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 .fqs_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);
164 /* rcuc/rcub kthread realtime priority */
165 #ifdef CONFIG_RCU_KTHREAD_PRIO
166 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
167 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
168 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
169 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
170 module_param(kthread_prio, int, 0644);
172 /* Delay in jiffies for grace-period initialization delays, debug only. */
174 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
175 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
176 module_param(gp_preinit_delay, int, 0644);
177 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
178 static const int gp_preinit_delay;
179 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
181 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
182 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
183 module_param(gp_init_delay, int, 0644);
184 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
185 static const int gp_init_delay;
186 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
188 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
189 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
190 module_param(gp_cleanup_delay, int, 0644);
191 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
192 static const int gp_cleanup_delay;
193 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
196 * Number of grace periods between delays, normalized by the duration of
197 * the delay. The longer the the delay, the more the grace periods between
198 * each delay. The reason for this normalization is that it means that,
199 * for non-zero delays, the overall slowdown of grace periods is constant
200 * regardless of the duration of the delay. This arrangement balances
201 * the need for long delays to increase some race probabilities with the
202 * need for fast grace periods to increase other race probabilities.
204 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
207 * Track the rcutorture test sequence number and the update version
208 * number within a given test. The rcutorture_testseq is incremented
209 * on every rcutorture module load and unload, so has an odd value
210 * when a test is running. The rcutorture_vernum is set to zero
211 * when rcutorture starts and is incremented on each rcutorture update.
212 * These variables enable correlating rcutorture output with the
213 * RCU tracing information.
215 unsigned long rcutorture_testseq;
216 unsigned long rcutorture_vernum;
219 * Compute the mask of online CPUs for the specified rcu_node structure.
220 * This will not be stable unless the rcu_node structure's ->lock is
221 * held, but the bit corresponding to the current CPU will be stable
224 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
226 return READ_ONCE(rnp->qsmaskinitnext);
230 * Return true if an RCU grace period is in progress. The READ_ONCE()s
231 * permit this function to be invoked without holding the root rcu_node
232 * structure's ->lock, but of course results can be subject to change.
234 static int rcu_gp_in_progress(struct rcu_state *rsp)
236 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
240 * Note a quiescent state. Because we do not need to know
241 * how many quiescent states passed, just if there was at least
242 * one since the start of the grace period, this just sets a flag.
243 * The caller must have disabled preemption.
245 void rcu_sched_qs(void)
247 if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) {
248 trace_rcu_grace_period(TPS("rcu_sched"),
249 __this_cpu_read(rcu_sched_data.gpnum),
251 __this_cpu_write(rcu_sched_data.passed_quiesce, 1);
257 if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) {
258 trace_rcu_grace_period(TPS("rcu_bh"),
259 __this_cpu_read(rcu_bh_data.gpnum),
261 __this_cpu_write(rcu_bh_data.passed_quiesce, 1);
265 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
267 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
268 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
269 .dynticks = ATOMIC_INIT(1),
270 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
271 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
272 .dynticks_idle = ATOMIC_INIT(1),
273 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
276 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
277 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
280 * Let the RCU core know that this CPU has gone through the scheduler,
281 * which is a quiescent state. This is called when the need for a
282 * quiescent state is urgent, so we burn an atomic operation and full
283 * memory barriers to let the RCU core know about it, regardless of what
284 * this CPU might (or might not) do in the near future.
286 * We inform the RCU core by emulating a zero-duration dyntick-idle
287 * period, which we in turn do by incrementing the ->dynticks counter
290 static void rcu_momentary_dyntick_idle(void)
293 struct rcu_data *rdp;
294 struct rcu_dynticks *rdtp;
296 struct rcu_state *rsp;
298 local_irq_save(flags);
301 * Yes, we can lose flag-setting operations. This is OK, because
302 * the flag will be set again after some delay.
304 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
305 raw_cpu_write(rcu_sched_qs_mask, 0);
307 /* Find the flavor that needs a quiescent state. */
308 for_each_rcu_flavor(rsp) {
309 rdp = raw_cpu_ptr(rsp->rda);
310 if (!(resched_mask & rsp->flavor_mask))
312 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
313 if (READ_ONCE(rdp->mynode->completed) !=
314 READ_ONCE(rdp->cond_resched_completed))
318 * Pretend to be momentarily idle for the quiescent state.
319 * This allows the grace-period kthread to record the
320 * quiescent state, with no need for this CPU to do anything
323 rdtp = this_cpu_ptr(&rcu_dynticks);
324 smp_mb__before_atomic(); /* Earlier stuff before QS. */
325 atomic_add(2, &rdtp->dynticks); /* QS. */
326 smp_mb__after_atomic(); /* Later stuff after QS. */
329 local_irq_restore(flags);
333 * Note a context switch. This is a quiescent state for RCU-sched,
334 * and requires special handling for preemptible RCU.
335 * The caller must have disabled preemption.
337 void rcu_note_context_switch(void)
339 trace_rcu_utilization(TPS("Start context switch"));
341 rcu_preempt_note_context_switch();
342 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
343 rcu_momentary_dyntick_idle();
344 trace_rcu_utilization(TPS("End context switch"));
346 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
349 * Register a quiescent state for all RCU flavors. If there is an
350 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
351 * dyntick-idle quiescent state visible to other CPUs (but only for those
352 * RCU flavors in desperate need of a quiescent state, which will normally
353 * be none of them). Either way, do a lightweight quiescent state for
356 void rcu_all_qs(void)
358 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
359 rcu_momentary_dyntick_idle();
360 this_cpu_inc(rcu_qs_ctr);
362 EXPORT_SYMBOL_GPL(rcu_all_qs);
364 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
365 static long qhimark = 10000; /* If this many pending, ignore blimit. */
366 static long qlowmark = 100; /* Once only this many pending, use blimit. */
368 module_param(blimit, long, 0444);
369 module_param(qhimark, long, 0444);
370 module_param(qlowmark, long, 0444);
372 static ulong jiffies_till_first_fqs = ULONG_MAX;
373 static ulong jiffies_till_next_fqs = ULONG_MAX;
375 module_param(jiffies_till_first_fqs, ulong, 0644);
376 module_param(jiffies_till_next_fqs, ulong, 0644);
379 * How long the grace period must be before we start recruiting
380 * quiescent-state help from rcu_note_context_switch().
382 static ulong jiffies_till_sched_qs = HZ / 20;
383 module_param(jiffies_till_sched_qs, ulong, 0644);
385 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
386 struct rcu_data *rdp);
387 static void force_qs_rnp(struct rcu_state *rsp,
388 int (*f)(struct rcu_data *rsp, bool *isidle,
389 unsigned long *maxj),
390 bool *isidle, unsigned long *maxj);
391 static void force_quiescent_state(struct rcu_state *rsp);
392 static int rcu_pending(void);
395 * Return the number of RCU batches started thus far for debug & stats.
397 unsigned long rcu_batches_started(void)
399 return rcu_state_p->gpnum;
401 EXPORT_SYMBOL_GPL(rcu_batches_started);
404 * Return the number of RCU-sched batches started thus far for debug & stats.
406 unsigned long rcu_batches_started_sched(void)
408 return rcu_sched_state.gpnum;
410 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
413 * Return the number of RCU BH batches started thus far for debug & stats.
415 unsigned long rcu_batches_started_bh(void)
417 return rcu_bh_state.gpnum;
419 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
422 * Return the number of RCU batches completed thus far for debug & stats.
424 unsigned long rcu_batches_completed(void)
426 return rcu_state_p->completed;
428 EXPORT_SYMBOL_GPL(rcu_batches_completed);
431 * Return the number of RCU-sched batches completed thus far for debug & stats.
433 unsigned long rcu_batches_completed_sched(void)
435 return rcu_sched_state.completed;
437 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
440 * Return the number of RCU BH batches completed thus far for debug & stats.
442 unsigned long rcu_batches_completed_bh(void)
444 return rcu_bh_state.completed;
446 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
449 * Force a quiescent state.
451 void rcu_force_quiescent_state(void)
453 force_quiescent_state(rcu_state_p);
455 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
458 * Force a quiescent state for RCU BH.
460 void rcu_bh_force_quiescent_state(void)
462 force_quiescent_state(&rcu_bh_state);
464 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
467 * Force a quiescent state for RCU-sched.
469 void rcu_sched_force_quiescent_state(void)
471 force_quiescent_state(&rcu_sched_state);
473 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
476 * Show the state of the grace-period kthreads.
478 void show_rcu_gp_kthreads(void)
480 struct rcu_state *rsp;
482 for_each_rcu_flavor(rsp) {
483 pr_info("%s: wait state: %d ->state: %#lx\n",
484 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
485 /* sched_show_task(rsp->gp_kthread); */
488 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
491 * Record the number of times rcutorture tests have been initiated and
492 * terminated. This information allows the debugfs tracing stats to be
493 * correlated to the rcutorture messages, even when the rcutorture module
494 * is being repeatedly loaded and unloaded. In other words, we cannot
495 * store this state in rcutorture itself.
497 void rcutorture_record_test_transition(void)
499 rcutorture_testseq++;
500 rcutorture_vernum = 0;
502 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
505 * Send along grace-period-related data for rcutorture diagnostics.
507 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
508 unsigned long *gpnum, unsigned long *completed)
510 struct rcu_state *rsp = NULL;
519 case RCU_SCHED_FLAVOR:
520 rsp = &rcu_sched_state;
526 *flags = READ_ONCE(rsp->gp_flags);
527 *gpnum = READ_ONCE(rsp->gpnum);
528 *completed = READ_ONCE(rsp->completed);
535 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
538 * Record the number of writer passes through the current rcutorture test.
539 * This is also used to correlate debugfs tracing stats with the rcutorture
542 void rcutorture_record_progress(unsigned long vernum)
546 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
549 * Does the CPU have callbacks ready to be invoked?
552 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
554 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
555 rdp->nxttail[RCU_DONE_TAIL] != NULL;
559 * Return the root node of the specified rcu_state structure.
561 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
563 return &rsp->node[0];
567 * Is there any need for future grace periods?
568 * Interrupts must be disabled. If the caller does not hold the root
569 * rnp_node structure's ->lock, the results are advisory only.
571 static int rcu_future_needs_gp(struct rcu_state *rsp)
573 struct rcu_node *rnp = rcu_get_root(rsp);
574 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
575 int *fp = &rnp->need_future_gp[idx];
577 return READ_ONCE(*fp);
581 * Does the current CPU require a not-yet-started grace period?
582 * The caller must have disabled interrupts to prevent races with
583 * normal callback registry.
586 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
590 if (rcu_gp_in_progress(rsp))
591 return 0; /* No, a grace period is already in progress. */
592 if (rcu_future_needs_gp(rsp))
593 return 1; /* Yes, a no-CBs CPU needs one. */
594 if (!rdp->nxttail[RCU_NEXT_TAIL])
595 return 0; /* No, this is a no-CBs (or offline) CPU. */
596 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
597 return 1; /* Yes, this CPU has newly registered callbacks. */
598 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
599 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
600 ULONG_CMP_LT(READ_ONCE(rsp->completed),
601 rdp->nxtcompleted[i]))
602 return 1; /* Yes, CBs for future grace period. */
603 return 0; /* No grace period needed. */
607 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
609 * If the new value of the ->dynticks_nesting counter now is zero,
610 * we really have entered idle, and must do the appropriate accounting.
611 * The caller must have disabled interrupts.
613 static void rcu_eqs_enter_common(long long oldval, bool user)
615 struct rcu_state *rsp;
616 struct rcu_data *rdp;
617 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
619 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
620 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
621 !user && !is_idle_task(current)) {
622 struct task_struct *idle __maybe_unused =
623 idle_task(smp_processor_id());
625 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
626 ftrace_dump(DUMP_ORIG);
627 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
628 current->pid, current->comm,
629 idle->pid, idle->comm); /* must be idle task! */
631 for_each_rcu_flavor(rsp) {
632 rdp = this_cpu_ptr(rsp->rda);
633 do_nocb_deferred_wakeup(rdp);
635 rcu_prepare_for_idle();
636 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
637 smp_mb__before_atomic(); /* See above. */
638 atomic_inc(&rdtp->dynticks);
639 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
640 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
641 atomic_read(&rdtp->dynticks) & 0x1);
642 rcu_dynticks_task_enter();
645 * It is illegal to enter an extended quiescent state while
646 * in an RCU read-side critical section.
648 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
649 "Illegal idle entry in RCU read-side critical section.");
650 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
651 "Illegal idle entry in RCU-bh read-side critical section.");
652 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
653 "Illegal idle entry in RCU-sched read-side critical section.");
657 * Enter an RCU extended quiescent state, which can be either the
658 * idle loop or adaptive-tickless usermode execution.
660 static void rcu_eqs_enter(bool user)
663 struct rcu_dynticks *rdtp;
665 rdtp = this_cpu_ptr(&rcu_dynticks);
666 oldval = rdtp->dynticks_nesting;
667 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
668 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
669 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
670 rdtp->dynticks_nesting = 0;
671 rcu_eqs_enter_common(oldval, user);
673 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
678 * rcu_idle_enter - inform RCU that current CPU is entering idle
680 * Enter idle mode, in other words, -leave- the mode in which RCU
681 * read-side critical sections can occur. (Though RCU read-side
682 * critical sections can occur in irq handlers in idle, a possibility
683 * handled by irq_enter() and irq_exit().)
685 * We crowbar the ->dynticks_nesting field to zero to allow for
686 * the possibility of usermode upcalls having messed up our count
687 * of interrupt nesting level during the prior busy period.
689 void rcu_idle_enter(void)
693 local_irq_save(flags);
694 rcu_eqs_enter(false);
695 rcu_sysidle_enter(0);
696 local_irq_restore(flags);
698 EXPORT_SYMBOL_GPL(rcu_idle_enter);
700 #ifdef CONFIG_RCU_USER_QS
702 * rcu_user_enter - inform RCU that we are resuming userspace.
704 * Enter RCU idle mode right before resuming userspace. No use of RCU
705 * is permitted between this call and rcu_user_exit(). This way the
706 * CPU doesn't need to maintain the tick for RCU maintenance purposes
707 * when the CPU runs in userspace.
709 void rcu_user_enter(void)
713 #endif /* CONFIG_RCU_USER_QS */
716 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
718 * Exit from an interrupt handler, which might possibly result in entering
719 * idle mode, in other words, leaving the mode in which read-side critical
720 * sections can occur.
722 * This code assumes that the idle loop never does anything that might
723 * result in unbalanced calls to irq_enter() and irq_exit(). If your
724 * architecture violates this assumption, RCU will give you what you
725 * deserve, good and hard. But very infrequently and irreproducibly.
727 * Use things like work queues to work around this limitation.
729 * You have been warned.
731 void rcu_irq_exit(void)
735 struct rcu_dynticks *rdtp;
737 local_irq_save(flags);
738 rdtp = this_cpu_ptr(&rcu_dynticks);
739 oldval = rdtp->dynticks_nesting;
740 rdtp->dynticks_nesting--;
741 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
742 rdtp->dynticks_nesting < 0);
743 if (rdtp->dynticks_nesting)
744 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
746 rcu_eqs_enter_common(oldval, true);
747 rcu_sysidle_enter(1);
748 local_irq_restore(flags);
752 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
754 * If the new value of the ->dynticks_nesting counter was previously zero,
755 * we really have exited idle, and must do the appropriate accounting.
756 * The caller must have disabled interrupts.
758 static void rcu_eqs_exit_common(long long oldval, int user)
760 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
762 rcu_dynticks_task_exit();
763 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
764 atomic_inc(&rdtp->dynticks);
765 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
766 smp_mb__after_atomic(); /* See above. */
767 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
768 !(atomic_read(&rdtp->dynticks) & 0x1));
769 rcu_cleanup_after_idle();
770 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
771 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
772 !user && !is_idle_task(current)) {
773 struct task_struct *idle __maybe_unused =
774 idle_task(smp_processor_id());
776 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
777 oldval, rdtp->dynticks_nesting);
778 ftrace_dump(DUMP_ORIG);
779 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
780 current->pid, current->comm,
781 idle->pid, idle->comm); /* must be idle task! */
786 * Exit an RCU extended quiescent state, which can be either the
787 * idle loop or adaptive-tickless usermode execution.
789 static void rcu_eqs_exit(bool user)
791 struct rcu_dynticks *rdtp;
794 rdtp = this_cpu_ptr(&rcu_dynticks);
795 oldval = rdtp->dynticks_nesting;
796 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
797 if (oldval & DYNTICK_TASK_NEST_MASK) {
798 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
800 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
801 rcu_eqs_exit_common(oldval, user);
806 * rcu_idle_exit - inform RCU that current CPU is leaving idle
808 * Exit idle mode, in other words, -enter- the mode in which RCU
809 * read-side critical sections can occur.
811 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
812 * allow for the possibility of usermode upcalls messing up our count
813 * of interrupt nesting level during the busy period that is just
816 void rcu_idle_exit(void)
820 local_irq_save(flags);
823 local_irq_restore(flags);
825 EXPORT_SYMBOL_GPL(rcu_idle_exit);
827 #ifdef CONFIG_RCU_USER_QS
829 * rcu_user_exit - inform RCU that we are exiting userspace.
831 * Exit RCU idle mode while entering the kernel because it can
832 * run a RCU read side critical section anytime.
834 void rcu_user_exit(void)
838 #endif /* CONFIG_RCU_USER_QS */
841 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
843 * Enter an interrupt handler, which might possibly result in exiting
844 * idle mode, in other words, entering the mode in which read-side critical
845 * sections can occur.
847 * Note that the Linux kernel is fully capable of entering an interrupt
848 * handler that it never exits, for example when doing upcalls to
849 * user mode! This code assumes that the idle loop never does upcalls to
850 * user mode. If your architecture does do upcalls from the idle loop (or
851 * does anything else that results in unbalanced calls to the irq_enter()
852 * and irq_exit() functions), RCU will give you what you deserve, good
853 * and hard. But very infrequently and irreproducibly.
855 * Use things like work queues to work around this limitation.
857 * You have been warned.
859 void rcu_irq_enter(void)
862 struct rcu_dynticks *rdtp;
865 local_irq_save(flags);
866 rdtp = this_cpu_ptr(&rcu_dynticks);
867 oldval = rdtp->dynticks_nesting;
868 rdtp->dynticks_nesting++;
869 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
870 rdtp->dynticks_nesting == 0);
872 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
874 rcu_eqs_exit_common(oldval, true);
876 local_irq_restore(flags);
880 * rcu_nmi_enter - inform RCU of entry to NMI context
882 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
883 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
884 * that the CPU is active. This implementation permits nested NMIs, as
885 * long as the nesting level does not overflow an int. (You will probably
886 * run out of stack space first.)
888 void rcu_nmi_enter(void)
890 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
893 /* Complain about underflow. */
894 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
897 * If idle from RCU viewpoint, atomically increment ->dynticks
898 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
899 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
900 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
901 * to be in the outermost NMI handler that interrupted an RCU-idle
902 * period (observation due to Andy Lutomirski).
904 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
905 smp_mb__before_atomic(); /* Force delay from prior write. */
906 atomic_inc(&rdtp->dynticks);
907 /* atomic_inc() before later RCU read-side crit sects */
908 smp_mb__after_atomic(); /* See above. */
909 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
912 rdtp->dynticks_nmi_nesting += incby;
917 * rcu_nmi_exit - inform RCU of exit from NMI context
919 * If we are returning from the outermost NMI handler that interrupted an
920 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
921 * to let the RCU grace-period handling know that the CPU is back to
924 void rcu_nmi_exit(void)
926 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
929 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
930 * (We are exiting an NMI handler, so RCU better be paying attention
933 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
934 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
937 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
938 * leave it in non-RCU-idle state.
940 if (rdtp->dynticks_nmi_nesting != 1) {
941 rdtp->dynticks_nmi_nesting -= 2;
945 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
946 rdtp->dynticks_nmi_nesting = 0;
947 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
948 smp_mb__before_atomic(); /* See above. */
949 atomic_inc(&rdtp->dynticks);
950 smp_mb__after_atomic(); /* Force delay to next write. */
951 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
955 * __rcu_is_watching - are RCU read-side critical sections safe?
957 * Return true if RCU is watching the running CPU, which means that
958 * this CPU can safely enter RCU read-side critical sections. Unlike
959 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
960 * least disabled preemption.
962 bool notrace __rcu_is_watching(void)
964 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
968 * rcu_is_watching - see if RCU thinks that the current CPU is idle
970 * If the current CPU is in its idle loop and is neither in an interrupt
971 * or NMI handler, return true.
973 bool notrace rcu_is_watching(void)
978 ret = __rcu_is_watching();
982 EXPORT_SYMBOL_GPL(rcu_is_watching);
984 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
987 * Is the current CPU online? Disable preemption to avoid false positives
988 * that could otherwise happen due to the current CPU number being sampled,
989 * this task being preempted, its old CPU being taken offline, resuming
990 * on some other CPU, then determining that its old CPU is now offline.
991 * It is OK to use RCU on an offline processor during initial boot, hence
992 * the check for rcu_scheduler_fully_active. Note also that it is OK
993 * for a CPU coming online to use RCU for one jiffy prior to marking itself
994 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
995 * offline to continue to use RCU for one jiffy after marking itself
996 * offline in the cpu_online_mask. This leniency is necessary given the
997 * non-atomic nature of the online and offline processing, for example,
998 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1001 * This is also why RCU internally marks CPUs online during the
1002 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1004 * Disable checking if in an NMI handler because we cannot safely report
1005 * errors from NMI handlers anyway.
1007 bool rcu_lockdep_current_cpu_online(void)
1009 struct rcu_data *rdp;
1010 struct rcu_node *rnp;
1016 rdp = this_cpu_ptr(&rcu_sched_data);
1018 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1019 !rcu_scheduler_fully_active;
1023 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1025 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1028 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1030 * If the current CPU is idle or running at a first-level (not nested)
1031 * interrupt from idle, return true. The caller must have at least
1032 * disabled preemption.
1034 static int rcu_is_cpu_rrupt_from_idle(void)
1036 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1040 * Snapshot the specified CPU's dynticks counter so that we can later
1041 * credit them with an implicit quiescent state. Return 1 if this CPU
1042 * is in dynticks idle mode, which is an extended quiescent state.
1044 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1045 bool *isidle, unsigned long *maxj)
1047 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1048 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1049 if ((rdp->dynticks_snap & 0x1) == 0) {
1050 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1053 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1054 rdp->mynode->gpnum))
1055 WRITE_ONCE(rdp->gpwrap, true);
1061 * Return true if the specified CPU has passed through a quiescent
1062 * state by virtue of being in or having passed through an dynticks
1063 * idle state since the last call to dyntick_save_progress_counter()
1064 * for this same CPU, or by virtue of having been offline.
1066 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1067 bool *isidle, unsigned long *maxj)
1073 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1074 snap = (unsigned int)rdp->dynticks_snap;
1077 * If the CPU passed through or entered a dynticks idle phase with
1078 * no active irq/NMI handlers, then we can safely pretend that the CPU
1079 * already acknowledged the request to pass through a quiescent
1080 * state. Either way, that CPU cannot possibly be in an RCU
1081 * read-side critical section that started before the beginning
1082 * of the current RCU grace period.
1084 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1085 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1086 rdp->dynticks_fqs++;
1091 * Check for the CPU being offline, but only if the grace period
1092 * is old enough. We don't need to worry about the CPU changing
1093 * state: If we see it offline even once, it has been through a
1096 * The reason for insisting that the grace period be at least
1097 * one jiffy old is that CPUs that are not quite online and that
1098 * have just gone offline can still execute RCU read-side critical
1101 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1102 return 0; /* Grace period is not old enough. */
1104 if (cpu_is_offline(rdp->cpu)) {
1105 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1111 * A CPU running for an extended time within the kernel can
1112 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1113 * even context-switching back and forth between a pair of
1114 * in-kernel CPU-bound tasks cannot advance grace periods.
1115 * So if the grace period is old enough, make the CPU pay attention.
1116 * Note that the unsynchronized assignments to the per-CPU
1117 * rcu_sched_qs_mask variable are safe. Yes, setting of
1118 * bits can be lost, but they will be set again on the next
1119 * force-quiescent-state pass. So lost bit sets do not result
1120 * in incorrect behavior, merely in a grace period lasting
1121 * a few jiffies longer than it might otherwise. Because
1122 * there are at most four threads involved, and because the
1123 * updates are only once every few jiffies, the probability of
1124 * lossage (and thus of slight grace-period extension) is
1127 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1128 * is set too high, we override with half of the RCU CPU stall
1131 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1132 if (ULONG_CMP_GE(jiffies,
1133 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1134 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1135 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1136 WRITE_ONCE(rdp->cond_resched_completed,
1137 READ_ONCE(rdp->mynode->completed));
1138 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1140 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1141 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1142 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
1143 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1144 /* Time to beat on that CPU again! */
1145 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1146 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1153 static void record_gp_stall_check_time(struct rcu_state *rsp)
1155 unsigned long j = jiffies;
1159 smp_wmb(); /* Record start time before stall time. */
1160 j1 = rcu_jiffies_till_stall_check();
1161 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1162 rsp->jiffies_resched = j + j1 / 2;
1163 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1167 * Complain about starvation of grace-period kthread.
1169 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1175 gpa = READ_ONCE(rsp->gp_activity);
1176 if (j - gpa > 2 * HZ)
1177 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n",
1179 rsp->gpnum, rsp->completed,
1180 rsp->gp_flags, rsp->gp_state,
1181 rsp->gp_kthread ? rsp->gp_kthread->state : 0);
1185 * Dump stacks of all tasks running on stalled CPUs.
1187 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1190 unsigned long flags;
1191 struct rcu_node *rnp;
1193 rcu_for_each_leaf_node(rsp, rnp) {
1194 raw_spin_lock_irqsave(&rnp->lock, flags);
1195 if (rnp->qsmask != 0) {
1196 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1197 if (rnp->qsmask & (1UL << cpu))
1198 dump_cpu_task(rnp->grplo + cpu);
1200 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1204 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1208 unsigned long flags;
1212 struct rcu_node *rnp = rcu_get_root(rsp);
1215 /* Only let one CPU complain about others per time interval. */
1217 raw_spin_lock_irqsave(&rnp->lock, flags);
1218 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1219 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1220 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1223 WRITE_ONCE(rsp->jiffies_stall,
1224 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1225 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1228 * OK, time to rat on our buddy...
1229 * See Documentation/RCU/stallwarn.txt for info on how to debug
1230 * RCU CPU stall warnings.
1232 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1234 print_cpu_stall_info_begin();
1235 rcu_for_each_leaf_node(rsp, rnp) {
1236 raw_spin_lock_irqsave(&rnp->lock, flags);
1237 ndetected += rcu_print_task_stall(rnp);
1238 if (rnp->qsmask != 0) {
1239 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1240 if (rnp->qsmask & (1UL << cpu)) {
1241 print_cpu_stall_info(rsp,
1246 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1249 print_cpu_stall_info_end();
1250 for_each_possible_cpu(cpu)
1251 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1252 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1253 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1254 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1256 rcu_dump_cpu_stacks(rsp);
1258 if (READ_ONCE(rsp->gpnum) != gpnum ||
1259 READ_ONCE(rsp->completed) == gpnum) {
1260 pr_err("INFO: Stall ended before state dump start\n");
1263 gpa = READ_ONCE(rsp->gp_activity);
1264 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1265 rsp->name, j - gpa, j, gpa,
1266 jiffies_till_next_fqs,
1267 rcu_get_root(rsp)->qsmask);
1268 /* In this case, the current CPU might be at fault. */
1269 sched_show_task(current);
1273 /* Complain about tasks blocking the grace period. */
1274 rcu_print_detail_task_stall(rsp);
1276 rcu_check_gp_kthread_starvation(rsp);
1278 force_quiescent_state(rsp); /* Kick them all. */
1281 static void print_cpu_stall(struct rcu_state *rsp)
1284 unsigned long flags;
1285 struct rcu_node *rnp = rcu_get_root(rsp);
1289 * OK, time to rat on ourselves...
1290 * See Documentation/RCU/stallwarn.txt for info on how to debug
1291 * RCU CPU stall warnings.
1293 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1294 print_cpu_stall_info_begin();
1295 print_cpu_stall_info(rsp, smp_processor_id());
1296 print_cpu_stall_info_end();
1297 for_each_possible_cpu(cpu)
1298 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1299 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1300 jiffies - rsp->gp_start,
1301 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1303 rcu_check_gp_kthread_starvation(rsp);
1305 rcu_dump_cpu_stacks(rsp);
1307 raw_spin_lock_irqsave(&rnp->lock, flags);
1308 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1309 WRITE_ONCE(rsp->jiffies_stall,
1310 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1311 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1314 * Attempt to revive the RCU machinery by forcing a context switch.
1316 * A context switch would normally allow the RCU state machine to make
1317 * progress and it could be we're stuck in kernel space without context
1318 * switches for an entirely unreasonable amount of time.
1320 resched_cpu(smp_processor_id());
1323 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1325 unsigned long completed;
1326 unsigned long gpnum;
1330 struct rcu_node *rnp;
1332 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1337 * Lots of memory barriers to reject false positives.
1339 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1340 * then rsp->gp_start, and finally rsp->completed. These values
1341 * are updated in the opposite order with memory barriers (or
1342 * equivalent) during grace-period initialization and cleanup.
1343 * Now, a false positive can occur if we get an new value of
1344 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1345 * the memory barriers, the only way that this can happen is if one
1346 * grace period ends and another starts between these two fetches.
1347 * Detect this by comparing rsp->completed with the previous fetch
1350 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1351 * and rsp->gp_start suffice to forestall false positives.
1353 gpnum = READ_ONCE(rsp->gpnum);
1354 smp_rmb(); /* Pick up ->gpnum first... */
1355 js = READ_ONCE(rsp->jiffies_stall);
1356 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1357 gps = READ_ONCE(rsp->gp_start);
1358 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1359 completed = READ_ONCE(rsp->completed);
1360 if (ULONG_CMP_GE(completed, gpnum) ||
1361 ULONG_CMP_LT(j, js) ||
1362 ULONG_CMP_GE(gps, js))
1363 return; /* No stall or GP completed since entering function. */
1365 if (rcu_gp_in_progress(rsp) &&
1366 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1368 /* We haven't checked in, so go dump stack. */
1369 print_cpu_stall(rsp);
1371 } else if (rcu_gp_in_progress(rsp) &&
1372 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1374 /* They had a few time units to dump stack, so complain. */
1375 print_other_cpu_stall(rsp, gpnum);
1380 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1382 * Set the stall-warning timeout way off into the future, thus preventing
1383 * any RCU CPU stall-warning messages from appearing in the current set of
1384 * RCU grace periods.
1386 * The caller must disable hard irqs.
1388 void rcu_cpu_stall_reset(void)
1390 struct rcu_state *rsp;
1392 for_each_rcu_flavor(rsp)
1393 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1397 * Initialize the specified rcu_data structure's default callback list
1398 * to empty. The default callback list is the one that is not used by
1399 * no-callbacks CPUs.
1401 static void init_default_callback_list(struct rcu_data *rdp)
1405 rdp->nxtlist = NULL;
1406 for (i = 0; i < RCU_NEXT_SIZE; i++)
1407 rdp->nxttail[i] = &rdp->nxtlist;
1411 * Initialize the specified rcu_data structure's callback list to empty.
1413 static void init_callback_list(struct rcu_data *rdp)
1415 if (init_nocb_callback_list(rdp))
1417 init_default_callback_list(rdp);
1421 * Determine the value that ->completed will have at the end of the
1422 * next subsequent grace period. This is used to tag callbacks so that
1423 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1424 * been dyntick-idle for an extended period with callbacks under the
1425 * influence of RCU_FAST_NO_HZ.
1427 * The caller must hold rnp->lock with interrupts disabled.
1429 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1430 struct rcu_node *rnp)
1433 * If RCU is idle, we just wait for the next grace period.
1434 * But we can only be sure that RCU is idle if we are looking
1435 * at the root rcu_node structure -- otherwise, a new grace
1436 * period might have started, but just not yet gotten around
1437 * to initializing the current non-root rcu_node structure.
1439 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1440 return rnp->completed + 1;
1443 * Otherwise, wait for a possible partial grace period and
1444 * then the subsequent full grace period.
1446 return rnp->completed + 2;
1450 * Trace-event helper function for rcu_start_future_gp() and
1451 * rcu_nocb_wait_gp().
1453 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1454 unsigned long c, const char *s)
1456 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1457 rnp->completed, c, rnp->level,
1458 rnp->grplo, rnp->grphi, s);
1462 * Start some future grace period, as needed to handle newly arrived
1463 * callbacks. The required future grace periods are recorded in each
1464 * rcu_node structure's ->need_future_gp field. Returns true if there
1465 * is reason to awaken the grace-period kthread.
1467 * The caller must hold the specified rcu_node structure's ->lock.
1469 static bool __maybe_unused
1470 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1471 unsigned long *c_out)
1476 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1479 * Pick up grace-period number for new callbacks. If this
1480 * grace period is already marked as needed, return to the caller.
1482 c = rcu_cbs_completed(rdp->rsp, rnp);
1483 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1484 if (rnp->need_future_gp[c & 0x1]) {
1485 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1490 * If either this rcu_node structure or the root rcu_node structure
1491 * believe that a grace period is in progress, then we must wait
1492 * for the one following, which is in "c". Because our request
1493 * will be noticed at the end of the current grace period, we don't
1494 * need to explicitly start one. We only do the lockless check
1495 * of rnp_root's fields if the current rcu_node structure thinks
1496 * there is no grace period in flight, and because we hold rnp->lock,
1497 * the only possible change is when rnp_root's two fields are
1498 * equal, in which case rnp_root->gpnum might be concurrently
1499 * incremented. But that is OK, as it will just result in our
1500 * doing some extra useless work.
1502 if (rnp->gpnum != rnp->completed ||
1503 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1504 rnp->need_future_gp[c & 0x1]++;
1505 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1510 * There might be no grace period in progress. If we don't already
1511 * hold it, acquire the root rcu_node structure's lock in order to
1512 * start one (if needed).
1514 if (rnp != rnp_root) {
1515 raw_spin_lock(&rnp_root->lock);
1516 smp_mb__after_unlock_lock();
1520 * Get a new grace-period number. If there really is no grace
1521 * period in progress, it will be smaller than the one we obtained
1522 * earlier. Adjust callbacks as needed. Note that even no-CBs
1523 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1525 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1526 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1527 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1528 rdp->nxtcompleted[i] = c;
1531 * If the needed for the required grace period is already
1532 * recorded, trace and leave.
1534 if (rnp_root->need_future_gp[c & 0x1]) {
1535 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1539 /* Record the need for the future grace period. */
1540 rnp_root->need_future_gp[c & 0x1]++;
1542 /* If a grace period is not already in progress, start one. */
1543 if (rnp_root->gpnum != rnp_root->completed) {
1544 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1546 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1547 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1550 if (rnp != rnp_root)
1551 raw_spin_unlock(&rnp_root->lock);
1559 * Clean up any old requests for the just-ended grace period. Also return
1560 * whether any additional grace periods have been requested. Also invoke
1561 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1562 * waiting for this grace period to complete.
1564 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1566 int c = rnp->completed;
1568 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1570 rcu_nocb_gp_cleanup(rsp, rnp);
1571 rnp->need_future_gp[c & 0x1] = 0;
1572 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1573 trace_rcu_future_gp(rnp, rdp, c,
1574 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1579 * Awaken the grace-period kthread for the specified flavor of RCU.
1580 * Don't do a self-awaken, and don't bother awakening when there is
1581 * nothing for the grace-period kthread to do (as in several CPUs
1582 * raced to awaken, and we lost), and finally don't try to awaken
1583 * a kthread that has not yet been created.
1585 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1587 if (current == rsp->gp_kthread ||
1588 !READ_ONCE(rsp->gp_flags) ||
1591 wake_up(&rsp->gp_wq);
1595 * If there is room, assign a ->completed number to any callbacks on
1596 * this CPU that have not already been assigned. Also accelerate any
1597 * callbacks that were previously assigned a ->completed number that has
1598 * since proven to be too conservative, which can happen if callbacks get
1599 * assigned a ->completed number while RCU is idle, but with reference to
1600 * a non-root rcu_node structure. This function is idempotent, so it does
1601 * not hurt to call it repeatedly. Returns an flag saying that we should
1602 * awaken the RCU grace-period kthread.
1604 * The caller must hold rnp->lock with interrupts disabled.
1606 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1607 struct rcu_data *rdp)
1613 /* If the CPU has no callbacks, nothing to do. */
1614 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1618 * Starting from the sublist containing the callbacks most
1619 * recently assigned a ->completed number and working down, find the
1620 * first sublist that is not assignable to an upcoming grace period.
1621 * Such a sublist has something in it (first two tests) and has
1622 * a ->completed number assigned that will complete sooner than
1623 * the ->completed number for newly arrived callbacks (last test).
1625 * The key point is that any later sublist can be assigned the
1626 * same ->completed number as the newly arrived callbacks, which
1627 * means that the callbacks in any of these later sublist can be
1628 * grouped into a single sublist, whether or not they have already
1629 * been assigned a ->completed number.
1631 c = rcu_cbs_completed(rsp, rnp);
1632 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1633 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1634 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1638 * If there are no sublist for unassigned callbacks, leave.
1639 * At the same time, advance "i" one sublist, so that "i" will
1640 * index into the sublist where all the remaining callbacks should
1643 if (++i >= RCU_NEXT_TAIL)
1647 * Assign all subsequent callbacks' ->completed number to the next
1648 * full grace period and group them all in the sublist initially
1651 for (; i <= RCU_NEXT_TAIL; i++) {
1652 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1653 rdp->nxtcompleted[i] = c;
1655 /* Record any needed additional grace periods. */
1656 ret = rcu_start_future_gp(rnp, rdp, NULL);
1658 /* Trace depending on how much we were able to accelerate. */
1659 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1660 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1662 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1667 * Move any callbacks whose grace period has completed to the
1668 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1669 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1670 * sublist. This function is idempotent, so it does not hurt to
1671 * invoke it repeatedly. As long as it is not invoked -too- often...
1672 * Returns true if the RCU grace-period kthread needs to be awakened.
1674 * The caller must hold rnp->lock with interrupts disabled.
1676 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1677 struct rcu_data *rdp)
1681 /* If the CPU has no callbacks, nothing to do. */
1682 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1686 * Find all callbacks whose ->completed numbers indicate that they
1687 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1689 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1690 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1692 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1694 /* Clean up any sublist tail pointers that were misordered above. */
1695 for (j = RCU_WAIT_TAIL; j < i; j++)
1696 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1698 /* Copy down callbacks to fill in empty sublists. */
1699 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1700 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1702 rdp->nxttail[j] = rdp->nxttail[i];
1703 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1706 /* Classify any remaining callbacks. */
1707 return rcu_accelerate_cbs(rsp, rnp, rdp);
1711 * Update CPU-local rcu_data state to record the beginnings and ends of
1712 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1713 * structure corresponding to the current CPU, and must have irqs disabled.
1714 * Returns true if the grace-period kthread needs to be awakened.
1716 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1717 struct rcu_data *rdp)
1721 /* Handle the ends of any preceding grace periods first. */
1722 if (rdp->completed == rnp->completed &&
1723 !unlikely(READ_ONCE(rdp->gpwrap))) {
1725 /* No grace period end, so just accelerate recent callbacks. */
1726 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1730 /* Advance callbacks. */
1731 ret = rcu_advance_cbs(rsp, rnp, rdp);
1733 /* Remember that we saw this grace-period completion. */
1734 rdp->completed = rnp->completed;
1735 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1738 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1740 * If the current grace period is waiting for this CPU,
1741 * set up to detect a quiescent state, otherwise don't
1742 * go looking for one.
1744 rdp->gpnum = rnp->gpnum;
1745 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1746 rdp->passed_quiesce = 0;
1747 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1748 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1749 zero_cpu_stall_ticks(rdp);
1750 WRITE_ONCE(rdp->gpwrap, false);
1755 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1757 unsigned long flags;
1759 struct rcu_node *rnp;
1761 local_irq_save(flags);
1763 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1764 rdp->completed == READ_ONCE(rnp->completed) &&
1765 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1766 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1767 local_irq_restore(flags);
1770 smp_mb__after_unlock_lock();
1771 needwake = __note_gp_changes(rsp, rnp, rdp);
1772 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1774 rcu_gp_kthread_wake(rsp);
1777 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1780 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1781 schedule_timeout_uninterruptible(delay);
1785 * Initialize a new grace period. Return 0 if no grace period required.
1787 static int rcu_gp_init(struct rcu_state *rsp)
1789 unsigned long oldmask;
1790 struct rcu_data *rdp;
1791 struct rcu_node *rnp = rcu_get_root(rsp);
1793 WRITE_ONCE(rsp->gp_activity, jiffies);
1794 raw_spin_lock_irq(&rnp->lock);
1795 smp_mb__after_unlock_lock();
1796 if (!READ_ONCE(rsp->gp_flags)) {
1797 /* Spurious wakeup, tell caller to go back to sleep. */
1798 raw_spin_unlock_irq(&rnp->lock);
1801 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1803 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1805 * Grace period already in progress, don't start another.
1806 * Not supposed to be able to happen.
1808 raw_spin_unlock_irq(&rnp->lock);
1812 /* Advance to a new grace period and initialize state. */
1813 record_gp_stall_check_time(rsp);
1814 /* Record GP times before starting GP, hence smp_store_release(). */
1815 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1816 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1817 raw_spin_unlock_irq(&rnp->lock);
1820 * Apply per-leaf buffered online and offline operations to the
1821 * rcu_node tree. Note that this new grace period need not wait
1822 * for subsequent online CPUs, and that quiescent-state forcing
1823 * will handle subsequent offline CPUs.
1825 rcu_for_each_leaf_node(rsp, rnp) {
1826 rcu_gp_slow(rsp, gp_preinit_delay);
1827 raw_spin_lock_irq(&rnp->lock);
1828 smp_mb__after_unlock_lock();
1829 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1830 !rnp->wait_blkd_tasks) {
1831 /* Nothing to do on this leaf rcu_node structure. */
1832 raw_spin_unlock_irq(&rnp->lock);
1836 /* Record old state, apply changes to ->qsmaskinit field. */
1837 oldmask = rnp->qsmaskinit;
1838 rnp->qsmaskinit = rnp->qsmaskinitnext;
1840 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1841 if (!oldmask != !rnp->qsmaskinit) {
1842 if (!oldmask) /* First online CPU for this rcu_node. */
1843 rcu_init_new_rnp(rnp);
1844 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1845 rnp->wait_blkd_tasks = true;
1846 else /* Last offline CPU and can propagate. */
1847 rcu_cleanup_dead_rnp(rnp);
1851 * If all waited-on tasks from prior grace period are
1852 * done, and if all this rcu_node structure's CPUs are
1853 * still offline, propagate up the rcu_node tree and
1854 * clear ->wait_blkd_tasks. Otherwise, if one of this
1855 * rcu_node structure's CPUs has since come back online,
1856 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1857 * checks for this, so just call it unconditionally).
1859 if (rnp->wait_blkd_tasks &&
1860 (!rcu_preempt_has_tasks(rnp) ||
1862 rnp->wait_blkd_tasks = false;
1863 rcu_cleanup_dead_rnp(rnp);
1866 raw_spin_unlock_irq(&rnp->lock);
1870 * Set the quiescent-state-needed bits in all the rcu_node
1871 * structures for all currently online CPUs in breadth-first order,
1872 * starting from the root rcu_node structure, relying on the layout
1873 * of the tree within the rsp->node[] array. Note that other CPUs
1874 * will access only the leaves of the hierarchy, thus seeing that no
1875 * grace period is in progress, at least until the corresponding
1876 * leaf node has been initialized. In addition, we have excluded
1877 * CPU-hotplug operations.
1879 * The grace period cannot complete until the initialization
1880 * process finishes, because this kthread handles both.
1882 rcu_for_each_node_breadth_first(rsp, rnp) {
1883 rcu_gp_slow(rsp, gp_init_delay);
1884 raw_spin_lock_irq(&rnp->lock);
1885 smp_mb__after_unlock_lock();
1886 rdp = this_cpu_ptr(rsp->rda);
1887 rcu_preempt_check_blocked_tasks(rnp);
1888 rnp->qsmask = rnp->qsmaskinit;
1889 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1890 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1891 WRITE_ONCE(rnp->completed, rsp->completed);
1892 if (rnp == rdp->mynode)
1893 (void)__note_gp_changes(rsp, rnp, rdp);
1894 rcu_preempt_boost_start_gp(rnp);
1895 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1896 rnp->level, rnp->grplo,
1897 rnp->grphi, rnp->qsmask);
1898 raw_spin_unlock_irq(&rnp->lock);
1899 cond_resched_rcu_qs();
1900 WRITE_ONCE(rsp->gp_activity, jiffies);
1907 * Helper function for wait_event_interruptible_timeout() wakeup
1908 * at force-quiescent-state time.
1910 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1912 struct rcu_node *rnp = rcu_get_root(rsp);
1914 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1915 *gfp = READ_ONCE(rsp->gp_flags);
1916 if (*gfp & RCU_GP_FLAG_FQS)
1919 /* The current grace period has completed. */
1920 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1927 * Do one round of quiescent-state forcing.
1929 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1931 int fqs_state = fqs_state_in;
1932 bool isidle = false;
1934 struct rcu_node *rnp = rcu_get_root(rsp);
1936 WRITE_ONCE(rsp->gp_activity, jiffies);
1938 if (fqs_state == RCU_SAVE_DYNTICK) {
1939 /* Collect dyntick-idle snapshots. */
1940 if (is_sysidle_rcu_state(rsp)) {
1942 maxj = jiffies - ULONG_MAX / 4;
1944 force_qs_rnp(rsp, dyntick_save_progress_counter,
1946 rcu_sysidle_report_gp(rsp, isidle, maxj);
1947 fqs_state = RCU_FORCE_QS;
1949 /* Handle dyntick-idle and offline CPUs. */
1951 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1953 /* Clear flag to prevent immediate re-entry. */
1954 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1955 raw_spin_lock_irq(&rnp->lock);
1956 smp_mb__after_unlock_lock();
1957 WRITE_ONCE(rsp->gp_flags,
1958 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
1959 raw_spin_unlock_irq(&rnp->lock);
1965 * Clean up after the old grace period.
1967 static void rcu_gp_cleanup(struct rcu_state *rsp)
1969 unsigned long gp_duration;
1970 bool needgp = false;
1972 struct rcu_data *rdp;
1973 struct rcu_node *rnp = rcu_get_root(rsp);
1975 WRITE_ONCE(rsp->gp_activity, jiffies);
1976 raw_spin_lock_irq(&rnp->lock);
1977 smp_mb__after_unlock_lock();
1978 gp_duration = jiffies - rsp->gp_start;
1979 if (gp_duration > rsp->gp_max)
1980 rsp->gp_max = gp_duration;
1983 * We know the grace period is complete, but to everyone else
1984 * it appears to still be ongoing. But it is also the case
1985 * that to everyone else it looks like there is nothing that
1986 * they can do to advance the grace period. It is therefore
1987 * safe for us to drop the lock in order to mark the grace
1988 * period as completed in all of the rcu_node structures.
1990 raw_spin_unlock_irq(&rnp->lock);
1993 * Propagate new ->completed value to rcu_node structures so
1994 * that other CPUs don't have to wait until the start of the next
1995 * grace period to process their callbacks. This also avoids
1996 * some nasty RCU grace-period initialization races by forcing
1997 * the end of the current grace period to be completely recorded in
1998 * all of the rcu_node structures before the beginning of the next
1999 * grace period is recorded in any of the rcu_node structures.
2001 rcu_for_each_node_breadth_first(rsp, rnp) {
2002 raw_spin_lock_irq(&rnp->lock);
2003 smp_mb__after_unlock_lock();
2004 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2005 WARN_ON_ONCE(rnp->qsmask);
2006 WRITE_ONCE(rnp->completed, rsp->gpnum);
2007 rdp = this_cpu_ptr(rsp->rda);
2008 if (rnp == rdp->mynode)
2009 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2010 /* smp_mb() provided by prior unlock-lock pair. */
2011 nocb += rcu_future_gp_cleanup(rsp, rnp);
2012 raw_spin_unlock_irq(&rnp->lock);
2013 cond_resched_rcu_qs();
2014 WRITE_ONCE(rsp->gp_activity, jiffies);
2015 rcu_gp_slow(rsp, gp_cleanup_delay);
2017 rnp = rcu_get_root(rsp);
2018 raw_spin_lock_irq(&rnp->lock);
2019 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
2020 rcu_nocb_gp_set(rnp, nocb);
2022 /* Declare grace period done. */
2023 WRITE_ONCE(rsp->completed, rsp->gpnum);
2024 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2025 rsp->fqs_state = RCU_GP_IDLE;
2026 rdp = this_cpu_ptr(rsp->rda);
2027 /* Advance CBs to reduce false positives below. */
2028 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2029 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2030 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2031 trace_rcu_grace_period(rsp->name,
2032 READ_ONCE(rsp->gpnum),
2035 raw_spin_unlock_irq(&rnp->lock);
2039 * Body of kthread that handles grace periods.
2041 static int __noreturn rcu_gp_kthread(void *arg)
2047 struct rcu_state *rsp = arg;
2048 struct rcu_node *rnp = rcu_get_root(rsp);
2050 rcu_bind_gp_kthread();
2053 /* Handle grace-period start. */
2055 trace_rcu_grace_period(rsp->name,
2056 READ_ONCE(rsp->gpnum),
2058 rsp->gp_state = RCU_GP_WAIT_GPS;
2059 wait_event_interruptible(rsp->gp_wq,
2060 READ_ONCE(rsp->gp_flags) &
2062 rsp->gp_state = RCU_GP_DONE_GPS;
2063 /* Locking provides needed memory barrier. */
2064 if (rcu_gp_init(rsp))
2066 cond_resched_rcu_qs();
2067 WRITE_ONCE(rsp->gp_activity, jiffies);
2068 WARN_ON(signal_pending(current));
2069 trace_rcu_grace_period(rsp->name,
2070 READ_ONCE(rsp->gpnum),
2074 /* Handle quiescent-state forcing. */
2075 fqs_state = RCU_SAVE_DYNTICK;
2076 j = jiffies_till_first_fqs;
2079 jiffies_till_first_fqs = HZ;
2084 rsp->jiffies_force_qs = jiffies + j;
2085 trace_rcu_grace_period(rsp->name,
2086 READ_ONCE(rsp->gpnum),
2088 rsp->gp_state = RCU_GP_WAIT_FQS;
2089 ret = wait_event_interruptible_timeout(rsp->gp_wq,
2090 rcu_gp_fqs_check_wake(rsp, &gf), j);
2091 rsp->gp_state = RCU_GP_DONE_FQS;
2092 /* Locking provides needed memory barriers. */
2093 /* If grace period done, leave loop. */
2094 if (!READ_ONCE(rnp->qsmask) &&
2095 !rcu_preempt_blocked_readers_cgp(rnp))
2097 /* If time for quiescent-state forcing, do it. */
2098 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2099 (gf & RCU_GP_FLAG_FQS)) {
2100 trace_rcu_grace_period(rsp->name,
2101 READ_ONCE(rsp->gpnum),
2103 fqs_state = rcu_gp_fqs(rsp, fqs_state);
2104 trace_rcu_grace_period(rsp->name,
2105 READ_ONCE(rsp->gpnum),
2107 cond_resched_rcu_qs();
2108 WRITE_ONCE(rsp->gp_activity, jiffies);
2110 /* Deal with stray signal. */
2111 cond_resched_rcu_qs();
2112 WRITE_ONCE(rsp->gp_activity, jiffies);
2113 WARN_ON(signal_pending(current));
2114 trace_rcu_grace_period(rsp->name,
2115 READ_ONCE(rsp->gpnum),
2118 j = jiffies_till_next_fqs;
2121 jiffies_till_next_fqs = HZ;
2124 jiffies_till_next_fqs = 1;
2128 /* Handle grace-period end. */
2129 rsp->gp_state = RCU_GP_CLEANUP;
2130 rcu_gp_cleanup(rsp);
2131 rsp->gp_state = RCU_GP_CLEANED;
2136 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2137 * in preparation for detecting the next grace period. The caller must hold
2138 * the root node's ->lock and hard irqs must be disabled.
2140 * Note that it is legal for a dying CPU (which is marked as offline) to
2141 * invoke this function. This can happen when the dying CPU reports its
2144 * Returns true if the grace-period kthread must be awakened.
2147 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2148 struct rcu_data *rdp)
2150 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2152 * Either we have not yet spawned the grace-period
2153 * task, this CPU does not need another grace period,
2154 * or a grace period is already in progress.
2155 * Either way, don't start a new grace period.
2159 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2160 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2164 * We can't do wakeups while holding the rnp->lock, as that
2165 * could cause possible deadlocks with the rq->lock. Defer
2166 * the wakeup to our caller.
2172 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2173 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2174 * is invoked indirectly from rcu_advance_cbs(), which would result in
2175 * endless recursion -- or would do so if it wasn't for the self-deadlock
2176 * that is encountered beforehand.
2178 * Returns true if the grace-period kthread needs to be awakened.
2180 static bool rcu_start_gp(struct rcu_state *rsp)
2182 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2183 struct rcu_node *rnp = rcu_get_root(rsp);
2187 * If there is no grace period in progress right now, any
2188 * callbacks we have up to this point will be satisfied by the
2189 * next grace period. Also, advancing the callbacks reduces the
2190 * probability of false positives from cpu_needs_another_gp()
2191 * resulting in pointless grace periods. So, advance callbacks
2192 * then start the grace period!
2194 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2195 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2200 * Report a full set of quiescent states to the specified rcu_state
2201 * data structure. This involves cleaning up after the prior grace
2202 * period and letting rcu_start_gp() start up the next grace period
2203 * if one is needed. Note that the caller must hold rnp->lock, which
2204 * is released before return.
2206 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2207 __releases(rcu_get_root(rsp)->lock)
2209 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2210 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2211 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2212 rcu_gp_kthread_wake(rsp);
2216 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2217 * Allows quiescent states for a group of CPUs to be reported at one go
2218 * to the specified rcu_node structure, though all the CPUs in the group
2219 * must be represented by the same rcu_node structure (which need not be a
2220 * leaf rcu_node structure, though it often will be). The gps parameter
2221 * is the grace-period snapshot, which means that the quiescent states
2222 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2223 * must be held upon entry, and it is released before return.
2226 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2227 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2228 __releases(rnp->lock)
2230 unsigned long oldmask = 0;
2231 struct rcu_node *rnp_c;
2233 /* Walk up the rcu_node hierarchy. */
2235 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2238 * Our bit has already been cleared, or the
2239 * relevant grace period is already over, so done.
2241 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2244 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2245 rnp->qsmask &= ~mask;
2246 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2247 mask, rnp->qsmask, rnp->level,
2248 rnp->grplo, rnp->grphi,
2250 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2252 /* Other bits still set at this level, so done. */
2253 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2256 mask = rnp->grpmask;
2257 if (rnp->parent == NULL) {
2259 /* No more levels. Exit loop holding root lock. */
2263 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2266 raw_spin_lock_irqsave(&rnp->lock, flags);
2267 smp_mb__after_unlock_lock();
2268 oldmask = rnp_c->qsmask;
2272 * Get here if we are the last CPU to pass through a quiescent
2273 * state for this grace period. Invoke rcu_report_qs_rsp()
2274 * to clean up and start the next grace period if one is needed.
2276 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2280 * Record a quiescent state for all tasks that were previously queued
2281 * on the specified rcu_node structure and that were blocking the current
2282 * RCU grace period. The caller must hold the specified rnp->lock with
2283 * irqs disabled, and this lock is released upon return, but irqs remain
2286 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2287 struct rcu_node *rnp, unsigned long flags)
2288 __releases(rnp->lock)
2292 struct rcu_node *rnp_p;
2294 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2295 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2296 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2297 return; /* Still need more quiescent states! */
2300 rnp_p = rnp->parent;
2301 if (rnp_p == NULL) {
2303 * Only one rcu_node structure in the tree, so don't
2304 * try to report up to its nonexistent parent!
2306 rcu_report_qs_rsp(rsp, flags);
2310 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2312 mask = rnp->grpmask;
2313 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2314 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
2315 smp_mb__after_unlock_lock();
2316 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2320 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2321 * structure. This must be either called from the specified CPU, or
2322 * called when the specified CPU is known to be offline (and when it is
2323 * also known that no other CPU is concurrently trying to help the offline
2324 * CPU). The lastcomp argument is used to make sure we are still in the
2325 * grace period of interest. We don't want to end the current grace period
2326 * based on quiescent states detected in an earlier grace period!
2329 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2331 unsigned long flags;
2334 struct rcu_node *rnp;
2337 raw_spin_lock_irqsave(&rnp->lock, flags);
2338 smp_mb__after_unlock_lock();
2339 if ((rdp->passed_quiesce == 0 &&
2340 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2341 rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2345 * The grace period in which this quiescent state was
2346 * recorded has ended, so don't report it upwards.
2347 * We will instead need a new quiescent state that lies
2348 * within the current grace period.
2350 rdp->passed_quiesce = 0; /* need qs for new gp. */
2351 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2352 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2355 mask = rdp->grpmask;
2356 if ((rnp->qsmask & mask) == 0) {
2357 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2359 rdp->qs_pending = 0;
2362 * This GP can't end until cpu checks in, so all of our
2363 * callbacks can be processed during the next GP.
2365 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2367 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2368 /* ^^^ Released rnp->lock */
2370 rcu_gp_kthread_wake(rsp);
2375 * Check to see if there is a new grace period of which this CPU
2376 * is not yet aware, and if so, set up local rcu_data state for it.
2377 * Otherwise, see if this CPU has just passed through its first
2378 * quiescent state for this grace period, and record that fact if so.
2381 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2383 /* Check for grace-period ends and beginnings. */
2384 note_gp_changes(rsp, rdp);
2387 * Does this CPU still need to do its part for current grace period?
2388 * If no, return and let the other CPUs do their part as well.
2390 if (!rdp->qs_pending)
2394 * Was there a quiescent state since the beginning of the grace
2395 * period? If no, then exit and wait for the next call.
2397 if (!rdp->passed_quiesce &&
2398 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2402 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2405 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2409 * Send the specified CPU's RCU callbacks to the orphanage. The
2410 * specified CPU must be offline, and the caller must hold the
2414 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2415 struct rcu_node *rnp, struct rcu_data *rdp)
2417 /* No-CBs CPUs do not have orphanable callbacks. */
2418 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2422 * Orphan the callbacks. First adjust the counts. This is safe
2423 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2424 * cannot be running now. Thus no memory barrier is required.
2426 if (rdp->nxtlist != NULL) {
2427 rsp->qlen_lazy += rdp->qlen_lazy;
2428 rsp->qlen += rdp->qlen;
2429 rdp->n_cbs_orphaned += rdp->qlen;
2431 WRITE_ONCE(rdp->qlen, 0);
2435 * Next, move those callbacks still needing a grace period to
2436 * the orphanage, where some other CPU will pick them up.
2437 * Some of the callbacks might have gone partway through a grace
2438 * period, but that is too bad. They get to start over because we
2439 * cannot assume that grace periods are synchronized across CPUs.
2440 * We don't bother updating the ->nxttail[] array yet, instead
2441 * we just reset the whole thing later on.
2443 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2444 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2445 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2446 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2450 * Then move the ready-to-invoke callbacks to the orphanage,
2451 * where some other CPU will pick them up. These will not be
2452 * required to pass though another grace period: They are done.
2454 if (rdp->nxtlist != NULL) {
2455 *rsp->orphan_donetail = rdp->nxtlist;
2456 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2460 * Finally, initialize the rcu_data structure's list to empty and
2461 * disallow further callbacks on this CPU.
2463 init_callback_list(rdp);
2464 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2468 * Adopt the RCU callbacks from the specified rcu_state structure's
2469 * orphanage. The caller must hold the ->orphan_lock.
2471 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2474 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2476 /* No-CBs CPUs are handled specially. */
2477 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2478 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2481 /* Do the accounting first. */
2482 rdp->qlen_lazy += rsp->qlen_lazy;
2483 rdp->qlen += rsp->qlen;
2484 rdp->n_cbs_adopted += rsp->qlen;
2485 if (rsp->qlen_lazy != rsp->qlen)
2486 rcu_idle_count_callbacks_posted();
2491 * We do not need a memory barrier here because the only way we
2492 * can get here if there is an rcu_barrier() in flight is if
2493 * we are the task doing the rcu_barrier().
2496 /* First adopt the ready-to-invoke callbacks. */
2497 if (rsp->orphan_donelist != NULL) {
2498 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2499 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2500 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2501 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2502 rdp->nxttail[i] = rsp->orphan_donetail;
2503 rsp->orphan_donelist = NULL;
2504 rsp->orphan_donetail = &rsp->orphan_donelist;
2507 /* And then adopt the callbacks that still need a grace period. */
2508 if (rsp->orphan_nxtlist != NULL) {
2509 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2510 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2511 rsp->orphan_nxtlist = NULL;
2512 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2517 * Trace the fact that this CPU is going offline.
2519 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2521 RCU_TRACE(unsigned long mask);
2522 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2523 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2525 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2528 RCU_TRACE(mask = rdp->grpmask);
2529 trace_rcu_grace_period(rsp->name,
2530 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2535 * All CPUs for the specified rcu_node structure have gone offline,
2536 * and all tasks that were preempted within an RCU read-side critical
2537 * section while running on one of those CPUs have since exited their RCU
2538 * read-side critical section. Some other CPU is reporting this fact with
2539 * the specified rcu_node structure's ->lock held and interrupts disabled.
2540 * This function therefore goes up the tree of rcu_node structures,
2541 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2542 * the leaf rcu_node structure's ->qsmaskinit field has already been
2545 * This function does check that the specified rcu_node structure has
2546 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2547 * prematurely. That said, invoking it after the fact will cost you
2548 * a needless lock acquisition. So once it has done its work, don't
2551 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2554 struct rcu_node *rnp = rnp_leaf;
2556 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2557 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2560 mask = rnp->grpmask;
2564 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2565 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2566 rnp->qsmaskinit &= ~mask;
2567 rnp->qsmask &= ~mask;
2568 if (rnp->qsmaskinit) {
2569 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2572 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2577 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2578 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2581 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
2583 unsigned long flags;
2585 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2586 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2588 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2591 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2592 mask = rdp->grpmask;
2593 raw_spin_lock_irqsave(&rnp->lock, flags);
2594 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2595 rnp->qsmaskinitnext &= ~mask;
2596 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2600 * The CPU has been completely removed, and some other CPU is reporting
2601 * this fact from process context. Do the remainder of the cleanup,
2602 * including orphaning the outgoing CPU's RCU callbacks, and also
2603 * adopting them. There can only be one CPU hotplug operation at a time,
2604 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2606 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2608 unsigned long flags;
2609 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2610 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2612 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2615 /* Adjust any no-longer-needed kthreads. */
2616 rcu_boost_kthread_setaffinity(rnp, -1);
2618 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2619 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2620 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2621 rcu_adopt_orphan_cbs(rsp, flags);
2622 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2624 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2625 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2626 cpu, rdp->qlen, rdp->nxtlist);
2630 * Invoke any RCU callbacks that have made it to the end of their grace
2631 * period. Thottle as specified by rdp->blimit.
2633 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2635 unsigned long flags;
2636 struct rcu_head *next, *list, **tail;
2637 long bl, count, count_lazy;
2640 /* If no callbacks are ready, just return. */
2641 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2642 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2643 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2644 need_resched(), is_idle_task(current),
2645 rcu_is_callbacks_kthread());
2650 * Extract the list of ready callbacks, disabling to prevent
2651 * races with call_rcu() from interrupt handlers.
2653 local_irq_save(flags);
2654 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2656 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2657 list = rdp->nxtlist;
2658 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2659 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2660 tail = rdp->nxttail[RCU_DONE_TAIL];
2661 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2662 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2663 rdp->nxttail[i] = &rdp->nxtlist;
2664 local_irq_restore(flags);
2666 /* Invoke callbacks. */
2667 count = count_lazy = 0;
2671 debug_rcu_head_unqueue(list);
2672 if (__rcu_reclaim(rsp->name, list))
2675 /* Stop only if limit reached and CPU has something to do. */
2676 if (++count >= bl &&
2678 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2682 local_irq_save(flags);
2683 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2684 is_idle_task(current),
2685 rcu_is_callbacks_kthread());
2687 /* Update count, and requeue any remaining callbacks. */
2689 *tail = rdp->nxtlist;
2690 rdp->nxtlist = list;
2691 for (i = 0; i < RCU_NEXT_SIZE; i++)
2692 if (&rdp->nxtlist == rdp->nxttail[i])
2693 rdp->nxttail[i] = tail;
2697 smp_mb(); /* List handling before counting for rcu_barrier(). */
2698 rdp->qlen_lazy -= count_lazy;
2699 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2700 rdp->n_cbs_invoked += count;
2702 /* Reinstate batch limit if we have worked down the excess. */
2703 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2704 rdp->blimit = blimit;
2706 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2707 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2708 rdp->qlen_last_fqs_check = 0;
2709 rdp->n_force_qs_snap = rsp->n_force_qs;
2710 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2711 rdp->qlen_last_fqs_check = rdp->qlen;
2712 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2714 local_irq_restore(flags);
2716 /* Re-invoke RCU core processing if there are callbacks remaining. */
2717 if (cpu_has_callbacks_ready_to_invoke(rdp))
2722 * Check to see if this CPU is in a non-context-switch quiescent state
2723 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2724 * Also schedule RCU core processing.
2726 * This function must be called from hardirq context. It is normally
2727 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2728 * false, there is no point in invoking rcu_check_callbacks().
2730 void rcu_check_callbacks(int user)
2732 trace_rcu_utilization(TPS("Start scheduler-tick"));
2733 increment_cpu_stall_ticks();
2734 if (user || rcu_is_cpu_rrupt_from_idle()) {
2737 * Get here if this CPU took its interrupt from user
2738 * mode or from the idle loop, and if this is not a
2739 * nested interrupt. In this case, the CPU is in
2740 * a quiescent state, so note it.
2742 * No memory barrier is required here because both
2743 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2744 * variables that other CPUs neither access nor modify,
2745 * at least not while the corresponding CPU is online.
2751 } else if (!in_softirq()) {
2754 * Get here if this CPU did not take its interrupt from
2755 * softirq, in other words, if it is not interrupting
2756 * a rcu_bh read-side critical section. This is an _bh
2757 * critical section, so note it.
2762 rcu_preempt_check_callbacks();
2766 rcu_note_voluntary_context_switch(current);
2767 trace_rcu_utilization(TPS("End scheduler-tick"));
2771 * Scan the leaf rcu_node structures, processing dyntick state for any that
2772 * have not yet encountered a quiescent state, using the function specified.
2773 * Also initiate boosting for any threads blocked on the root rcu_node.
2775 * The caller must have suppressed start of new grace periods.
2777 static void force_qs_rnp(struct rcu_state *rsp,
2778 int (*f)(struct rcu_data *rsp, bool *isidle,
2779 unsigned long *maxj),
2780 bool *isidle, unsigned long *maxj)
2784 unsigned long flags;
2786 struct rcu_node *rnp;
2788 rcu_for_each_leaf_node(rsp, rnp) {
2789 cond_resched_rcu_qs();
2791 raw_spin_lock_irqsave(&rnp->lock, flags);
2792 smp_mb__after_unlock_lock();
2793 if (rnp->qsmask == 0) {
2794 if (rcu_state_p == &rcu_sched_state ||
2795 rsp != rcu_state_p ||
2796 rcu_preempt_blocked_readers_cgp(rnp)) {
2798 * No point in scanning bits because they
2799 * are all zero. But we might need to
2800 * priority-boost blocked readers.
2802 rcu_initiate_boost(rnp, flags);
2803 /* rcu_initiate_boost() releases rnp->lock */
2807 (rnp->parent->qsmask & rnp->grpmask)) {
2809 * Race between grace-period
2810 * initialization and task exiting RCU
2811 * read-side critical section: Report.
2813 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2814 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2820 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2821 if ((rnp->qsmask & bit) != 0) {
2822 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2827 /* Idle/offline CPUs, report (releases rnp->lock. */
2828 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2830 /* Nothing to do here, so just drop the lock. */
2831 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2837 * Force quiescent states on reluctant CPUs, and also detect which
2838 * CPUs are in dyntick-idle mode.
2840 static void force_quiescent_state(struct rcu_state *rsp)
2842 unsigned long flags;
2844 struct rcu_node *rnp;
2845 struct rcu_node *rnp_old = NULL;
2847 /* Funnel through hierarchy to reduce memory contention. */
2848 rnp = __this_cpu_read(rsp->rda->mynode);
2849 for (; rnp != NULL; rnp = rnp->parent) {
2850 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2851 !raw_spin_trylock(&rnp->fqslock);
2852 if (rnp_old != NULL)
2853 raw_spin_unlock(&rnp_old->fqslock);
2855 rsp->n_force_qs_lh++;
2860 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2862 /* Reached the root of the rcu_node tree, acquire lock. */
2863 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2864 smp_mb__after_unlock_lock();
2865 raw_spin_unlock(&rnp_old->fqslock);
2866 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2867 rsp->n_force_qs_lh++;
2868 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2869 return; /* Someone beat us to it. */
2871 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2872 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2873 rcu_gp_kthread_wake(rsp);
2877 * This does the RCU core processing work for the specified rcu_state
2878 * and rcu_data structures. This may be called only from the CPU to
2879 * whom the rdp belongs.
2882 __rcu_process_callbacks(struct rcu_state *rsp)
2884 unsigned long flags;
2886 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2888 WARN_ON_ONCE(rdp->beenonline == 0);
2890 /* Update RCU state based on any recent quiescent states. */
2891 rcu_check_quiescent_state(rsp, rdp);
2893 /* Does this CPU require a not-yet-started grace period? */
2894 local_irq_save(flags);
2895 if (cpu_needs_another_gp(rsp, rdp)) {
2896 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2897 needwake = rcu_start_gp(rsp);
2898 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2900 rcu_gp_kthread_wake(rsp);
2902 local_irq_restore(flags);
2905 /* If there are callbacks ready, invoke them. */
2906 if (cpu_has_callbacks_ready_to_invoke(rdp))
2907 invoke_rcu_callbacks(rsp, rdp);
2909 /* Do any needed deferred wakeups of rcuo kthreads. */
2910 do_nocb_deferred_wakeup(rdp);
2914 * Do RCU core processing for the current CPU.
2916 static void rcu_process_callbacks(struct softirq_action *unused)
2918 struct rcu_state *rsp;
2920 if (cpu_is_offline(smp_processor_id()))
2922 trace_rcu_utilization(TPS("Start RCU core"));
2923 for_each_rcu_flavor(rsp)
2924 __rcu_process_callbacks(rsp);
2925 trace_rcu_utilization(TPS("End RCU core"));
2929 * Schedule RCU callback invocation. If the specified type of RCU
2930 * does not support RCU priority boosting, just do a direct call,
2931 * otherwise wake up the per-CPU kernel kthread. Note that because we
2932 * are running on the current CPU with softirqs disabled, the
2933 * rcu_cpu_kthread_task cannot disappear out from under us.
2935 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2937 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2939 if (likely(!rsp->boost)) {
2940 rcu_do_batch(rsp, rdp);
2943 invoke_rcu_callbacks_kthread();
2946 static void invoke_rcu_core(void)
2948 if (cpu_online(smp_processor_id()))
2949 raise_softirq(RCU_SOFTIRQ);
2953 * Handle any core-RCU processing required by a call_rcu() invocation.
2955 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2956 struct rcu_head *head, unsigned long flags)
2961 * If called from an extended quiescent state, invoke the RCU
2962 * core in order to force a re-evaluation of RCU's idleness.
2964 if (!rcu_is_watching())
2967 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2968 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2972 * Force the grace period if too many callbacks or too long waiting.
2973 * Enforce hysteresis, and don't invoke force_quiescent_state()
2974 * if some other CPU has recently done so. Also, don't bother
2975 * invoking force_quiescent_state() if the newly enqueued callback
2976 * is the only one waiting for a grace period to complete.
2978 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2980 /* Are we ignoring a completed grace period? */
2981 note_gp_changes(rsp, rdp);
2983 /* Start a new grace period if one not already started. */
2984 if (!rcu_gp_in_progress(rsp)) {
2985 struct rcu_node *rnp_root = rcu_get_root(rsp);
2987 raw_spin_lock(&rnp_root->lock);
2988 smp_mb__after_unlock_lock();
2989 needwake = rcu_start_gp(rsp);
2990 raw_spin_unlock(&rnp_root->lock);
2992 rcu_gp_kthread_wake(rsp);
2994 /* Give the grace period a kick. */
2995 rdp->blimit = LONG_MAX;
2996 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2997 *rdp->nxttail[RCU_DONE_TAIL] != head)
2998 force_quiescent_state(rsp);
2999 rdp->n_force_qs_snap = rsp->n_force_qs;
3000 rdp->qlen_last_fqs_check = rdp->qlen;
3006 * RCU callback function to leak a callback.
3008 static void rcu_leak_callback(struct rcu_head *rhp)
3013 * Helper function for call_rcu() and friends. The cpu argument will
3014 * normally be -1, indicating "currently running CPU". It may specify
3015 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3016 * is expected to specify a CPU.
3019 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
3020 struct rcu_state *rsp, int cpu, bool lazy)
3022 unsigned long flags;
3023 struct rcu_data *rdp;
3025 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
3026 if (debug_rcu_head_queue(head)) {
3027 /* Probable double call_rcu(), so leak the callback. */
3028 WRITE_ONCE(head->func, rcu_leak_callback);
3029 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3036 * Opportunistically note grace-period endings and beginnings.
3037 * Note that we might see a beginning right after we see an
3038 * end, but never vice versa, since this CPU has to pass through
3039 * a quiescent state betweentimes.
3041 local_irq_save(flags);
3042 rdp = this_cpu_ptr(rsp->rda);
3044 /* Add the callback to our list. */
3045 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3049 rdp = per_cpu_ptr(rsp->rda, cpu);
3050 if (likely(rdp->mynode)) {
3051 /* Post-boot, so this should be for a no-CBs CPU. */
3052 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3053 WARN_ON_ONCE(offline);
3054 /* Offline CPU, _call_rcu() illegal, leak callback. */
3055 local_irq_restore(flags);
3059 * Very early boot, before rcu_init(). Initialize if needed
3060 * and then drop through to queue the callback.
3063 WARN_ON_ONCE(!rcu_is_watching());
3064 if (!likely(rdp->nxtlist))
3065 init_default_callback_list(rdp);
3067 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3071 rcu_idle_count_callbacks_posted();
3072 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3073 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3074 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3076 if (__is_kfree_rcu_offset((unsigned long)func))
3077 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3078 rdp->qlen_lazy, rdp->qlen);
3080 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3082 /* Go handle any RCU core processing required. */
3083 __call_rcu_core(rsp, rdp, head, flags);
3084 local_irq_restore(flags);
3088 * Queue an RCU-sched callback for invocation after a grace period.
3090 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
3092 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3094 EXPORT_SYMBOL_GPL(call_rcu_sched);
3097 * Queue an RCU callback for invocation after a quicker grace period.
3099 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
3101 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3103 EXPORT_SYMBOL_GPL(call_rcu_bh);
3106 * Queue an RCU callback for lazy invocation after a grace period.
3107 * This will likely be later named something like "call_rcu_lazy()",
3108 * but this change will require some way of tagging the lazy RCU
3109 * callbacks in the list of pending callbacks. Until then, this
3110 * function may only be called from __kfree_rcu().
3112 void kfree_call_rcu(struct rcu_head *head,
3113 void (*func)(struct rcu_head *rcu))
3115 __call_rcu(head, func, rcu_state_p, -1, 1);
3117 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3120 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3121 * any blocking grace-period wait automatically implies a grace period
3122 * if there is only one CPU online at any point time during execution
3123 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3124 * occasionally incorrectly indicate that there are multiple CPUs online
3125 * when there was in fact only one the whole time, as this just adds
3126 * some overhead: RCU still operates correctly.
3128 static inline int rcu_blocking_is_gp(void)
3132 might_sleep(); /* Check for RCU read-side critical section. */
3134 ret = num_online_cpus() <= 1;
3140 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3142 * Control will return to the caller some time after a full rcu-sched
3143 * grace period has elapsed, in other words after all currently executing
3144 * rcu-sched read-side critical sections have completed. These read-side
3145 * critical sections are delimited by rcu_read_lock_sched() and
3146 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3147 * local_irq_disable(), and so on may be used in place of
3148 * rcu_read_lock_sched().
3150 * This means that all preempt_disable code sequences, including NMI and
3151 * non-threaded hardware-interrupt handlers, in progress on entry will
3152 * have completed before this primitive returns. However, this does not
3153 * guarantee that softirq handlers will have completed, since in some
3154 * kernels, these handlers can run in process context, and can block.
3156 * Note that this guarantee implies further memory-ordering guarantees.
3157 * On systems with more than one CPU, when synchronize_sched() returns,
3158 * each CPU is guaranteed to have executed a full memory barrier since the
3159 * end of its last RCU-sched read-side critical section whose beginning
3160 * preceded the call to synchronize_sched(). In addition, each CPU having
3161 * an RCU read-side critical section that extends beyond the return from
3162 * synchronize_sched() is guaranteed to have executed a full memory barrier
3163 * after the beginning of synchronize_sched() and before the beginning of
3164 * that RCU read-side critical section. Note that these guarantees include
3165 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3166 * that are executing in the kernel.
3168 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3169 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3170 * to have executed a full memory barrier during the execution of
3171 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3172 * again only if the system has more than one CPU).
3174 * This primitive provides the guarantees made by the (now removed)
3175 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3176 * guarantees that rcu_read_lock() sections will have completed.
3177 * In "classic RCU", these two guarantees happen to be one and
3178 * the same, but can differ in realtime RCU implementations.
3180 void synchronize_sched(void)
3182 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
3183 !lock_is_held(&rcu_lock_map) &&
3184 !lock_is_held(&rcu_sched_lock_map),
3185 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3186 if (rcu_blocking_is_gp())
3188 if (rcu_gp_is_expedited())
3189 synchronize_sched_expedited();
3191 wait_rcu_gp(call_rcu_sched);
3193 EXPORT_SYMBOL_GPL(synchronize_sched);
3196 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3198 * Control will return to the caller some time after a full rcu_bh grace
3199 * period has elapsed, in other words after all currently executing rcu_bh
3200 * read-side critical sections have completed. RCU read-side critical
3201 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3202 * and may be nested.
3204 * See the description of synchronize_sched() for more detailed information
3205 * on memory ordering guarantees.
3207 void synchronize_rcu_bh(void)
3209 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
3210 !lock_is_held(&rcu_lock_map) &&
3211 !lock_is_held(&rcu_sched_lock_map),
3212 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3213 if (rcu_blocking_is_gp())
3215 if (rcu_gp_is_expedited())
3216 synchronize_rcu_bh_expedited();
3218 wait_rcu_gp(call_rcu_bh);
3220 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3223 * get_state_synchronize_rcu - Snapshot current RCU state
3225 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3226 * to determine whether or not a full grace period has elapsed in the
3229 unsigned long get_state_synchronize_rcu(void)
3232 * Any prior manipulation of RCU-protected data must happen
3233 * before the load from ->gpnum.
3238 * Make sure this load happens before the purportedly
3239 * time-consuming work between get_state_synchronize_rcu()
3240 * and cond_synchronize_rcu().
3242 return smp_load_acquire(&rcu_state_p->gpnum);
3244 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3247 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3249 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3251 * If a full RCU grace period has elapsed since the earlier call to
3252 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3253 * synchronize_rcu() to wait for a full grace period.
3255 * Yes, this function does not take counter wrap into account. But
3256 * counter wrap is harmless. If the counter wraps, we have waited for
3257 * more than 2 billion grace periods (and way more on a 64-bit system!),
3258 * so waiting for one additional grace period should be just fine.
3260 void cond_synchronize_rcu(unsigned long oldstate)
3262 unsigned long newstate;
3265 * Ensure that this load happens before any RCU-destructive
3266 * actions the caller might carry out after we return.
3268 newstate = smp_load_acquire(&rcu_state_p->completed);
3269 if (ULONG_CMP_GE(oldstate, newstate))
3272 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3274 /* Adjust sequence number for start of update-side operation. */
3275 static void rcu_seq_start(unsigned long *sp)
3277 WRITE_ONCE(*sp, *sp + 1);
3278 smp_mb(); /* Ensure update-side operation after counter increment. */
3279 WARN_ON_ONCE(!(*sp & 0x1));
3282 /* Adjust sequence number for end of update-side operation. */
3283 static void rcu_seq_end(unsigned long *sp)
3285 smp_mb(); /* Ensure update-side operation before counter increment. */
3286 WRITE_ONCE(*sp, *sp + 1);
3287 WARN_ON_ONCE(*sp & 0x1);
3290 /* Take a snapshot of the update side's sequence number. */
3291 static unsigned long rcu_seq_snap(unsigned long *sp)
3295 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3296 s = (READ_ONCE(*sp) + 3) & ~0x1;
3297 smp_mb(); /* Above access must not bleed into critical section. */
3302 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3303 * full update-side operation has occurred.
3305 static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3307 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3310 /* Wrapper functions for expedited grace periods. */
3311 static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
3313 rcu_seq_start(&rsp->expedited_sequence);
3315 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
3317 rcu_seq_end(&rsp->expedited_sequence);
3318 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3320 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
3322 return rcu_seq_snap(&rsp->expedited_sequence);
3324 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
3326 return rcu_seq_done(&rsp->expedited_sequence, s);
3329 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3330 static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp,
3331 struct rcu_data *rdp,
3332 atomic_long_t *stat, unsigned long s)
3334 if (rcu_exp_gp_seq_done(rsp, s)) {
3336 mutex_unlock(&rnp->exp_funnel_mutex);
3338 mutex_unlock(&rdp->exp_funnel_mutex);
3339 /* Ensure test happens before caller kfree(). */
3340 smp_mb__before_atomic(); /* ^^^ */
3341 atomic_long_inc(stat);
3348 * Funnel-lock acquisition for expedited grace periods. Returns a
3349 * pointer to the root rcu_node structure, or NULL if some other
3350 * task did the expedited grace period for us.
3352 static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
3354 struct rcu_data *rdp;
3355 struct rcu_node *rnp0;
3356 struct rcu_node *rnp1 = NULL;
3359 * Each pass through the following loop works its way
3360 * up the rcu_node tree, returning if others have done the
3361 * work or otherwise falls through holding the root rnp's
3362 * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure
3363 * can be inexact, as it is just promoting locality and is not
3364 * strictly needed for correctness.
3366 rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
3367 if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s))
3369 mutex_lock(&rdp->exp_funnel_mutex);
3371 for (; rnp0 != NULL; rnp0 = rnp0->parent) {
3372 if (sync_exp_work_done(rsp, rnp1, rdp,
3373 &rsp->expedited_workdone2, s))
3375 mutex_lock(&rnp0->exp_funnel_mutex);
3377 mutex_unlock(&rnp1->exp_funnel_mutex);
3379 mutex_unlock(&rdp->exp_funnel_mutex);
3382 if (sync_exp_work_done(rsp, rnp1, rdp,
3383 &rsp->expedited_workdone3, s))
3388 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3389 static int synchronize_sched_expedited_cpu_stop(void *data)
3391 struct rcu_data *rdp = data;
3392 struct rcu_state *rsp = rdp->rsp;
3394 /* We are here: If we are last, do the wakeup. */
3395 rdp->exp_done = true;
3396 if (atomic_dec_and_test(&rsp->expedited_need_qs))
3397 wake_up(&rsp->expedited_wq);
3401 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
3404 unsigned long jiffies_stall;
3405 unsigned long jiffies_start;
3406 struct rcu_data *rdp;
3409 jiffies_stall = rcu_jiffies_till_stall_check();
3410 jiffies_start = jiffies;
3413 ret = wait_event_interruptible_timeout(
3415 !atomic_read(&rsp->expedited_need_qs),
3420 /* Hit a signal, disable CPU stall warnings. */
3421 wait_event(rsp->expedited_wq,
3422 !atomic_read(&rsp->expedited_need_qs));
3425 pr_err("INFO: %s detected expedited stalls on CPUs: {",
3427 for_each_online_cpu(cpu) {
3428 rdp = per_cpu_ptr(rsp->rda, cpu);
3432 pr_cont(" %d", cpu);
3434 pr_cont(" } %lu jiffies s: %lu\n",
3435 jiffies - jiffies_start, rsp->expedited_sequence);
3436 for_each_online_cpu(cpu) {
3437 rdp = per_cpu_ptr(rsp->rda, cpu);
3443 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
3448 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3450 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3451 * approach to force the grace period to end quickly. This consumes
3452 * significant time on all CPUs and is unfriendly to real-time workloads,
3453 * so is thus not recommended for any sort of common-case code. In fact,
3454 * if you are using synchronize_sched_expedited() in a loop, please
3455 * restructure your code to batch your updates, and then use a single
3456 * synchronize_sched() instead.
3458 * This implementation can be thought of as an application of sequence
3459 * locking to expedited grace periods, but using the sequence counter to
3460 * determine when someone else has already done the work instead of for
3463 void synchronize_sched_expedited(void)
3467 struct rcu_node *rnp;
3468 struct rcu_state *rsp = &rcu_sched_state;
3470 /* Take a snapshot of the sequence number. */
3471 s = rcu_exp_gp_seq_snap(rsp);
3473 if (!try_get_online_cpus()) {
3474 /* CPU hotplug operation in flight, fall back to normal GP. */
3475 wait_rcu_gp(call_rcu_sched);
3476 atomic_long_inc(&rsp->expedited_normal);
3479 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3481 rnp = exp_funnel_lock(rsp, s);
3484 return; /* Someone else did our work for us. */
3487 rcu_exp_gp_seq_start(rsp);
3489 /* Stop each CPU that is online, non-idle, and not us. */
3490 init_waitqueue_head(&rsp->expedited_wq);
3491 atomic_set(&rsp->expedited_need_qs, 1); /* Extra count avoids race. */
3492 for_each_online_cpu(cpu) {
3493 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3494 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3496 rdp->exp_done = false;
3498 /* Skip our CPU and any idle CPUs. */
3499 if (raw_smp_processor_id() == cpu ||
3500 !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3502 atomic_inc(&rsp->expedited_need_qs);
3503 stop_one_cpu_nowait(cpu, synchronize_sched_expedited_cpu_stop,
3504 rdp, &rdp->exp_stop_work);
3507 /* Remove extra count and, if necessary, wait for CPUs to stop. */
3508 if (!atomic_dec_and_test(&rsp->expedited_need_qs))
3509 synchronize_sched_expedited_wait(rsp);
3511 rcu_exp_gp_seq_end(rsp);
3512 mutex_unlock(&rnp->exp_funnel_mutex);
3516 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3519 * Check to see if there is any immediate RCU-related work to be done
3520 * by the current CPU, for the specified type of RCU, returning 1 if so.
3521 * The checks are in order of increasing expense: checks that can be
3522 * carried out against CPU-local state are performed first. However,
3523 * we must check for CPU stalls first, else we might not get a chance.
3525 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3527 struct rcu_node *rnp = rdp->mynode;
3529 rdp->n_rcu_pending++;
3531 /* Check for CPU stalls, if enabled. */
3532 check_cpu_stall(rsp, rdp);
3534 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3535 if (rcu_nohz_full_cpu(rsp))
3538 /* Is the RCU core waiting for a quiescent state from this CPU? */
3539 if (rcu_scheduler_fully_active &&
3540 rdp->qs_pending && !rdp->passed_quiesce &&
3541 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3542 rdp->n_rp_qs_pending++;
3543 } else if (rdp->qs_pending &&
3544 (rdp->passed_quiesce ||
3545 rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3546 rdp->n_rp_report_qs++;
3550 /* Does this CPU have callbacks ready to invoke? */
3551 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3552 rdp->n_rp_cb_ready++;
3556 /* Has RCU gone idle with this CPU needing another grace period? */
3557 if (cpu_needs_another_gp(rsp, rdp)) {
3558 rdp->n_rp_cpu_needs_gp++;
3562 /* Has another RCU grace period completed? */
3563 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3564 rdp->n_rp_gp_completed++;
3568 /* Has a new RCU grace period started? */
3569 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3570 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3571 rdp->n_rp_gp_started++;
3575 /* Does this CPU need a deferred NOCB wakeup? */
3576 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3577 rdp->n_rp_nocb_defer_wakeup++;
3582 rdp->n_rp_need_nothing++;
3587 * Check to see if there is any immediate RCU-related work to be done
3588 * by the current CPU, returning 1 if so. This function is part of the
3589 * RCU implementation; it is -not- an exported member of the RCU API.
3591 static int rcu_pending(void)
3593 struct rcu_state *rsp;
3595 for_each_rcu_flavor(rsp)
3596 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3602 * Return true if the specified CPU has any callback. If all_lazy is
3603 * non-NULL, store an indication of whether all callbacks are lazy.
3604 * (If there are no callbacks, all of them are deemed to be lazy.)
3606 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3610 struct rcu_data *rdp;
3611 struct rcu_state *rsp;
3613 for_each_rcu_flavor(rsp) {
3614 rdp = this_cpu_ptr(rsp->rda);
3618 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3629 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3630 * the compiler is expected to optimize this away.
3632 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3633 int cpu, unsigned long done)
3635 trace_rcu_barrier(rsp->name, s, cpu,
3636 atomic_read(&rsp->barrier_cpu_count), done);
3640 * RCU callback function for _rcu_barrier(). If we are last, wake
3641 * up the task executing _rcu_barrier().
3643 static void rcu_barrier_callback(struct rcu_head *rhp)
3645 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3646 struct rcu_state *rsp = rdp->rsp;
3648 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3649 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
3650 complete(&rsp->barrier_completion);
3652 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
3657 * Called with preemption disabled, and from cross-cpu IRQ context.
3659 static void rcu_barrier_func(void *type)
3661 struct rcu_state *rsp = type;
3662 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3664 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
3665 atomic_inc(&rsp->barrier_cpu_count);
3666 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3670 * Orchestrate the specified type of RCU barrier, waiting for all
3671 * RCU callbacks of the specified type to complete.
3673 static void _rcu_barrier(struct rcu_state *rsp)
3676 struct rcu_data *rdp;
3677 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3679 _rcu_barrier_trace(rsp, "Begin", -1, s);
3681 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3682 mutex_lock(&rsp->barrier_mutex);
3684 /* Did someone else do our work for us? */
3685 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3686 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
3687 smp_mb(); /* caller's subsequent code after above check. */
3688 mutex_unlock(&rsp->barrier_mutex);
3692 /* Mark the start of the barrier operation. */
3693 rcu_seq_start(&rsp->barrier_sequence);
3694 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
3697 * Initialize the count to one rather than to zero in order to
3698 * avoid a too-soon return to zero in case of a short grace period
3699 * (or preemption of this task). Exclude CPU-hotplug operations
3700 * to ensure that no offline CPU has callbacks queued.
3702 init_completion(&rsp->barrier_completion);
3703 atomic_set(&rsp->barrier_cpu_count, 1);
3707 * Force each CPU with callbacks to register a new callback.
3708 * When that callback is invoked, we will know that all of the
3709 * corresponding CPU's preceding callbacks have been invoked.
3711 for_each_possible_cpu(cpu) {
3712 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3714 rdp = per_cpu_ptr(rsp->rda, cpu);
3715 if (rcu_is_nocb_cpu(cpu)) {
3716 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3717 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3718 rsp->barrier_sequence);
3720 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3721 rsp->barrier_sequence);
3722 smp_mb__before_atomic();
3723 atomic_inc(&rsp->barrier_cpu_count);
3724 __call_rcu(&rdp->barrier_head,
3725 rcu_barrier_callback, rsp, cpu, 0);
3727 } else if (READ_ONCE(rdp->qlen)) {
3728 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3729 rsp->barrier_sequence);
3730 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3732 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3733 rsp->barrier_sequence);
3739 * Now that we have an rcu_barrier_callback() callback on each
3740 * CPU, and thus each counted, remove the initial count.
3742 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3743 complete(&rsp->barrier_completion);
3745 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3746 wait_for_completion(&rsp->barrier_completion);
3748 /* Mark the end of the barrier operation. */
3749 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
3750 rcu_seq_end(&rsp->barrier_sequence);
3752 /* Other rcu_barrier() invocations can now safely proceed. */
3753 mutex_unlock(&rsp->barrier_mutex);
3757 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3759 void rcu_barrier_bh(void)
3761 _rcu_barrier(&rcu_bh_state);
3763 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3766 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3768 void rcu_barrier_sched(void)
3770 _rcu_barrier(&rcu_sched_state);
3772 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3775 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3776 * first CPU in a given leaf rcu_node structure coming online. The caller
3777 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3780 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3783 struct rcu_node *rnp = rnp_leaf;
3786 mask = rnp->grpmask;
3790 raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */
3791 rnp->qsmaskinit |= mask;
3792 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
3797 * Do boot-time initialization of a CPU's per-CPU RCU data.
3800 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3802 unsigned long flags;
3803 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3804 struct rcu_node *rnp = rcu_get_root(rsp);
3806 /* Set up local state, ensuring consistent view of global state. */
3807 raw_spin_lock_irqsave(&rnp->lock, flags);
3808 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3809 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3810 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3811 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3814 mutex_init(&rdp->exp_funnel_mutex);
3815 rcu_boot_init_nocb_percpu_data(rdp);
3816 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3820 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3821 * offline event can be happening at a given time. Note also that we
3822 * can accept some slop in the rsp->completed access due to the fact
3823 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3826 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3828 unsigned long flags;
3830 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3831 struct rcu_node *rnp = rcu_get_root(rsp);
3833 /* Set up local state, ensuring consistent view of global state. */
3834 raw_spin_lock_irqsave(&rnp->lock, flags);
3835 rdp->beenonline = 1; /* We have now been online. */
3836 rdp->qlen_last_fqs_check = 0;
3837 rdp->n_force_qs_snap = rsp->n_force_qs;
3838 rdp->blimit = blimit;
3840 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3841 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3842 rcu_sysidle_init_percpu_data(rdp->dynticks);
3843 atomic_set(&rdp->dynticks->dynticks,
3844 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3845 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3848 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3849 * propagation up the rcu_node tree will happen at the beginning
3850 * of the next grace period.
3853 mask = rdp->grpmask;
3854 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3855 smp_mb__after_unlock_lock();
3856 rnp->qsmaskinitnext |= mask;
3857 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3858 rdp->completed = rnp->completed;
3859 rdp->passed_quiesce = false;
3860 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
3861 rdp->qs_pending = false;
3862 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3863 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3866 static void rcu_prepare_cpu(int cpu)
3868 struct rcu_state *rsp;
3870 for_each_rcu_flavor(rsp)
3871 rcu_init_percpu_data(cpu, rsp);
3875 * Handle CPU online/offline notification events.
3877 int rcu_cpu_notify(struct notifier_block *self,
3878 unsigned long action, void *hcpu)
3880 long cpu = (long)hcpu;
3881 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3882 struct rcu_node *rnp = rdp->mynode;
3883 struct rcu_state *rsp;
3886 case CPU_UP_PREPARE:
3887 case CPU_UP_PREPARE_FROZEN:
3888 rcu_prepare_cpu(cpu);
3889 rcu_prepare_kthreads(cpu);
3890 rcu_spawn_all_nocb_kthreads(cpu);
3893 case CPU_DOWN_FAILED:
3894 rcu_boost_kthread_setaffinity(rnp, -1);
3896 case CPU_DOWN_PREPARE:
3897 rcu_boost_kthread_setaffinity(rnp, cpu);
3900 case CPU_DYING_FROZEN:
3901 for_each_rcu_flavor(rsp)
3902 rcu_cleanup_dying_cpu(rsp);
3904 case CPU_DYING_IDLE:
3905 for_each_rcu_flavor(rsp) {
3906 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3910 case CPU_DEAD_FROZEN:
3911 case CPU_UP_CANCELED:
3912 case CPU_UP_CANCELED_FROZEN:
3913 for_each_rcu_flavor(rsp) {
3914 rcu_cleanup_dead_cpu(cpu, rsp);
3915 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3924 static int rcu_pm_notify(struct notifier_block *self,
3925 unsigned long action, void *hcpu)
3928 case PM_HIBERNATION_PREPARE:
3929 case PM_SUSPEND_PREPARE:
3930 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3933 case PM_POST_HIBERNATION:
3934 case PM_POST_SUSPEND:
3935 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3936 rcu_unexpedite_gp();
3945 * Spawn the kthreads that handle each RCU flavor's grace periods.
3947 static int __init rcu_spawn_gp_kthread(void)
3949 unsigned long flags;
3950 int kthread_prio_in = kthread_prio;
3951 struct rcu_node *rnp;
3952 struct rcu_state *rsp;
3953 struct sched_param sp;
3954 struct task_struct *t;
3956 /* Force priority into range. */
3957 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3959 else if (kthread_prio < 0)
3961 else if (kthread_prio > 99)
3963 if (kthread_prio != kthread_prio_in)
3964 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3965 kthread_prio, kthread_prio_in);
3967 rcu_scheduler_fully_active = 1;
3968 for_each_rcu_flavor(rsp) {
3969 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3971 rnp = rcu_get_root(rsp);
3972 raw_spin_lock_irqsave(&rnp->lock, flags);
3973 rsp->gp_kthread = t;
3975 sp.sched_priority = kthread_prio;
3976 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3979 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3981 rcu_spawn_nocb_kthreads();
3982 rcu_spawn_boost_kthreads();
3985 early_initcall(rcu_spawn_gp_kthread);
3988 * This function is invoked towards the end of the scheduler's initialization
3989 * process. Before this is called, the idle task might contain
3990 * RCU read-side critical sections (during which time, this idle
3991 * task is booting the system). After this function is called, the
3992 * idle tasks are prohibited from containing RCU read-side critical
3993 * sections. This function also enables RCU lockdep checking.
3995 void rcu_scheduler_starting(void)
3997 WARN_ON(num_online_cpus() != 1);
3998 WARN_ON(nr_context_switches() > 0);
3999 rcu_scheduler_active = 1;
4003 * Compute the per-level fanout, either using the exact fanout specified
4004 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4006 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4010 if (rcu_fanout_exact) {
4011 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4012 for (i = rcu_num_lvls - 2; i >= 0; i--)
4013 levelspread[i] = RCU_FANOUT;
4019 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4021 levelspread[i] = (cprv + ccur - 1) / ccur;
4028 * Helper function for rcu_init() that initializes one rcu_state structure.
4030 static void __init rcu_init_one(struct rcu_state *rsp,
4031 struct rcu_data __percpu *rda)
4033 static const char * const buf[] = RCU_NODE_NAME_INIT;
4034 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4035 static const char * const exp[] = RCU_EXP_NAME_INIT;
4036 static u8 fl_mask = 0x1;
4038 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4039 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4043 struct rcu_node *rnp;
4045 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4047 /* Silence gcc 4.8 false positive about array index out of range. */
4048 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4049 panic("rcu_init_one: rcu_num_lvls out of range");
4051 /* Initialize the level-tracking arrays. */
4053 for (i = 0; i < rcu_num_lvls; i++)
4054 levelcnt[i] = num_rcu_lvl[i];
4055 for (i = 1; i < rcu_num_lvls; i++)
4056 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4057 rcu_init_levelspread(levelspread, levelcnt);
4058 rsp->flavor_mask = fl_mask;
4061 /* Initialize the elements themselves, starting from the leaves. */
4063 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4064 cpustride *= levelspread[i];
4065 rnp = rsp->level[i];
4066 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4067 raw_spin_lock_init(&rnp->lock);
4068 lockdep_set_class_and_name(&rnp->lock,
4069 &rcu_node_class[i], buf[i]);
4070 raw_spin_lock_init(&rnp->fqslock);
4071 lockdep_set_class_and_name(&rnp->fqslock,
4072 &rcu_fqs_class[i], fqs[i]);
4073 rnp->gpnum = rsp->gpnum;
4074 rnp->completed = rsp->completed;
4076 rnp->qsmaskinit = 0;
4077 rnp->grplo = j * cpustride;
4078 rnp->grphi = (j + 1) * cpustride - 1;
4079 if (rnp->grphi >= nr_cpu_ids)
4080 rnp->grphi = nr_cpu_ids - 1;
4086 rnp->grpnum = j % levelspread[i - 1];
4087 rnp->grpmask = 1UL << rnp->grpnum;
4088 rnp->parent = rsp->level[i - 1] +
4089 j / levelspread[i - 1];
4092 INIT_LIST_HEAD(&rnp->blkd_tasks);
4093 rcu_init_one_nocb(rnp);
4094 mutex_init(&rnp->exp_funnel_mutex);
4095 lockdep_set_class_and_name(&rnp->exp_funnel_mutex,
4096 &rcu_exp_class[i], exp[i]);
4100 init_waitqueue_head(&rsp->gp_wq);
4101 rnp = rsp->level[rcu_num_lvls - 1];
4102 for_each_possible_cpu(i) {
4103 while (i > rnp->grphi)
4105 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4106 rcu_boot_init_percpu_data(i, rsp);
4108 list_add(&rsp->flavors, &rcu_struct_flavors);
4112 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4113 * replace the definitions in tree.h because those are needed to size
4114 * the ->node array in the rcu_state structure.
4116 static void __init rcu_init_geometry(void)
4120 int rcu_capacity[RCU_NUM_LVLS];
4123 * Initialize any unspecified boot parameters.
4124 * The default values of jiffies_till_first_fqs and
4125 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4126 * value, which is a function of HZ, then adding one for each
4127 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4129 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4130 if (jiffies_till_first_fqs == ULONG_MAX)
4131 jiffies_till_first_fqs = d;
4132 if (jiffies_till_next_fqs == ULONG_MAX)
4133 jiffies_till_next_fqs = d;
4135 /* If the compile-time values are accurate, just leave. */
4136 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4137 nr_cpu_ids == NR_CPUS)
4139 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4140 rcu_fanout_leaf, nr_cpu_ids);
4143 * The boot-time rcu_fanout_leaf parameter is only permitted
4144 * to increase the leaf-level fanout, not decrease it. Of course,
4145 * the leaf-level fanout cannot exceed the number of bits in
4146 * the rcu_node masks. Complain and fall back to the compile-
4147 * time values if these limits are exceeded.
4149 if (rcu_fanout_leaf < RCU_FANOUT_LEAF ||
4150 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4151 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4157 * Compute number of nodes that can be handled an rcu_node tree
4158 * with the given number of levels.
4160 rcu_capacity[0] = rcu_fanout_leaf;
4161 for (i = 1; i < RCU_NUM_LVLS; i++)
4162 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4165 * The tree must be able to accommodate the configured number of CPUs.
4166 * If this limit is exceeded than we have a serious problem elsewhere.
4168 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1])
4169 panic("rcu_init_geometry: rcu_capacity[] is too small");
4171 /* Calculate the number of levels in the tree. */
4172 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4174 rcu_num_lvls = i + 1;
4176 /* Calculate the number of rcu_nodes at each level of the tree. */
4177 for (i = 0; i < rcu_num_lvls; i++) {
4178 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4179 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4182 /* Calculate the total number of rcu_node structures. */
4184 for (i = 0; i < rcu_num_lvls; i++)
4185 rcu_num_nodes += num_rcu_lvl[i];
4189 * Dump out the structure of the rcu_node combining tree associated
4190 * with the rcu_state structure referenced by rsp.
4192 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4195 struct rcu_node *rnp;
4197 pr_info("rcu_node tree layout dump\n");
4199 rcu_for_each_node_breadth_first(rsp, rnp) {
4200 if (rnp->level != level) {
4205 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4210 void __init rcu_init(void)
4214 rcu_early_boot_tests();
4216 rcu_bootup_announce();
4217 rcu_init_geometry();
4218 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
4219 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
4221 rcu_dump_rcu_node_tree(&rcu_sched_state);
4222 __rcu_init_preempt();
4223 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4226 * We don't need protection against CPU-hotplug here because
4227 * this is called early in boot, before either interrupts
4228 * or the scheduler are operational.
4230 cpu_notifier(rcu_cpu_notify, 0);
4231 pm_notifier(rcu_pm_notify, 0);
4232 for_each_online_cpu(cpu)
4233 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
4236 #include "tree_plugin.h"