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];
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
83 # define DEFINE_RCU_TPS(sname) \
84 static char sname##_varname[] = #sname; \
85 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
86 # define RCU_STATE_NAME(sname) sname##_varname
88 # define DEFINE_RCU_TPS(sname)
89 # define RCU_STATE_NAME(sname) __stringify(sname)
92 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
93 DEFINE_RCU_TPS(sname) \
94 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
95 struct rcu_state sname##_state = { \
96 .level = { &sname##_state.node[0] }, \
97 .rda = &sname##_data, \
99 .fqs_state = RCU_GP_IDLE, \
100 .gpnum = 0UL - 300UL, \
101 .completed = 0UL - 300UL, \
102 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
103 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
104 .orphan_donetail = &sname##_state.orphan_donelist, \
105 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
106 .name = RCU_STATE_NAME(sname), \
110 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
111 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
113 static struct rcu_state *const rcu_state_p;
114 static struct rcu_data __percpu *const rcu_data_p;
115 LIST_HEAD(rcu_struct_flavors);
117 /* Dump rcu_node combining tree at boot to verify correct setup. */
118 static bool dump_tree;
119 module_param(dump_tree, bool, 0444);
120 /* Control rcu_node-tree auto-balancing at boot time. */
121 static bool rcu_fanout_exact;
122 module_param(rcu_fanout_exact, bool, 0444);
123 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
124 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
125 module_param(rcu_fanout_leaf, int, 0444);
126 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
127 static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
134 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
137 * The rcu_scheduler_active variable transitions from zero to one just
138 * before the first task is spawned. So when this variable is zero, RCU
139 * can assume that there is but one task, allowing RCU to (for example)
140 * optimize synchronize_sched() to a simple barrier(). When this variable
141 * is one, RCU must actually do all the hard work required to detect real
142 * grace periods. This variable is also used to suppress boot-time false
143 * positives from lockdep-RCU error checking.
145 int rcu_scheduler_active __read_mostly;
146 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
149 * The rcu_scheduler_fully_active variable transitions from zero to one
150 * during the early_initcall() processing, which is after the scheduler
151 * is capable of creating new tasks. So RCU processing (for example,
152 * creating tasks for RCU priority boosting) must be delayed until after
153 * rcu_scheduler_fully_active transitions from zero to one. We also
154 * currently delay invocation of any RCU callbacks until after this point.
156 * It might later prove better for people registering RCU callbacks during
157 * early boot to take responsibility for these callbacks, but one step at
160 static int rcu_scheduler_fully_active __read_mostly;
162 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
163 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
164 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
165 static void invoke_rcu_core(void);
166 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
168 /* rcuc/rcub kthread realtime priority */
169 #ifdef CONFIG_RCU_KTHREAD_PRIO
170 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
171 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
172 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
173 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
174 module_param(kthread_prio, int, 0644);
176 /* Delay in jiffies for grace-period initialization delays, debug only. */
178 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
179 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
180 module_param(gp_preinit_delay, int, 0644);
181 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
182 static const int gp_preinit_delay;
183 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
185 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
186 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
187 module_param(gp_init_delay, int, 0644);
188 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
189 static const int gp_init_delay;
190 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
192 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
193 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
194 module_param(gp_cleanup_delay, int, 0644);
195 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
196 static const int gp_cleanup_delay;
197 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
200 * Number of grace periods between delays, normalized by the duration of
201 * the delay. The longer the the delay, the more the grace periods between
202 * each delay. The reason for this normalization is that it means that,
203 * for non-zero delays, the overall slowdown of grace periods is constant
204 * regardless of the duration of the delay. This arrangement balances
205 * the need for long delays to increase some race probabilities with the
206 * need for fast grace periods to increase other race probabilities.
208 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
211 * Track the rcutorture test sequence number and the update version
212 * number within a given test. The rcutorture_testseq is incremented
213 * on every rcutorture module load and unload, so has an odd value
214 * when a test is running. The rcutorture_vernum is set to zero
215 * when rcutorture starts and is incremented on each rcutorture update.
216 * These variables enable correlating rcutorture output with the
217 * RCU tracing information.
219 unsigned long rcutorture_testseq;
220 unsigned long rcutorture_vernum;
223 * Compute the mask of online CPUs for the specified rcu_node structure.
224 * This will not be stable unless the rcu_node structure's ->lock is
225 * held, but the bit corresponding to the current CPU will be stable
228 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
230 return READ_ONCE(rnp->qsmaskinitnext);
234 * Return true if an RCU grace period is in progress. The READ_ONCE()s
235 * permit this function to be invoked without holding the root rcu_node
236 * structure's ->lock, but of course results can be subject to change.
238 static int rcu_gp_in_progress(struct rcu_state *rsp)
240 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
244 * Note a quiescent state. Because we do not need to know
245 * how many quiescent states passed, just if there was at least
246 * one since the start of the grace period, this just sets a flag.
247 * The caller must have disabled preemption.
249 void rcu_sched_qs(void)
251 if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) {
252 trace_rcu_grace_period(TPS("rcu_sched"),
253 __this_cpu_read(rcu_sched_data.gpnum),
255 __this_cpu_write(rcu_sched_data.passed_quiesce, 1);
261 if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) {
262 trace_rcu_grace_period(TPS("rcu_bh"),
263 __this_cpu_read(rcu_bh_data.gpnum),
265 __this_cpu_write(rcu_bh_data.passed_quiesce, 1);
269 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
271 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
272 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
273 .dynticks = ATOMIC_INIT(1),
274 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
275 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
276 .dynticks_idle = ATOMIC_INIT(1),
277 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
280 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
281 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
284 * Let the RCU core know that this CPU has gone through the scheduler,
285 * which is a quiescent state. This is called when the need for a
286 * quiescent state is urgent, so we burn an atomic operation and full
287 * memory barriers to let the RCU core know about it, regardless of what
288 * this CPU might (or might not) do in the near future.
290 * We inform the RCU core by emulating a zero-duration dyntick-idle
291 * period, which we in turn do by incrementing the ->dynticks counter
294 static void rcu_momentary_dyntick_idle(void)
297 struct rcu_data *rdp;
298 struct rcu_dynticks *rdtp;
300 struct rcu_state *rsp;
302 local_irq_save(flags);
305 * Yes, we can lose flag-setting operations. This is OK, because
306 * the flag will be set again after some delay.
308 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
309 raw_cpu_write(rcu_sched_qs_mask, 0);
311 /* Find the flavor that needs a quiescent state. */
312 for_each_rcu_flavor(rsp) {
313 rdp = raw_cpu_ptr(rsp->rda);
314 if (!(resched_mask & rsp->flavor_mask))
316 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
317 if (READ_ONCE(rdp->mynode->completed) !=
318 READ_ONCE(rdp->cond_resched_completed))
322 * Pretend to be momentarily idle for the quiescent state.
323 * This allows the grace-period kthread to record the
324 * quiescent state, with no need for this CPU to do anything
327 rdtp = this_cpu_ptr(&rcu_dynticks);
328 smp_mb__before_atomic(); /* Earlier stuff before QS. */
329 atomic_add(2, &rdtp->dynticks); /* QS. */
330 smp_mb__after_atomic(); /* Later stuff after QS. */
333 local_irq_restore(flags);
337 * Note a context switch. This is a quiescent state for RCU-sched,
338 * and requires special handling for preemptible RCU.
339 * The caller must have disabled preemption.
341 void rcu_note_context_switch(void)
343 trace_rcu_utilization(TPS("Start context switch"));
345 rcu_preempt_note_context_switch();
346 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
347 rcu_momentary_dyntick_idle();
348 trace_rcu_utilization(TPS("End context switch"));
350 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
353 * Register a quiescent state for all RCU flavors. If there is an
354 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
355 * dyntick-idle quiescent state visible to other CPUs (but only for those
356 * RCU flavors in desperate need of a quiescent state, which will normally
357 * be none of them). Either way, do a lightweight quiescent state for
360 void rcu_all_qs(void)
362 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
363 rcu_momentary_dyntick_idle();
364 this_cpu_inc(rcu_qs_ctr);
366 EXPORT_SYMBOL_GPL(rcu_all_qs);
368 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
369 static long qhimark = 10000; /* If this many pending, ignore blimit. */
370 static long qlowmark = 100; /* Once only this many pending, use blimit. */
372 module_param(blimit, long, 0444);
373 module_param(qhimark, long, 0444);
374 module_param(qlowmark, long, 0444);
376 static ulong jiffies_till_first_fqs = ULONG_MAX;
377 static ulong jiffies_till_next_fqs = ULONG_MAX;
379 module_param(jiffies_till_first_fqs, ulong, 0644);
380 module_param(jiffies_till_next_fqs, ulong, 0644);
383 * How long the grace period must be before we start recruiting
384 * quiescent-state help from rcu_note_context_switch().
386 static ulong jiffies_till_sched_qs = HZ / 20;
387 module_param(jiffies_till_sched_qs, ulong, 0644);
389 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
390 struct rcu_data *rdp);
391 static void force_qs_rnp(struct rcu_state *rsp,
392 int (*f)(struct rcu_data *rsp, bool *isidle,
393 unsigned long *maxj),
394 bool *isidle, unsigned long *maxj);
395 static void force_quiescent_state(struct rcu_state *rsp);
396 static int rcu_pending(void);
399 * Return the number of RCU batches started thus far for debug & stats.
401 unsigned long rcu_batches_started(void)
403 return rcu_state_p->gpnum;
405 EXPORT_SYMBOL_GPL(rcu_batches_started);
408 * Return the number of RCU-sched batches started thus far for debug & stats.
410 unsigned long rcu_batches_started_sched(void)
412 return rcu_sched_state.gpnum;
414 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
417 * Return the number of RCU BH batches started thus far for debug & stats.
419 unsigned long rcu_batches_started_bh(void)
421 return rcu_bh_state.gpnum;
423 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
426 * Return the number of RCU batches completed thus far for debug & stats.
428 unsigned long rcu_batches_completed(void)
430 return rcu_state_p->completed;
432 EXPORT_SYMBOL_GPL(rcu_batches_completed);
435 * Return the number of RCU-sched batches completed thus far for debug & stats.
437 unsigned long rcu_batches_completed_sched(void)
439 return rcu_sched_state.completed;
441 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
444 * Return the number of RCU BH batches completed thus far for debug & stats.
446 unsigned long rcu_batches_completed_bh(void)
448 return rcu_bh_state.completed;
450 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
453 * Force a quiescent state.
455 void rcu_force_quiescent_state(void)
457 force_quiescent_state(rcu_state_p);
459 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
462 * Force a quiescent state for RCU BH.
464 void rcu_bh_force_quiescent_state(void)
466 force_quiescent_state(&rcu_bh_state);
468 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
471 * Force a quiescent state for RCU-sched.
473 void rcu_sched_force_quiescent_state(void)
475 force_quiescent_state(&rcu_sched_state);
477 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
480 * Show the state of the grace-period kthreads.
482 void show_rcu_gp_kthreads(void)
484 struct rcu_state *rsp;
486 for_each_rcu_flavor(rsp) {
487 pr_info("%s: wait state: %d ->state: %#lx\n",
488 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
489 /* sched_show_task(rsp->gp_kthread); */
492 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
495 * Record the number of times rcutorture tests have been initiated and
496 * terminated. This information allows the debugfs tracing stats to be
497 * correlated to the rcutorture messages, even when the rcutorture module
498 * is being repeatedly loaded and unloaded. In other words, we cannot
499 * store this state in rcutorture itself.
501 void rcutorture_record_test_transition(void)
503 rcutorture_testseq++;
504 rcutorture_vernum = 0;
506 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
509 * Send along grace-period-related data for rcutorture diagnostics.
511 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
512 unsigned long *gpnum, unsigned long *completed)
514 struct rcu_state *rsp = NULL;
523 case RCU_SCHED_FLAVOR:
524 rsp = &rcu_sched_state;
530 *flags = READ_ONCE(rsp->gp_flags);
531 *gpnum = READ_ONCE(rsp->gpnum);
532 *completed = READ_ONCE(rsp->completed);
539 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
542 * Record the number of writer passes through the current rcutorture test.
543 * This is also used to correlate debugfs tracing stats with the rcutorture
546 void rcutorture_record_progress(unsigned long vernum)
550 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
553 * Does the CPU have callbacks ready to be invoked?
556 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
558 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
559 rdp->nxttail[RCU_DONE_TAIL] != NULL;
563 * Return the root node of the specified rcu_state structure.
565 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
567 return &rsp->node[0];
571 * Is there any need for future grace periods?
572 * Interrupts must be disabled. If the caller does not hold the root
573 * rnp_node structure's ->lock, the results are advisory only.
575 static int rcu_future_needs_gp(struct rcu_state *rsp)
577 struct rcu_node *rnp = rcu_get_root(rsp);
578 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
579 int *fp = &rnp->need_future_gp[idx];
581 return READ_ONCE(*fp);
585 * Does the current CPU require a not-yet-started grace period?
586 * The caller must have disabled interrupts to prevent races with
587 * normal callback registry.
590 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
594 if (rcu_gp_in_progress(rsp))
595 return 0; /* No, a grace period is already in progress. */
596 if (rcu_future_needs_gp(rsp))
597 return 1; /* Yes, a no-CBs CPU needs one. */
598 if (!rdp->nxttail[RCU_NEXT_TAIL])
599 return 0; /* No, this is a no-CBs (or offline) CPU. */
600 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
601 return 1; /* Yes, this CPU has newly registered callbacks. */
602 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
603 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
604 ULONG_CMP_LT(READ_ONCE(rsp->completed),
605 rdp->nxtcompleted[i]))
606 return 1; /* Yes, CBs for future grace period. */
607 return 0; /* No grace period needed. */
611 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
613 * If the new value of the ->dynticks_nesting counter now is zero,
614 * we really have entered idle, and must do the appropriate accounting.
615 * The caller must have disabled interrupts.
617 static void rcu_eqs_enter_common(long long oldval, bool user)
619 struct rcu_state *rsp;
620 struct rcu_data *rdp;
621 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
623 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
624 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
625 !user && !is_idle_task(current)) {
626 struct task_struct *idle __maybe_unused =
627 idle_task(smp_processor_id());
629 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
630 ftrace_dump(DUMP_ORIG);
631 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
632 current->pid, current->comm,
633 idle->pid, idle->comm); /* must be idle task! */
635 for_each_rcu_flavor(rsp) {
636 rdp = this_cpu_ptr(rsp->rda);
637 do_nocb_deferred_wakeup(rdp);
639 rcu_prepare_for_idle();
640 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
641 smp_mb__before_atomic(); /* See above. */
642 atomic_inc(&rdtp->dynticks);
643 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
644 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
645 atomic_read(&rdtp->dynticks) & 0x1);
646 rcu_dynticks_task_enter();
649 * It is illegal to enter an extended quiescent state while
650 * in an RCU read-side critical section.
652 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
653 "Illegal idle entry in RCU read-side critical section.");
654 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
655 "Illegal idle entry in RCU-bh read-side critical section.");
656 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
657 "Illegal idle entry in RCU-sched read-side critical section.");
661 * Enter an RCU extended quiescent state, which can be either the
662 * idle loop or adaptive-tickless usermode execution.
664 static void rcu_eqs_enter(bool user)
667 struct rcu_dynticks *rdtp;
669 rdtp = this_cpu_ptr(&rcu_dynticks);
670 oldval = rdtp->dynticks_nesting;
671 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
672 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
673 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
674 rdtp->dynticks_nesting = 0;
675 rcu_eqs_enter_common(oldval, user);
677 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
682 * rcu_idle_enter - inform RCU that current CPU is entering idle
684 * Enter idle mode, in other words, -leave- the mode in which RCU
685 * read-side critical sections can occur. (Though RCU read-side
686 * critical sections can occur in irq handlers in idle, a possibility
687 * handled by irq_enter() and irq_exit().)
689 * We crowbar the ->dynticks_nesting field to zero to allow for
690 * the possibility of usermode upcalls having messed up our count
691 * of interrupt nesting level during the prior busy period.
693 void rcu_idle_enter(void)
697 local_irq_save(flags);
698 rcu_eqs_enter(false);
699 rcu_sysidle_enter(0);
700 local_irq_restore(flags);
702 EXPORT_SYMBOL_GPL(rcu_idle_enter);
704 #ifdef CONFIG_NO_HZ_FULL
706 * rcu_user_enter - inform RCU that we are resuming userspace.
708 * Enter RCU idle mode right before resuming userspace. No use of RCU
709 * is permitted between this call and rcu_user_exit(). This way the
710 * CPU doesn't need to maintain the tick for RCU maintenance purposes
711 * when the CPU runs in userspace.
713 void rcu_user_enter(void)
717 #endif /* CONFIG_NO_HZ_FULL */
720 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
722 * Exit from an interrupt handler, which might possibly result in entering
723 * idle mode, in other words, leaving the mode in which read-side critical
724 * sections can occur.
726 * This code assumes that the idle loop never does anything that might
727 * result in unbalanced calls to irq_enter() and irq_exit(). If your
728 * architecture violates this assumption, RCU will give you what you
729 * deserve, good and hard. But very infrequently and irreproducibly.
731 * Use things like work queues to work around this limitation.
733 * You have been warned.
735 void rcu_irq_exit(void)
739 struct rcu_dynticks *rdtp;
741 local_irq_save(flags);
742 rdtp = this_cpu_ptr(&rcu_dynticks);
743 oldval = rdtp->dynticks_nesting;
744 rdtp->dynticks_nesting--;
745 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
746 rdtp->dynticks_nesting < 0);
747 if (rdtp->dynticks_nesting)
748 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
750 rcu_eqs_enter_common(oldval, true);
751 rcu_sysidle_enter(1);
752 local_irq_restore(flags);
756 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
758 * If the new value of the ->dynticks_nesting counter was previously zero,
759 * we really have exited idle, and must do the appropriate accounting.
760 * The caller must have disabled interrupts.
762 static void rcu_eqs_exit_common(long long oldval, int user)
764 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
766 rcu_dynticks_task_exit();
767 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
768 atomic_inc(&rdtp->dynticks);
769 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
770 smp_mb__after_atomic(); /* See above. */
771 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
772 !(atomic_read(&rdtp->dynticks) & 0x1));
773 rcu_cleanup_after_idle();
774 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
775 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
776 !user && !is_idle_task(current)) {
777 struct task_struct *idle __maybe_unused =
778 idle_task(smp_processor_id());
780 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
781 oldval, rdtp->dynticks_nesting);
782 ftrace_dump(DUMP_ORIG);
783 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
784 current->pid, current->comm,
785 idle->pid, idle->comm); /* must be idle task! */
790 * Exit an RCU extended quiescent state, which can be either the
791 * idle loop or adaptive-tickless usermode execution.
793 static void rcu_eqs_exit(bool user)
795 struct rcu_dynticks *rdtp;
798 rdtp = this_cpu_ptr(&rcu_dynticks);
799 oldval = rdtp->dynticks_nesting;
800 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
801 if (oldval & DYNTICK_TASK_NEST_MASK) {
802 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
804 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
805 rcu_eqs_exit_common(oldval, user);
810 * rcu_idle_exit - inform RCU that current CPU is leaving idle
812 * Exit idle mode, in other words, -enter- the mode in which RCU
813 * read-side critical sections can occur.
815 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
816 * allow for the possibility of usermode upcalls messing up our count
817 * of interrupt nesting level during the busy period that is just
820 void rcu_idle_exit(void)
824 local_irq_save(flags);
827 local_irq_restore(flags);
829 EXPORT_SYMBOL_GPL(rcu_idle_exit);
831 #ifdef CONFIG_NO_HZ_FULL
833 * rcu_user_exit - inform RCU that we are exiting userspace.
835 * Exit RCU idle mode while entering the kernel because it can
836 * run a RCU read side critical section anytime.
838 void rcu_user_exit(void)
842 #endif /* CONFIG_NO_HZ_FULL */
845 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
847 * Enter an interrupt handler, which might possibly result in exiting
848 * idle mode, in other words, entering the mode in which read-side critical
849 * sections can occur.
851 * Note that the Linux kernel is fully capable of entering an interrupt
852 * handler that it never exits, for example when doing upcalls to
853 * user mode! This code assumes that the idle loop never does upcalls to
854 * user mode. If your architecture does do upcalls from the idle loop (or
855 * does anything else that results in unbalanced calls to the irq_enter()
856 * and irq_exit() functions), RCU will give you what you deserve, good
857 * and hard. But very infrequently and irreproducibly.
859 * Use things like work queues to work around this limitation.
861 * You have been warned.
863 void rcu_irq_enter(void)
866 struct rcu_dynticks *rdtp;
869 local_irq_save(flags);
870 rdtp = this_cpu_ptr(&rcu_dynticks);
871 oldval = rdtp->dynticks_nesting;
872 rdtp->dynticks_nesting++;
873 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
874 rdtp->dynticks_nesting == 0);
876 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
878 rcu_eqs_exit_common(oldval, true);
880 local_irq_restore(flags);
884 * rcu_nmi_enter - inform RCU of entry to NMI context
886 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
887 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
888 * that the CPU is active. This implementation permits nested NMIs, as
889 * long as the nesting level does not overflow an int. (You will probably
890 * run out of stack space first.)
892 void rcu_nmi_enter(void)
894 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
897 /* Complain about underflow. */
898 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
901 * If idle from RCU viewpoint, atomically increment ->dynticks
902 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
903 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
904 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
905 * to be in the outermost NMI handler that interrupted an RCU-idle
906 * period (observation due to Andy Lutomirski).
908 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
909 smp_mb__before_atomic(); /* Force delay from prior write. */
910 atomic_inc(&rdtp->dynticks);
911 /* atomic_inc() before later RCU read-side crit sects */
912 smp_mb__after_atomic(); /* See above. */
913 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
916 rdtp->dynticks_nmi_nesting += incby;
921 * rcu_nmi_exit - inform RCU of exit from NMI context
923 * If we are returning from the outermost NMI handler that interrupted an
924 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
925 * to let the RCU grace-period handling know that the CPU is back to
928 void rcu_nmi_exit(void)
930 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
933 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
934 * (We are exiting an NMI handler, so RCU better be paying attention
937 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
938 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
941 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
942 * leave it in non-RCU-idle state.
944 if (rdtp->dynticks_nmi_nesting != 1) {
945 rdtp->dynticks_nmi_nesting -= 2;
949 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
950 rdtp->dynticks_nmi_nesting = 0;
951 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
952 smp_mb__before_atomic(); /* See above. */
953 atomic_inc(&rdtp->dynticks);
954 smp_mb__after_atomic(); /* Force delay to next write. */
955 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
959 * __rcu_is_watching - are RCU read-side critical sections safe?
961 * Return true if RCU is watching the running CPU, which means that
962 * this CPU can safely enter RCU read-side critical sections. Unlike
963 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
964 * least disabled preemption.
966 bool notrace __rcu_is_watching(void)
968 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
972 * rcu_is_watching - see if RCU thinks that the current CPU is idle
974 * If the current CPU is in its idle loop and is neither in an interrupt
975 * or NMI handler, return true.
977 bool notrace rcu_is_watching(void)
981 preempt_disable_notrace();
982 ret = __rcu_is_watching();
983 preempt_enable_notrace();
986 EXPORT_SYMBOL_GPL(rcu_is_watching);
988 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
991 * Is the current CPU online? Disable preemption to avoid false positives
992 * that could otherwise happen due to the current CPU number being sampled,
993 * this task being preempted, its old CPU being taken offline, resuming
994 * on some other CPU, then determining that its old CPU is now offline.
995 * It is OK to use RCU on an offline processor during initial boot, hence
996 * the check for rcu_scheduler_fully_active. Note also that it is OK
997 * for a CPU coming online to use RCU for one jiffy prior to marking itself
998 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
999 * offline to continue to use RCU for one jiffy after marking itself
1000 * offline in the cpu_online_mask. This leniency is necessary given the
1001 * non-atomic nature of the online and offline processing, for example,
1002 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1005 * This is also why RCU internally marks CPUs online during the
1006 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1008 * Disable checking if in an NMI handler because we cannot safely report
1009 * errors from NMI handlers anyway.
1011 bool rcu_lockdep_current_cpu_online(void)
1013 struct rcu_data *rdp;
1014 struct rcu_node *rnp;
1020 rdp = this_cpu_ptr(&rcu_sched_data);
1022 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1023 !rcu_scheduler_fully_active;
1027 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1029 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1032 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1034 * If the current CPU is idle or running at a first-level (not nested)
1035 * interrupt from idle, return true. The caller must have at least
1036 * disabled preemption.
1038 static int rcu_is_cpu_rrupt_from_idle(void)
1040 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1044 * Snapshot the specified CPU's dynticks counter so that we can later
1045 * credit them with an implicit quiescent state. Return 1 if this CPU
1046 * is in dynticks idle mode, which is an extended quiescent state.
1048 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1049 bool *isidle, unsigned long *maxj)
1051 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1052 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1053 if ((rdp->dynticks_snap & 0x1) == 0) {
1054 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1057 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1058 rdp->mynode->gpnum))
1059 WRITE_ONCE(rdp->gpwrap, true);
1065 * Return true if the specified CPU has passed through a quiescent
1066 * state by virtue of being in or having passed through an dynticks
1067 * idle state since the last call to dyntick_save_progress_counter()
1068 * for this same CPU, or by virtue of having been offline.
1070 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1071 bool *isidle, unsigned long *maxj)
1077 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1078 snap = (unsigned int)rdp->dynticks_snap;
1081 * If the CPU passed through or entered a dynticks idle phase with
1082 * no active irq/NMI handlers, then we can safely pretend that the CPU
1083 * already acknowledged the request to pass through a quiescent
1084 * state. Either way, that CPU cannot possibly be in an RCU
1085 * read-side critical section that started before the beginning
1086 * of the current RCU grace period.
1088 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1089 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1090 rdp->dynticks_fqs++;
1095 * Check for the CPU being offline, but only if the grace period
1096 * is old enough. We don't need to worry about the CPU changing
1097 * state: If we see it offline even once, it has been through a
1100 * The reason for insisting that the grace period be at least
1101 * one jiffy old is that CPUs that are not quite online and that
1102 * have just gone offline can still execute RCU read-side critical
1105 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1106 return 0; /* Grace period is not old enough. */
1108 if (cpu_is_offline(rdp->cpu)) {
1109 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1115 * A CPU running for an extended time within the kernel can
1116 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1117 * even context-switching back and forth between a pair of
1118 * in-kernel CPU-bound tasks cannot advance grace periods.
1119 * So if the grace period is old enough, make the CPU pay attention.
1120 * Note that the unsynchronized assignments to the per-CPU
1121 * rcu_sched_qs_mask variable are safe. Yes, setting of
1122 * bits can be lost, but they will be set again on the next
1123 * force-quiescent-state pass. So lost bit sets do not result
1124 * in incorrect behavior, merely in a grace period lasting
1125 * a few jiffies longer than it might otherwise. Because
1126 * there are at most four threads involved, and because the
1127 * updates are only once every few jiffies, the probability of
1128 * lossage (and thus of slight grace-period extension) is
1131 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1132 * is set too high, we override with half of the RCU CPU stall
1135 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1136 if (ULONG_CMP_GE(jiffies,
1137 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1138 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1139 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1140 WRITE_ONCE(rdp->cond_resched_completed,
1141 READ_ONCE(rdp->mynode->completed));
1142 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1144 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1145 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1146 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
1147 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1148 /* Time to beat on that CPU again! */
1149 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1150 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1157 static void record_gp_stall_check_time(struct rcu_state *rsp)
1159 unsigned long j = jiffies;
1163 smp_wmb(); /* Record start time before stall time. */
1164 j1 = rcu_jiffies_till_stall_check();
1165 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1166 rsp->jiffies_resched = j + j1 / 2;
1167 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1171 * Complain about starvation of grace-period kthread.
1173 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1179 gpa = READ_ONCE(rsp->gp_activity);
1180 if (j - gpa > 2 * HZ)
1181 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x\n",
1183 rsp->gpnum, rsp->completed, rsp->gp_flags);
1187 * Dump stacks of all tasks running on stalled CPUs.
1189 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1192 unsigned long flags;
1193 struct rcu_node *rnp;
1195 rcu_for_each_leaf_node(rsp, rnp) {
1196 raw_spin_lock_irqsave(&rnp->lock, flags);
1197 if (rnp->qsmask != 0) {
1198 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1199 if (rnp->qsmask & (1UL << cpu))
1200 dump_cpu_task(rnp->grplo + cpu);
1202 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1206 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1210 unsigned long flags;
1214 struct rcu_node *rnp = rcu_get_root(rsp);
1217 /* Only let one CPU complain about others per time interval. */
1219 raw_spin_lock_irqsave(&rnp->lock, flags);
1220 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1221 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1222 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1225 WRITE_ONCE(rsp->jiffies_stall,
1226 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1227 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1230 * OK, time to rat on our buddy...
1231 * See Documentation/RCU/stallwarn.txt for info on how to debug
1232 * RCU CPU stall warnings.
1234 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1236 print_cpu_stall_info_begin();
1237 rcu_for_each_leaf_node(rsp, rnp) {
1238 raw_spin_lock_irqsave(&rnp->lock, flags);
1239 ndetected += rcu_print_task_stall(rnp);
1240 if (rnp->qsmask != 0) {
1241 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1242 if (rnp->qsmask & (1UL << cpu)) {
1243 print_cpu_stall_info(rsp,
1248 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1251 print_cpu_stall_info_end();
1252 for_each_possible_cpu(cpu)
1253 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1254 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1255 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1256 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1258 rcu_dump_cpu_stacks(rsp);
1260 if (READ_ONCE(rsp->gpnum) != gpnum ||
1261 READ_ONCE(rsp->completed) == gpnum) {
1262 pr_err("INFO: Stall ended before state dump start\n");
1265 gpa = READ_ONCE(rsp->gp_activity);
1266 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1267 rsp->name, j - gpa, j, gpa,
1268 jiffies_till_next_fqs,
1269 rcu_get_root(rsp)->qsmask);
1270 /* In this case, the current CPU might be at fault. */
1271 sched_show_task(current);
1275 /* Complain about tasks blocking the grace period. */
1276 rcu_print_detail_task_stall(rsp);
1278 rcu_check_gp_kthread_starvation(rsp);
1280 force_quiescent_state(rsp); /* Kick them all. */
1283 static void print_cpu_stall(struct rcu_state *rsp)
1286 unsigned long flags;
1287 struct rcu_node *rnp = rcu_get_root(rsp);
1291 * OK, time to rat on ourselves...
1292 * See Documentation/RCU/stallwarn.txt for info on how to debug
1293 * RCU CPU stall warnings.
1295 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1296 print_cpu_stall_info_begin();
1297 print_cpu_stall_info(rsp, smp_processor_id());
1298 print_cpu_stall_info_end();
1299 for_each_possible_cpu(cpu)
1300 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1301 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1302 jiffies - rsp->gp_start,
1303 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1305 rcu_check_gp_kthread_starvation(rsp);
1307 rcu_dump_cpu_stacks(rsp);
1309 raw_spin_lock_irqsave(&rnp->lock, flags);
1310 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1311 WRITE_ONCE(rsp->jiffies_stall,
1312 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1313 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1316 * Attempt to revive the RCU machinery by forcing a context switch.
1318 * A context switch would normally allow the RCU state machine to make
1319 * progress and it could be we're stuck in kernel space without context
1320 * switches for an entirely unreasonable amount of time.
1322 resched_cpu(smp_processor_id());
1325 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1327 unsigned long completed;
1328 unsigned long gpnum;
1332 struct rcu_node *rnp;
1334 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1339 * Lots of memory barriers to reject false positives.
1341 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1342 * then rsp->gp_start, and finally rsp->completed. These values
1343 * are updated in the opposite order with memory barriers (or
1344 * equivalent) during grace-period initialization and cleanup.
1345 * Now, a false positive can occur if we get an new value of
1346 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1347 * the memory barriers, the only way that this can happen is if one
1348 * grace period ends and another starts between these two fetches.
1349 * Detect this by comparing rsp->completed with the previous fetch
1352 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1353 * and rsp->gp_start suffice to forestall false positives.
1355 gpnum = READ_ONCE(rsp->gpnum);
1356 smp_rmb(); /* Pick up ->gpnum first... */
1357 js = READ_ONCE(rsp->jiffies_stall);
1358 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1359 gps = READ_ONCE(rsp->gp_start);
1360 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1361 completed = READ_ONCE(rsp->completed);
1362 if (ULONG_CMP_GE(completed, gpnum) ||
1363 ULONG_CMP_LT(j, js) ||
1364 ULONG_CMP_GE(gps, js))
1365 return; /* No stall or GP completed since entering function. */
1367 if (rcu_gp_in_progress(rsp) &&
1368 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1370 /* We haven't checked in, so go dump stack. */
1371 print_cpu_stall(rsp);
1373 } else if (rcu_gp_in_progress(rsp) &&
1374 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1376 /* They had a few time units to dump stack, so complain. */
1377 print_other_cpu_stall(rsp, gpnum);
1382 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1384 * Set the stall-warning timeout way off into the future, thus preventing
1385 * any RCU CPU stall-warning messages from appearing in the current set of
1386 * RCU grace periods.
1388 * The caller must disable hard irqs.
1390 void rcu_cpu_stall_reset(void)
1392 struct rcu_state *rsp;
1394 for_each_rcu_flavor(rsp)
1395 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1399 * Initialize the specified rcu_data structure's default callback list
1400 * to empty. The default callback list is the one that is not used by
1401 * no-callbacks CPUs.
1403 static void init_default_callback_list(struct rcu_data *rdp)
1407 rdp->nxtlist = NULL;
1408 for (i = 0; i < RCU_NEXT_SIZE; i++)
1409 rdp->nxttail[i] = &rdp->nxtlist;
1413 * Initialize the specified rcu_data structure's callback list to empty.
1415 static void init_callback_list(struct rcu_data *rdp)
1417 if (init_nocb_callback_list(rdp))
1419 init_default_callback_list(rdp);
1423 * Determine the value that ->completed will have at the end of the
1424 * next subsequent grace period. This is used to tag callbacks so that
1425 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1426 * been dyntick-idle for an extended period with callbacks under the
1427 * influence of RCU_FAST_NO_HZ.
1429 * The caller must hold rnp->lock with interrupts disabled.
1431 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1432 struct rcu_node *rnp)
1435 * If RCU is idle, we just wait for the next grace period.
1436 * But we can only be sure that RCU is idle if we are looking
1437 * at the root rcu_node structure -- otherwise, a new grace
1438 * period might have started, but just not yet gotten around
1439 * to initializing the current non-root rcu_node structure.
1441 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1442 return rnp->completed + 1;
1445 * Otherwise, wait for a possible partial grace period and
1446 * then the subsequent full grace period.
1448 return rnp->completed + 2;
1452 * Trace-event helper function for rcu_start_future_gp() and
1453 * rcu_nocb_wait_gp().
1455 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1456 unsigned long c, const char *s)
1458 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1459 rnp->completed, c, rnp->level,
1460 rnp->grplo, rnp->grphi, s);
1464 * Start some future grace period, as needed to handle newly arrived
1465 * callbacks. The required future grace periods are recorded in each
1466 * rcu_node structure's ->need_future_gp field. Returns true if there
1467 * is reason to awaken the grace-period kthread.
1469 * The caller must hold the specified rcu_node structure's ->lock.
1471 static bool __maybe_unused
1472 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1473 unsigned long *c_out)
1478 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1481 * Pick up grace-period number for new callbacks. If this
1482 * grace period is already marked as needed, return to the caller.
1484 c = rcu_cbs_completed(rdp->rsp, rnp);
1485 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1486 if (rnp->need_future_gp[c & 0x1]) {
1487 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1492 * If either this rcu_node structure or the root rcu_node structure
1493 * believe that a grace period is in progress, then we must wait
1494 * for the one following, which is in "c". Because our request
1495 * will be noticed at the end of the current grace period, we don't
1496 * need to explicitly start one. We only do the lockless check
1497 * of rnp_root's fields if the current rcu_node structure thinks
1498 * there is no grace period in flight, and because we hold rnp->lock,
1499 * the only possible change is when rnp_root's two fields are
1500 * equal, in which case rnp_root->gpnum might be concurrently
1501 * incremented. But that is OK, as it will just result in our
1502 * doing some extra useless work.
1504 if (rnp->gpnum != rnp->completed ||
1505 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1506 rnp->need_future_gp[c & 0x1]++;
1507 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1512 * There might be no grace period in progress. If we don't already
1513 * hold it, acquire the root rcu_node structure's lock in order to
1514 * start one (if needed).
1516 if (rnp != rnp_root) {
1517 raw_spin_lock(&rnp_root->lock);
1518 smp_mb__after_unlock_lock();
1522 * Get a new grace-period number. If there really is no grace
1523 * period in progress, it will be smaller than the one we obtained
1524 * earlier. Adjust callbacks as needed. Note that even no-CBs
1525 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1527 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1528 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1529 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1530 rdp->nxtcompleted[i] = c;
1533 * If the needed for the required grace period is already
1534 * recorded, trace and leave.
1536 if (rnp_root->need_future_gp[c & 0x1]) {
1537 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1541 /* Record the need for the future grace period. */
1542 rnp_root->need_future_gp[c & 0x1]++;
1544 /* If a grace period is not already in progress, start one. */
1545 if (rnp_root->gpnum != rnp_root->completed) {
1546 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1548 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1549 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1552 if (rnp != rnp_root)
1553 raw_spin_unlock(&rnp_root->lock);
1561 * Clean up any old requests for the just-ended grace period. Also return
1562 * whether any additional grace periods have been requested. Also invoke
1563 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1564 * waiting for this grace period to complete.
1566 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1568 int c = rnp->completed;
1570 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1572 rcu_nocb_gp_cleanup(rsp, rnp);
1573 rnp->need_future_gp[c & 0x1] = 0;
1574 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1575 trace_rcu_future_gp(rnp, rdp, c,
1576 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1581 * Awaken the grace-period kthread for the specified flavor of RCU.
1582 * Don't do a self-awaken, and don't bother awakening when there is
1583 * nothing for the grace-period kthread to do (as in several CPUs
1584 * raced to awaken, and we lost), and finally don't try to awaken
1585 * a kthread that has not yet been created.
1587 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1589 if (current == rsp->gp_kthread ||
1590 !READ_ONCE(rsp->gp_flags) ||
1593 wake_up(&rsp->gp_wq);
1597 * If there is room, assign a ->completed number to any callbacks on
1598 * this CPU that have not already been assigned. Also accelerate any
1599 * callbacks that were previously assigned a ->completed number that has
1600 * since proven to be too conservative, which can happen if callbacks get
1601 * assigned a ->completed number while RCU is idle, but with reference to
1602 * a non-root rcu_node structure. This function is idempotent, so it does
1603 * not hurt to call it repeatedly. Returns an flag saying that we should
1604 * awaken the RCU grace-period kthread.
1606 * The caller must hold rnp->lock with interrupts disabled.
1608 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1609 struct rcu_data *rdp)
1615 /* If the CPU has no callbacks, nothing to do. */
1616 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1620 * Starting from the sublist containing the callbacks most
1621 * recently assigned a ->completed number and working down, find the
1622 * first sublist that is not assignable to an upcoming grace period.
1623 * Such a sublist has something in it (first two tests) and has
1624 * a ->completed number assigned that will complete sooner than
1625 * the ->completed number for newly arrived callbacks (last test).
1627 * The key point is that any later sublist can be assigned the
1628 * same ->completed number as the newly arrived callbacks, which
1629 * means that the callbacks in any of these later sublist can be
1630 * grouped into a single sublist, whether or not they have already
1631 * been assigned a ->completed number.
1633 c = rcu_cbs_completed(rsp, rnp);
1634 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1635 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1636 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1640 * If there are no sublist for unassigned callbacks, leave.
1641 * At the same time, advance "i" one sublist, so that "i" will
1642 * index into the sublist where all the remaining callbacks should
1645 if (++i >= RCU_NEXT_TAIL)
1649 * Assign all subsequent callbacks' ->completed number to the next
1650 * full grace period and group them all in the sublist initially
1653 for (; i <= RCU_NEXT_TAIL; i++) {
1654 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1655 rdp->nxtcompleted[i] = c;
1657 /* Record any needed additional grace periods. */
1658 ret = rcu_start_future_gp(rnp, rdp, NULL);
1660 /* Trace depending on how much we were able to accelerate. */
1661 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1662 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1664 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1669 * Move any callbacks whose grace period has completed to the
1670 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1671 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1672 * sublist. This function is idempotent, so it does not hurt to
1673 * invoke it repeatedly. As long as it is not invoked -too- often...
1674 * Returns true if the RCU grace-period kthread needs to be awakened.
1676 * The caller must hold rnp->lock with interrupts disabled.
1678 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1679 struct rcu_data *rdp)
1683 /* If the CPU has no callbacks, nothing to do. */
1684 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1688 * Find all callbacks whose ->completed numbers indicate that they
1689 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1691 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1692 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1694 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1696 /* Clean up any sublist tail pointers that were misordered above. */
1697 for (j = RCU_WAIT_TAIL; j < i; j++)
1698 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1700 /* Copy down callbacks to fill in empty sublists. */
1701 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1702 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1704 rdp->nxttail[j] = rdp->nxttail[i];
1705 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1708 /* Classify any remaining callbacks. */
1709 return rcu_accelerate_cbs(rsp, rnp, rdp);
1713 * Update CPU-local rcu_data state to record the beginnings and ends of
1714 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1715 * structure corresponding to the current CPU, and must have irqs disabled.
1716 * Returns true if the grace-period kthread needs to be awakened.
1718 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1719 struct rcu_data *rdp)
1723 /* Handle the ends of any preceding grace periods first. */
1724 if (rdp->completed == rnp->completed &&
1725 !unlikely(READ_ONCE(rdp->gpwrap))) {
1727 /* No grace period end, so just accelerate recent callbacks. */
1728 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1732 /* Advance callbacks. */
1733 ret = rcu_advance_cbs(rsp, rnp, rdp);
1735 /* Remember that we saw this grace-period completion. */
1736 rdp->completed = rnp->completed;
1737 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1740 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1742 * If the current grace period is waiting for this CPU,
1743 * set up to detect a quiescent state, otherwise don't
1744 * go looking for one.
1746 rdp->gpnum = rnp->gpnum;
1747 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1748 rdp->passed_quiesce = 0;
1749 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1750 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1751 zero_cpu_stall_ticks(rdp);
1752 WRITE_ONCE(rdp->gpwrap, false);
1757 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1759 unsigned long flags;
1761 struct rcu_node *rnp;
1763 local_irq_save(flags);
1765 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1766 rdp->completed == READ_ONCE(rnp->completed) &&
1767 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1768 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1769 local_irq_restore(flags);
1772 smp_mb__after_unlock_lock();
1773 needwake = __note_gp_changes(rsp, rnp, rdp);
1774 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1776 rcu_gp_kthread_wake(rsp);
1779 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1782 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1783 schedule_timeout_uninterruptible(delay);
1787 * Initialize a new grace period. Return 0 if no grace period required.
1789 static int rcu_gp_init(struct rcu_state *rsp)
1791 unsigned long oldmask;
1792 struct rcu_data *rdp;
1793 struct rcu_node *rnp = rcu_get_root(rsp);
1795 WRITE_ONCE(rsp->gp_activity, jiffies);
1796 raw_spin_lock_irq(&rnp->lock);
1797 smp_mb__after_unlock_lock();
1798 if (!READ_ONCE(rsp->gp_flags)) {
1799 /* Spurious wakeup, tell caller to go back to sleep. */
1800 raw_spin_unlock_irq(&rnp->lock);
1803 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1805 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1807 * Grace period already in progress, don't start another.
1808 * Not supposed to be able to happen.
1810 raw_spin_unlock_irq(&rnp->lock);
1814 /* Advance to a new grace period and initialize state. */
1815 record_gp_stall_check_time(rsp);
1816 /* Record GP times before starting GP, hence smp_store_release(). */
1817 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1818 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1819 raw_spin_unlock_irq(&rnp->lock);
1822 * Apply per-leaf buffered online and offline operations to the
1823 * rcu_node tree. Note that this new grace period need not wait
1824 * for subsequent online CPUs, and that quiescent-state forcing
1825 * will handle subsequent offline CPUs.
1827 rcu_for_each_leaf_node(rsp, rnp) {
1828 rcu_gp_slow(rsp, gp_preinit_delay);
1829 raw_spin_lock_irq(&rnp->lock);
1830 smp_mb__after_unlock_lock();
1831 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1832 !rnp->wait_blkd_tasks) {
1833 /* Nothing to do on this leaf rcu_node structure. */
1834 raw_spin_unlock_irq(&rnp->lock);
1838 /* Record old state, apply changes to ->qsmaskinit field. */
1839 oldmask = rnp->qsmaskinit;
1840 rnp->qsmaskinit = rnp->qsmaskinitnext;
1842 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1843 if (!oldmask != !rnp->qsmaskinit) {
1844 if (!oldmask) /* First online CPU for this rcu_node. */
1845 rcu_init_new_rnp(rnp);
1846 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1847 rnp->wait_blkd_tasks = true;
1848 else /* Last offline CPU and can propagate. */
1849 rcu_cleanup_dead_rnp(rnp);
1853 * If all waited-on tasks from prior grace period are
1854 * done, and if all this rcu_node structure's CPUs are
1855 * still offline, propagate up the rcu_node tree and
1856 * clear ->wait_blkd_tasks. Otherwise, if one of this
1857 * rcu_node structure's CPUs has since come back online,
1858 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1859 * checks for this, so just call it unconditionally).
1861 if (rnp->wait_blkd_tasks &&
1862 (!rcu_preempt_has_tasks(rnp) ||
1864 rnp->wait_blkd_tasks = false;
1865 rcu_cleanup_dead_rnp(rnp);
1868 raw_spin_unlock_irq(&rnp->lock);
1872 * Set the quiescent-state-needed bits in all the rcu_node
1873 * structures for all currently online CPUs in breadth-first order,
1874 * starting from the root rcu_node structure, relying on the layout
1875 * of the tree within the rsp->node[] array. Note that other CPUs
1876 * will access only the leaves of the hierarchy, thus seeing that no
1877 * grace period is in progress, at least until the corresponding
1878 * leaf node has been initialized. In addition, we have excluded
1879 * CPU-hotplug operations.
1881 * The grace period cannot complete until the initialization
1882 * process finishes, because this kthread handles both.
1884 rcu_for_each_node_breadth_first(rsp, rnp) {
1885 rcu_gp_slow(rsp, gp_init_delay);
1886 raw_spin_lock_irq(&rnp->lock);
1887 smp_mb__after_unlock_lock();
1888 rdp = this_cpu_ptr(rsp->rda);
1889 rcu_preempt_check_blocked_tasks(rnp);
1890 rnp->qsmask = rnp->qsmaskinit;
1891 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1892 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1893 WRITE_ONCE(rnp->completed, rsp->completed);
1894 if (rnp == rdp->mynode)
1895 (void)__note_gp_changes(rsp, rnp, rdp);
1896 rcu_preempt_boost_start_gp(rnp);
1897 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1898 rnp->level, rnp->grplo,
1899 rnp->grphi, rnp->qsmask);
1900 raw_spin_unlock_irq(&rnp->lock);
1901 cond_resched_rcu_qs();
1902 WRITE_ONCE(rsp->gp_activity, jiffies);
1909 * Do one round of quiescent-state forcing.
1911 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1913 int fqs_state = fqs_state_in;
1914 bool isidle = false;
1916 struct rcu_node *rnp = rcu_get_root(rsp);
1918 WRITE_ONCE(rsp->gp_activity, jiffies);
1920 if (fqs_state == RCU_SAVE_DYNTICK) {
1921 /* Collect dyntick-idle snapshots. */
1922 if (is_sysidle_rcu_state(rsp)) {
1924 maxj = jiffies - ULONG_MAX / 4;
1926 force_qs_rnp(rsp, dyntick_save_progress_counter,
1928 rcu_sysidle_report_gp(rsp, isidle, maxj);
1929 fqs_state = RCU_FORCE_QS;
1931 /* Handle dyntick-idle and offline CPUs. */
1933 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1935 /* Clear flag to prevent immediate re-entry. */
1936 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1937 raw_spin_lock_irq(&rnp->lock);
1938 smp_mb__after_unlock_lock();
1939 WRITE_ONCE(rsp->gp_flags,
1940 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
1941 raw_spin_unlock_irq(&rnp->lock);
1947 * Clean up after the old grace period.
1949 static void rcu_gp_cleanup(struct rcu_state *rsp)
1951 unsigned long gp_duration;
1952 bool needgp = false;
1954 struct rcu_data *rdp;
1955 struct rcu_node *rnp = rcu_get_root(rsp);
1957 WRITE_ONCE(rsp->gp_activity, jiffies);
1958 raw_spin_lock_irq(&rnp->lock);
1959 smp_mb__after_unlock_lock();
1960 gp_duration = jiffies - rsp->gp_start;
1961 if (gp_duration > rsp->gp_max)
1962 rsp->gp_max = gp_duration;
1965 * We know the grace period is complete, but to everyone else
1966 * it appears to still be ongoing. But it is also the case
1967 * that to everyone else it looks like there is nothing that
1968 * they can do to advance the grace period. It is therefore
1969 * safe for us to drop the lock in order to mark the grace
1970 * period as completed in all of the rcu_node structures.
1972 raw_spin_unlock_irq(&rnp->lock);
1975 * Propagate new ->completed value to rcu_node structures so
1976 * that other CPUs don't have to wait until the start of the next
1977 * grace period to process their callbacks. This also avoids
1978 * some nasty RCU grace-period initialization races by forcing
1979 * the end of the current grace period to be completely recorded in
1980 * all of the rcu_node structures before the beginning of the next
1981 * grace period is recorded in any of the rcu_node structures.
1983 rcu_for_each_node_breadth_first(rsp, rnp) {
1984 raw_spin_lock_irq(&rnp->lock);
1985 smp_mb__after_unlock_lock();
1986 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
1987 WARN_ON_ONCE(rnp->qsmask);
1988 WRITE_ONCE(rnp->completed, rsp->gpnum);
1989 rdp = this_cpu_ptr(rsp->rda);
1990 if (rnp == rdp->mynode)
1991 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
1992 /* smp_mb() provided by prior unlock-lock pair. */
1993 nocb += rcu_future_gp_cleanup(rsp, rnp);
1994 raw_spin_unlock_irq(&rnp->lock);
1995 cond_resched_rcu_qs();
1996 WRITE_ONCE(rsp->gp_activity, jiffies);
1997 rcu_gp_slow(rsp, gp_cleanup_delay);
1999 rnp = rcu_get_root(rsp);
2000 raw_spin_lock_irq(&rnp->lock);
2001 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
2002 rcu_nocb_gp_set(rnp, nocb);
2004 /* Declare grace period done. */
2005 WRITE_ONCE(rsp->completed, rsp->gpnum);
2006 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2007 rsp->fqs_state = RCU_GP_IDLE;
2008 rdp = this_cpu_ptr(rsp->rda);
2009 /* Advance CBs to reduce false positives below. */
2010 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2011 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2012 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2013 trace_rcu_grace_period(rsp->name,
2014 READ_ONCE(rsp->gpnum),
2017 raw_spin_unlock_irq(&rnp->lock);
2021 * Body of kthread that handles grace periods.
2023 static int __noreturn rcu_gp_kthread(void *arg)
2029 struct rcu_state *rsp = arg;
2030 struct rcu_node *rnp = rcu_get_root(rsp);
2032 rcu_bind_gp_kthread();
2035 /* Handle grace-period start. */
2037 trace_rcu_grace_period(rsp->name,
2038 READ_ONCE(rsp->gpnum),
2040 rsp->gp_state = RCU_GP_WAIT_GPS;
2041 wait_event_interruptible(rsp->gp_wq,
2042 READ_ONCE(rsp->gp_flags) &
2044 /* Locking provides needed memory barrier. */
2045 if (rcu_gp_init(rsp))
2047 cond_resched_rcu_qs();
2048 WRITE_ONCE(rsp->gp_activity, jiffies);
2049 WARN_ON(signal_pending(current));
2050 trace_rcu_grace_period(rsp->name,
2051 READ_ONCE(rsp->gpnum),
2055 /* Handle quiescent-state forcing. */
2056 fqs_state = RCU_SAVE_DYNTICK;
2057 j = jiffies_till_first_fqs;
2060 jiffies_till_first_fqs = HZ;
2065 rsp->jiffies_force_qs = jiffies + j;
2066 trace_rcu_grace_period(rsp->name,
2067 READ_ONCE(rsp->gpnum),
2069 rsp->gp_state = RCU_GP_WAIT_FQS;
2070 ret = wait_event_interruptible_timeout(rsp->gp_wq,
2071 ((gf = READ_ONCE(rsp->gp_flags)) &
2073 (!READ_ONCE(rnp->qsmask) &&
2074 !rcu_preempt_blocked_readers_cgp(rnp)),
2076 /* Locking provides needed memory barriers. */
2077 /* If grace period done, leave loop. */
2078 if (!READ_ONCE(rnp->qsmask) &&
2079 !rcu_preempt_blocked_readers_cgp(rnp))
2081 /* If time for quiescent-state forcing, do it. */
2082 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2083 (gf & RCU_GP_FLAG_FQS)) {
2084 trace_rcu_grace_period(rsp->name,
2085 READ_ONCE(rsp->gpnum),
2087 fqs_state = rcu_gp_fqs(rsp, fqs_state);
2088 trace_rcu_grace_period(rsp->name,
2089 READ_ONCE(rsp->gpnum),
2091 cond_resched_rcu_qs();
2092 WRITE_ONCE(rsp->gp_activity, jiffies);
2094 /* Deal with stray signal. */
2095 cond_resched_rcu_qs();
2096 WRITE_ONCE(rsp->gp_activity, jiffies);
2097 WARN_ON(signal_pending(current));
2098 trace_rcu_grace_period(rsp->name,
2099 READ_ONCE(rsp->gpnum),
2102 j = jiffies_till_next_fqs;
2105 jiffies_till_next_fqs = HZ;
2108 jiffies_till_next_fqs = 1;
2112 /* Handle grace-period end. */
2113 rcu_gp_cleanup(rsp);
2118 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2119 * in preparation for detecting the next grace period. The caller must hold
2120 * the root node's ->lock and hard irqs must be disabled.
2122 * Note that it is legal for a dying CPU (which is marked as offline) to
2123 * invoke this function. This can happen when the dying CPU reports its
2126 * Returns true if the grace-period kthread must be awakened.
2129 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2130 struct rcu_data *rdp)
2132 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2134 * Either we have not yet spawned the grace-period
2135 * task, this CPU does not need another grace period,
2136 * or a grace period is already in progress.
2137 * Either way, don't start a new grace period.
2141 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2142 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2146 * We can't do wakeups while holding the rnp->lock, as that
2147 * could cause possible deadlocks with the rq->lock. Defer
2148 * the wakeup to our caller.
2154 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2155 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2156 * is invoked indirectly from rcu_advance_cbs(), which would result in
2157 * endless recursion -- or would do so if it wasn't for the self-deadlock
2158 * that is encountered beforehand.
2160 * Returns true if the grace-period kthread needs to be awakened.
2162 static bool rcu_start_gp(struct rcu_state *rsp)
2164 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2165 struct rcu_node *rnp = rcu_get_root(rsp);
2169 * If there is no grace period in progress right now, any
2170 * callbacks we have up to this point will be satisfied by the
2171 * next grace period. Also, advancing the callbacks reduces the
2172 * probability of false positives from cpu_needs_another_gp()
2173 * resulting in pointless grace periods. So, advance callbacks
2174 * then start the grace period!
2176 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2177 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2182 * Report a full set of quiescent states to the specified rcu_state
2183 * data structure. This involves cleaning up after the prior grace
2184 * period and letting rcu_start_gp() start up the next grace period
2185 * if one is needed. Note that the caller must hold rnp->lock, which
2186 * is released before return.
2188 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2189 __releases(rcu_get_root(rsp)->lock)
2191 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2192 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2193 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2194 rcu_gp_kthread_wake(rsp);
2198 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2199 * Allows quiescent states for a group of CPUs to be reported at one go
2200 * to the specified rcu_node structure, though all the CPUs in the group
2201 * must be represented by the same rcu_node structure (which need not be a
2202 * leaf rcu_node structure, though it often will be). The gps parameter
2203 * is the grace-period snapshot, which means that the quiescent states
2204 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2205 * must be held upon entry, and it is released before return.
2208 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2209 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2210 __releases(rnp->lock)
2212 unsigned long oldmask = 0;
2213 struct rcu_node *rnp_c;
2215 /* Walk up the rcu_node hierarchy. */
2217 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2220 * Our bit has already been cleared, or the
2221 * relevant grace period is already over, so done.
2223 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2226 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2227 rnp->qsmask &= ~mask;
2228 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2229 mask, rnp->qsmask, rnp->level,
2230 rnp->grplo, rnp->grphi,
2232 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2234 /* Other bits still set at this level, so done. */
2235 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2238 mask = rnp->grpmask;
2239 if (rnp->parent == NULL) {
2241 /* No more levels. Exit loop holding root lock. */
2245 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2248 raw_spin_lock_irqsave(&rnp->lock, flags);
2249 smp_mb__after_unlock_lock();
2250 oldmask = rnp_c->qsmask;
2254 * Get here if we are the last CPU to pass through a quiescent
2255 * state for this grace period. Invoke rcu_report_qs_rsp()
2256 * to clean up and start the next grace period if one is needed.
2258 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2262 * Record a quiescent state for all tasks that were previously queued
2263 * on the specified rcu_node structure and that were blocking the current
2264 * RCU grace period. The caller must hold the specified rnp->lock with
2265 * irqs disabled, and this lock is released upon return, but irqs remain
2268 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2269 struct rcu_node *rnp, unsigned long flags)
2270 __releases(rnp->lock)
2274 struct rcu_node *rnp_p;
2276 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2277 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2278 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2279 return; /* Still need more quiescent states! */
2282 rnp_p = rnp->parent;
2283 if (rnp_p == NULL) {
2285 * Only one rcu_node structure in the tree, so don't
2286 * try to report up to its nonexistent parent!
2288 rcu_report_qs_rsp(rsp, flags);
2292 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2294 mask = rnp->grpmask;
2295 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2296 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
2297 smp_mb__after_unlock_lock();
2298 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2302 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2303 * structure. This must be either called from the specified CPU, or
2304 * called when the specified CPU is known to be offline (and when it is
2305 * also known that no other CPU is concurrently trying to help the offline
2306 * CPU). The lastcomp argument is used to make sure we are still in the
2307 * grace period of interest. We don't want to end the current grace period
2308 * based on quiescent states detected in an earlier grace period!
2311 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2313 unsigned long flags;
2316 struct rcu_node *rnp;
2319 raw_spin_lock_irqsave(&rnp->lock, flags);
2320 smp_mb__after_unlock_lock();
2321 if ((rdp->passed_quiesce == 0 &&
2322 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2323 rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2327 * The grace period in which this quiescent state was
2328 * recorded has ended, so don't report it upwards.
2329 * We will instead need a new quiescent state that lies
2330 * within the current grace period.
2332 rdp->passed_quiesce = 0; /* need qs for new gp. */
2333 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2334 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2337 mask = rdp->grpmask;
2338 if ((rnp->qsmask & mask) == 0) {
2339 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2341 rdp->qs_pending = 0;
2344 * This GP can't end until cpu checks in, so all of our
2345 * callbacks can be processed during the next GP.
2347 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2349 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2350 /* ^^^ Released rnp->lock */
2352 rcu_gp_kthread_wake(rsp);
2357 * Check to see if there is a new grace period of which this CPU
2358 * is not yet aware, and if so, set up local rcu_data state for it.
2359 * Otherwise, see if this CPU has just passed through its first
2360 * quiescent state for this grace period, and record that fact if so.
2363 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2365 /* Check for grace-period ends and beginnings. */
2366 note_gp_changes(rsp, rdp);
2369 * Does this CPU still need to do its part for current grace period?
2370 * If no, return and let the other CPUs do their part as well.
2372 if (!rdp->qs_pending)
2376 * Was there a quiescent state since the beginning of the grace
2377 * period? If no, then exit and wait for the next call.
2379 if (!rdp->passed_quiesce &&
2380 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2384 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2387 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2391 * Send the specified CPU's RCU callbacks to the orphanage. The
2392 * specified CPU must be offline, and the caller must hold the
2396 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2397 struct rcu_node *rnp, struct rcu_data *rdp)
2399 /* No-CBs CPUs do not have orphanable callbacks. */
2400 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2404 * Orphan the callbacks. First adjust the counts. This is safe
2405 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2406 * cannot be running now. Thus no memory barrier is required.
2408 if (rdp->nxtlist != NULL) {
2409 rsp->qlen_lazy += rdp->qlen_lazy;
2410 rsp->qlen += rdp->qlen;
2411 rdp->n_cbs_orphaned += rdp->qlen;
2413 WRITE_ONCE(rdp->qlen, 0);
2417 * Next, move those callbacks still needing a grace period to
2418 * the orphanage, where some other CPU will pick them up.
2419 * Some of the callbacks might have gone partway through a grace
2420 * period, but that is too bad. They get to start over because we
2421 * cannot assume that grace periods are synchronized across CPUs.
2422 * We don't bother updating the ->nxttail[] array yet, instead
2423 * we just reset the whole thing later on.
2425 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2426 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2427 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2428 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2432 * Then move the ready-to-invoke callbacks to the orphanage,
2433 * where some other CPU will pick them up. These will not be
2434 * required to pass though another grace period: They are done.
2436 if (rdp->nxtlist != NULL) {
2437 *rsp->orphan_donetail = rdp->nxtlist;
2438 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2442 * Finally, initialize the rcu_data structure's list to empty and
2443 * disallow further callbacks on this CPU.
2445 init_callback_list(rdp);
2446 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2450 * Adopt the RCU callbacks from the specified rcu_state structure's
2451 * orphanage. The caller must hold the ->orphan_lock.
2453 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2456 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2458 /* No-CBs CPUs are handled specially. */
2459 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2460 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2463 /* Do the accounting first. */
2464 rdp->qlen_lazy += rsp->qlen_lazy;
2465 rdp->qlen += rsp->qlen;
2466 rdp->n_cbs_adopted += rsp->qlen;
2467 if (rsp->qlen_lazy != rsp->qlen)
2468 rcu_idle_count_callbacks_posted();
2473 * We do not need a memory barrier here because the only way we
2474 * can get here if there is an rcu_barrier() in flight is if
2475 * we are the task doing the rcu_barrier().
2478 /* First adopt the ready-to-invoke callbacks. */
2479 if (rsp->orphan_donelist != NULL) {
2480 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2481 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2482 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2483 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2484 rdp->nxttail[i] = rsp->orphan_donetail;
2485 rsp->orphan_donelist = NULL;
2486 rsp->orphan_donetail = &rsp->orphan_donelist;
2489 /* And then adopt the callbacks that still need a grace period. */
2490 if (rsp->orphan_nxtlist != NULL) {
2491 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2492 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2493 rsp->orphan_nxtlist = NULL;
2494 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2499 * Trace the fact that this CPU is going offline.
2501 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2503 RCU_TRACE(unsigned long mask);
2504 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2505 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2507 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2510 RCU_TRACE(mask = rdp->grpmask);
2511 trace_rcu_grace_period(rsp->name,
2512 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2517 * All CPUs for the specified rcu_node structure have gone offline,
2518 * and all tasks that were preempted within an RCU read-side critical
2519 * section while running on one of those CPUs have since exited their RCU
2520 * read-side critical section. Some other CPU is reporting this fact with
2521 * the specified rcu_node structure's ->lock held and interrupts disabled.
2522 * This function therefore goes up the tree of rcu_node structures,
2523 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2524 * the leaf rcu_node structure's ->qsmaskinit field has already been
2527 * This function does check that the specified rcu_node structure has
2528 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2529 * prematurely. That said, invoking it after the fact will cost you
2530 * a needless lock acquisition. So once it has done its work, don't
2533 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2536 struct rcu_node *rnp = rnp_leaf;
2538 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2539 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2542 mask = rnp->grpmask;
2546 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2547 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2548 rnp->qsmaskinit &= ~mask;
2549 rnp->qsmask &= ~mask;
2550 if (rnp->qsmaskinit) {
2551 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2554 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2559 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2560 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2563 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
2565 unsigned long flags;
2567 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2568 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2570 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2573 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2574 mask = rdp->grpmask;
2575 raw_spin_lock_irqsave(&rnp->lock, flags);
2576 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2577 rnp->qsmaskinitnext &= ~mask;
2578 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2582 * The CPU has been completely removed, and some other CPU is reporting
2583 * this fact from process context. Do the remainder of the cleanup,
2584 * including orphaning the outgoing CPU's RCU callbacks, and also
2585 * adopting them. There can only be one CPU hotplug operation at a time,
2586 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2588 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2590 unsigned long flags;
2591 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2592 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2594 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2597 /* Adjust any no-longer-needed kthreads. */
2598 rcu_boost_kthread_setaffinity(rnp, -1);
2600 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2601 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2602 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2603 rcu_adopt_orphan_cbs(rsp, flags);
2604 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2606 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2607 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2608 cpu, rdp->qlen, rdp->nxtlist);
2612 * Invoke any RCU callbacks that have made it to the end of their grace
2613 * period. Thottle as specified by rdp->blimit.
2615 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2617 unsigned long flags;
2618 struct rcu_head *next, *list, **tail;
2619 long bl, count, count_lazy;
2622 /* If no callbacks are ready, just return. */
2623 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2624 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2625 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2626 need_resched(), is_idle_task(current),
2627 rcu_is_callbacks_kthread());
2632 * Extract the list of ready callbacks, disabling to prevent
2633 * races with call_rcu() from interrupt handlers.
2635 local_irq_save(flags);
2636 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2638 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2639 list = rdp->nxtlist;
2640 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2641 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2642 tail = rdp->nxttail[RCU_DONE_TAIL];
2643 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2644 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2645 rdp->nxttail[i] = &rdp->nxtlist;
2646 local_irq_restore(flags);
2648 /* Invoke callbacks. */
2649 count = count_lazy = 0;
2653 debug_rcu_head_unqueue(list);
2654 if (__rcu_reclaim(rsp->name, list))
2657 /* Stop only if limit reached and CPU has something to do. */
2658 if (++count >= bl &&
2660 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2664 local_irq_save(flags);
2665 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2666 is_idle_task(current),
2667 rcu_is_callbacks_kthread());
2669 /* Update count, and requeue any remaining callbacks. */
2671 *tail = rdp->nxtlist;
2672 rdp->nxtlist = list;
2673 for (i = 0; i < RCU_NEXT_SIZE; i++)
2674 if (&rdp->nxtlist == rdp->nxttail[i])
2675 rdp->nxttail[i] = tail;
2679 smp_mb(); /* List handling before counting for rcu_barrier(). */
2680 rdp->qlen_lazy -= count_lazy;
2681 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2682 rdp->n_cbs_invoked += count;
2684 /* Reinstate batch limit if we have worked down the excess. */
2685 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2686 rdp->blimit = blimit;
2688 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2689 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2690 rdp->qlen_last_fqs_check = 0;
2691 rdp->n_force_qs_snap = rsp->n_force_qs;
2692 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2693 rdp->qlen_last_fqs_check = rdp->qlen;
2694 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2696 local_irq_restore(flags);
2698 /* Re-invoke RCU core processing if there are callbacks remaining. */
2699 if (cpu_has_callbacks_ready_to_invoke(rdp))
2704 * Check to see if this CPU is in a non-context-switch quiescent state
2705 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2706 * Also schedule RCU core processing.
2708 * This function must be called from hardirq context. It is normally
2709 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2710 * false, there is no point in invoking rcu_check_callbacks().
2712 void rcu_check_callbacks(int user)
2714 trace_rcu_utilization(TPS("Start scheduler-tick"));
2715 increment_cpu_stall_ticks();
2716 if (user || rcu_is_cpu_rrupt_from_idle()) {
2719 * Get here if this CPU took its interrupt from user
2720 * mode or from the idle loop, and if this is not a
2721 * nested interrupt. In this case, the CPU is in
2722 * a quiescent state, so note it.
2724 * No memory barrier is required here because both
2725 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2726 * variables that other CPUs neither access nor modify,
2727 * at least not while the corresponding CPU is online.
2733 } else if (!in_softirq()) {
2736 * Get here if this CPU did not take its interrupt from
2737 * softirq, in other words, if it is not interrupting
2738 * a rcu_bh read-side critical section. This is an _bh
2739 * critical section, so note it.
2744 rcu_preempt_check_callbacks();
2748 rcu_note_voluntary_context_switch(current);
2749 trace_rcu_utilization(TPS("End scheduler-tick"));
2753 * Scan the leaf rcu_node structures, processing dyntick state for any that
2754 * have not yet encountered a quiescent state, using the function specified.
2755 * Also initiate boosting for any threads blocked on the root rcu_node.
2757 * The caller must have suppressed start of new grace periods.
2759 static void force_qs_rnp(struct rcu_state *rsp,
2760 int (*f)(struct rcu_data *rsp, bool *isidle,
2761 unsigned long *maxj),
2762 bool *isidle, unsigned long *maxj)
2766 unsigned long flags;
2768 struct rcu_node *rnp;
2770 rcu_for_each_leaf_node(rsp, rnp) {
2771 cond_resched_rcu_qs();
2773 raw_spin_lock_irqsave(&rnp->lock, flags);
2774 smp_mb__after_unlock_lock();
2775 if (rnp->qsmask == 0) {
2776 if (rcu_state_p == &rcu_sched_state ||
2777 rsp != rcu_state_p ||
2778 rcu_preempt_blocked_readers_cgp(rnp)) {
2780 * No point in scanning bits because they
2781 * are all zero. But we might need to
2782 * priority-boost blocked readers.
2784 rcu_initiate_boost(rnp, flags);
2785 /* rcu_initiate_boost() releases rnp->lock */
2789 (rnp->parent->qsmask & rnp->grpmask)) {
2791 * Race between grace-period
2792 * initialization and task exiting RCU
2793 * read-side critical section: Report.
2795 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2796 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2802 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2803 if ((rnp->qsmask & bit) != 0) {
2804 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2809 /* Idle/offline CPUs, report (releases rnp->lock. */
2810 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2812 /* Nothing to do here, so just drop the lock. */
2813 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2819 * Force quiescent states on reluctant CPUs, and also detect which
2820 * CPUs are in dyntick-idle mode.
2822 static void force_quiescent_state(struct rcu_state *rsp)
2824 unsigned long flags;
2826 struct rcu_node *rnp;
2827 struct rcu_node *rnp_old = NULL;
2829 /* Funnel through hierarchy to reduce memory contention. */
2830 rnp = __this_cpu_read(rsp->rda->mynode);
2831 for (; rnp != NULL; rnp = rnp->parent) {
2832 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2833 !raw_spin_trylock(&rnp->fqslock);
2834 if (rnp_old != NULL)
2835 raw_spin_unlock(&rnp_old->fqslock);
2837 rsp->n_force_qs_lh++;
2842 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2844 /* Reached the root of the rcu_node tree, acquire lock. */
2845 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2846 smp_mb__after_unlock_lock();
2847 raw_spin_unlock(&rnp_old->fqslock);
2848 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2849 rsp->n_force_qs_lh++;
2850 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2851 return; /* Someone beat us to it. */
2853 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2854 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2855 rcu_gp_kthread_wake(rsp);
2859 * This does the RCU core processing work for the specified rcu_state
2860 * and rcu_data structures. This may be called only from the CPU to
2861 * whom the rdp belongs.
2864 __rcu_process_callbacks(struct rcu_state *rsp)
2866 unsigned long flags;
2868 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2870 WARN_ON_ONCE(rdp->beenonline == 0);
2872 /* Update RCU state based on any recent quiescent states. */
2873 rcu_check_quiescent_state(rsp, rdp);
2875 /* Does this CPU require a not-yet-started grace period? */
2876 local_irq_save(flags);
2877 if (cpu_needs_another_gp(rsp, rdp)) {
2878 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2879 needwake = rcu_start_gp(rsp);
2880 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2882 rcu_gp_kthread_wake(rsp);
2884 local_irq_restore(flags);
2887 /* If there are callbacks ready, invoke them. */
2888 if (cpu_has_callbacks_ready_to_invoke(rdp))
2889 invoke_rcu_callbacks(rsp, rdp);
2891 /* Do any needed deferred wakeups of rcuo kthreads. */
2892 do_nocb_deferred_wakeup(rdp);
2896 * Do RCU core processing for the current CPU.
2898 static void rcu_process_callbacks(struct softirq_action *unused)
2900 struct rcu_state *rsp;
2902 if (cpu_is_offline(smp_processor_id()))
2904 trace_rcu_utilization(TPS("Start RCU core"));
2905 for_each_rcu_flavor(rsp)
2906 __rcu_process_callbacks(rsp);
2907 trace_rcu_utilization(TPS("End RCU core"));
2911 * Schedule RCU callback invocation. If the specified type of RCU
2912 * does not support RCU priority boosting, just do a direct call,
2913 * otherwise wake up the per-CPU kernel kthread. Note that because we
2914 * are running on the current CPU with softirqs disabled, the
2915 * rcu_cpu_kthread_task cannot disappear out from under us.
2917 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2919 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2921 if (likely(!rsp->boost)) {
2922 rcu_do_batch(rsp, rdp);
2925 invoke_rcu_callbacks_kthread();
2928 static void invoke_rcu_core(void)
2930 if (cpu_online(smp_processor_id()))
2931 raise_softirq(RCU_SOFTIRQ);
2935 * Handle any core-RCU processing required by a call_rcu() invocation.
2937 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2938 struct rcu_head *head, unsigned long flags)
2943 * If called from an extended quiescent state, invoke the RCU
2944 * core in order to force a re-evaluation of RCU's idleness.
2946 if (!rcu_is_watching())
2949 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2950 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2954 * Force the grace period if too many callbacks or too long waiting.
2955 * Enforce hysteresis, and don't invoke force_quiescent_state()
2956 * if some other CPU has recently done so. Also, don't bother
2957 * invoking force_quiescent_state() if the newly enqueued callback
2958 * is the only one waiting for a grace period to complete.
2960 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2962 /* Are we ignoring a completed grace period? */
2963 note_gp_changes(rsp, rdp);
2965 /* Start a new grace period if one not already started. */
2966 if (!rcu_gp_in_progress(rsp)) {
2967 struct rcu_node *rnp_root = rcu_get_root(rsp);
2969 raw_spin_lock(&rnp_root->lock);
2970 smp_mb__after_unlock_lock();
2971 needwake = rcu_start_gp(rsp);
2972 raw_spin_unlock(&rnp_root->lock);
2974 rcu_gp_kthread_wake(rsp);
2976 /* Give the grace period a kick. */
2977 rdp->blimit = LONG_MAX;
2978 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2979 *rdp->nxttail[RCU_DONE_TAIL] != head)
2980 force_quiescent_state(rsp);
2981 rdp->n_force_qs_snap = rsp->n_force_qs;
2982 rdp->qlen_last_fqs_check = rdp->qlen;
2988 * RCU callback function to leak a callback.
2990 static void rcu_leak_callback(struct rcu_head *rhp)
2995 * Helper function for call_rcu() and friends. The cpu argument will
2996 * normally be -1, indicating "currently running CPU". It may specify
2997 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2998 * is expected to specify a CPU.
3001 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
3002 struct rcu_state *rsp, int cpu, bool lazy)
3004 unsigned long flags;
3005 struct rcu_data *rdp;
3007 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
3008 if (debug_rcu_head_queue(head)) {
3009 /* Probable double call_rcu(), so leak the callback. */
3010 WRITE_ONCE(head->func, rcu_leak_callback);
3011 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3018 * Opportunistically note grace-period endings and beginnings.
3019 * Note that we might see a beginning right after we see an
3020 * end, but never vice versa, since this CPU has to pass through
3021 * a quiescent state betweentimes.
3023 local_irq_save(flags);
3024 rdp = this_cpu_ptr(rsp->rda);
3026 /* Add the callback to our list. */
3027 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3031 rdp = per_cpu_ptr(rsp->rda, cpu);
3032 if (likely(rdp->mynode)) {
3033 /* Post-boot, so this should be for a no-CBs CPU. */
3034 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3035 WARN_ON_ONCE(offline);
3036 /* Offline CPU, _call_rcu() illegal, leak callback. */
3037 local_irq_restore(flags);
3041 * Very early boot, before rcu_init(). Initialize if needed
3042 * and then drop through to queue the callback.
3045 WARN_ON_ONCE(!rcu_is_watching());
3046 if (!likely(rdp->nxtlist))
3047 init_default_callback_list(rdp);
3049 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3053 rcu_idle_count_callbacks_posted();
3054 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3055 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3056 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3058 if (__is_kfree_rcu_offset((unsigned long)func))
3059 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3060 rdp->qlen_lazy, rdp->qlen);
3062 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3064 /* Go handle any RCU core processing required. */
3065 __call_rcu_core(rsp, rdp, head, flags);
3066 local_irq_restore(flags);
3070 * Queue an RCU-sched callback for invocation after a grace period.
3072 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
3074 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3076 EXPORT_SYMBOL_GPL(call_rcu_sched);
3079 * Queue an RCU callback for invocation after a quicker grace period.
3081 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
3083 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3085 EXPORT_SYMBOL_GPL(call_rcu_bh);
3088 * Queue an RCU callback for lazy invocation after a grace period.
3089 * This will likely be later named something like "call_rcu_lazy()",
3090 * but this change will require some way of tagging the lazy RCU
3091 * callbacks in the list of pending callbacks. Until then, this
3092 * function may only be called from __kfree_rcu().
3094 void kfree_call_rcu(struct rcu_head *head,
3095 void (*func)(struct rcu_head *rcu))
3097 __call_rcu(head, func, rcu_state_p, -1, 1);
3099 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3102 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3103 * any blocking grace-period wait automatically implies a grace period
3104 * if there is only one CPU online at any point time during execution
3105 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3106 * occasionally incorrectly indicate that there are multiple CPUs online
3107 * when there was in fact only one the whole time, as this just adds
3108 * some overhead: RCU still operates correctly.
3110 static inline int rcu_blocking_is_gp(void)
3114 might_sleep(); /* Check for RCU read-side critical section. */
3116 ret = num_online_cpus() <= 1;
3122 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3124 * Control will return to the caller some time after a full rcu-sched
3125 * grace period has elapsed, in other words after all currently executing
3126 * rcu-sched read-side critical sections have completed. These read-side
3127 * critical sections are delimited by rcu_read_lock_sched() and
3128 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3129 * local_irq_disable(), and so on may be used in place of
3130 * rcu_read_lock_sched().
3132 * This means that all preempt_disable code sequences, including NMI and
3133 * non-threaded hardware-interrupt handlers, in progress on entry will
3134 * have completed before this primitive returns. However, this does not
3135 * guarantee that softirq handlers will have completed, since in some
3136 * kernels, these handlers can run in process context, and can block.
3138 * Note that this guarantee implies further memory-ordering guarantees.
3139 * On systems with more than one CPU, when synchronize_sched() returns,
3140 * each CPU is guaranteed to have executed a full memory barrier since the
3141 * end of its last RCU-sched read-side critical section whose beginning
3142 * preceded the call to synchronize_sched(). In addition, each CPU having
3143 * an RCU read-side critical section that extends beyond the return from
3144 * synchronize_sched() is guaranteed to have executed a full memory barrier
3145 * after the beginning of synchronize_sched() and before the beginning of
3146 * that RCU read-side critical section. Note that these guarantees include
3147 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3148 * that are executing in the kernel.
3150 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3151 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3152 * to have executed a full memory barrier during the execution of
3153 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3154 * again only if the system has more than one CPU).
3156 * This primitive provides the guarantees made by the (now removed)
3157 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3158 * guarantees that rcu_read_lock() sections will have completed.
3159 * In "classic RCU", these two guarantees happen to be one and
3160 * the same, but can differ in realtime RCU implementations.
3162 void synchronize_sched(void)
3164 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
3165 !lock_is_held(&rcu_lock_map) &&
3166 !lock_is_held(&rcu_sched_lock_map),
3167 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3168 if (rcu_blocking_is_gp())
3170 if (rcu_gp_is_expedited())
3171 synchronize_sched_expedited();
3173 wait_rcu_gp(call_rcu_sched);
3175 EXPORT_SYMBOL_GPL(synchronize_sched);
3178 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3180 * Control will return to the caller some time after a full rcu_bh grace
3181 * period has elapsed, in other words after all currently executing rcu_bh
3182 * read-side critical sections have completed. RCU read-side critical
3183 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3184 * and may be nested.
3186 * See the description of synchronize_sched() for more detailed information
3187 * on memory ordering guarantees.
3189 void synchronize_rcu_bh(void)
3191 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
3192 !lock_is_held(&rcu_lock_map) &&
3193 !lock_is_held(&rcu_sched_lock_map),
3194 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3195 if (rcu_blocking_is_gp())
3197 if (rcu_gp_is_expedited())
3198 synchronize_rcu_bh_expedited();
3200 wait_rcu_gp(call_rcu_bh);
3202 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3205 * get_state_synchronize_rcu - Snapshot current RCU state
3207 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3208 * to determine whether or not a full grace period has elapsed in the
3211 unsigned long get_state_synchronize_rcu(void)
3214 * Any prior manipulation of RCU-protected data must happen
3215 * before the load from ->gpnum.
3220 * Make sure this load happens before the purportedly
3221 * time-consuming work between get_state_synchronize_rcu()
3222 * and cond_synchronize_rcu().
3224 return smp_load_acquire(&rcu_state_p->gpnum);
3226 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3229 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3231 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3233 * If a full RCU grace period has elapsed since the earlier call to
3234 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3235 * synchronize_rcu() to wait for a full grace period.
3237 * Yes, this function does not take counter wrap into account. But
3238 * counter wrap is harmless. If the counter wraps, we have waited for
3239 * more than 2 billion grace periods (and way more on a 64-bit system!),
3240 * so waiting for one additional grace period should be just fine.
3242 void cond_synchronize_rcu(unsigned long oldstate)
3244 unsigned long newstate;
3247 * Ensure that this load happens before any RCU-destructive
3248 * actions the caller might carry out after we return.
3250 newstate = smp_load_acquire(&rcu_state_p->completed);
3251 if (ULONG_CMP_GE(oldstate, newstate))
3254 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3257 * get_state_synchronize_sched - Snapshot current RCU-sched state
3259 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3260 * to determine whether or not a full grace period has elapsed in the
3263 unsigned long get_state_synchronize_sched(void)
3266 * Any prior manipulation of RCU-protected data must happen
3267 * before the load from ->gpnum.
3272 * Make sure this load happens before the purportedly
3273 * time-consuming work between get_state_synchronize_sched()
3274 * and cond_synchronize_sched().
3276 return smp_load_acquire(&rcu_sched_state.gpnum);
3278 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3281 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3283 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3285 * If a full RCU-sched grace period has elapsed since the earlier call to
3286 * get_state_synchronize_sched(), just return. Otherwise, invoke
3287 * synchronize_sched() to wait for a full grace period.
3289 * Yes, this function does not take counter wrap into account. But
3290 * counter wrap is harmless. If the counter wraps, we have waited for
3291 * more than 2 billion grace periods (and way more on a 64-bit system!),
3292 * so waiting for one additional grace period should be just fine.
3294 void cond_synchronize_sched(unsigned long oldstate)
3296 unsigned long newstate;
3299 * Ensure that this load happens before any RCU-destructive
3300 * actions the caller might carry out after we return.
3302 newstate = smp_load_acquire(&rcu_sched_state.completed);
3303 if (ULONG_CMP_GE(oldstate, newstate))
3304 synchronize_sched();
3306 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3308 static int synchronize_sched_expedited_cpu_stop(void *data)
3311 * There must be a full memory barrier on each affected CPU
3312 * between the time that try_stop_cpus() is called and the
3313 * time that it returns.
3315 * In the current initial implementation of cpu_stop, the
3316 * above condition is already met when the control reaches
3317 * this point and the following smp_mb() is not strictly
3318 * necessary. Do smp_mb() anyway for documentation and
3319 * robustness against future implementation changes.
3321 smp_mb(); /* See above comment block. */
3326 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3328 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3329 * approach to force the grace period to end quickly. This consumes
3330 * significant time on all CPUs and is unfriendly to real-time workloads,
3331 * so is thus not recommended for any sort of common-case code. In fact,
3332 * if you are using synchronize_sched_expedited() in a loop, please
3333 * restructure your code to batch your updates, and then use a single
3334 * synchronize_sched() instead.
3336 * This implementation can be thought of as an application of ticket
3337 * locking to RCU, with sync_sched_expedited_started and
3338 * sync_sched_expedited_done taking on the roles of the halves
3339 * of the ticket-lock word. Each task atomically increments
3340 * sync_sched_expedited_started upon entry, snapshotting the old value,
3341 * then attempts to stop all the CPUs. If this succeeds, then each
3342 * CPU will have executed a context switch, resulting in an RCU-sched
3343 * grace period. We are then done, so we use atomic_cmpxchg() to
3344 * update sync_sched_expedited_done to match our snapshot -- but
3345 * only if someone else has not already advanced past our snapshot.
3347 * On the other hand, if try_stop_cpus() fails, we check the value
3348 * of sync_sched_expedited_done. If it has advanced past our
3349 * initial snapshot, then someone else must have forced a grace period
3350 * some time after we took our snapshot. In this case, our work is
3351 * done for us, and we can simply return. Otherwise, we try again,
3352 * but keep our initial snapshot for purposes of checking for someone
3353 * doing our work for us.
3355 * If we fail too many times in a row, we fall back to synchronize_sched().
3357 void synchronize_sched_expedited(void)
3362 long firstsnap, s, snap;
3364 struct rcu_state *rsp = &rcu_sched_state;
3367 * If we are in danger of counter wrap, just do synchronize_sched().
3368 * By allowing sync_sched_expedited_started to advance no more than
3369 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
3370 * that more than 3.5 billion CPUs would be required to force a
3371 * counter wrap on a 32-bit system. Quite a few more CPUs would of
3372 * course be required on a 64-bit system.
3374 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
3375 (ulong)atomic_long_read(&rsp->expedited_done) +
3377 wait_rcu_gp(call_rcu_sched);
3378 atomic_long_inc(&rsp->expedited_wrap);
3383 * Take a ticket. Note that atomic_inc_return() implies a
3384 * full memory barrier.
3386 snap = atomic_long_inc_return(&rsp->expedited_start);
3388 if (!try_get_online_cpus()) {
3389 /* CPU hotplug operation in flight, fall back to normal GP. */
3390 wait_rcu_gp(call_rcu_sched);
3391 atomic_long_inc(&rsp->expedited_normal);
3394 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
3396 /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */
3397 cma = zalloc_cpumask_var(&cm, GFP_KERNEL);
3399 cpumask_copy(cm, cpu_online_mask);
3400 cpumask_clear_cpu(raw_smp_processor_id(), cm);
3401 for_each_cpu(cpu, cm) {
3402 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3404 if (!(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3405 cpumask_clear_cpu(cpu, cm);
3407 if (cpumask_weight(cm) == 0)
3412 * Each pass through the following loop attempts to force a
3413 * context switch on each CPU.
3415 while (try_stop_cpus(cma ? cm : cpu_online_mask,
3416 synchronize_sched_expedited_cpu_stop,
3419 atomic_long_inc(&rsp->expedited_tryfail);
3421 /* Check to see if someone else did our work for us. */
3422 s = atomic_long_read(&rsp->expedited_done);
3423 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3424 /* ensure test happens before caller kfree */
3425 smp_mb__before_atomic(); /* ^^^ */
3426 atomic_long_inc(&rsp->expedited_workdone1);
3427 free_cpumask_var(cm);
3431 /* No joy, try again later. Or just synchronize_sched(). */
3432 if (trycount++ < 10) {
3433 udelay(trycount * num_online_cpus());
3435 wait_rcu_gp(call_rcu_sched);
3436 atomic_long_inc(&rsp->expedited_normal);
3437 free_cpumask_var(cm);
3441 /* Recheck to see if someone else did our work for us. */
3442 s = atomic_long_read(&rsp->expedited_done);
3443 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
3444 /* ensure test happens before caller kfree */
3445 smp_mb__before_atomic(); /* ^^^ */
3446 atomic_long_inc(&rsp->expedited_workdone2);
3447 free_cpumask_var(cm);
3452 * Refetching sync_sched_expedited_started allows later
3453 * callers to piggyback on our grace period. We retry
3454 * after they started, so our grace period works for them,
3455 * and they started after our first try, so their grace
3456 * period works for us.
3458 if (!try_get_online_cpus()) {
3459 /* CPU hotplug operation in flight, use normal GP. */
3460 wait_rcu_gp(call_rcu_sched);
3461 atomic_long_inc(&rsp->expedited_normal);
3462 free_cpumask_var(cm);
3465 snap = atomic_long_read(&rsp->expedited_start);
3466 smp_mb(); /* ensure read is before try_stop_cpus(). */
3468 atomic_long_inc(&rsp->expedited_stoppedcpus);
3471 free_cpumask_var(cm);
3474 * Everyone up to our most recent fetch is covered by our grace
3475 * period. Update the counter, but only if our work is still
3476 * relevant -- which it won't be if someone who started later
3477 * than we did already did their update.
3480 atomic_long_inc(&rsp->expedited_done_tries);
3481 s = atomic_long_read(&rsp->expedited_done);
3482 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
3483 /* ensure test happens before caller kfree */
3484 smp_mb__before_atomic(); /* ^^^ */
3485 atomic_long_inc(&rsp->expedited_done_lost);
3488 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
3489 atomic_long_inc(&rsp->expedited_done_exit);
3493 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3496 * Check to see if there is any immediate RCU-related work to be done
3497 * by the current CPU, for the specified type of RCU, returning 1 if so.
3498 * The checks are in order of increasing expense: checks that can be
3499 * carried out against CPU-local state are performed first. However,
3500 * we must check for CPU stalls first, else we might not get a chance.
3502 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3504 struct rcu_node *rnp = rdp->mynode;
3506 rdp->n_rcu_pending++;
3508 /* Check for CPU stalls, if enabled. */
3509 check_cpu_stall(rsp, rdp);
3511 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3512 if (rcu_nohz_full_cpu(rsp))
3515 /* Is the RCU core waiting for a quiescent state from this CPU? */
3516 if (rcu_scheduler_fully_active &&
3517 rdp->qs_pending && !rdp->passed_quiesce &&
3518 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3519 rdp->n_rp_qs_pending++;
3520 } else if (rdp->qs_pending &&
3521 (rdp->passed_quiesce ||
3522 rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3523 rdp->n_rp_report_qs++;
3527 /* Does this CPU have callbacks ready to invoke? */
3528 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3529 rdp->n_rp_cb_ready++;
3533 /* Has RCU gone idle with this CPU needing another grace period? */
3534 if (cpu_needs_another_gp(rsp, rdp)) {
3535 rdp->n_rp_cpu_needs_gp++;
3539 /* Has another RCU grace period completed? */
3540 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3541 rdp->n_rp_gp_completed++;
3545 /* Has a new RCU grace period started? */
3546 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3547 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3548 rdp->n_rp_gp_started++;
3552 /* Does this CPU need a deferred NOCB wakeup? */
3553 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3554 rdp->n_rp_nocb_defer_wakeup++;
3559 rdp->n_rp_need_nothing++;
3564 * Check to see if there is any immediate RCU-related work to be done
3565 * by the current CPU, returning 1 if so. This function is part of the
3566 * RCU implementation; it is -not- an exported member of the RCU API.
3568 static int rcu_pending(void)
3570 struct rcu_state *rsp;
3572 for_each_rcu_flavor(rsp)
3573 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3579 * Return true if the specified CPU has any callback. If all_lazy is
3580 * non-NULL, store an indication of whether all callbacks are lazy.
3581 * (If there are no callbacks, all of them are deemed to be lazy.)
3583 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3587 struct rcu_data *rdp;
3588 struct rcu_state *rsp;
3590 for_each_rcu_flavor(rsp) {
3591 rdp = this_cpu_ptr(rsp->rda);
3595 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3606 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3607 * the compiler is expected to optimize this away.
3609 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3610 int cpu, unsigned long done)
3612 trace_rcu_barrier(rsp->name, s, cpu,
3613 atomic_read(&rsp->barrier_cpu_count), done);
3617 * RCU callback function for _rcu_barrier(). If we are last, wake
3618 * up the task executing _rcu_barrier().
3620 static void rcu_barrier_callback(struct rcu_head *rhp)
3622 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3623 struct rcu_state *rsp = rdp->rsp;
3625 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3626 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
3627 complete(&rsp->barrier_completion);
3629 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
3634 * Called with preemption disabled, and from cross-cpu IRQ context.
3636 static void rcu_barrier_func(void *type)
3638 struct rcu_state *rsp = type;
3639 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3641 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
3642 atomic_inc(&rsp->barrier_cpu_count);
3643 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3647 * Orchestrate the specified type of RCU barrier, waiting for all
3648 * RCU callbacks of the specified type to complete.
3650 static void _rcu_barrier(struct rcu_state *rsp)
3653 struct rcu_data *rdp;
3654 unsigned long snap = READ_ONCE(rsp->n_barrier_done);
3655 unsigned long snap_done;
3657 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3659 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3660 mutex_lock(&rsp->barrier_mutex);
3663 * Ensure that all prior references, including to ->n_barrier_done,
3664 * are ordered before the _rcu_barrier() machinery.
3666 smp_mb(); /* See above block comment. */
3669 * Recheck ->n_barrier_done to see if others did our work for us.
3670 * This means checking ->n_barrier_done for an even-to-odd-to-even
3671 * transition. The "if" expression below therefore rounds the old
3672 * value up to the next even number and adds two before comparing.
3674 snap_done = rsp->n_barrier_done;
3675 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3678 * If the value in snap is odd, we needed to wait for the current
3679 * rcu_barrier() to complete, then wait for the next one, in other
3680 * words, we need the value of snap_done to be three larger than
3681 * the value of snap. On the other hand, if the value in snap is
3682 * even, we only had to wait for the next rcu_barrier() to complete,
3683 * in other words, we need the value of snap_done to be only two
3684 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3685 * this for us (thank you, Linus!).
3687 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3688 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3689 smp_mb(); /* caller's subsequent code after above check. */
3690 mutex_unlock(&rsp->barrier_mutex);
3695 * Increment ->n_barrier_done to avoid duplicate work. Use
3696 * WRITE_ONCE() to prevent the compiler from speculating
3697 * the increment to precede the early-exit check.
3699 WRITE_ONCE(rsp->n_barrier_done, rsp->n_barrier_done + 1);
3700 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3701 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3702 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3705 * Initialize the count to one rather than to zero in order to
3706 * avoid a too-soon return to zero in case of a short grace period
3707 * (or preemption of this task). Exclude CPU-hotplug operations
3708 * to ensure that no offline CPU has callbacks queued.
3710 init_completion(&rsp->barrier_completion);
3711 atomic_set(&rsp->barrier_cpu_count, 1);
3715 * Force each CPU with callbacks to register a new callback.
3716 * When that callback is invoked, we will know that all of the
3717 * corresponding CPU's preceding callbacks have been invoked.
3719 for_each_possible_cpu(cpu) {
3720 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3722 rdp = per_cpu_ptr(rsp->rda, cpu);
3723 if (rcu_is_nocb_cpu(cpu)) {
3724 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3725 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3726 rsp->n_barrier_done);
3728 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3729 rsp->n_barrier_done);
3730 smp_mb__before_atomic();
3731 atomic_inc(&rsp->barrier_cpu_count);
3732 __call_rcu(&rdp->barrier_head,
3733 rcu_barrier_callback, rsp, cpu, 0);
3735 } else if (READ_ONCE(rdp->qlen)) {
3736 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3737 rsp->n_barrier_done);
3738 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3740 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3741 rsp->n_barrier_done);
3747 * Now that we have an rcu_barrier_callback() callback on each
3748 * CPU, and thus each counted, remove the initial count.
3750 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3751 complete(&rsp->barrier_completion);
3753 /* Increment ->n_barrier_done to prevent duplicate work. */
3754 smp_mb(); /* Keep increment after above mechanism. */
3755 WRITE_ONCE(rsp->n_barrier_done, rsp->n_barrier_done + 1);
3756 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3757 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3758 smp_mb(); /* Keep increment before caller's subsequent code. */
3760 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3761 wait_for_completion(&rsp->barrier_completion);
3763 /* Other rcu_barrier() invocations can now safely proceed. */
3764 mutex_unlock(&rsp->barrier_mutex);
3768 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3770 void rcu_barrier_bh(void)
3772 _rcu_barrier(&rcu_bh_state);
3774 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3777 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3779 void rcu_barrier_sched(void)
3781 _rcu_barrier(&rcu_sched_state);
3783 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3786 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3787 * first CPU in a given leaf rcu_node structure coming online. The caller
3788 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3791 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3794 struct rcu_node *rnp = rnp_leaf;
3797 mask = rnp->grpmask;
3801 raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */
3802 rnp->qsmaskinit |= mask;
3803 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
3808 * Do boot-time initialization of a CPU's per-CPU RCU data.
3811 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3813 unsigned long flags;
3814 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3815 struct rcu_node *rnp = rcu_get_root(rsp);
3817 /* Set up local state, ensuring consistent view of global state. */
3818 raw_spin_lock_irqsave(&rnp->lock, flags);
3819 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3820 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3821 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3822 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3825 rcu_boot_init_nocb_percpu_data(rdp);
3826 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3830 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3831 * offline event can be happening at a given time. Note also that we
3832 * can accept some slop in the rsp->completed access due to the fact
3833 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3836 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3838 unsigned long flags;
3840 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3841 struct rcu_node *rnp = rcu_get_root(rsp);
3843 /* Set up local state, ensuring consistent view of global state. */
3844 raw_spin_lock_irqsave(&rnp->lock, flags);
3845 rdp->beenonline = 1; /* We have now been online. */
3846 rdp->qlen_last_fqs_check = 0;
3847 rdp->n_force_qs_snap = rsp->n_force_qs;
3848 rdp->blimit = blimit;
3850 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3851 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3852 rcu_sysidle_init_percpu_data(rdp->dynticks);
3853 atomic_set(&rdp->dynticks->dynticks,
3854 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3855 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3858 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3859 * propagation up the rcu_node tree will happen at the beginning
3860 * of the next grace period.
3863 mask = rdp->grpmask;
3864 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3865 smp_mb__after_unlock_lock();
3866 rnp->qsmaskinitnext |= mask;
3867 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3868 rdp->completed = rnp->completed;
3869 rdp->passed_quiesce = false;
3870 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
3871 rdp->qs_pending = false;
3872 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3873 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3876 static void rcu_prepare_cpu(int cpu)
3878 struct rcu_state *rsp;
3880 for_each_rcu_flavor(rsp)
3881 rcu_init_percpu_data(cpu, rsp);
3885 * Handle CPU online/offline notification events.
3887 int rcu_cpu_notify(struct notifier_block *self,
3888 unsigned long action, void *hcpu)
3890 long cpu = (long)hcpu;
3891 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3892 struct rcu_node *rnp = rdp->mynode;
3893 struct rcu_state *rsp;
3896 case CPU_UP_PREPARE:
3897 case CPU_UP_PREPARE_FROZEN:
3898 rcu_prepare_cpu(cpu);
3899 rcu_prepare_kthreads(cpu);
3900 rcu_spawn_all_nocb_kthreads(cpu);
3903 case CPU_DOWN_FAILED:
3904 rcu_boost_kthread_setaffinity(rnp, -1);
3906 case CPU_DOWN_PREPARE:
3907 rcu_boost_kthread_setaffinity(rnp, cpu);
3910 case CPU_DYING_FROZEN:
3911 for_each_rcu_flavor(rsp)
3912 rcu_cleanup_dying_cpu(rsp);
3914 case CPU_DYING_IDLE:
3915 for_each_rcu_flavor(rsp) {
3916 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3920 case CPU_DEAD_FROZEN:
3921 case CPU_UP_CANCELED:
3922 case CPU_UP_CANCELED_FROZEN:
3923 for_each_rcu_flavor(rsp) {
3924 rcu_cleanup_dead_cpu(cpu, rsp);
3925 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3934 static int rcu_pm_notify(struct notifier_block *self,
3935 unsigned long action, void *hcpu)
3938 case PM_HIBERNATION_PREPARE:
3939 case PM_SUSPEND_PREPARE:
3940 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3943 case PM_POST_HIBERNATION:
3944 case PM_POST_SUSPEND:
3945 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3946 rcu_unexpedite_gp();
3955 * Spawn the kthreads that handle each RCU flavor's grace periods.
3957 static int __init rcu_spawn_gp_kthread(void)
3959 unsigned long flags;
3960 int kthread_prio_in = kthread_prio;
3961 struct rcu_node *rnp;
3962 struct rcu_state *rsp;
3963 struct sched_param sp;
3964 struct task_struct *t;
3966 /* Force priority into range. */
3967 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3969 else if (kthread_prio < 0)
3971 else if (kthread_prio > 99)
3973 if (kthread_prio != kthread_prio_in)
3974 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3975 kthread_prio, kthread_prio_in);
3977 rcu_scheduler_fully_active = 1;
3978 for_each_rcu_flavor(rsp) {
3979 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3981 rnp = rcu_get_root(rsp);
3982 raw_spin_lock_irqsave(&rnp->lock, flags);
3983 rsp->gp_kthread = t;
3985 sp.sched_priority = kthread_prio;
3986 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3989 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3991 rcu_spawn_nocb_kthreads();
3992 rcu_spawn_boost_kthreads();
3995 early_initcall(rcu_spawn_gp_kthread);
3998 * This function is invoked towards the end of the scheduler's initialization
3999 * process. Before this is called, the idle task might contain
4000 * RCU read-side critical sections (during which time, this idle
4001 * task is booting the system). After this function is called, the
4002 * idle tasks are prohibited from containing RCU read-side critical
4003 * sections. This function also enables RCU lockdep checking.
4005 void rcu_scheduler_starting(void)
4007 WARN_ON(num_online_cpus() != 1);
4008 WARN_ON(nr_context_switches() > 0);
4009 rcu_scheduler_active = 1;
4013 * Compute the per-level fanout, either using the exact fanout specified
4014 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4016 static void __init rcu_init_levelspread(struct rcu_state *rsp)
4020 if (rcu_fanout_exact) {
4021 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4022 for (i = rcu_num_lvls - 2; i >= 0; i--)
4023 rsp->levelspread[i] = RCU_FANOUT;
4029 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4030 ccur = rsp->levelcnt[i];
4031 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
4038 * Helper function for rcu_init() that initializes one rcu_state structure.
4040 static void __init rcu_init_one(struct rcu_state *rsp,
4041 struct rcu_data __percpu *rda)
4043 static const char * const buf[] = {
4047 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
4048 static const char * const fqs[] = {
4052 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
4053 static u8 fl_mask = 0x1;
4057 struct rcu_node *rnp;
4059 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4061 /* Silence gcc 4.8 false positive about array index out of range. */
4062 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4063 panic("rcu_init_one: rcu_num_lvls out of range");
4065 /* Initialize the level-tracking arrays. */
4067 for (i = 0; i < rcu_num_lvls; i++)
4068 rsp->levelcnt[i] = num_rcu_lvl[i];
4069 for (i = 1; i < rcu_num_lvls; i++)
4070 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
4071 rcu_init_levelspread(rsp);
4072 rsp->flavor_mask = fl_mask;
4075 /* Initialize the elements themselves, starting from the leaves. */
4077 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4078 cpustride *= rsp->levelspread[i];
4079 rnp = rsp->level[i];
4080 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
4081 raw_spin_lock_init(&rnp->lock);
4082 lockdep_set_class_and_name(&rnp->lock,
4083 &rcu_node_class[i], buf[i]);
4084 raw_spin_lock_init(&rnp->fqslock);
4085 lockdep_set_class_and_name(&rnp->fqslock,
4086 &rcu_fqs_class[i], fqs[i]);
4087 rnp->gpnum = rsp->gpnum;
4088 rnp->completed = rsp->completed;
4090 rnp->qsmaskinit = 0;
4091 rnp->grplo = j * cpustride;
4092 rnp->grphi = (j + 1) * cpustride - 1;
4093 if (rnp->grphi >= nr_cpu_ids)
4094 rnp->grphi = nr_cpu_ids - 1;
4100 rnp->grpnum = j % rsp->levelspread[i - 1];
4101 rnp->grpmask = 1UL << rnp->grpnum;
4102 rnp->parent = rsp->level[i - 1] +
4103 j / rsp->levelspread[i - 1];
4106 INIT_LIST_HEAD(&rnp->blkd_tasks);
4107 rcu_init_one_nocb(rnp);
4111 init_waitqueue_head(&rsp->gp_wq);
4112 rnp = rsp->level[rcu_num_lvls - 1];
4113 for_each_possible_cpu(i) {
4114 while (i > rnp->grphi)
4116 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4117 rcu_boot_init_percpu_data(i, rsp);
4119 list_add(&rsp->flavors, &rcu_struct_flavors);
4123 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4124 * replace the definitions in tree.h because those are needed to size
4125 * the ->node array in the rcu_state structure.
4127 static void __init rcu_init_geometry(void)
4133 int rcu_capacity[MAX_RCU_LVLS + 1];
4136 * Initialize any unspecified boot parameters.
4137 * The default values of jiffies_till_first_fqs and
4138 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4139 * value, which is a function of HZ, then adding one for each
4140 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4142 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4143 if (jiffies_till_first_fqs == ULONG_MAX)
4144 jiffies_till_first_fqs = d;
4145 if (jiffies_till_next_fqs == ULONG_MAX)
4146 jiffies_till_next_fqs = d;
4148 /* If the compile-time values are accurate, just leave. */
4149 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4150 nr_cpu_ids == NR_CPUS)
4152 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4153 rcu_fanout_leaf, nr_cpu_ids);
4156 * Compute number of nodes that can be handled an rcu_node tree
4157 * with the given number of levels. Setting rcu_capacity[0] makes
4158 * some of the arithmetic easier.
4160 rcu_capacity[0] = 1;
4161 rcu_capacity[1] = rcu_fanout_leaf;
4162 for (i = 2; i <= MAX_RCU_LVLS; i++)
4163 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4166 * The boot-time rcu_fanout_leaf parameter is only permitted
4167 * to increase the leaf-level fanout, not decrease it. Of course,
4168 * the leaf-level fanout cannot exceed the number of bits in
4169 * the rcu_node masks. Finally, the tree must be able to accommodate
4170 * the configured number of CPUs. Complain and fall back to the
4171 * compile-time values if these limits are exceeded.
4173 if (rcu_fanout_leaf < RCU_FANOUT_LEAF ||
4174 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
4175 n > rcu_capacity[MAX_RCU_LVLS]) {
4180 /* Calculate the number of rcu_nodes at each level of the tree. */
4181 for (i = 1; i <= MAX_RCU_LVLS; i++)
4182 if (n <= rcu_capacity[i]) {
4183 for (j = 0; j <= i; j++)
4185 DIV_ROUND_UP(n, rcu_capacity[i - j]);
4187 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
4192 /* Calculate the total number of rcu_node structures. */
4194 for (i = 0; i <= MAX_RCU_LVLS; i++)
4195 rcu_num_nodes += num_rcu_lvl[i];
4200 * Dump out the structure of the rcu_node combining tree associated
4201 * with the rcu_state structure referenced by rsp.
4203 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4206 struct rcu_node *rnp;
4208 pr_info("rcu_node tree layout dump\n");
4210 rcu_for_each_node_breadth_first(rsp, rnp) {
4211 if (rnp->level != level) {
4216 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4221 void __init rcu_init(void)
4225 rcu_early_boot_tests();
4227 rcu_bootup_announce();
4228 rcu_init_geometry();
4229 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
4230 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
4232 rcu_dump_rcu_node_tree(&rcu_sched_state);
4233 __rcu_init_preempt();
4234 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4237 * We don't need protection against CPU-hotplug here because
4238 * this is called early in boot, before either interrupts
4239 * or the scheduler are operational.
4241 cpu_notifier(rcu_cpu_notify, 0);
4242 pm_notifier(rcu_pm_notify, 0);
4243 for_each_online_cpu(cpu)
4244 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
4247 #include "tree_plugin.h"