaken, and we lost), and finally don't try to awaken a kthread that
1618 * has not yet been created. If all those checks are passed, track some
1619 * debug information and awaken.
1621 * So why do the self-wakeup when in an interrupt or softirq handler
1622 * in the grace-period kthread's context? Because the kthread might have
1623 * been interrupted just as it was going to sleep, and just after the final
1624 * pre-sleep check of the awaken condition. In this case, a wakeup really
1625 * is required, and is therefore supplied.
1627 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1629 if ((current == rsp->gp_kthread &&
1630 !in_interrupt() && !in_serving_softirq()) ||
1631 !READ_ONCE(rsp->gp_flags) ||
1634 wake_up(&rsp->gp_wq);
1638 * If there is room, assign a ->completed number to any callbacks on
1639 * this CPU that have not already been assigned. Also accelerate any
1640 * callbacks that were previously assigned a ->completed number that has
1641 * since proven to be too conservative, which can happen if callbacks get
1642 * assigned a ->completed number while RCU is idle, but with reference to
1643 * a non-root rcu_node structure. This function is idempotent, so it does
1644 * not hurt to call it repeatedly. Returns an flag saying that we should
1645 * awaken the RCU grace-period kthread.
1647 * The caller must hold rnp->lock with interrupts disabled.
1649 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1650 struct rcu_data *rdp)
1656 /* If the CPU has no callbacks, nothing to do. */
1657 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1661 * Starting from the sublist containing the callbacks most
1662 * recently assigned a ->completed number and working down, find the
1663 * first sublist that is not assignable to an upcoming grace period.
1664 * Such a sublist has something in it (first two tests) and has
1665 * a ->completed number assigned that will complete sooner than
1666 * the ->completed number for newly arrived callbacks (last test).
1668 * The key point is that any later sublist can be assigned the
1669 * same ->completed number as the newly arrived callbacks, which
1670 * means that the callbacks in any of these later sublist can be
1671 * grouped into a single sublist, whether or not they have already
1672 * been assigned a ->completed number.
1674 c = rcu_cbs_completed(rsp, rnp);
1675 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1676 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1677 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1681 * If there are no sublist for unassigned callbacks, leave.
1682 * At the same time, advance "i" one sublist, so that "i" will
1683 * index into the sublist where all the remaining callbacks should
1686 if (++i >= RCU_NEXT_TAIL)
1690 * Assign all subsequent callbacks' ->completed number to the next
1691 * full grace period and group them all in the sublist initially
1694 for (; i <= RCU_NEXT_TAIL; i++) {
1695 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1696 rdp->nxtcompleted[i] = c;
1698 /* Record any needed additional grace periods. */
1699 ret = rcu_start_future_gp(rnp, rdp, NULL);
1701 /* Trace depending on how much we were able to accelerate. */
1702 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1703 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1705 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1710 * Move any callbacks whose grace period has completed to the
1711 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1712 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1713 * sublist. This function is idempotent, so it does not hurt to
1714 * invoke it repeatedly. As long as it is not invoked -too- often...
1715 * Returns true if the RCU grace-period kthread needs to be awakened.
1717 * The caller must hold rnp->lock with interrupts disabled.
1719 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1720 struct rcu_data *rdp)
1724 /* If the CPU has no callbacks, nothing to do. */
1725 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1729 * Find all callbacks whose ->completed numbers indicate that they
1730 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1732 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1733 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1735 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1737 /* Clean up any sublist tail pointers that were misordered above. */
1738 for (j = RCU_WAIT_TAIL; j < i; j++)
1739 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1741 /* Copy down callbacks to fill in empty sublists. */
1742 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1743 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1745 rdp->nxttail[j] = rdp->nxttail[i];
1746 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1749 /* Classify any remaining callbacks. */
1750 return rcu_accelerate_cbs(rsp, rnp, rdp);
1754 * Update CPU-local rcu_data state to record the beginnings and ends of
1755 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1756 * structure corresponding to the current CPU, and must have irqs disabled.
1757 * Returns true if the grace-period kthread needs to be awakened.
1759 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1760 struct rcu_data *rdp)
1764 /* Handle the ends of any preceding grace periods first. */
1765 if (rdp->completed == rnp->completed &&
1766 !unlikely(READ_ONCE(rdp->gpwrap))) {
1768 /* No grace period end, so just accelerate recent callbacks. */
1769 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1773 /* Advance callbacks. */
1774 ret = rcu_advance_cbs(rsp, rnp, rdp);
1776 /* Remember that we saw this grace-period completion. */
1777 rdp->completed = rnp->completed;
1778 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1781 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1783 * If the current grace period is waiting for this CPU,
1784 * set up to detect a quiescent state, otherwise don't
1785 * go looking for one.
1787 rdp->gpnum = rnp->gpnum;
1788 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1789 rdp->cpu_no_qs.b.norm = true;
1790 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1791 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1792 zero_cpu_stall_ticks(rdp);
1793 WRITE_ONCE(rdp->gpwrap, false);
1798 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1800 unsigned long flags;
1802 struct rcu_node *rnp;
1804 local_irq_save(flags);
1806 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1807 rdp->completed == READ_ONCE(rnp->completed) &&
1808 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1809 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1810 local_irq_restore(flags);
1813 smp_mb__after_unlock_lock();
1814 needwake = __note_gp_changes(rsp, rnp, rdp);
1815 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1817 rcu_gp_kthread_wake(rsp);
1820 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1823 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1824 schedule_timeout_uninterruptible(delay);
1828 * Initialize a new grace period. Return 0 if no grace period required.
1830 static int rcu_gp_init(struct rcu_state *rsp)
1832 unsigned long oldmask;
1833 struct rcu_data *rdp;
1834 struct rcu_node *rnp = rcu_get_root(rsp);
1836 WRITE_ONCE(rsp->gp_activity, jiffies);
1837 raw_spin_lock_irq(&rnp->lock);
1838 smp_mb__after_unlock_lock();
1839 if (!READ_ONCE(rsp->gp_flags)) {
1840 /* Spurious wakeup, tell caller to go back to sleep. */
1841 raw_spin_unlock_irq(&rnp->lock);
1844 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1846 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1848 * Grace period already in progress, don't start another.
1849 * Not supposed to be able to happen.
1851 raw_spin_unlock_irq(&rnp->lock);
1855 /* Advance to a new grace period and initialize state. */
1856 record_gp_stall_check_time(rsp);
1857 /* Record GP times before starting GP, hence smp_store_release(). */
1858 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1859 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1860 raw_spin_unlock_irq(&rnp->lock);
1863 * Apply per-leaf buffered online and offline operations to the
1864 * rcu_node tree. Note that this new grace period need not wait
1865 * for subsequent online CPUs, and that quiescent-state forcing
1866 * will handle subsequent offline CPUs.
1868 rcu_for_each_leaf_node(rsp, rnp) {
1869 rcu_gp_slow(rsp, gp_preinit_delay);
1870 raw_spin_lock_irq(&rnp->lock);
1871 smp_mb__after_unlock_lock();
1872 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1873 !rnp->wait_blkd_tasks) {
1874 /* Nothing to do on this leaf rcu_node structure. */
1875 raw_spin_unlock_irq(&rnp->lock);
1879 /* Record old state, apply changes to ->qsmaskinit field. */
1880 oldmask = rnp->qsmaskinit;
1881 rnp->qsmaskinit = rnp->qsmaskinitnext;
1883 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1884 if (!oldmask != !rnp->qsmaskinit) {
1885 if (!oldmask) /* First online CPU for this rcu_node. */
1886 rcu_init_new_rnp(rnp);
1887 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1888 rnp->wait_blkd_tasks = true;
1889 else /* Last offline CPU and can propagate. */
1890 rcu_cleanup_dead_rnp(rnp);
1894 * If all waited-on tasks from prior grace period are
1895 * done, and if all this rcu_node structure's CPUs are
1896 * still offline, propagate up the rcu_node tree and
1897 * clear ->wait_blkd_tasks. Otherwise, if one of this
1898 * rcu_node structure's CPUs has since come back online,
1899 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1900 * checks for this, so just call it unconditionally).
1902 if (rnp->wait_blkd_tasks &&
1903 (!rcu_preempt_has_tasks(rnp) ||
1905 rnp->wait_blkd_tasks = false;
1906 rcu_cleanup_dead_rnp(rnp);
1909 raw_spin_unlock_irq(&rnp->lock);
1913 * Set the quiescent-state-needed bits in all the rcu_node
1914 * structures for all currently online CPUs in breadth-first order,
1915 * starting from the root rcu_node structure, relying on the layout
1916 * of the tree within the rsp->node[] array. Note that other CPUs
1917 * will access only the leaves of the hierarchy, thus seeing that no
1918 * grace period is in progress, at least until the corresponding
1919 * leaf node has been initialized. In addition, we have excluded
1920 * CPU-hotplug operations.
1922 * The grace period cannot complete until the initialization
1923 * process finishes, because this kthread handles both.
1925 rcu_for_each_node_breadth_first(rsp, rnp) {
1926 rcu_gp_slow(rsp, gp_init_delay);
1927 raw_spin_lock_irq(&rnp->lock);
1928 smp_mb__after_unlock_lock();
1929 rdp = this_cpu_ptr(rsp->rda);
1930 rcu_preempt_check_blocked_tasks(rnp);
1931 rnp->qsmask = rnp->qsmaskinit;
1932 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1933 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1934 WRITE_ONCE(rnp->completed, rsp->completed);
1935 if (rnp == rdp->mynode)
1936 (void)__note_gp_changes(rsp, rnp, rdp);
1937 rcu_preempt_boost_start_gp(rnp);
1938 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1939 rnp->level, rnp->grplo,
1940 rnp->grphi, rnp->qsmask);
1941 raw_spin_unlock_irq(&rnp->lock);
1942 cond_resched_rcu_qs();
1943 WRITE_ONCE(rsp->gp_activity, jiffies);
1950 * Helper function for wait_event_interruptible_timeout() wakeup
1951 * at force-quiescent-state time.
1953 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1955 struct rcu_node *rnp = rcu_get_root(rsp);
1957 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1958 *gfp = READ_ONCE(rsp->gp_flags);
1959 if (*gfp & RCU_GP_FLAG_FQS)
1962 /* The current grace period has completed. */
1963 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1970 * Do one round of quiescent-state forcing.
1972 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
1974 bool isidle = false;
1976 struct rcu_node *rnp = rcu_get_root(rsp);
1978 WRITE_ONCE(rsp->gp_activity, jiffies);
1981 /* Collect dyntick-idle snapshots. */
1982 if (is_sysidle_rcu_state(rsp)) {
1984 maxj = jiffies - ULONG_MAX / 4;
1986 force_qs_rnp(rsp, dyntick_save_progress_counter,
1988 rcu_sysidle_report_gp(rsp, isidle, maxj);
1990 /* Handle dyntick-idle and offline CPUs. */
1992 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1994 /* Clear flag to prevent immediate re-entry. */
1995 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1996 raw_spin_lock_irq(&rnp->lock);
1997 smp_mb__after_unlock_lock();
1998 WRITE_ONCE(rsp->gp_flags,
1999 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2000 raw_spin_unlock_irq(&rnp->lock);
2005 * Clean up after the old grace period.
2007 static void rcu_gp_cleanup(struct rcu_state *rsp)
2009 unsigned long gp_duration;
2010 bool needgp = false;
2012 struct rcu_data *rdp;
2013 struct rcu_node *rnp = rcu_get_root(rsp);
2015 WRITE_ONCE(rsp->gp_activity, jiffies);
2016 raw_spin_lock_irq(&rnp->lock);
2017 smp_mb__after_unlock_lock();
2018 gp_duration = jiffies - rsp->gp_start;
2019 if (gp_duration > rsp->gp_max)
2020 rsp->gp_max = gp_duration;
2023 * We know the grace period is complete, but to everyone else
2024 * it appears to still be ongoing. But it is also the case
2025 * that to everyone else it looks like there is nothing that
2026 * they can do to advance the grace period. It is therefore
2027 * safe for us to drop the lock in order to mark the grace
2028 * period as completed in all of the rcu_node structures.
2030 raw_spin_unlock_irq(&rnp->lock);
2033 * Propagate new ->completed value to rcu_node structures so
2034 * that other CPUs don't have to wait until the start of the next
2035 * grace period to process their callbacks. This also avoids
2036 * some nasty RCU grace-period initialization races by forcing
2037 * the end of the current grace period to be completely recorded in
2038 * all of the rcu_node structures before the beginning of the next
2039 * grace period is recorded in any of the rcu_node structures.
2041 rcu_for_each_node_breadth_first(rsp, rnp) {
2042 raw_spin_lock_irq(&rnp->lock);
2043 smp_mb__after_unlock_lock();
2044 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2045 WARN_ON_ONCE(rnp->qsmask);
2046 WRITE_ONCE(rnp->completed, rsp->gpnum);
2047 rdp = this_cpu_ptr(rsp->rda);
2048 if (rnp == rdp->mynode)
2049 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2050 /* smp_mb() provided by prior unlock-lock pair. */
2051 nocb += rcu_future_gp_cleanup(rsp, rnp);
2052 raw_spin_unlock_irq(&rnp->lock);
2053 cond_resched_rcu_qs();
2054 WRITE_ONCE(rsp->gp_activity, jiffies);
2055 rcu_gp_slow(rsp, gp_cleanup_delay);
2057 rnp = rcu_get_root(rsp);
2058 raw_spin_lock_irq(&rnp->lock);
2059 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
2060 rcu_nocb_gp_set(rnp, nocb);
2062 /* Declare grace period done. */
2063 WRITE_ONCE(rsp->completed, rsp->gpnum);
2064 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2065 rsp->gp_state = RCU_GP_IDLE;
2066 rdp = this_cpu_ptr(rsp->rda);
2067 /* Advance CBs to reduce false positives below. */
2068 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2069 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2070 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2071 trace_rcu_grace_period(rsp->name,
2072 READ_ONCE(rsp->gpnum),
2075 raw_spin_unlock_irq(&rnp->lock);
2079 * Body of kthread that handles grace periods.
2081 static int __noreturn rcu_gp_kthread(void *arg)
2087 struct rcu_state *rsp = arg;
2088 struct rcu_node *rnp = rcu_get_root(rsp);
2090 rcu_bind_gp_kthread();
2093 /* Handle grace-period start. */
2095 trace_rcu_grace_period(rsp->name,
2096 READ_ONCE(rsp->gpnum),
2098 rsp->gp_state = RCU_GP_WAIT_GPS;
2099 wait_event_interruptible(rsp->gp_wq,
2100 READ_ONCE(rsp->gp_flags) &
2102 rsp->gp_state = RCU_GP_DONE_GPS;
2103 /* Locking provides needed memory barrier. */
2104 if (rcu_gp_init(rsp))
2106 cond_resched_rcu_qs();
2107 WRITE_ONCE(rsp->gp_activity, jiffies);
2108 WARN_ON(signal_pending(current));
2109 trace_rcu_grace_period(rsp->name,
2110 READ_ONCE(rsp->gpnum),
2114 /* Handle quiescent-state forcing. */
2115 first_gp_fqs = true;
2116 j = jiffies_till_first_fqs;
2119 jiffies_till_first_fqs = HZ;
2124 rsp->jiffies_force_qs = jiffies + j;
2125 trace_rcu_grace_period(rsp->name,
2126 READ_ONCE(rsp->gpnum),
2128 rsp->gp_state = RCU_GP_WAIT_FQS;
2129 ret = wait_event_interruptible_timeout(rsp->gp_wq,
2130 rcu_gp_fqs_check_wake(rsp, &gf), j);
2131 rsp->gp_state = RCU_GP_DOING_FQS;
2132 /* Locking provides needed memory barriers. */
2133 /* If grace period done, leave loop. */
2134 if (!READ_ONCE(rnp->qsmask) &&
2135 !rcu_preempt_blocked_readers_cgp(rnp))
2137 /* If time for quiescent-state forcing, do it. */
2138 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2139 (gf & RCU_GP_FLAG_FQS)) {
2140 trace_rcu_grace_period(rsp->name,
2141 READ_ONCE(rsp->gpnum),
2143 rcu_gp_fqs(rsp, first_gp_fqs);
2144 first_gp_fqs = false;
2145 trace_rcu_grace_period(rsp->name,
2146 READ_ONCE(rsp->gpnum),
2148 cond_resched_rcu_qs();
2149 WRITE_ONCE(rsp->gp_activity, jiffies);
2151 /* Deal with stray signal. */
2152 cond_resched_rcu_qs();
2153 WRITE_ONCE(rsp->gp_activity, jiffies);
2154 WARN_ON(signal_pending(current));
2155 trace_rcu_grace_period(rsp->name,
2156 READ_ONCE(rsp->gpnum),
2159 j = jiffies_till_next_fqs;
2162 jiffies_till_next_fqs = HZ;
2165 jiffies_till_next_fqs = 1;
2169 /* Handle grace-period end. */
2170 rsp->gp_state = RCU_GP_CLEANUP;
2171 rcu_gp_cleanup(rsp);
2172 rsp->gp_state = RCU_GP_CLEANED;
2177 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2178 * in preparation for detecting the next grace period. The caller must hold
2179 * the root node's ->lock and hard irqs must be disabled.
2181 * Note that it is legal for a dying CPU (which is marked as offline) to
2182 * invoke this function. This can happen when the dying CPU reports its
2185 * Returns true if the grace-period kthread must be awakened.
2188 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2189 struct rcu_data *rdp)
2191 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2193 * Either we have not yet spawned the grace-period
2194 * task, this CPU does not need another grace period,
2195 * or a grace period is already in progress.
2196 * Either way, don't start a new grace period.
2200 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2201 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2205 * We can't do wakeups while holding the rnp->lock, as that
2206 * could cause possible deadlocks with the rq->lock. Defer
2207 * the wakeup to our caller.
2213 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2214 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2215 * is invoked indirectly from rcu_advance_cbs(), which would result in
2216 * endless recursion -- or would do so if it wasn't for the self-deadlock
2217 * that is encountered beforehand.
2219 * Returns true if the grace-period kthread needs to be awakened.
2221 static bool rcu_start_gp(struct rcu_state *rsp)
2223 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2224 struct rcu_node *rnp = rcu_get_root(rsp);
2228 * If there is no grace period in progress right now, any
2229 * callbacks we have up to this point will be satisfied by the
2230 * next grace period. Also, advancing the callbacks reduces the
2231 * probability of false positives from cpu_needs_another_gp()
2232 * resulting in pointless grace periods. So, advance callbacks
2233 * then start the grace period!
2235 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2236 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2241 * Report a full set of quiescent states to the specified rcu_state
2242 * data structure. This involves cleaning up after the prior grace
2243 * period and letting rcu_start_gp() start up the next grace period
2244 * if one is needed. Note that the caller must hold rnp->lock, which
2245 * is released before return.
2247 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2248 __releases(rcu_get_root(rsp)->lock)
2250 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2251 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2252 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2253 rcu_gp_kthread_wake(rsp);
2257 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2258 * Allows quiescent states for a group of CPUs to be reported at one go
2259 * to the specified rcu_node structure, though all the CPUs in the group
2260 * must be represented by the same rcu_node structure (which need not be a
2261 * leaf rcu_node structure, though it often will be). The gps parameter
2262 * is the grace-period snapshot, which means that the quiescent states
2263 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2264 * must be held upon entry, and it is released before return.
2267 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2268 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2269 __releases(rnp->lock)
2271 unsigned long oldmask = 0;
2272 struct rcu_node *rnp_c;
2274 /* Walk up the rcu_node hierarchy. */
2276 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2279 * Our bit has already been cleared, or the
2280 * relevant grace period is already over, so done.
2282 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2285 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2286 rnp->qsmask &= ~mask;
2287 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2288 mask, rnp->qsmask, rnp->level,
2289 rnp->grplo, rnp->grphi,
2291 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2293 /* Other bits still set at this level, so done. */
2294 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2297 mask = rnp->grpmask;
2298 if (rnp->parent == NULL) {
2300 /* No more levels. Exit loop holding root lock. */
2304 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2307 raw_spin_lock_irqsave(&rnp->lock, flags);
2308 smp_mb__after_unlock_lock();
2309 oldmask = rnp_c->qsmask;
2313 * Get here if we are the last CPU to pass through a quiescent
2314 * state for this grace period. Invoke rcu_report_qs_rsp()
2315 * to clean up and start the next grace period if one is needed.
2317 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2321 * Record a quiescent state for all tasks that were previously queued
2322 * on the specified rcu_node structure and that were blocking the current
2323 * RCU grace period. The caller must hold the specified rnp->lock with
2324 * irqs disabled, and this lock is released upon return, but irqs remain
2327 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2328 struct rcu_node *rnp, unsigned long flags)
2329 __releases(rnp->lock)
2333 struct rcu_node *rnp_p;
2335 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2336 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2337 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2338 return; /* Still need more quiescent states! */
2341 rnp_p = rnp->parent;
2342 if (rnp_p == NULL) {
2344 * Only one rcu_node structure in the tree, so don't
2345 * try to report up to its nonexistent parent!
2347 rcu_report_qs_rsp(rsp, flags);
2351 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2353 mask = rnp->grpmask;
2354 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2355 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
2356 smp_mb__after_unlock_lock();
2357 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2361 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2362 * structure. This must be either called from the specified CPU, or
2363 * called when the specified CPU is known to be offline (and when it is
2364 * also known that no other CPU is concurrently trying to help the offline
2365 * CPU). The lastcomp argument is used to make sure we are still in the
2366 * grace period of interest. We don't want to end the current grace period
2367 * based on quiescent states detected in an earlier grace period!
2370 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2372 unsigned long flags;
2375 struct rcu_node *rnp;
2378 raw_spin_lock_irqsave(&rnp->lock, flags);
2379 smp_mb__after_unlock_lock();
2380 if ((rdp->cpu_no_qs.b.norm &&
2381 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2382 rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2386 * The grace period in which this quiescent state was
2387 * recorded has ended, so don't report it upwards.
2388 * We will instead need a new quiescent state that lies
2389 * within the current grace period.
2391 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2392 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2393 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2396 mask = rdp->grpmask;
2397 if ((rnp->qsmask & mask) == 0) {
2398 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2400 rdp->core_needs_qs = 0;
2403 * This GP can't end until cpu checks in, so all of our
2404 * callbacks can be processed during the next GP.
2406 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2408 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2409 /* ^^^ Released rnp->lock */
2411 rcu_gp_kthread_wake(rsp);
2416 * Check to see if there is a new grace period of which this CPU
2417 * is not yet aware, and if so, set up local rcu_data state for it.
2418 * Otherwise, see if this CPU has just passed through its first
2419 * quiescent state for this grace period, and record that fact if so.
2422 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2424 /* Check for grace-period ends and beginnings. */
2425 note_gp_changes(rsp, rdp);
2428 * Does this CPU still need to do its part for current grace period?
2429 * If no, return and let the other CPUs do their part as well.
2431 if (!rdp->core_needs_qs)
2435 * Was there a quiescent state since the beginning of the grace
2436 * period? If no, then exit and wait for the next call.
2438 if (rdp->cpu_no_qs.b.norm &&
2439 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2443 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2446 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2450 * Send the specified CPU's RCU callbacks to the orphanage. The
2451 * specified CPU must be offline, and the caller must hold the
2455 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2456 struct rcu_node *rnp, struct rcu_data *rdp)
2458 /* No-CBs CPUs do not have orphanable callbacks. */
2459 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2463 * Orphan the callbacks. First adjust the counts. This is safe
2464 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2465 * cannot be running now. Thus no memory barrier is required.
2467 if (rdp->nxtlist != NULL) {
2468 rsp->qlen_lazy += rdp->qlen_lazy;
2469 rsp->qlen += rdp->qlen;
2470 rdp->n_cbs_orphaned += rdp->qlen;
2472 WRITE_ONCE(rdp->qlen, 0);
2476 * Next, move those callbacks still needing a grace period to
2477 * the orphanage, where some other CPU will pick them up.
2478 * Some of the callbacks might have gone partway through a grace
2479 * period, but that is too bad. They get to start over because we
2480 * cannot assume that grace periods are synchronized across CPUs.
2481 * We don't bother updating the ->nxttail[] array yet, instead
2482 * we just reset the whole thing later on.
2484 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2485 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2486 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2487 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2491 * Then move the ready-to-invoke callbacks to the orphanage,
2492 * where some other CPU will pick them up. These will not be
2493 * required to pass though another grace period: They are done.
2495 if (rdp->nxtlist != NULL) {
2496 *rsp->orphan_donetail = rdp->nxtlist;
2497 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2501 * Finally, initialize the rcu_data structure's list to empty and
2502 * disallow further callbacks on this CPU.
2504 init_callback_list(rdp);
2505 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2509 * Adopt the RCU callbacks from the specified rcu_state structure's
2510 * orphanage. The caller must hold the ->orphan_lock.
2512 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2515 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2517 /* No-CBs CPUs are handled specially. */
2518 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2519 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2522 /* Do the accounting first. */
2523 rdp->qlen_lazy += rsp->qlen_lazy;
2524 rdp->qlen += rsp->qlen;
2525 rdp->n_cbs_adopted += rsp->qlen;
2526 if (rsp->qlen_lazy != rsp->qlen)
2527 rcu_idle_count_callbacks_posted();
2532 * We do not need a memory barrier here because the only way we
2533 * can get here if there is an rcu_barrier() in flight is if
2534 * we are the task doing the rcu_barrier().
2537 /* First adopt the ready-to-invoke callbacks. */
2538 if (rsp->orphan_donelist != NULL) {
2539 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2540 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2541 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2542 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2543 rdp->nxttail[i] = rsp->orphan_donetail;
2544 rsp->orphan_donelist = NULL;
2545 rsp->orphan_donetail = &rsp->orphan_donelist;
2548 /* And then adopt the callbacks that still need a grace period. */
2549 if (rsp->orphan_nxtlist != NULL) {
2550 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2551 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2552 rsp->orphan_nxtlist = NULL;
2553 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2558 * Trace the fact that this CPU is going offline.
2560 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2562 RCU_TRACE(unsigned long mask);
2563 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2564 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2566 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2569 RCU_TRACE(mask = rdp->grpmask);
2570 trace_rcu_grace_period(rsp->name,
2571 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2576 * All CPUs for the specified rcu_node structure have gone offline,
2577 * and all tasks that were preempted within an RCU read-side critical
2578 * section while running on one of those CPUs have since exited their RCU
2579 * read-side critical section. Some other CPU is reporting this fact with
2580 * the specified rcu_node structure's ->lock held and interrupts disabled.
2581 * This function therefore goes up the tree of rcu_node structures,
2582 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2583 * the leaf rcu_node structure's ->qsmaskinit field has already been
2586 * This function does check that the specified rcu_node structure has
2587 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2588 * prematurely. That said, invoking it after the fact will cost you
2589 * a needless lock acquisition. So once it has done its work, don't
2592 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2595 struct rcu_node *rnp = rnp_leaf;
2597 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2598 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2601 mask = rnp->grpmask;
2605 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2606 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2607 rnp->qsmaskinit &= ~mask;
2608 rnp->qsmask &= ~mask;
2609 if (rnp->qsmaskinit) {
2610 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2613 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2618 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2619 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2622 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
2624 unsigned long flags;
2626 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2627 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2629 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2632 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2633 mask = rdp->grpmask;
2634 raw_spin_lock_irqsave(&rnp->lock, flags);
2635 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2636 rnp->qsmaskinitnext &= ~mask;
2637 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2641 * The CPU has been completely removed, and some other CPU is reporting
2642 * this fact from process context. Do the remainder of the cleanup,
2643 * including orphaning the outgoing CPU's RCU callbacks, and also
2644 * adopting them. There can only be one CPU hotplug operation at a time,
2645 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2647 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2649 unsigned long flags;
2650 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2651 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2653 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2656 /* Adjust any no-longer-needed kthreads. */
2657 rcu_boost_kthread_setaffinity(rnp, -1);
2659 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2660 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2661 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2662 rcu_adopt_orphan_cbs(rsp, flags);
2663 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2665 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2666 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2667 cpu, rdp->qlen, rdp->nxtlist);
2671 * Invoke any RCU callbacks that have made it to the end of their grace
2672 * period. Thottle as specified by rdp->blimit.
2674 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2676 unsigned long flags;
2677 struct rcu_head *next, *list, **tail;
2678 long bl, count, count_lazy;
2681 /* If no callbacks are ready, just return. */
2682 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2683 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2684 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2685 need_resched(), is_idle_task(current),
2686 rcu_is_callbacks_kthread());
2691 * Extract the list of ready callbacks, disabling to prevent
2692 * races with call_rcu() from interrupt handlers.
2694 local_irq_save(flags);
2695 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2697 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2698 list = rdp->nxtlist;
2699 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2700 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2701 tail = rdp->nxttail[RCU_DONE_TAIL];
2702 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2703 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2704 rdp->nxttail[i] = &rdp->nxtlist;
2705 local_irq_restore(flags);
2707 /* Invoke callbacks. */
2708 count = count_lazy = 0;
2712 debug_rcu_head_unqueue(list);
2713 if (__rcu_reclaim(rsp->name, list))
2716 /* Stop only if limit reached and CPU has something to do. */
2717 if (++count >= bl &&
2719 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2723 local_irq_save(flags);
2724 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2725 is_idle_task(current),
2726 rcu_is_callbacks_kthread());
2728 /* Update count, and requeue any remaining callbacks. */
2730 *tail = rdp->nxtlist;
2731 rdp->nxtlist = list;
2732 for (i = 0; i < RCU_NEXT_SIZE; i++)
2733 if (&rdp->nxtlist == rdp->nxttail[i])
2734 rdp->nxttail[i] = tail;
2738 smp_mb(); /* List handling before counting for rcu_barrier(). */
2739 rdp->qlen_lazy -= count_lazy;
2740 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2741 rdp->n_cbs_invoked += count;
2743 /* Reinstate batch limit if we have worked down the excess. */
2744 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2745 rdp->blimit = blimit;
2747 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2748 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2749 rdp->qlen_last_fqs_check = 0;
2750 rdp->n_force_qs_snap = rsp->n_force_qs;
2751 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2752 rdp->qlen_last_fqs_check = rdp->qlen;
2753 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2755 local_irq_restore(flags);
2757 /* Re-invoke RCU core processing if there are callbacks remaining. */
2758 if (cpu_has_callbacks_ready_to_invoke(rdp))
2763 * Check to see if this CPU is in a non-context-switch quiescent state
2764 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2765 * Also schedule RCU core processing.
2767 * This function must be called from hardirq context. It is normally
2768 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2769 * false, there is no point in invoking rcu_check_callbacks().
2771 void rcu_check_callbacks(int user)
2773 trace_rcu_utilization(TPS("Start scheduler-tick"));
2774 increment_cpu_stall_ticks();
2775 if (user || rcu_is_cpu_rrupt_from_idle()) {
2778 * Get here if this CPU took its interrupt from user
2779 * mode or from the idle loop, and if this is not a
2780 * nested interrupt. In this case, the CPU is in
2781 * a quiescent state, so note it.
2783 * No memory barrier is required here because both
2784 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2785 * variables that other CPUs neither access nor modify,
2786 * at least not while the corresponding CPU is online.
2792 } else if (!in_softirq()) {
2795 * Get here if this CPU did not take its interrupt from
2796 * softirq, in other words, if it is not interrupting
2797 * a rcu_bh read-side critical section. This is an _bh
2798 * critical section, so note it.
2803 rcu_preempt_check_callbacks();
2807 rcu_note_voluntary_context_switch(current);
2808 trace_rcu_utilization(TPS("End scheduler-tick"));
2812 * Scan the leaf rcu_node structures, processing dyntick state for any that
2813 * have not yet encountered a quiescent state, using the function specified.
2814 * Also initiate boosting for any threads blocked on the root rcu_node.
2816 * The caller must have suppressed start of new grace periods.
2818 static void force_qs_rnp(struct rcu_state *rsp,
2819 int (*f)(struct rcu_data *rsp, bool *isidle,
2820 unsigned long *maxj),
2821 bool *isidle, unsigned long *maxj)
2825 unsigned long flags;
2827 struct rcu_node *rnp;
2829 rcu_for_each_leaf_node(rsp, rnp) {
2830 cond_resched_rcu_qs();
2832 raw_spin_lock_irqsave(&rnp->lock, flags);
2833 smp_mb__after_unlock_lock();
2834 if (rnp->qsmask == 0) {
2835 if (rcu_state_p == &rcu_sched_state ||
2836 rsp != rcu_state_p ||
2837 rcu_preempt_blocked_readers_cgp(rnp)) {
2839 * No point in scanning bits because they
2840 * are all zero. But we might need to
2841 * priority-boost blocked readers.
2843 rcu_initiate_boost(rnp, flags);
2844 /* rcu_initiate_boost() releases rnp->lock */
2848 (rnp->parent->qsmask & rnp->grpmask)) {
2850 * Race between grace-period
2851 * initialization and task exiting RCU
2852 * read-side critical section: Report.
2854 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2855 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2861 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2862 if ((rnp->qsmask & bit) != 0) {
2863 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2868 /* Idle/offline CPUs, report (releases rnp->lock. */
2869 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2871 /* Nothing to do here, so just drop the lock. */
2872 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2878 * Force quiescent states on reluctant CPUs, and also detect which
2879 * CPUs are in dyntick-idle mode.
2881 static void force_quiescent_state(struct rcu_state *rsp)
2883 unsigned long flags;
2885 struct rcu_node *rnp;
2886 struct rcu_node *rnp_old = NULL;
2888 /* Funnel through hierarchy to reduce memory contention. */
2889 rnp = __this_cpu_read(rsp->rda->mynode);
2890 for (; rnp != NULL; rnp = rnp->parent) {
2891 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2892 !raw_spin_trylock(&rnp->fqslock);
2893 if (rnp_old != NULL)
2894 raw_spin_unlock(&rnp_old->fqslock);
2896 rsp->n_force_qs_lh++;
2901 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2903 /* Reached the root of the rcu_node tree, acquire lock. */
2904 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2905 smp_mb__after_unlock_lock();
2906 raw_spin_unlock(&rnp_old->fqslock);
2907 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2908 rsp->n_force_qs_lh++;
2909 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2910 return; /* Someone beat us to it. */
2912 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2913 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2914 rcu_gp_kthread_wake(rsp);
2918 * This does the RCU core processing work for the specified rcu_state
2919 * and rcu_data structures. This may be called only from the CPU to
2920 * whom the rdp belongs.
2923 __rcu_process_callbacks(struct rcu_state *rsp)
2925 unsigned long flags;
2927 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2929 WARN_ON_ONCE(rdp->beenonline == 0);
2931 /* Update RCU state based on any recent quiescent states. */
2932 rcu_check_quiescent_state(rsp, rdp);
2934 /* Does this CPU require a not-yet-started grace period? */
2935 local_irq_save(flags);
2936 if (cpu_needs_another_gp(rsp, rdp)) {
2937 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2938 needwake = rcu_start_gp(rsp);
2939 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2941 rcu_gp_kthread_wake(rsp);
2943 local_irq_restore(flags);
2946 /* If there are callbacks ready, invoke them. */
2947 if (cpu_has_callbacks_ready_to_invoke(rdp))
2948 invoke_rcu_callbacks(rsp, rdp);
2950 /* Do any needed deferred wakeups of rcuo kthreads. */
2951 do_nocb_deferred_wakeup(rdp);
2955 * Do RCU core processing for the current CPU.
2957 static void rcu_process_callbacks(struct softirq_action *unused)
2959 struct rcu_state *rsp;
2961 if (cpu_is_offline(smp_processor_id()))
2963 trace_rcu_utilization(TPS("Start RCU core"));
2964 for_each_rcu_flavor(rsp)
2965 __rcu_process_callbacks(rsp);
2966 trace_rcu_utilization(TPS("End RCU core"));
2970 * Schedule RCU callback invocation. If the specified type of RCU
2971 * does not support RCU priority boosting, just do a direct call,
2972 * otherwise wake up the per-CPU kernel kthread. Note that because we
2973 * are running on the current CPU with softirqs disabled, the
2974 * rcu_cpu_kthread_task cannot disappear out from under us.
2976 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2978 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2980 if (likely(!rsp->boost)) {
2981 rcu_do_batch(rsp, rdp);
2984 invoke_rcu_callbacks_kthread();
2987 static void invoke_rcu_core(void)
2989 if (cpu_online(smp_processor_id()))
2990 raise_softirq(RCU_SOFTIRQ);
2994 * Handle any core-RCU processing required by a call_rcu() invocation.
2996 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2997 struct rcu_head *head, unsigned long flags)
3002 * If called from an extended quiescent state, invoke the RCU
3003 * core in order to force a re-evaluation of RCU's idleness.
3005 if (!rcu_is_watching())
3008 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3009 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3013 * Force the grace period if too many callbacks or too long waiting.
3014 * Enforce hysteresis, and don't invoke force_quiescent_state()
3015 * if some other CPU has recently done so. Also, don't bother
3016 * invoking force_quiescent_state() if the newly enqueued callback
3017 * is the only one waiting for a grace period to complete.
3019 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
3021 /* Are we ignoring a completed grace period? */
3022 note_gp_changes(rsp, rdp);
3024 /* Start a new grace period if one not already started. */
3025 if (!rcu_gp_in_progress(rsp)) {
3026 struct rcu_node *rnp_root = rcu_get_root(rsp);
3028 raw_spin_lock(&rnp_root->lock);
3029 smp_mb__after_unlock_lock();
3030 needwake = rcu_start_gp(rsp);
3031 raw_spin_unlock(&rnp_root->lock);
3033 rcu_gp_kthread_wake(rsp);
3035 /* Give the grace period a kick. */
3036 rdp->blimit = LONG_MAX;
3037 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3038 *rdp->nxttail[RCU_DONE_TAIL] != head)
3039 force_quiescent_state(rsp);
3040 rdp->n_force_qs_snap = rsp->n_force_qs;
3041 rdp->qlen_last_fqs_check = rdp->qlen;
3047 * RCU callback function to leak a callback.
3049 static void rcu_leak_callback(struct rcu_head *rhp)
3054 * Helper function for call_rcu() and friends. The cpu argument will
3055 * normally be -1, indicating "currently running CPU". It may specify
3056 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3057 * is expected to specify a CPU.
3060 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3061 struct rcu_state *rsp, int cpu, bool lazy)
3063 unsigned long flags;
3064 struct rcu_data *rdp;
3066 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
3067 if (debug_rcu_head_queue(head)) {
3068 /* Probable double call_rcu(), so leak the callback. */
3069 WRITE_ONCE(head->func, rcu_leak_callback);
3070 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3077 * Opportunistically note grace-period endings and beginnings.
3078 * Note that we might see a beginning right after we see an
3079 * end, but never vice versa, since this CPU has to pass through
3080 * a quiescent state betweentimes.
3082 local_irq_save(flags);
3083 rdp = this_cpu_ptr(rsp->rda);
3085 /* Add the callback to our list. */
3086 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3090 rdp = per_cpu_ptr(rsp->rda, cpu);
3091 if (likely(rdp->mynode)) {
3092 /* Post-boot, so this should be for a no-CBs CPU. */
3093 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3094 WARN_ON_ONCE(offline);
3095 /* Offline CPU, _call_rcu() illegal, leak callback. */
3096 local_irq_restore(flags);
3100 * Very early boot, before rcu_init(). Initialize if needed
3101 * and then drop through to queue the callback.
3104 WARN_ON_ONCE(!rcu_is_watching());
3105 if (!likely(rdp->nxtlist))
3106 init_default_callback_list(rdp);
3108 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3112 rcu_idle_count_callbacks_posted();
3113 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3114 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3115 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3117 if (__is_kfree_rcu_offset((unsigned long)func))
3118 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3119 rdp->qlen_lazy, rdp->qlen);
3121 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3123 /* Go handle any RCU core processing required. */
3124 __call_rcu_core(rsp, rdp, head, flags);
3125 local_irq_restore(flags);
3129 * Queue an RCU-sched callback for invocation after a grace period.
3131 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3133 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3135 EXPORT_SYMBOL_GPL(call_rcu_sched);
3138 * Queue an RCU callback for invocation after a quicker grace period.
3140 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3142 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3144 EXPORT_SYMBOL_GPL(call_rcu_bh);
3147 * Queue an RCU callback for lazy invocation after a grace period.
3148 * This will likely be later named something like "call_rcu_lazy()",
3149 * but this change will require some way of tagging the lazy RCU
3150 * callbacks in the list of pending callbacks. Until then, this
3151 * function may only be called from __kfree_rcu().
3153 void kfree_call_rcu(struct rcu_head *head,
3154 rcu_callback_t func)
3156 __call_rcu(head, func, rcu_state_p, -1, 1);
3158 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3161 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3162 * any blocking grace-period wait automatically implies a grace period
3163 * if there is only one CPU online at any point time during execution
3164 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3165 * occasionally incorrectly indicate that there are multiple CPUs online
3166 * when there was in fact only one the whole time, as this just adds
3167 * some overhead: RCU still operates correctly.
3169 static inline int rcu_blocking_is_gp(void)
3173 might_sleep(); /* Check for RCU read-side critical section. */
3175 ret = num_online_cpus() <= 1;
3181 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3183 * Control will return to the caller some time after a full rcu-sched
3184 * grace period has elapsed, in other words after all currently executing
3185 * rcu-sched read-side critical sections have completed. These read-side
3186 * critical sections are delimited by rcu_read_lock_sched() and
3187 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3188 * local_irq_disable(), and so on may be used in place of
3189 * rcu_read_lock_sched().
3191 * This means that all preempt_disable code sequences, including NMI and
3192 * non-threaded hardware-interrupt handlers, in progress on entry will
3193 * have completed before this primitive returns. However, this does not
3194 * guarantee that softirq handlers will have completed, since in some
3195 * kernels, these handlers can run in process context, and can block.
3197 * Note that this guarantee implies further memory-ordering guarantees.
3198 * On systems with more than one CPU, when synchronize_sched() returns,
3199 * each CPU is guaranteed to have executed a full memory barrier since the
3200 * end of its last RCU-sched read-side critical section whose beginning
3201 * preceded the call to synchronize_sched(). In addition, each CPU having
3202 * an RCU read-side critical section that extends beyond the return from
3203 * synchronize_sched() is guaranteed to have executed a full memory barrier
3204 * after the beginning of synchronize_sched() and before the beginning of
3205 * that RCU read-side critical section. Note that these guarantees include
3206 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3207 * that are executing in the kernel.
3209 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3210 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3211 * to have executed a full memory barrier during the execution of
3212 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3213 * again only if the system has more than one CPU).
3215 * This primitive provides the guarantees made by the (now removed)
3216 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3217 * guarantees that rcu_read_lock() sections will have completed.
3218 * In "classic RCU", these two guarantees happen to be one and
3219 * the same, but can differ in realtime RCU implementations.
3221 void synchronize_sched(void)
3223 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3224 lock_is_held(&rcu_lock_map) ||
3225 lock_is_held(&rcu_sched_lock_map),
3226 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3227 if (rcu_blocking_is_gp())
3229 if (rcu_gp_is_expedited())
3230 synchronize_sched_expedited();
3232 wait_rcu_gp(call_rcu_sched);
3234 EXPORT_SYMBOL_GPL(synchronize_sched);
3237 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3239 * Control will return to the caller some time after a full rcu_bh grace
3240 * period has elapsed, in other words after all currently executing rcu_bh
3241 * read-side critical sections have completed. RCU read-side critical
3242 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3243 * and may be nested.
3245 * See the description of synchronize_sched() for more detailed information
3246 * on memory ordering guarantees.
3248 void synchronize_rcu_bh(void)
3250 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3251 lock_is_held(&rcu_lock_map) ||
3252 lock_is_held(&rcu_sched_lock_map),
3253 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3254 if (rcu_blocking_is_gp())
3256 if (rcu_gp_is_expedited())
3257 synchronize_rcu_bh_expedited();
3259 wait_rcu_gp(call_rcu_bh);
3261 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3264 * get_state_synchronize_rcu - Snapshot current RCU state
3266 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3267 * to determine whether or not a full grace period has elapsed in the
3270 unsigned long get_state_synchronize_rcu(void)
3273 * Any prior manipulation of RCU-protected data must happen
3274 * before the load from ->gpnum.
3279 * Make sure this load happens before the purportedly
3280 * time-consuming work between get_state_synchronize_rcu()
3281 * and cond_synchronize_rcu().
3283 return smp_load_acquire(&rcu_state_p->gpnum);
3285 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3288 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3290 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3292 * If a full RCU grace period has elapsed since the earlier call to
3293 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3294 * synchronize_rcu() to wait for a full grace period.
3296 * Yes, this function does not take counter wrap into account. But
3297 * counter wrap is harmless. If the counter wraps, we have waited for
3298 * more than 2 billion grace periods (and way more on a 64-bit system!),
3299 * so waiting for one additional grace period should be just fine.
3301 void cond_synchronize_rcu(unsigned long oldstate)
3303 unsigned long newstate;
3306 * Ensure that this load happens before any RCU-destructive
3307 * actions the caller might carry out after we return.
3309 newstate = smp_load_acquire(&rcu_state_p->completed);
3310 if (ULONG_CMP_GE(oldstate, newstate))
3313 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3316 * get_state_synchronize_sched - Snapshot current RCU-sched state
3318 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3319 * to determine whether or not a full grace period has elapsed in the
3322 unsigned long get_state_synchronize_sched(void)
3325 * Any prior manipulation of RCU-protected data must happen
3326 * before the load from ->gpnum.
3331 * Make sure this load happens before the purportedly
3332 * time-consuming work between get_state_synchronize_sched()
3333 * and cond_synchronize_sched().
3335 return smp_load_acquire(&rcu_sched_state.gpnum);
3337 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3340 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3342 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3344 * If a full RCU-sched grace period has elapsed since the earlier call to
3345 * get_state_synchronize_sched(), just return. Otherwise, invoke
3346 * synchronize_sched() to wait for a full grace period.
3348 * Yes, this function does not take counter wrap into account. But
3349 * counter wrap is harmless. If the counter wraps, we have waited for
3350 * more than 2 billion grace periods (and way more on a 64-bit system!),
3351 * so waiting for one additional grace period should be just fine.
3353 void cond_synchronize_sched(unsigned long oldstate)
3355 unsigned long newstate;
3358 * Ensure that this load happens before any RCU-destructive
3359 * actions the caller might carry out after we return.
3361 newstate = smp_load_acquire(&rcu_sched_state.completed);
3362 if (ULONG_CMP_GE(oldstate, newstate))
3363 synchronize_sched();
3365 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3367 /* Adjust sequence number for start of update-side operation. */
3368 static void rcu_seq_start(unsigned long *sp)
3370 WRITE_ONCE(*sp, *sp + 1);
3371 smp_mb(); /* Ensure update-side operation after counter increment. */
3372 WARN_ON_ONCE(!(*sp & 0x1));
3375 /* Adjust sequence number for end of update-side operation. */
3376 static void rcu_seq_end(unsigned long *sp)
3378 smp_mb(); /* Ensure update-side operation before counter increment. */
3379 WRITE_ONCE(*sp, *sp + 1);
3380 WARN_ON_ONCE(*sp & 0x1);
3383 /* Take a snapshot of the update side's sequence number. */
3384 static unsigned long rcu_seq_snap(unsigned long *sp)
3388 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3389 s = (READ_ONCE(*sp) + 3) & ~0x1;
3390 smp_mb(); /* Above access must not bleed into critical section. */
3395 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3396 * full update-side operation has occurred.
3398 static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3400 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3403 /* Wrapper functions for expedited grace periods. */
3404 static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
3406 rcu_seq_start(&rsp->expedited_sequence);
3408 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
3410 rcu_seq_end(&rsp->expedited_sequence);
3411 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3413 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
3415 return rcu_seq_snap(&rsp->expedited_sequence);
3417 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
3419 return rcu_seq_done(&rsp->expedited_sequence, s);
3423 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3424 * recent CPU-online activity. Note that these masks are not cleared
3425 * when CPUs go offline, so they reflect the union of all CPUs that have
3426 * ever been online. This means that this function normally takes its
3427 * no-work-to-do fastpath.
3429 static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
3432 unsigned long flags;
3434 unsigned long oldmask;
3435 int ncpus = READ_ONCE(rsp->ncpus);
3436 struct rcu_node *rnp;
3437 struct rcu_node *rnp_up;
3439 /* If no new CPUs onlined since last time, nothing to do. */
3440 if (likely(ncpus == rsp->ncpus_snap))
3442 rsp->ncpus_snap = ncpus;
3445 * Each pass through the following loop propagates newly onlined
3446 * CPUs for the current rcu_node structure up the rcu_node tree.
3448 rcu_for_each_leaf_node(rsp, rnp) {
3449 raw_spin_lock_irqsave(&rnp->lock, flags);
3450 smp_mb__after_unlock_lock();
3451 if (rnp->expmaskinit == rnp->expmaskinitnext) {
3452 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3453 continue; /* No new CPUs, nothing to do. */
3456 /* Update this node's mask, track old value for propagation. */
3457 oldmask = rnp->expmaskinit;
3458 rnp->expmaskinit = rnp->expmaskinitnext;
3459 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3461 /* If was already nonzero, nothing to propagate. */
3465 /* Propagate the new CPU up the tree. */
3466 mask = rnp->grpmask;
3467 rnp_up = rnp->parent;
3470 raw_spin_lock_irqsave(&rnp_up->lock, flags);
3471 smp_mb__after_unlock_lock();
3472 if (rnp_up->expmaskinit)
3474 rnp_up->expmaskinit |= mask;
3475 raw_spin_unlock_irqrestore(&rnp_up->lock, flags);
3478 mask = rnp_up->grpmask;
3479 rnp_up = rnp_up->parent;
3485 * Reset the ->expmask values in the rcu_node tree in preparation for
3486 * a new expedited grace period.
3488 static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
3490 unsigned long flags;
3491 struct rcu_node *rnp;
3493 sync_exp_reset_tree_hotplug(rsp);
3494 rcu_for_each_node_breadth_first(rsp, rnp) {
3495 raw_spin_lock_irqsave(&rnp->lock, flags);
3496 smp_mb__after_unlock_lock();
3497 WARN_ON_ONCE(rnp->expmask);
3498 rnp->expmask = rnp->expmaskinit;
3499 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3504 * Return non-zero if there is no RCU expedited grace period in progress
3505 * for the specified rcu_node structure, in other words, if all CPUs and
3506 * tasks covered by the specified rcu_node structure have done their bit
3507 * for the current expedited grace period. Works only for preemptible
3508 * RCU -- other RCU implementation use other means.
3510 * Caller must hold the root rcu_node's exp_funnel_mutex.
3512 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
3514 return rnp->exp_tasks == NULL &&
3515 READ_ONCE(rnp->expmask) == 0;
3519 * Report the exit from RCU read-side critical section for the last task
3520 * that queued itself during or before the current expedited preemptible-RCU
3521 * grace period. This event is reported either to the rcu_node structure on
3522 * which the task was queued or to one of that rcu_node structure's ancestors,
3523 * recursively up the tree. (Calm down, calm down, we do the recursion
3526 * Caller must hold the root rcu_node's exp_funnel_mutex and the
3527 * specified rcu_node structure's ->lock.
3529 static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
3530 bool wake, unsigned long flags)
3531 __releases(rnp->lock)
3536 if (!sync_rcu_preempt_exp_done(rnp)) {
3538 rcu_initiate_boost(rnp, flags);
3540 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3543 if (rnp->parent == NULL) {
3544 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3546 smp_mb(); /* EGP done before wake_up(). */
3547 wake_up(&rsp->expedited_wq);
3551 mask = rnp->grpmask;
3552 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
3554 raw_spin_lock(&rnp->lock); /* irqs already disabled */
3555 smp_mb__after_unlock_lock();
3556 WARN_ON_ONCE(!(rnp->expmask & mask));
3557 rnp->expmask &= ~mask;
3562 * Report expedited quiescent state for specified node. This is a
3563 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3565 * Caller must hold the root rcu_node's exp_funnel_mutex.
3567 static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
3568 struct rcu_node *rnp, bool wake)
3570 unsigned long flags;
3572 raw_spin_lock_irqsave(&rnp->lock, flags);
3573 smp_mb__after_unlock_lock();
3574 __rcu_report_exp_rnp(rsp, rnp, wake, flags);
3578 * Report expedited quiescent state for multiple CPUs, all covered by the
3579 * specified leaf rcu_node structure. Caller must hold the root
3580 * rcu_node's exp_funnel_mutex.
3582 static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
3583 unsigned long mask, bool wake)
3585 unsigned long flags;
3587 raw_spin_lock_irqsave(&rnp->lock, flags);
3588 smp_mb__after_unlock_lock();
3589 if (!(rnp->expmask & mask)) {
3590 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3593 rnp->expmask &= ~mask;
3594 __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
3598 * Report expedited quiescent state for specified rcu_data (CPU).
3599 * Caller must hold the root rcu_node's exp_funnel_mutex.
3601 static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
3604 rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
3607 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3608 static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp,
3609 struct rcu_data *rdp,
3610 atomic_long_t *stat, unsigned long s)
3612 if (rcu_exp_gp_seq_done(rsp, s)) {
3614 mutex_unlock(&rnp->exp_funnel_mutex);
3616 mutex_unlock(&rdp->exp_funnel_mutex);
3617 /* Ensure test happens before caller kfree(). */
3618 smp_mb__before_atomic(); /* ^^^ */
3619 atomic_long_inc(stat);
3626 * Funnel-lock acquisition for expedited grace periods. Returns a
3627 * pointer to the root rcu_node structure, or NULL if some other
3628 * task did the expedited grace period for us.
3630 static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
3632 struct rcu_data *rdp;
3633 struct rcu_node *rnp0;
3634 struct rcu_node *rnp1 = NULL;
3637 * First try directly acquiring the root lock in order to reduce
3638 * latency in the common case where expedited grace periods are
3639 * rare. We check mutex_is_locked() to avoid pathological levels of
3640 * memory contention on ->exp_funnel_mutex in the heavy-load case.
3642 rnp0 = rcu_get_root(rsp);
3643 if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) {
3644 if (mutex_trylock(&rnp0->exp_funnel_mutex)) {
3645 if (sync_exp_work_done(rsp, rnp0, NULL,
3646 &rsp->expedited_workdone0, s))
3653 * Each pass through the following loop works its way
3654 * up the rcu_node tree, returning if others have done the
3655 * work or otherwise falls through holding the root rnp's
3656 * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure
3657 * can be inexact, as it is just promoting locality and is not
3658 * strictly needed for correctness.
3660 rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
3661 if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s))
3663 mutex_lock(&rdp->exp_funnel_mutex);
3665 for (; rnp0 != NULL; rnp0 = rnp0->parent) {
3666 if (sync_exp_work_done(rsp, rnp1, rdp,
3667 &rsp->expedited_workdone2, s))
3669 mutex_lock(&rnp0->exp_funnel_mutex);
3671 mutex_unlock(&rnp1->exp_funnel_mutex);
3673 mutex_unlock(&rdp->exp_funnel_mutex);
3676 if (sync_exp_work_done(rsp, rnp1, rdp,
3677 &rsp->expedited_workdone3, s))
3682 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3683 static void sync_sched_exp_handler(void *data)
3685 struct rcu_data *rdp;
3686 struct rcu_node *rnp;
3687 struct rcu_state *rsp = data;
3689 rdp = this_cpu_ptr(rsp->rda);
3691 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
3692 __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
3694 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
3695 resched_cpu(smp_processor_id());
3698 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
3699 static void sync_sched_exp_online_cleanup(int cpu)
3701 struct rcu_data *rdp;
3703 struct rcu_node *rnp;
3704 struct rcu_state *rsp = &rcu_sched_state;
3706 rdp = per_cpu_ptr(rsp->rda, cpu);
3708 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
3710 ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
3715 * Select the nodes that the upcoming expedited grace period needs
3718 static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
3719 smp_call_func_t func)
3722 unsigned long flags;
3724 unsigned long mask_ofl_test;
3725 unsigned long mask_ofl_ipi;
3727 struct rcu_node *rnp;
3729 sync_exp_reset_tree(rsp);
3730 rcu_for_each_leaf_node(rsp, rnp) {
3731 raw_spin_lock_irqsave(&rnp->lock, flags);
3732 smp_mb__after_unlock_lock();
3734 /* Each pass checks a CPU for identity, offline, and idle. */
3736 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
3737 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3738 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3740 if (raw_smp_processor_id() == cpu ||
3741 !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3742 mask_ofl_test |= rdp->grpmask;
3744 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
3747 * Need to wait for any blocked tasks as well. Note that
3748 * additional blocking tasks will also block the expedited
3749 * GP until such time as the ->expmask bits are cleared.
3751 if (rcu_preempt_has_tasks(rnp))
3752 rnp->exp_tasks = rnp->blkd_tasks.next;
3753 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3755 /* IPI the remaining CPUs for expedited quiescent state. */
3757 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3758 if (!(mask_ofl_ipi & mask))
3761 ret = smp_call_function_single(cpu, func, rsp, 0);
3763 mask_ofl_ipi &= ~mask;
3765 /* Failed, raced with offline. */
3766 raw_spin_lock_irqsave(&rnp->lock, flags);
3767 if (cpu_online(cpu) &&
3768 (rnp->expmask & mask)) {
3769 raw_spin_unlock_irqrestore(&rnp->lock,
3771 schedule_timeout_uninterruptible(1);
3772 if (cpu_online(cpu) &&
3773 (rnp->expmask & mask))
3775 raw_spin_lock_irqsave(&rnp->lock,
3778 if (!(rnp->expmask & mask))
3779 mask_ofl_ipi &= ~mask;
3780 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3783 /* Report quiescent states for those that went offline. */
3784 mask_ofl_test |= mask_ofl_ipi;
3786 rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
3790 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
3793 unsigned long jiffies_stall;
3794 unsigned long jiffies_start;
3796 struct rcu_node *rnp;
3797 struct rcu_node *rnp_root = rcu_get_root(rsp);
3800 jiffies_stall = rcu_jiffies_till_stall_check();
3801 jiffies_start = jiffies;
3804 ret = wait_event_interruptible_timeout(
3806 sync_rcu_preempt_exp_done(rnp_root),
3811 /* Hit a signal, disable CPU stall warnings. */
3812 wait_event(rsp->expedited_wq,
3813 sync_rcu_preempt_exp_done(rnp_root));
3816 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3818 rcu_for_each_leaf_node(rsp, rnp) {
3819 (void)rcu_print_task_exp_stall(rnp);
3821 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3822 struct rcu_data *rdp;
3824 if (!(rnp->expmask & mask))
3826 rdp = per_cpu_ptr(rsp->rda, cpu);
3827 pr_cont(" %d-%c%c%c", cpu,
3828 "O."[!!cpu_online(cpu)],
3829 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
3830 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
3834 pr_cont(" } %lu jiffies s: %lu\n",
3835 jiffies - jiffies_start, rsp->expedited_sequence);
3836 rcu_for_each_leaf_node(rsp, rnp) {
3838 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3839 if (!(rnp->expmask & mask))
3844 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
3849 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3851 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3852 * approach to force the grace period to end quickly. This consumes
3853 * significant time on all CPUs and is unfriendly to real-time workloads,
3854 * so is thus not recommended for any sort of common-case code. In fact,
3855 * if you are using synchronize_sched_expedited() in a loop, please
3856 * restructure your code to batch your updates, and then use a single
3857 * synchronize_sched() instead.
3859 * This implementation can be thought of as an application of sequence
3860 * locking to expedited grace periods, but using the sequence counter to
3861 * determine when someone else has already done the work instead of for
3864 void synchronize_sched_expedited(void)
3867 struct rcu_node *rnp;
3868 struct rcu_state *rsp = &rcu_sched_state;
3870 /* Take a snapshot of the sequence number. */
3871 s = rcu_exp_gp_seq_snap(rsp);
3873 rnp = exp_funnel_lock(rsp, s);
3875 return; /* Someone else did our work for us. */
3877 rcu_exp_gp_seq_start(rsp);
3878 sync_rcu_exp_select_cpus(rsp, sync_sched_exp_handler);
3879 synchronize_sched_expedited_wait(rsp);
3881 rcu_exp_gp_seq_end(rsp);
3882 mutex_unlock(&rnp->exp_funnel_mutex);
3884 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3887 * Check to see if there is any immediate RCU-related work to be done
3888 * by the current CPU, for the specified type of RCU, returning 1 if so.
3889 * The checks are in order of increasing expense: checks that can be
3890 * carried out against CPU-local state are performed first. However,
3891 * we must check for CPU stalls first, else we might not get a chance.
3893 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3895 struct rcu_node *rnp = rdp->mynode;
3897 rdp->n_rcu_pending++;
3899 /* Check for CPU stalls, if enabled. */
3900 check_cpu_stall(rsp, rdp);
3902 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3903 if (rcu_nohz_full_cpu(rsp))
3906 /* Is the RCU core waiting for a quiescent state from this CPU? */
3907 if (rcu_scheduler_fully_active &&
3908 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3909 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3910 rdp->n_rp_core_needs_qs++;
3911 } else if (rdp->core_needs_qs &&
3912 (!rdp->cpu_no_qs.b.norm ||
3913 rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3914 rdp->n_rp_report_qs++;
3918 /* Does this CPU have callbacks ready to invoke? */
3919 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3920 rdp->n_rp_cb_ready++;
3924 /* Has RCU gone idle with this CPU needing another grace period? */
3925 if (cpu_needs_another_gp(rsp, rdp)) {
3926 rdp->n_rp_cpu_needs_gp++;
3930 /* Has another RCU grace period completed? */
3931 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3932 rdp->n_rp_gp_completed++;
3936 /* Has a new RCU grace period started? */
3937 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3938 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3939 rdp->n_rp_gp_started++;
3943 /* Does this CPU need a deferred NOCB wakeup? */
3944 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3945 rdp->n_rp_nocb_defer_wakeup++;
3950 rdp->n_rp_need_nothing++;
3955 * Check to see if there is any immediate RCU-related work to be done
3956 * by the current CPU, returning 1 if so. This function is part of the
3957 * RCU implementation; it is -not- an exported member of the RCU API.
3959 static int rcu_pending(void)
3961 struct rcu_state *rsp;
3963 for_each_rcu_flavor(rsp)
3964 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3970 * Return true if the specified CPU has any callback. If all_lazy is
3971 * non-NULL, store an indication of whether all callbacks are lazy.
3972 * (If there are no callbacks, all of them are deemed to be lazy.)
3974 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3978 struct rcu_data *rdp;
3979 struct rcu_state *rsp;
3981 for_each_rcu_flavor(rsp) {
3982 rdp = this_cpu_ptr(rsp->rda);
3986 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3997 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3998 * the compiler is expected to optimize this away.
4000 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
4001 int cpu, unsigned long done)
4003 trace_rcu_barrier(rsp->name, s, cpu,
4004 atomic_read(&rsp->barrier_cpu_count), done);
4008 * RCU callback function for _rcu_barrier(). If we are last, wake
4009 * up the task executing _rcu_barrier().
4011 static void rcu_barrier_callback(struct rcu_head *rhp)
4013 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
4014 struct rcu_state *rsp = rdp->rsp;
4016 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
4017 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
4018 complete(&rsp->barrier_completion);
4020 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
4025 * Called with preemption disabled, and from cross-cpu IRQ context.
4027 static void rcu_barrier_func(void *type)
4029 struct rcu_state *rsp = type;
4030 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
4032 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
4033 atomic_inc(&rsp->barrier_cpu_count);
4034 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
4038 * Orchestrate the specified type of RCU barrier, waiting for all
4039 * RCU callbacks of the specified type to complete.
4041 static void _rcu_barrier(struct rcu_state *rsp)
4044 struct rcu_data *rdp;
4045 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
4047 _rcu_barrier_trace(rsp, "Begin", -1, s);
4049 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4050 mutex_lock(&rsp->barrier_mutex);
4052 /* Did someone else do our work for us? */
4053 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
4054 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
4055 smp_mb(); /* caller's subsequent code after above check. */
4056 mutex_unlock(&rsp->barrier_mutex);
4060 /* Mark the start of the barrier operation. */
4061 rcu_seq_start(&rsp->barrier_sequence);
4062 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
4065 * Initialize the count to one rather than to zero in order to
4066 * avoid a too-soon return to zero in case of a short grace period
4067 * (or preemption of this task). Exclude CPU-hotplug operations
4068 * to ensure that no offline CPU has callbacks queued.
4070 init_completion(&rsp->barrier_completion);
4071 atomic_set(&rsp->barrier_cpu_count, 1);
4075 * Force each CPU with callbacks to register a new callback.
4076 * When that callback is invoked, we will know that all of the
4077 * corresponding CPU's preceding callbacks have been invoked.
4079 for_each_possible_cpu(cpu) {
4080 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
4082 rdp = per_cpu_ptr(rsp->rda, cpu);
4083 if (rcu_is_nocb_cpu(cpu)) {
4084 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
4085 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
4086 rsp->barrier_sequence);
4088 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
4089 rsp->barrier_sequence);
4090 smp_mb__before_atomic();
4091 atomic_inc(&rsp->barrier_cpu_count);
4092 __call_rcu(&rdp->barrier_head,
4093 rcu_barrier_callback, rsp, cpu, 0);
4095 } else if (READ_ONCE(rdp->qlen)) {
4096 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
4097 rsp->barrier_sequence);
4098 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
4100 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
4101 rsp->barrier_sequence);
4107 * Now that we have an rcu_barrier_callback() callback on each
4108 * CPU, and thus each counted, remove the initial count.
4110 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
4111 complete(&rsp->barrier_completion);
4113 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4114 wait_for_completion(&rsp->barrier_completion);
4116 /* Mark the end of the barrier operation. */
4117 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
4118 rcu_seq_end(&rsp->barrier_sequence);
4120 /* Other rcu_barrier() invocations can now safely proceed. */
4121 mutex_unlock(&rsp->barrier_mutex);
4125 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4127 void rcu_barrier_bh(void)
4129 _rcu_barrier(&rcu_bh_state);
4131 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
4134 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4136 void rcu_barrier_sched(void)
4138 _rcu_barrier(&rcu_sched_state);
4140 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
4143 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4144 * first CPU in a given leaf rcu_node structure coming online. The caller
4145 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4148 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4151 struct rcu_node *rnp = rnp_leaf;
4154 mask = rnp->grpmask;
4158 raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */
4159 rnp->qsmaskinit |= mask;
4160 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
4165 * Do boot-time initialization of a CPU's per-CPU RCU data.
4168 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
4170 unsigned long flags;
4171 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4172 struct rcu_node *rnp = rcu_get_root(rsp);
4174 /* Set up local state, ensuring consistent view of global state. */
4175 raw_spin_lock_irqsave(&rnp->lock, flags);
4176 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
4177 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
4178 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
4179 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
4182 mutex_init(&rdp->exp_funnel_mutex);
4183 rcu_boot_init_nocb_percpu_data(rdp);
4184 raw_spin_unlock_irqrestore(&rnp->lock, flags);
4188 * Initialize a CPU's per-CPU RCU data. Note that only one online or
4189 * offline event can be happening at a given time. Note also that we
4190 * can accept some slop in the rsp->completed access due to the fact
4191 * that this CPU cannot possibly have any RCU callbacks in flight yet.
4194 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
4196 unsigned long flags;
4198 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4199 struct rcu_node *rnp = rcu_get_root(rsp);
4201 /* Set up local state, ensuring consistent view of global state. */
4202 raw_spin_lock_irqsave(&rnp->lock, flags);
4203 rdp->qlen_last_fqs_check = 0;
4204 rdp->n_force_qs_snap = rsp->n_force_qs;
4205 rdp->blimit = blimit;
4207 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
4208 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
4209 rcu_sysidle_init_percpu_data(rdp->dynticks);
4210 atomic_set(&rdp->dynticks->dynticks,
4211 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
4212 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
4215 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4216 * propagation up the rcu_node tree will happen at the beginning
4217 * of the next grace period.
4220 mask = rdp->grpmask;
4221 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
4222 smp_mb__after_unlock_lock();
4223 rnp->qsmaskinitnext |= mask;
4224 rnp->expmaskinitnext |= mask;
4225 if (!rdp->beenonline)
4226 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
4227 rdp->beenonline = true; /* We have now been online. */
4228 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
4229 rdp->completed = rnp->completed;
4230 rdp->cpu_no_qs.b.norm = true;
4231 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
4232 rdp->core_needs_qs = false;
4233 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
4234 raw_spin_unlock_irqrestore(&rnp->lock, flags);
4237 static void rcu_prepare_cpu(int cpu)
4239 struct rcu_state *rsp;
4241 for_each_rcu_flavor(rsp)
4242 rcu_init_percpu_data(cpu, rsp);
4246 * Handle CPU online/offline notification events.
4248 int rcu_cpu_notify(struct notifier_block *self,
4249 unsigned long action, void *hcpu)
4251 long cpu = (long)hcpu;
4252 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
4253 struct rcu_node *rnp = rdp->mynode;
4254 struct rcu_state *rsp;
4257 case CPU_UP_PREPARE:
4258 case CPU_UP_PREPARE_FROZEN:
4259 rcu_prepare_cpu(cpu);
4260 rcu_prepare_kthreads(cpu);
4261 rcu_spawn_all_nocb_kthreads(cpu);
4264 case CPU_DOWN_FAILED:
4265 sync_sched_exp_online_cleanup(cpu);
4266 rcu_boost_kthread_setaffinity(rnp, -1);
4268 case CPU_DOWN_PREPARE:
4269 rcu_boost_kthread_setaffinity(rnp, cpu);
4272 case CPU_DYING_FROZEN:
4273 for_each_rcu_flavor(rsp)
4274 rcu_cleanup_dying_cpu(rsp);
4276 case CPU_DYING_IDLE:
4277 /* QS for any half-done expedited RCU-sched GP. */
4279 rcu_report_exp_rdp(&rcu_sched_state,
4280 this_cpu_ptr(rcu_sched_state.rda), true);
4283 for_each_rcu_flavor(rsp) {
4284 rcu_cleanup_dying_idle_cpu(cpu, rsp);
4288 case CPU_DEAD_FROZEN:
4289 case CPU_UP_CANCELED:
4290 case CPU_UP_CANCELED_FROZEN:
4291 for_each_rcu_flavor(rsp) {
4292 rcu_cleanup_dead_cpu(cpu, rsp);
4293 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
4302 static int rcu_pm_notify(struct notifier_block *self,
4303 unsigned long action, void *hcpu)
4306 case PM_HIBERNATION_PREPARE:
4307 case PM_SUSPEND_PREPARE:
4308 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4311 case PM_POST_HIBERNATION:
4312 case PM_POST_SUSPEND:
4313 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4314 rcu_unexpedite_gp();
4323 * Spawn the kthreads that handle each RCU flavor's grace periods.
4325 static int __init rcu_spawn_gp_kthread(void)
4327 unsigned long flags;
4328 int kthread_prio_in = kthread_prio;
4329 struct rcu_node *rnp;
4330 struct rcu_state *rsp;
4331 struct sched_param sp;
4332 struct task_struct *t;
4334 /* Force priority into range. */
4335 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4337 else if (kthread_prio < 0)
4339 else if (kthread_prio > 99)
4341 if (kthread_prio != kthread_prio_in)
4342 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4343 kthread_prio, kthread_prio_in);
4345 rcu_scheduler_fully_active = 1;
4346 for_each_rcu_flavor(rsp) {
4347 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4349 rnp = rcu_get_root(rsp);
4350 raw_spin_lock_irqsave(&rnp->lock, flags);
4351 rsp->gp_kthread = t;
4353 sp.sched_priority = kthread_prio;
4354 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4357 raw_spin_unlock_irqrestore(&rnp->lock, flags);
4359 rcu_spawn_nocb_kthreads();
4360 rcu_spawn_boost_kthreads();
4363 early_initcall(rcu_spawn_gp_kthread);
4366 * This function is invoked towards the end of the scheduler's initialization
4367 * process. Before this is called, the idle task might contain
4368 * RCU read-side critical sections (during which time, this idle
4369 * task is booting the system). After this function is called, the
4370 * idle tasks are prohibited from containing RCU read-side critical
4371 * sections. This function also enables RCU lockdep checking.
4373 void rcu_scheduler_starting(void)
4375 WARN_ON(num_online_cpus() != 1);
4376 WARN_ON(nr_context_switches() > 0);
4377 rcu_scheduler_active = 1;
4381 * Compute the per-level fanout, either using the exact fanout specified
4382 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4384 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4388 if (rcu_fanout_exact) {
4389 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4390 for (i = rcu_num_lvls - 2; i >= 0; i--)
4391 levelspread[i] = RCU_FANOUT;
4397 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4399 levelspread[i] = (cprv + ccur - 1) / ccur;
4406 * Helper function for rcu_init() that initializes one rcu_state structure.
4408 static void __init rcu_init_one(struct rcu_state *rsp,
4409 struct rcu_data __percpu *rda)
4411 static const char * const buf[] = RCU_NODE_NAME_INIT;
4412 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4413 static const char * const exp[] = RCU_EXP_NAME_INIT;
4414 static u8 fl_mask = 0x1;
4416 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4417 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4421 struct rcu_node *rnp;
4423 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4425 /* Silence gcc 4.8 false positive about array index out of range. */
4426 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4427 panic("rcu_init_one: rcu_num_lvls out of range");
4429 /* Initialize the level-tracking arrays. */
4431 for (i = 0; i < rcu_num_lvls; i++)
4432 levelcnt[i] = num_rcu_lvl[i];
4433 for (i = 1; i < rcu_num_lvls; i++)
4434 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4435 rcu_init_levelspread(levelspread, levelcnt);
4436 rsp->flavor_mask = fl_mask;
4439 /* Initialize the elements themselves, starting from the leaves. */
4441 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4442 cpustride *= levelspread[i];
4443 rnp = rsp->level[i];
4444 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4445 raw_spin_lock_init(&rnp->lock);
4446 lockdep_set_class_and_name(&rnp->lock,
4447 &rcu_node_class[i], buf[i]);
4448 raw_spin_lock_init(&rnp->fqslock);
4449 lockdep_set_class_and_name(&rnp->fqslock,
4450 &rcu_fqs_class[i], fqs[i]);
4451 rnp->gpnum = rsp->gpnum;
4452 rnp->completed = rsp->completed;
4454 rnp->qsmaskinit = 0;
4455 rnp->grplo = j * cpustride;
4456 rnp->grphi = (j + 1) * cpustride - 1;
4457 if (rnp->grphi >= nr_cpu_ids)
4458 rnp->grphi = nr_cpu_ids - 1;
4464 rnp->grpnum = j % levelspread[i - 1];
4465 rnp->grpmask = 1UL << rnp->grpnum;
4466 rnp->parent = rsp->level[i - 1] +
4467 j / levelspread[i - 1];
4470 INIT_LIST_HEAD(&rnp->blkd_tasks);
4471 rcu_init_one_nocb(rnp);
4472 mutex_init(&rnp->exp_funnel_mutex);
4473 lockdep_set_class_and_name(&rnp->exp_funnel_mutex,
4474 &rcu_exp_class[i], exp[i]);
4478 init_waitqueue_head(&rsp->gp_wq);
4479 init_waitqueue_head(&rsp->expedited_wq);
4480 rnp = rsp->level[rcu_num_lvls - 1];
4481 for_each_possible_cpu(i) {
4482 while (i > rnp->grphi)
4484 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4485 rcu_boot_init_percpu_data(i, rsp);
4487 list_add(&rsp->flavors, &rcu_struct_flavors);
4491 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4492 * replace the definitions in tree.h because those are needed to size
4493 * the ->node array in the rcu_state structure.
4495 static void __init rcu_init_geometry(void)
4499 int rcu_capacity[RCU_NUM_LVLS];
4502 * Initialize any unspecified boot parameters.
4503 * The default values of jiffies_till_first_fqs and
4504 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4505 * value, which is a function of HZ, then adding one for each
4506 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4508 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4509 if (jiffies_till_first_fqs == ULONG_MAX)
4510 jiffies_till_first_fqs = d;
4511 if (jiffies_till_next_fqs == ULONG_MAX)
4512 jiffies_till_next_fqs = d;
4514 /* If the compile-time values are accurate, just leave. */
4515 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4516 nr_cpu_ids == NR_CPUS)
4518 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4519 rcu_fanout_leaf, nr_cpu_ids);
4522 * The boot-time rcu_fanout_leaf parameter must be at least two
4523 * and cannot exceed the number of bits in the rcu_node masks.
4524 * Complain and fall back to the compile-time values if this
4525 * limit is exceeded.
4527 if (rcu_fanout_leaf < 2 ||
4528 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4529 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4535 * Compute number of nodes that can be handled an rcu_node tree
4536 * with the given number of levels.
4538 rcu_capacity[0] = rcu_fanout_leaf;
4539 for (i = 1; i < RCU_NUM_LVLS; i++)
4540 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4543 * The tree must be able to accommodate the configured number of CPUs.
4544 * If this limit is exceeded, fall back to the compile-time values.
4546 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4547 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4552 /* Calculate the number of levels in the tree. */
4553 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4555 rcu_num_lvls = i + 1;
4557 /* Calculate the number of rcu_nodes at each level of the tree. */
4558 for (i = 0; i < rcu_num_lvls; i++) {
4559 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4560 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4563 /* Calculate the total number of rcu_node structures. */
4565 for (i = 0; i < rcu_num_lvls; i++)
4566 rcu_num_nodes += num_rcu_lvl[i];
4570 * Dump out the structure of the rcu_node combining tree associated
4571 * with the rcu_state structure referenced by rsp.
4573 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4576 struct rcu_node *rnp;
4578 pr_info("rcu_node tree layout dump\n");
4580 rcu_for_each_node_breadth_first(rsp, rnp) {
4581 if (rnp->level != level) {
4586 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4591 void __init rcu_init(void)
4595 rcu_early_boot_tests();
4597 rcu_bootup_announce();
4598 rcu_init_geometry();
4599 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
4600 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
4602 rcu_dump_rcu_node_tree(&rcu_sched_state);
4603 __rcu_init_preempt();
4604 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4607 * We don't need protection against CPU-hotplug here because
4608 * this is called early in boot, before either interrupts
4609 * or the scheduler are operational.
4611 cpu_notifier(rcu_cpu_notify, 0);
4612 pm_notifier(rcu_pm_notify, 0);
4613 for_each_online_cpu(cpu)
4614 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
4617 #include "tree_plugin.h"