2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/trace_events.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
72 * In order to export the rcu_state name to the tracing tools, it
73 * needs to be added in the __tracepoint_string section.
74 * This requires defining a separate variable tp_<sname>_varname
75 * that points to the string being used, and this will allow
76 * the tracing userspace tools to be able to decipher the string
77 * address to the matching string.
80 # define DEFINE_RCU_TPS(sname) \
81 static char sname##_varname[] = #sname; \
82 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
83 # define RCU_STATE_NAME(sname) sname##_varname
85 # define DEFINE_RCU_TPS(sname)
86 # define RCU_STATE_NAME(sname) __stringify(sname)
89 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
90 DEFINE_RCU_TPS(sname) \
91 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
92 struct rcu_state sname##_state = { \
93 .level = { &sname##_state.node[0] }, \
94 .rda = &sname##_data, \
96 .gp_state = RCU_GP_IDLE, \
97 .gpnum = 0UL - 300UL, \
98 .completed = 0UL - 300UL, \
99 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
100 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
101 .orphan_donetail = &sname##_state.orphan_donelist, \
102 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
103 .name = RCU_STATE_NAME(sname), \
107 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
110 static struct rcu_state *const rcu_state_p;
111 LIST_HEAD(rcu_struct_flavors);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree;
115 module_param(dump_tree, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact;
118 module_param(rcu_fanout_exact, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121 module_param(rcu_fanout_leaf, int, 0444);
122 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123 /* Number of rcu_nodes at specified level. */
124 static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
128 * The rcu_scheduler_active variable transitions from zero to one just
129 * before the first task is spawned. So when this variable is zero, RCU
130 * can assume that there is but one task, allowing RCU to (for example)
131 * optimize synchronize_sched() to a simple barrier(). When this variable
132 * is one, RCU must actually do all the hard work required to detect real
133 * grace periods. This variable is also used to suppress boot-time false
134 * positives from lockdep-RCU error checking.
136 int rcu_scheduler_active __read_mostly;
137 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
140 * The rcu_scheduler_fully_active variable transitions from zero to one
141 * during the early_initcall() processing, which is after the scheduler
142 * is capable of creating new tasks. So RCU processing (for example,
143 * creating tasks for RCU priority boosting) must be delayed until after
144 * rcu_scheduler_fully_active transitions from zero to one. We also
145 * currently delay invocation of any RCU callbacks until after this point.
147 * It might later prove better for people registering RCU callbacks during
148 * early boot to take responsibility for these callbacks, but one step at
151 static int rcu_scheduler_fully_active __read_mostly;
153 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
154 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
155 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
156 static void invoke_rcu_core(void);
157 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
158 static void rcu_report_exp_rdp(struct rcu_state *rsp,
159 struct rcu_data *rdp, bool wake);
161 /* rcuc/rcub kthread realtime priority */
162 #ifdef CONFIG_RCU_KTHREAD_PRIO
163 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
164 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
165 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
166 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
167 module_param(kthread_prio, int, 0644);
169 /* Delay in jiffies for grace-period initialization delays, debug only. */
171 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
172 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
173 module_param(gp_preinit_delay, int, 0644);
174 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
175 static const int gp_preinit_delay;
176 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
178 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
179 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
180 module_param(gp_init_delay, int, 0644);
181 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
182 static const int gp_init_delay;
183 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
185 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
186 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
187 module_param(gp_cleanup_delay, int, 0644);
188 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
189 static const int gp_cleanup_delay;
190 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
193 * Number of grace periods between delays, normalized by the duration of
194 * the delay. The longer the the delay, the more the grace periods between
195 * each delay. The reason for this normalization is that it means that,
196 * for non-zero delays, the overall slowdown of grace periods is constant
197 * regardless of the duration of the delay. This arrangement balances
198 * the need for long delays to increase some race probabilities with the
199 * need for fast grace periods to increase other race probabilities.
201 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
204 * Track the rcutorture test sequence number and the update version
205 * number within a given test. The rcutorture_testseq is incremented
206 * on every rcutorture module load and unload, so has an odd value
207 * when a test is running. The rcutorture_vernum is set to zero
208 * when rcutorture starts and is incremented on each rcutorture update.
209 * These variables enable correlating rcutorture output with the
210 * RCU tracing information.
212 unsigned long rcutorture_testseq;
213 unsigned long rcutorture_vernum;
216 * Compute the mask of online CPUs for the specified rcu_node structure.
217 * This will not be stable unless the rcu_node structure's ->lock is
218 * held, but the bit corresponding to the current CPU will be stable
221 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
223 return READ_ONCE(rnp->qsmaskinitnext);
227 * Return true if an RCU grace period is in progress. The READ_ONCE()s
228 * permit this function to be invoked without holding the root rcu_node
229 * structure's ->lock, but of course results can be subject to change.
231 static int rcu_gp_in_progress(struct rcu_state *rsp)
233 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
237 * Note a quiescent state. Because we do not need to know
238 * how many quiescent states passed, just if there was at least
239 * one since the start of the grace period, this just sets a flag.
240 * The caller must have disabled preemption.
242 void rcu_sched_qs(void)
244 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
246 trace_rcu_grace_period(TPS("rcu_sched"),
247 __this_cpu_read(rcu_sched_data.gpnum),
249 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
250 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
252 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
253 rcu_report_exp_rdp(&rcu_sched_state,
254 this_cpu_ptr(&rcu_sched_data), true);
259 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
260 trace_rcu_grace_period(TPS("rcu_bh"),
261 __this_cpu_read(rcu_bh_data.gpnum),
263 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
267 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
269 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
270 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
271 .dynticks = ATOMIC_INIT(1),
272 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
273 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
274 .dynticks_idle = ATOMIC_INIT(1),
275 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
278 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
279 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
282 * Let the RCU core know that this CPU has gone through the scheduler,
283 * which is a quiescent state. This is called when the need for a
284 * quiescent state is urgent, so we burn an atomic operation and full
285 * memory barriers to let the RCU core know about it, regardless of what
286 * this CPU might (or might not) do in the near future.
288 * We inform the RCU core by emulating a zero-duration dyntick-idle
289 * period, which we in turn do by incrementing the ->dynticks counter
292 * The caller must have disabled interrupts.
294 static void rcu_momentary_dyntick_idle(void)
296 struct rcu_data *rdp;
297 struct rcu_dynticks *rdtp;
299 struct rcu_state *rsp;
302 * Yes, we can lose flag-setting operations. This is OK, because
303 * the flag will be set again after some delay.
305 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
306 raw_cpu_write(rcu_sched_qs_mask, 0);
308 /* Find the flavor that needs a quiescent state. */
309 for_each_rcu_flavor(rsp) {
310 rdp = raw_cpu_ptr(rsp->rda);
311 if (!(resched_mask & rsp->flavor_mask))
313 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
314 if (READ_ONCE(rdp->mynode->completed) !=
315 READ_ONCE(rdp->cond_resched_completed))
319 * Pretend to be momentarily idle for the quiescent state.
320 * This allows the grace-period kthread to record the
321 * quiescent state, with no need for this CPU to do anything
324 rdtp = this_cpu_ptr(&rcu_dynticks);
325 smp_mb__before_atomic(); /* Earlier stuff before QS. */
326 atomic_add(2, &rdtp->dynticks); /* QS. */
327 smp_mb__after_atomic(); /* Later stuff after QS. */
333 * Note a context switch. This is a quiescent state for RCU-sched,
334 * and requires special handling for preemptible RCU.
335 * The caller must have disabled interrupts.
337 void rcu_note_context_switch(void)
339 barrier(); /* Avoid RCU read-side critical sections leaking down. */
340 trace_rcu_utilization(TPS("Start context switch"));
342 rcu_preempt_note_context_switch();
343 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
344 rcu_momentary_dyntick_idle();
345 trace_rcu_utilization(TPS("End context switch"));
346 barrier(); /* Avoid RCU read-side critical sections leaking up. */
348 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
351 * Register a quiescent state for all RCU flavors. If there is an
352 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
353 * dyntick-idle quiescent state visible to other CPUs (but only for those
354 * RCU flavors in desperate need of a quiescent state, which will normally
355 * be none of them). Either way, do a lightweight quiescent state for
358 * The barrier() calls are redundant in the common case when this is
359 * called externally, but just in case this is called from within this
363 void rcu_all_qs(void)
367 barrier(); /* Avoid RCU read-side critical sections leaking down. */
368 if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) {
369 local_irq_save(flags);
370 rcu_momentary_dyntick_idle();
371 local_irq_restore(flags);
373 this_cpu_inc(rcu_qs_ctr);
374 barrier(); /* Avoid RCU read-side critical sections leaking up. */
376 EXPORT_SYMBOL_GPL(rcu_all_qs);
378 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
379 static long qhimark = 10000; /* If this many pending, ignore blimit. */
380 static long qlowmark = 100; /* Once only this many pending, use blimit. */
382 module_param(blimit, long, 0444);
383 module_param(qhimark, long, 0444);
384 module_param(qlowmark, long, 0444);
386 static ulong jiffies_till_first_fqs = ULONG_MAX;
387 static ulong jiffies_till_next_fqs = ULONG_MAX;
389 module_param(jiffies_till_first_fqs, ulong, 0644);
390 module_param(jiffies_till_next_fqs, ulong, 0644);
393 * How long the grace period must be before we start recruiting
394 * quiescent-state help from rcu_note_context_switch().
396 static ulong jiffies_till_sched_qs = HZ / 20;
397 module_param(jiffies_till_sched_qs, ulong, 0644);
399 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
400 struct rcu_data *rdp);
401 static void force_qs_rnp(struct rcu_state *rsp,
402 int (*f)(struct rcu_data *rsp, bool *isidle,
403 unsigned long *maxj),
404 bool *isidle, unsigned long *maxj);
405 static void force_quiescent_state(struct rcu_state *rsp);
406 static int rcu_pending(void);
409 * Return the number of RCU batches started thus far for debug & stats.
411 unsigned long rcu_batches_started(void)
413 return rcu_state_p->gpnum;
415 EXPORT_SYMBOL_GPL(rcu_batches_started);
418 * Return the number of RCU-sched batches started thus far for debug & stats.
420 unsigned long rcu_batches_started_sched(void)
422 return rcu_sched_state.gpnum;
424 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
427 * Return the number of RCU BH batches started thus far for debug & stats.
429 unsigned long rcu_batches_started_bh(void)
431 return rcu_bh_state.gpnum;
433 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
436 * Return the number of RCU batches completed thus far for debug & stats.
438 unsigned long rcu_batches_completed(void)
440 return rcu_state_p->completed;
442 EXPORT_SYMBOL_GPL(rcu_batches_completed);
445 * Return the number of RCU-sched batches completed thus far for debug & stats.
447 unsigned long rcu_batches_completed_sched(void)
449 return rcu_sched_state.completed;
451 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
454 * Return the number of RCU BH batches completed thus far for debug & stats.
456 unsigned long rcu_batches_completed_bh(void)
458 return rcu_bh_state.completed;
460 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
463 * Force a quiescent state.
465 void rcu_force_quiescent_state(void)
467 force_quiescent_state(rcu_state_p);
469 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
472 * Force a quiescent state for RCU BH.
474 void rcu_bh_force_quiescent_state(void)
476 force_quiescent_state(&rcu_bh_state);
478 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
481 * Force a quiescent state for RCU-sched.
483 void rcu_sched_force_quiescent_state(void)
485 force_quiescent_state(&rcu_sched_state);
487 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
490 * Show the state of the grace-period kthreads.
492 void show_rcu_gp_kthreads(void)
494 struct rcu_state *rsp;
496 for_each_rcu_flavor(rsp) {
497 pr_info("%s: wait state: %d ->state: %#lx\n",
498 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
499 /* sched_show_task(rsp->gp_kthread); */
502 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
505 * Record the number of times rcutorture tests have been initiated and
506 * terminated. This information allows the debugfs tracing stats to be
507 * correlated to the rcutorture messages, even when the rcutorture module
508 * is being repeatedly loaded and unloaded. In other words, we cannot
509 * store this state in rcutorture itself.
511 void rcutorture_record_test_transition(void)
513 rcutorture_testseq++;
514 rcutorture_vernum = 0;
516 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
519 * Send along grace-period-related data for rcutorture diagnostics.
521 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
522 unsigned long *gpnum, unsigned long *completed)
524 struct rcu_state *rsp = NULL;
533 case RCU_SCHED_FLAVOR:
534 rsp = &rcu_sched_state;
540 *flags = READ_ONCE(rsp->gp_flags);
541 *gpnum = READ_ONCE(rsp->gpnum);
542 *completed = READ_ONCE(rsp->completed);
549 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
552 * Record the number of writer passes through the current rcutorture test.
553 * This is also used to correlate debugfs tracing stats with the rcutorture
556 void rcutorture_record_progress(unsigned long vernum)
560 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
563 * Does the CPU have callbacks ready to be invoked?
566 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
568 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
569 rdp->nxttail[RCU_DONE_TAIL] != NULL;
573 * Return the root node of the specified rcu_state structure.
575 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
577 return &rsp->node[0];
581 * Is there any need for future grace periods?
582 * Interrupts must be disabled. If the caller does not hold the root
583 * rnp_node structure's ->lock, the results are advisory only.
585 static int rcu_future_needs_gp(struct rcu_state *rsp)
587 struct rcu_node *rnp = rcu_get_root(rsp);
588 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
589 int *fp = &rnp->need_future_gp[idx];
591 return READ_ONCE(*fp);
595 * Does the current CPU require a not-yet-started grace period?
596 * The caller must have disabled interrupts to prevent races with
597 * normal callback registry.
600 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
604 if (rcu_gp_in_progress(rsp))
605 return false; /* No, a grace period is already in progress. */
606 if (rcu_future_needs_gp(rsp))
607 return true; /* Yes, a no-CBs CPU needs one. */
608 if (!rdp->nxttail[RCU_NEXT_TAIL])
609 return false; /* No, this is a no-CBs (or offline) CPU. */
610 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
611 return true; /* Yes, CPU has newly registered callbacks. */
612 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
613 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
614 ULONG_CMP_LT(READ_ONCE(rsp->completed),
615 rdp->nxtcompleted[i]))
616 return true; /* Yes, CBs for future grace period. */
617 return false; /* No grace period needed. */
621 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
623 * If the new value of the ->dynticks_nesting counter now is zero,
624 * we really have entered idle, and must do the appropriate accounting.
625 * The caller must have disabled interrupts.
627 static void rcu_eqs_enter_common(long long oldval, bool user)
629 struct rcu_state *rsp;
630 struct rcu_data *rdp;
631 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
633 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
634 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
635 !user && !is_idle_task(current)) {
636 struct task_struct *idle __maybe_unused =
637 idle_task(smp_processor_id());
639 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
640 ftrace_dump(DUMP_ORIG);
641 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
642 current->pid, current->comm,
643 idle->pid, idle->comm); /* must be idle task! */
645 for_each_rcu_flavor(rsp) {
646 rdp = this_cpu_ptr(rsp->rda);
647 do_nocb_deferred_wakeup(rdp);
649 rcu_prepare_for_idle();
650 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
651 smp_mb__before_atomic(); /* See above. */
652 atomic_inc(&rdtp->dynticks);
653 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
654 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
655 atomic_read(&rdtp->dynticks) & 0x1);
656 rcu_dynticks_task_enter();
659 * It is illegal to enter an extended quiescent state while
660 * in an RCU read-side critical section.
662 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
663 "Illegal idle entry in RCU read-side critical section.");
664 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
665 "Illegal idle entry in RCU-bh read-side critical section.");
666 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
667 "Illegal idle entry in RCU-sched read-side critical section.");
671 * Enter an RCU extended quiescent state, which can be either the
672 * idle loop or adaptive-tickless usermode execution.
674 static void rcu_eqs_enter(bool user)
677 struct rcu_dynticks *rdtp;
679 rdtp = this_cpu_ptr(&rcu_dynticks);
680 oldval = rdtp->dynticks_nesting;
681 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
682 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
683 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
684 rdtp->dynticks_nesting = 0;
685 rcu_eqs_enter_common(oldval, user);
687 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
692 * rcu_idle_enter - inform RCU that current CPU is entering idle
694 * Enter idle mode, in other words, -leave- the mode in which RCU
695 * read-side critical sections can occur. (Though RCU read-side
696 * critical sections can occur in irq handlers in idle, a possibility
697 * handled by irq_enter() and irq_exit().)
699 * We crowbar the ->dynticks_nesting field to zero to allow for
700 * the possibility of usermode upcalls having messed up our count
701 * of interrupt nesting level during the prior busy period.
703 void rcu_idle_enter(void)
707 local_irq_save(flags);
708 rcu_eqs_enter(false);
709 rcu_sysidle_enter(0);
710 local_irq_restore(flags);
712 EXPORT_SYMBOL_GPL(rcu_idle_enter);
714 #ifdef CONFIG_NO_HZ_FULL
716 * rcu_user_enter - inform RCU that we are resuming userspace.
718 * Enter RCU idle mode right before resuming userspace. No use of RCU
719 * is permitted between this call and rcu_user_exit(). This way the
720 * CPU doesn't need to maintain the tick for RCU maintenance purposes
721 * when the CPU runs in userspace.
723 void rcu_user_enter(void)
727 #endif /* CONFIG_NO_HZ_FULL */
730 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
732 * Exit from an interrupt handler, which might possibly result in entering
733 * idle mode, in other words, leaving the mode in which read-side critical
734 * sections can occur. The caller must have disabled interrupts.
736 * This code assumes that the idle loop never does anything that might
737 * result in unbalanced calls to irq_enter() and irq_exit(). If your
738 * architecture violates this assumption, RCU will give you what you
739 * deserve, good and hard. But very infrequently and irreproducibly.
741 * Use things like work queues to work around this limitation.
743 * You have been warned.
745 void rcu_irq_exit(void)
748 struct rcu_dynticks *rdtp;
750 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
751 rdtp = this_cpu_ptr(&rcu_dynticks);
752 oldval = rdtp->dynticks_nesting;
753 rdtp->dynticks_nesting--;
754 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
755 rdtp->dynticks_nesting < 0);
756 if (rdtp->dynticks_nesting)
757 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
759 rcu_eqs_enter_common(oldval, true);
760 rcu_sysidle_enter(1);
764 * Wrapper for rcu_irq_exit() where interrupts are enabled.
766 void rcu_irq_exit_irqson(void)
770 local_irq_save(flags);
772 local_irq_restore(flags);
776 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
778 * If the new value of the ->dynticks_nesting counter was previously zero,
779 * we really have exited idle, and must do the appropriate accounting.
780 * The caller must have disabled interrupts.
782 static void rcu_eqs_exit_common(long long oldval, int user)
784 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
786 rcu_dynticks_task_exit();
787 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
788 atomic_inc(&rdtp->dynticks);
789 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
790 smp_mb__after_atomic(); /* See above. */
791 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
792 !(atomic_read(&rdtp->dynticks) & 0x1));
793 rcu_cleanup_after_idle();
794 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
795 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
796 !user && !is_idle_task(current)) {
797 struct task_struct *idle __maybe_unused =
798 idle_task(smp_processor_id());
800 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
801 oldval, rdtp->dynticks_nesting);
802 ftrace_dump(DUMP_ORIG);
803 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
804 current->pid, current->comm,
805 idle->pid, idle->comm); /* must be idle task! */
810 * Exit an RCU extended quiescent state, which can be either the
811 * idle loop or adaptive-tickless usermode execution.
813 static void rcu_eqs_exit(bool user)
815 struct rcu_dynticks *rdtp;
818 rdtp = this_cpu_ptr(&rcu_dynticks);
819 oldval = rdtp->dynticks_nesting;
820 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
821 if (oldval & DYNTICK_TASK_NEST_MASK) {
822 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
824 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
825 rcu_eqs_exit_common(oldval, user);
830 * rcu_idle_exit - inform RCU that current CPU is leaving idle
832 * Exit idle mode, in other words, -enter- the mode in which RCU
833 * read-side critical sections can occur.
835 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
836 * allow for the possibility of usermode upcalls messing up our count
837 * of interrupt nesting level during the busy period that is just
840 void rcu_idle_exit(void)
844 local_irq_save(flags);
847 local_irq_restore(flags);
849 EXPORT_SYMBOL_GPL(rcu_idle_exit);
851 #ifdef CONFIG_NO_HZ_FULL
853 * rcu_user_exit - inform RCU that we are exiting userspace.
855 * Exit RCU idle mode while entering the kernel because it can
856 * run a RCU read side critical section anytime.
858 void rcu_user_exit(void)
862 #endif /* CONFIG_NO_HZ_FULL */
865 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
867 * Enter an interrupt handler, which might possibly result in exiting
868 * idle mode, in other words, entering the mode in which read-side critical
869 * sections can occur. The caller must have disabled interrupts.
871 * Note that the Linux kernel is fully capable of entering an interrupt
872 * handler that it never exits, for example when doing upcalls to
873 * user mode! This code assumes that the idle loop never does upcalls to
874 * user mode. If your architecture does do upcalls from the idle loop (or
875 * does anything else that results in unbalanced calls to the irq_enter()
876 * and irq_exit() functions), RCU will give you what you deserve, good
877 * and hard. But very infrequently and irreproducibly.
879 * Use things like work queues to work around this limitation.
881 * You have been warned.
883 void rcu_irq_enter(void)
885 struct rcu_dynticks *rdtp;
888 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
889 rdtp = this_cpu_ptr(&rcu_dynticks);
890 oldval = rdtp->dynticks_nesting;
891 rdtp->dynticks_nesting++;
892 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
893 rdtp->dynticks_nesting == 0);
895 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
897 rcu_eqs_exit_common(oldval, true);
902 * Wrapper for rcu_irq_enter() where interrupts are enabled.
904 void rcu_irq_enter_irqson(void)
908 local_irq_save(flags);
910 local_irq_restore(flags);
914 * rcu_nmi_enter - inform RCU of entry to NMI context
916 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
917 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
918 * that the CPU is active. This implementation permits nested NMIs, as
919 * long as the nesting level does not overflow an int. (You will probably
920 * run out of stack space first.)
922 void rcu_nmi_enter(void)
924 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
927 /* Complain about underflow. */
928 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
931 * If idle from RCU viewpoint, atomically increment ->dynticks
932 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
933 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
934 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
935 * to be in the outermost NMI handler that interrupted an RCU-idle
936 * period (observation due to Andy Lutomirski).
938 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
939 smp_mb__before_atomic(); /* Force delay from prior write. */
940 atomic_inc(&rdtp->dynticks);
941 /* atomic_inc() before later RCU read-side crit sects */
942 smp_mb__after_atomic(); /* See above. */
943 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
946 rdtp->dynticks_nmi_nesting += incby;
951 * rcu_nmi_exit - inform RCU of exit from NMI context
953 * If we are returning from the outermost NMI handler that interrupted an
954 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
955 * to let the RCU grace-period handling know that the CPU is back to
958 void rcu_nmi_exit(void)
960 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
963 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
964 * (We are exiting an NMI handler, so RCU better be paying attention
967 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
968 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
971 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
972 * leave it in non-RCU-idle state.
974 if (rdtp->dynticks_nmi_nesting != 1) {
975 rdtp->dynticks_nmi_nesting -= 2;
979 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
980 rdtp->dynticks_nmi_nesting = 0;
981 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
982 smp_mb__before_atomic(); /* See above. */
983 atomic_inc(&rdtp->dynticks);
984 smp_mb__after_atomic(); /* Force delay to next write. */
985 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
989 * __rcu_is_watching - are RCU read-side critical sections safe?
991 * Return true if RCU is watching the running CPU, which means that
992 * this CPU can safely enter RCU read-side critical sections. Unlike
993 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
994 * least disabled preemption.
996 bool notrace __rcu_is_watching(void)
998 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
1002 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1004 * If the current CPU is in its idle loop and is neither in an interrupt
1005 * or NMI handler, return true.
1007 bool notrace rcu_is_watching(void)
1011 preempt_disable_notrace();
1012 ret = __rcu_is_watching();
1013 preempt_enable_notrace();
1016 EXPORT_SYMBOL_GPL(rcu_is_watching);
1018 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1021 * Is the current CPU online? Disable preemption to avoid false positives
1022 * that could otherwise happen due to the current CPU number being sampled,
1023 * this task being preempted, its old CPU being taken offline, resuming
1024 * on some other CPU, then determining that its old CPU is now offline.
1025 * It is OK to use RCU on an offline processor during initial boot, hence
1026 * the check for rcu_scheduler_fully_active. Note also that it is OK
1027 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1028 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1029 * offline to continue to use RCU for one jiffy after marking itself
1030 * offline in the cpu_online_mask. This leniency is necessary given the
1031 * non-atomic nature of the online and offline processing, for example,
1032 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1035 * This is also why RCU internally marks CPUs online during the
1036 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1038 * Disable checking if in an NMI handler because we cannot safely report
1039 * errors from NMI handlers anyway.
1041 bool rcu_lockdep_current_cpu_online(void)
1043 struct rcu_data *rdp;
1044 struct rcu_node *rnp;
1050 rdp = this_cpu_ptr(&rcu_sched_data);
1052 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1053 !rcu_scheduler_fully_active;
1057 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1059 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1062 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1064 * If the current CPU is idle or running at a first-level (not nested)
1065 * interrupt from idle, return true. The caller must have at least
1066 * disabled preemption.
1068 static int rcu_is_cpu_rrupt_from_idle(void)
1070 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1074 * Snapshot the specified CPU's dynticks counter so that we can later
1075 * credit them with an implicit quiescent state. Return 1 if this CPU
1076 * is in dynticks idle mode, which is an extended quiescent state.
1078 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1079 bool *isidle, unsigned long *maxj)
1081 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1082 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1083 if ((rdp->dynticks_snap & 0x1) == 0) {
1084 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1085 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1086 rdp->mynode->gpnum))
1087 WRITE_ONCE(rdp->gpwrap, true);
1094 * Return true if the specified CPU has passed through a quiescent
1095 * state by virtue of being in or having passed through an dynticks
1096 * idle state since the last call to dyntick_save_progress_counter()
1097 * for this same CPU, or by virtue of having been offline.
1099 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1100 bool *isidle, unsigned long *maxj)
1106 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1107 snap = (unsigned int)rdp->dynticks_snap;
1110 * If the CPU passed through or entered a dynticks idle phase with
1111 * no active irq/NMI handlers, then we can safely pretend that the CPU
1112 * already acknowledged the request to pass through a quiescent
1113 * state. Either way, that CPU cannot possibly be in an RCU
1114 * read-side critical section that started before the beginning
1115 * of the current RCU grace period.
1117 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1118 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1119 rdp->dynticks_fqs++;
1124 * Check for the CPU being offline, but only if the grace period
1125 * is old enough. We don't need to worry about the CPU changing
1126 * state: If we see it offline even once, it has been through a
1129 * The reason for insisting that the grace period be at least
1130 * one jiffy old is that CPUs that are not quite online and that
1131 * have just gone offline can still execute RCU read-side critical
1134 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1135 return 0; /* Grace period is not old enough. */
1137 if (cpu_is_offline(rdp->cpu)) {
1138 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1144 * A CPU running for an extended time within the kernel can
1145 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1146 * even context-switching back and forth between a pair of
1147 * in-kernel CPU-bound tasks cannot advance grace periods.
1148 * So if the grace period is old enough, make the CPU pay attention.
1149 * Note that the unsynchronized assignments to the per-CPU
1150 * rcu_sched_qs_mask variable are safe. Yes, setting of
1151 * bits can be lost, but they will be set again on the next
1152 * force-quiescent-state pass. So lost bit sets do not result
1153 * in incorrect behavior, merely in a grace period lasting
1154 * a few jiffies longer than it might otherwise. Because
1155 * there are at most four threads involved, and because the
1156 * updates are only once every few jiffies, the probability of
1157 * lossage (and thus of slight grace-period extension) is
1160 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1161 * is set too high, we override with half of the RCU CPU stall
1164 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1165 if (ULONG_CMP_GE(jiffies,
1166 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1167 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1168 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1169 WRITE_ONCE(rdp->cond_resched_completed,
1170 READ_ONCE(rdp->mynode->completed));
1171 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1173 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1175 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1178 /* And if it has been a really long time, kick the CPU as well. */
1179 if (ULONG_CMP_GE(jiffies,
1180 rdp->rsp->gp_start + 2 * jiffies_till_sched_qs) ||
1181 ULONG_CMP_GE(jiffies, rdp->rsp->gp_start + jiffies_till_sched_qs))
1182 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1187 static void record_gp_stall_check_time(struct rcu_state *rsp)
1189 unsigned long j = jiffies;
1193 smp_wmb(); /* Record start time before stall time. */
1194 j1 = rcu_jiffies_till_stall_check();
1195 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1196 rsp->jiffies_resched = j + j1 / 2;
1197 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1201 * Convert a ->gp_state value to a character string.
1203 static const char *gp_state_getname(short gs)
1205 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1207 return gp_state_names[gs];
1211 * Complain about starvation of grace-period kthread.
1213 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1219 gpa = READ_ONCE(rsp->gp_activity);
1220 if (j - gpa > 2 * HZ) {
1221 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1223 rsp->gpnum, rsp->completed,
1225 gp_state_getname(rsp->gp_state), rsp->gp_state,
1226 rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
1227 if (rsp->gp_kthread)
1228 sched_show_task(rsp->gp_kthread);
1233 * Dump stacks of all tasks running on stalled CPUs.
1235 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1238 unsigned long flags;
1239 struct rcu_node *rnp;
1241 rcu_for_each_leaf_node(rsp, rnp) {
1242 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1243 if (rnp->qsmask != 0) {
1244 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1245 if (rnp->qsmask & (1UL << cpu))
1246 dump_cpu_task(rnp->grplo + cpu);
1248 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1252 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1256 unsigned long flags;
1260 struct rcu_node *rnp = rcu_get_root(rsp);
1263 /* Only let one CPU complain about others per time interval. */
1265 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1266 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1267 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1268 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1271 WRITE_ONCE(rsp->jiffies_stall,
1272 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1273 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1276 * OK, time to rat on our buddy...
1277 * See Documentation/RCU/stallwarn.txt for info on how to debug
1278 * RCU CPU stall warnings.
1280 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1282 print_cpu_stall_info_begin();
1283 rcu_for_each_leaf_node(rsp, rnp) {
1284 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1285 ndetected += rcu_print_task_stall(rnp);
1286 if (rnp->qsmask != 0) {
1287 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1288 if (rnp->qsmask & (1UL << cpu)) {
1289 print_cpu_stall_info(rsp,
1294 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1297 print_cpu_stall_info_end();
1298 for_each_possible_cpu(cpu)
1299 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1300 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1301 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1302 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1304 rcu_dump_cpu_stacks(rsp);
1306 if (READ_ONCE(rsp->gpnum) != gpnum ||
1307 READ_ONCE(rsp->completed) == gpnum) {
1308 pr_err("INFO: Stall ended before state dump start\n");
1311 gpa = READ_ONCE(rsp->gp_activity);
1312 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1313 rsp->name, j - gpa, j, gpa,
1314 jiffies_till_next_fqs,
1315 rcu_get_root(rsp)->qsmask);
1316 /* In this case, the current CPU might be at fault. */
1317 sched_show_task(current);
1321 /* Complain about tasks blocking the grace period. */
1322 rcu_print_detail_task_stall(rsp);
1324 rcu_check_gp_kthread_starvation(rsp);
1326 force_quiescent_state(rsp); /* Kick them all. */
1329 static void print_cpu_stall(struct rcu_state *rsp)
1332 unsigned long flags;
1333 struct rcu_node *rnp = rcu_get_root(rsp);
1337 * OK, time to rat on ourselves...
1338 * See Documentation/RCU/stallwarn.txt for info on how to debug
1339 * RCU CPU stall warnings.
1341 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1342 print_cpu_stall_info_begin();
1343 print_cpu_stall_info(rsp, smp_processor_id());
1344 print_cpu_stall_info_end();
1345 for_each_possible_cpu(cpu)
1346 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1347 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1348 jiffies - rsp->gp_start,
1349 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1351 rcu_check_gp_kthread_starvation(rsp);
1353 rcu_dump_cpu_stacks(rsp);
1355 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1356 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1357 WRITE_ONCE(rsp->jiffies_stall,
1358 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1359 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1362 * Attempt to revive the RCU machinery by forcing a context switch.
1364 * A context switch would normally allow the RCU state machine to make
1365 * progress and it could be we're stuck in kernel space without context
1366 * switches for an entirely unreasonable amount of time.
1368 resched_cpu(smp_processor_id());
1371 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1373 unsigned long completed;
1374 unsigned long gpnum;
1378 struct rcu_node *rnp;
1380 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1385 * Lots of memory barriers to reject false positives.
1387 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1388 * then rsp->gp_start, and finally rsp->completed. These values
1389 * are updated in the opposite order with memory barriers (or
1390 * equivalent) during grace-period initialization and cleanup.
1391 * Now, a false positive can occur if we get an new value of
1392 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1393 * the memory barriers, the only way that this can happen is if one
1394 * grace period ends and another starts between these two fetches.
1395 * Detect this by comparing rsp->completed with the previous fetch
1398 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1399 * and rsp->gp_start suffice to forestall false positives.
1401 gpnum = READ_ONCE(rsp->gpnum);
1402 smp_rmb(); /* Pick up ->gpnum first... */
1403 js = READ_ONCE(rsp->jiffies_stall);
1404 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1405 gps = READ_ONCE(rsp->gp_start);
1406 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1407 completed = READ_ONCE(rsp->completed);
1408 if (ULONG_CMP_GE(completed, gpnum) ||
1409 ULONG_CMP_LT(j, js) ||
1410 ULONG_CMP_GE(gps, js))
1411 return; /* No stall or GP completed since entering function. */
1413 if (rcu_gp_in_progress(rsp) &&
1414 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1416 /* We haven't checked in, so go dump stack. */
1417 print_cpu_stall(rsp);
1419 } else if (rcu_gp_in_progress(rsp) &&
1420 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1422 /* They had a few time units to dump stack, so complain. */
1423 print_other_cpu_stall(rsp, gpnum);
1428 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1430 * Set the stall-warning timeout way off into the future, thus preventing
1431 * any RCU CPU stall-warning messages from appearing in the current set of
1432 * RCU grace periods.
1434 * The caller must disable hard irqs.
1436 void rcu_cpu_stall_reset(void)
1438 struct rcu_state *rsp;
1440 for_each_rcu_flavor(rsp)
1441 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1445 * Initialize the specified rcu_data structure's default callback list
1446 * to empty. The default callback list is the one that is not used by
1447 * no-callbacks CPUs.
1449 static void init_default_callback_list(struct rcu_data *rdp)
1453 rdp->nxtlist = NULL;
1454 for (i = 0; i < RCU_NEXT_SIZE; i++)
1455 rdp->nxttail[i] = &rdp->nxtlist;
1459 * Initialize the specified rcu_data structure's callback list to empty.
1461 static void init_callback_list(struct rcu_data *rdp)
1463 if (init_nocb_callback_list(rdp))
1465 init_default_callback_list(rdp);
1469 * Determine the value that ->completed will have at the end of the
1470 * next subsequent grace period. This is used to tag callbacks so that
1471 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1472 * been dyntick-idle for an extended period with callbacks under the
1473 * influence of RCU_FAST_NO_HZ.
1475 * The caller must hold rnp->lock with interrupts disabled.
1477 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1478 struct rcu_node *rnp)
1481 * If RCU is idle, we just wait for the next grace period.
1482 * But we can only be sure that RCU is idle if we are looking
1483 * at the root rcu_node structure -- otherwise, a new grace
1484 * period might have started, but just not yet gotten around
1485 * to initializing the current non-root rcu_node structure.
1487 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1488 return rnp->completed + 1;
1491 * Otherwise, wait for a possible partial grace period and
1492 * then the subsequent full grace period.
1494 return rnp->completed + 2;
1498 * Trace-event helper function for rcu_start_future_gp() and
1499 * rcu_nocb_wait_gp().
1501 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1502 unsigned long c, const char *s)
1504 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1505 rnp->completed, c, rnp->level,
1506 rnp->grplo, rnp->grphi, s);
1510 * Start some future grace period, as needed to handle newly arrived
1511 * callbacks. The required future grace periods are recorded in each
1512 * rcu_node structure's ->need_future_gp field. Returns true if there
1513 * is reason to awaken the grace-period kthread.
1515 * The caller must hold the specified rcu_node structure's ->lock.
1517 static bool __maybe_unused
1518 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1519 unsigned long *c_out)
1524 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1527 * Pick up grace-period number for new callbacks. If this
1528 * grace period is already marked as needed, return to the caller.
1530 c = rcu_cbs_completed(rdp->rsp, rnp);
1531 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1532 if (rnp->need_future_gp[c & 0x1]) {
1533 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1538 * If either this rcu_node structure or the root rcu_node structure
1539 * believe that a grace period is in progress, then we must wait
1540 * for the one following, which is in "c". Because our request
1541 * will be noticed at the end of the current grace period, we don't
1542 * need to explicitly start one. We only do the lockless check
1543 * of rnp_root's fields if the current rcu_node structure thinks
1544 * there is no grace period in flight, and because we hold rnp->lock,
1545 * the only possible change is when rnp_root's two fields are
1546 * equal, in which case rnp_root->gpnum might be concurrently
1547 * incremented. But that is OK, as it will just result in our
1548 * doing some extra useless work.
1550 if (rnp->gpnum != rnp->completed ||
1551 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1552 rnp->need_future_gp[c & 0x1]++;
1553 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1558 * There might be no grace period in progress. If we don't already
1559 * hold it, acquire the root rcu_node structure's lock in order to
1560 * start one (if needed).
1562 if (rnp != rnp_root)
1563 raw_spin_lock_rcu_node(rnp_root);
1566 * Get a new grace-period number. If there really is no grace
1567 * period in progress, it will be smaller than the one we obtained
1568 * earlier. Adjust callbacks as needed. Note that even no-CBs
1569 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1571 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1572 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1573 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1574 rdp->nxtcompleted[i] = c;
1577 * If the needed for the required grace period is already
1578 * recorded, trace and leave.
1580 if (rnp_root->need_future_gp[c & 0x1]) {
1581 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1585 /* Record the need for the future grace period. */
1586 rnp_root->need_future_gp[c & 0x1]++;
1588 /* If a grace period is not already in progress, start one. */
1589 if (rnp_root->gpnum != rnp_root->completed) {
1590 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1592 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1593 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1596 if (rnp != rnp_root)
1597 raw_spin_unlock_rcu_node(rnp_root);
1605 * Clean up any old requests for the just-ended grace period. Also return
1606 * whether any additional grace periods have been requested. Also invoke
1607 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1608 * waiting for this grace period to complete.
1610 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1612 int c = rnp->completed;
1614 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1616 rcu_nocb_gp_cleanup(rsp, rnp);
1617 rnp->need_future_gp[c & 0x1] = 0;
1618 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1619 trace_rcu_future_gp(rnp, rdp, c,
1620 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1625 * Awaken the grace-period kthread for the specified flavor of RCU.
1626 * Don't do a self-awaken, and don't bother awakening when there is
1627 * nothing for the grace-period kthread to do (as in several CPUs
1628 * raced to awaken, and we lost), and finally don't try to awaken
1629 * a kthread that has not yet been created.
1631 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1633 if (current == rsp->gp_kthread ||
1634 !READ_ONCE(rsp->gp_flags) ||
1637 wake_up(&rsp->gp_wq);
1641 * If there is room, assign a ->completed number to any callbacks on
1642 * this CPU that have not already been assigned. Also accelerate any
1643 * callbacks that were previously assigned a ->completed number that has
1644 * since proven to be too conservative, which can happen if callbacks get
1645 * assigned a ->completed number while RCU is idle, but with reference to
1646 * a non-root rcu_node structure. This function is idempotent, so it does
1647 * not hurt to call it repeatedly. Returns an flag saying that we should
1648 * awaken the RCU grace-period kthread.
1650 * The caller must hold rnp->lock with interrupts disabled.
1652 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1653 struct rcu_data *rdp)
1659 /* If the CPU has no callbacks, nothing to do. */
1660 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1664 * Starting from the sublist containing the callbacks most
1665 * recently assigned a ->completed number and working down, find the
1666 * first sublist that is not assignable to an upcoming grace period.
1667 * Such a sublist has something in it (first two tests) and has
1668 * a ->completed number assigned that will complete sooner than
1669 * the ->completed number for newly arrived callbacks (last test).
1671 * The key point is that any later sublist can be assigned the
1672 * same ->completed number as the newly arrived callbacks, which
1673 * means that the callbacks in any of these later sublist can be
1674 * grouped into a single sublist, whether or not they have already
1675 * been assigned a ->completed number.
1677 c = rcu_cbs_completed(rsp, rnp);
1678 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1679 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1680 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1684 * If there are no sublist for unassigned callbacks, leave.
1685 * At the same time, advance "i" one sublist, so that "i" will
1686 * index into the sublist where all the remaining callbacks should
1689 if (++i >= RCU_NEXT_TAIL)
1693 * Assign all subsequent callbacks' ->completed number to the next
1694 * full grace period and group them all in the sublist initially
1697 for (; i <= RCU_NEXT_TAIL; i++) {
1698 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1699 rdp->nxtcompleted[i] = c;
1701 /* Record any needed additional grace periods. */
1702 ret = rcu_start_future_gp(rnp, rdp, NULL);
1704 /* Trace depending on how much we were able to accelerate. */
1705 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1706 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1708 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1713 * Move any callbacks whose grace period has completed to the
1714 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1715 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1716 * sublist. This function is idempotent, so it does not hurt to
1717 * invoke it repeatedly. As long as it is not invoked -too- often...
1718 * Returns true if the RCU grace-period kthread needs to be awakened.
1720 * The caller must hold rnp->lock with interrupts disabled.
1722 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1723 struct rcu_data *rdp)
1727 /* If the CPU has no callbacks, nothing to do. */
1728 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1732 * Find all callbacks whose ->completed numbers indicate that they
1733 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1735 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1736 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1738 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1740 /* Clean up any sublist tail pointers that were misordered above. */
1741 for (j = RCU_WAIT_TAIL; j < i; j++)
1742 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1744 /* Copy down callbacks to fill in empty sublists. */
1745 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1746 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1748 rdp->nxttail[j] = rdp->nxttail[i];
1749 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1752 /* Classify any remaining callbacks. */
1753 return rcu_accelerate_cbs(rsp, rnp, rdp);
1757 * Update CPU-local rcu_data state to record the beginnings and ends of
1758 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1759 * structure corresponding to the current CPU, and must have irqs disabled.
1760 * Returns true if the grace-period kthread needs to be awakened.
1762 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1763 struct rcu_data *rdp)
1767 /* Handle the ends of any preceding grace periods first. */
1768 if (rdp->completed == rnp->completed &&
1769 !unlikely(READ_ONCE(rdp->gpwrap))) {
1771 /* No grace period end, so just accelerate recent callbacks. */
1772 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1776 /* Advance callbacks. */
1777 ret = rcu_advance_cbs(rsp, rnp, rdp);
1779 /* Remember that we saw this grace-period completion. */
1780 rdp->completed = rnp->completed;
1781 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1784 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1786 * If the current grace period is waiting for this CPU,
1787 * set up to detect a quiescent state, otherwise don't
1788 * go looking for one.
1790 rdp->gpnum = rnp->gpnum;
1791 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1792 rdp->cpu_no_qs.b.norm = true;
1793 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1794 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1795 zero_cpu_stall_ticks(rdp);
1796 WRITE_ONCE(rdp->gpwrap, false);
1801 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1803 unsigned long flags;
1805 struct rcu_node *rnp;
1807 local_irq_save(flags);
1809 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1810 rdp->completed == READ_ONCE(rnp->completed) &&
1811 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1812 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1813 local_irq_restore(flags);
1816 needwake = __note_gp_changes(rsp, rnp, rdp);
1817 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1819 rcu_gp_kthread_wake(rsp);
1822 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1825 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1826 schedule_timeout_uninterruptible(delay);
1830 * Initialize a new grace period. Return false if no grace period required.
1832 static bool rcu_gp_init(struct rcu_state *rsp)
1834 unsigned long oldmask;
1835 struct rcu_data *rdp;
1836 struct rcu_node *rnp = rcu_get_root(rsp);
1838 WRITE_ONCE(rsp->gp_activity, jiffies);
1839 raw_spin_lock_irq_rcu_node(rnp);
1840 if (!READ_ONCE(rsp->gp_flags)) {
1841 /* Spurious wakeup, tell caller to go back to sleep. */
1842 raw_spin_unlock_irq_rcu_node(rnp);
1845 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1847 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1849 * Grace period already in progress, don't start another.
1850 * Not supposed to be able to happen.
1852 raw_spin_unlock_irq_rcu_node(rnp);
1856 /* Advance to a new grace period and initialize state. */
1857 record_gp_stall_check_time(rsp);
1858 /* Record GP times before starting GP, hence smp_store_release(). */
1859 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1860 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1861 raw_spin_unlock_irq_rcu_node(rnp);
1864 * Apply per-leaf buffered online and offline operations to the
1865 * rcu_node tree. Note that this new grace period need not wait
1866 * for subsequent online CPUs, and that quiescent-state forcing
1867 * will handle subsequent offline CPUs.
1869 rcu_for_each_leaf_node(rsp, rnp) {
1870 rcu_gp_slow(rsp, gp_preinit_delay);
1871 raw_spin_lock_irq_rcu_node(rnp);
1872 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1873 !rnp->wait_blkd_tasks) {
1874 /* Nothing to do on this leaf rcu_node structure. */
1875 raw_spin_unlock_irq_rcu_node(rnp);
1879 /* Record old state, apply changes to ->qsmaskinit field. */
1880 oldmask = rnp->qsmaskinit;
1881 rnp->qsmaskinit = rnp->qsmaskinitnext;
1883 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1884 if (!oldmask != !rnp->qsmaskinit) {
1885 if (!oldmask) /* First online CPU for this rcu_node. */
1886 rcu_init_new_rnp(rnp);
1887 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1888 rnp->wait_blkd_tasks = true;
1889 else /* Last offline CPU and can propagate. */
1890 rcu_cleanup_dead_rnp(rnp);
1894 * If all waited-on tasks from prior grace period are
1895 * done, and if all this rcu_node structure's CPUs are
1896 * still offline, propagate up the rcu_node tree and
1897 * clear ->wait_blkd_tasks. Otherwise, if one of this
1898 * rcu_node structure's CPUs has since come back online,
1899 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1900 * checks for this, so just call it unconditionally).
1902 if (rnp->wait_blkd_tasks &&
1903 (!rcu_preempt_has_tasks(rnp) ||
1905 rnp->wait_blkd_tasks = false;
1906 rcu_cleanup_dead_rnp(rnp);
1909 raw_spin_unlock_irq_rcu_node(rnp);
1913 * Set the quiescent-state-needed bits in all the rcu_node
1914 * structures for all currently online CPUs in breadth-first order,
1915 * starting from the root rcu_node structure, relying on the layout
1916 * of the tree within the rsp->node[] array. Note that other CPUs
1917 * will access only the leaves of the hierarchy, thus seeing that no
1918 * grace period is in progress, at least until the corresponding
1919 * leaf node has been initialized. In addition, we have excluded
1920 * CPU-hotplug operations.
1922 * The grace period cannot complete until the initialization
1923 * process finishes, because this kthread handles both.
1925 rcu_for_each_node_breadth_first(rsp, rnp) {
1926 rcu_gp_slow(rsp, gp_init_delay);
1927 raw_spin_lock_irq_rcu_node(rnp);
1928 rdp = this_cpu_ptr(rsp->rda);
1929 rcu_preempt_check_blocked_tasks(rnp);
1930 rnp->qsmask = rnp->qsmaskinit;
1931 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1932 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1933 WRITE_ONCE(rnp->completed, rsp->completed);
1934 if (rnp == rdp->mynode)
1935 (void)__note_gp_changes(rsp, rnp, rdp);
1936 rcu_preempt_boost_start_gp(rnp);
1937 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1938 rnp->level, rnp->grplo,
1939 rnp->grphi, rnp->qsmask);
1940 raw_spin_unlock_irq_rcu_node(rnp);
1941 cond_resched_rcu_qs();
1942 WRITE_ONCE(rsp->gp_activity, jiffies);
1949 * Helper function for wait_event_interruptible_timeout() wakeup
1950 * at force-quiescent-state time.
1952 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1954 struct rcu_node *rnp = rcu_get_root(rsp);
1956 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1957 *gfp = READ_ONCE(rsp->gp_flags);
1958 if (*gfp & RCU_GP_FLAG_FQS)
1961 /* The current grace period has completed. */
1962 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1969 * Do one round of quiescent-state forcing.
1971 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
1973 bool isidle = false;
1975 struct rcu_node *rnp = rcu_get_root(rsp);
1977 WRITE_ONCE(rsp->gp_activity, jiffies);
1980 /* Collect dyntick-idle snapshots. */
1981 if (is_sysidle_rcu_state(rsp)) {
1983 maxj = jiffies - ULONG_MAX / 4;
1985 force_qs_rnp(rsp, dyntick_save_progress_counter,
1987 rcu_sysidle_report_gp(rsp, isidle, maxj);
1989 /* Handle dyntick-idle and offline CPUs. */
1991 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1993 /* Clear flag to prevent immediate re-entry. */
1994 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1995 raw_spin_lock_irq_rcu_node(rnp);
1996 WRITE_ONCE(rsp->gp_flags,
1997 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
1998 raw_spin_unlock_irq_rcu_node(rnp);
2003 * Clean up after the old grace period.
2005 static void rcu_gp_cleanup(struct rcu_state *rsp)
2007 unsigned long gp_duration;
2008 bool needgp = false;
2010 struct rcu_data *rdp;
2011 struct rcu_node *rnp = rcu_get_root(rsp);
2013 WRITE_ONCE(rsp->gp_activity, jiffies);
2014 raw_spin_lock_irq_rcu_node(rnp);
2015 gp_duration = jiffies - rsp->gp_start;
2016 if (gp_duration > rsp->gp_max)
2017 rsp->gp_max = gp_duration;
2020 * We know the grace period is complete, but to everyone else
2021 * it appears to still be ongoing. But it is also the case
2022 * that to everyone else it looks like there is nothing that
2023 * they can do to advance the grace period. It is therefore
2024 * safe for us to drop the lock in order to mark the grace
2025 * period as completed in all of the rcu_node structures.
2027 raw_spin_unlock_irq_rcu_node(rnp);
2030 * Propagate new ->completed value to rcu_node structures so
2031 * that other CPUs don't have to wait until the start of the next
2032 * grace period to process their callbacks. This also avoids
2033 * some nasty RCU grace-period initialization races by forcing
2034 * the end of the current grace period to be completely recorded in
2035 * all of the rcu_node structures before the beginning of the next
2036 * grace period is recorded in any of the rcu_node structures.
2038 rcu_for_each_node_breadth_first(rsp, rnp) {
2039 raw_spin_lock_irq_rcu_node(rnp);
2040 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2041 WARN_ON_ONCE(rnp->qsmask);
2042 WRITE_ONCE(rnp->completed, rsp->gpnum);
2043 rdp = this_cpu_ptr(rsp->rda);
2044 if (rnp == rdp->mynode)
2045 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2046 /* smp_mb() provided by prior unlock-lock pair. */
2047 nocb += rcu_future_gp_cleanup(rsp, rnp);
2048 raw_spin_unlock_irq_rcu_node(rnp);
2049 cond_resched_rcu_qs();
2050 WRITE_ONCE(rsp->gp_activity, jiffies);
2051 rcu_gp_slow(rsp, gp_cleanup_delay);
2053 rnp = rcu_get_root(rsp);
2054 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2055 rcu_nocb_gp_set(rnp, nocb);
2057 /* Declare grace period done. */
2058 WRITE_ONCE(rsp->completed, rsp->gpnum);
2059 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2060 rsp->gp_state = RCU_GP_IDLE;
2061 rdp = this_cpu_ptr(rsp->rda);
2062 /* Advance CBs to reduce false positives below. */
2063 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2064 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2065 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2066 trace_rcu_grace_period(rsp->name,
2067 READ_ONCE(rsp->gpnum),
2070 raw_spin_unlock_irq_rcu_node(rnp);
2074 * Body of kthread that handles grace periods.
2076 static int __noreturn rcu_gp_kthread(void *arg)
2082 struct rcu_state *rsp = arg;
2083 struct rcu_node *rnp = rcu_get_root(rsp);
2085 rcu_bind_gp_kthread();
2088 /* Handle grace-period start. */
2090 trace_rcu_grace_period(rsp->name,
2091 READ_ONCE(rsp->gpnum),
2093 rsp->gp_state = RCU_GP_WAIT_GPS;
2094 wait_event_interruptible(rsp->gp_wq,
2095 READ_ONCE(rsp->gp_flags) &
2097 rsp->gp_state = RCU_GP_DONE_GPS;
2098 /* Locking provides needed memory barrier. */
2099 if (rcu_gp_init(rsp))
2101 cond_resched_rcu_qs();
2102 WRITE_ONCE(rsp->gp_activity, jiffies);
2103 WARN_ON(signal_pending(current));
2104 trace_rcu_grace_period(rsp->name,
2105 READ_ONCE(rsp->gpnum),
2109 /* Handle quiescent-state forcing. */
2110 first_gp_fqs = true;
2111 j = jiffies_till_first_fqs;
2114 jiffies_till_first_fqs = HZ;
2119 rsp->jiffies_force_qs = jiffies + j;
2120 trace_rcu_grace_period(rsp->name,
2121 READ_ONCE(rsp->gpnum),
2123 rsp->gp_state = RCU_GP_WAIT_FQS;
2124 ret = wait_event_interruptible_timeout(rsp->gp_wq,
2125 rcu_gp_fqs_check_wake(rsp, &gf), j);
2126 rsp->gp_state = RCU_GP_DOING_FQS;
2127 /* Locking provides needed memory barriers. */
2128 /* If grace period done, leave loop. */
2129 if (!READ_ONCE(rnp->qsmask) &&
2130 !rcu_preempt_blocked_readers_cgp(rnp))
2132 /* If time for quiescent-state forcing, do it. */
2133 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2134 (gf & RCU_GP_FLAG_FQS)) {
2135 trace_rcu_grace_period(rsp->name,
2136 READ_ONCE(rsp->gpnum),
2138 rcu_gp_fqs(rsp, first_gp_fqs);
2139 first_gp_fqs = false;
2140 trace_rcu_grace_period(rsp->name,
2141 READ_ONCE(rsp->gpnum),
2143 cond_resched_rcu_qs();
2144 WRITE_ONCE(rsp->gp_activity, jiffies);
2146 /* Deal with stray signal. */
2147 cond_resched_rcu_qs();
2148 WRITE_ONCE(rsp->gp_activity, jiffies);
2149 WARN_ON(signal_pending(current));
2150 trace_rcu_grace_period(rsp->name,
2151 READ_ONCE(rsp->gpnum),
2154 j = jiffies_till_next_fqs;
2157 jiffies_till_next_fqs = HZ;
2160 jiffies_till_next_fqs = 1;
2164 /* Handle grace-period end. */
2165 rsp->gp_state = RCU_GP_CLEANUP;
2166 rcu_gp_cleanup(rsp);
2167 rsp->gp_state = RCU_GP_CLEANED;
2172 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2173 * in preparation for detecting the next grace period. The caller must hold
2174 * the root node's ->lock and hard irqs must be disabled.
2176 * Note that it is legal for a dying CPU (which is marked as offline) to
2177 * invoke this function. This can happen when the dying CPU reports its
2180 * Returns true if the grace-period kthread must be awakened.
2183 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2184 struct rcu_data *rdp)
2186 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2188 * Either we have not yet spawned the grace-period
2189 * task, this CPU does not need another grace period,
2190 * or a grace period is already in progress.
2191 * Either way, don't start a new grace period.
2195 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2196 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2200 * We can't do wakeups while holding the rnp->lock, as that
2201 * could cause possible deadlocks with the rq->lock. Defer
2202 * the wakeup to our caller.
2208 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2209 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2210 * is invoked indirectly from rcu_advance_cbs(), which would result in
2211 * endless recursion -- or would do so if it wasn't for the self-deadlock
2212 * that is encountered beforehand.
2214 * Returns true if the grace-period kthread needs to be awakened.
2216 static bool rcu_start_gp(struct rcu_state *rsp)
2218 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2219 struct rcu_node *rnp = rcu_get_root(rsp);
2223 * If there is no grace period in progress right now, any
2224 * callbacks we have up to this point will be satisfied by the
2225 * next grace period. Also, advancing the callbacks reduces the
2226 * probability of false positives from cpu_needs_another_gp()
2227 * resulting in pointless grace periods. So, advance callbacks
2228 * then start the grace period!
2230 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2231 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2236 * Report a full set of quiescent states to the specified rcu_state data
2237 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2238 * kthread if another grace period is required. Whether we wake
2239 * the grace-period kthread or it awakens itself for the next round
2240 * of quiescent-state forcing, that kthread will clean up after the
2241 * just-completed grace period. Note that the caller must hold rnp->lock,
2242 * which is released before return.
2244 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2245 __releases(rcu_get_root(rsp)->lock)
2247 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2248 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2249 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2250 rcu_gp_kthread_wake(rsp);
2254 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2255 * Allows quiescent states for a group of CPUs to be reported at one go
2256 * to the specified rcu_node structure, though all the CPUs in the group
2257 * must be represented by the same rcu_node structure (which need not be a
2258 * leaf rcu_node structure, though it often will be). The gps parameter
2259 * is the grace-period snapshot, which means that the quiescent states
2260 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2261 * must be held upon entry, and it is released before return.
2264 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2265 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2266 __releases(rnp->lock)
2268 unsigned long oldmask = 0;
2269 struct rcu_node *rnp_c;
2271 /* Walk up the rcu_node hierarchy. */
2273 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2276 * Our bit has already been cleared, or the
2277 * relevant grace period is already over, so done.
2279 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2282 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2283 rnp->qsmask &= ~mask;
2284 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2285 mask, rnp->qsmask, rnp->level,
2286 rnp->grplo, rnp->grphi,
2288 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2290 /* Other bits still set at this level, so done. */
2291 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2294 mask = rnp->grpmask;
2295 if (rnp->parent == NULL) {
2297 /* No more levels. Exit loop holding root lock. */
2301 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2304 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2305 oldmask = rnp_c->qsmask;
2309 * Get here if we are the last CPU to pass through a quiescent
2310 * state for this grace period. Invoke rcu_report_qs_rsp()
2311 * to clean up and start the next grace period if one is needed.
2313 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2317 * Record a quiescent state for all tasks that were previously queued
2318 * on the specified rcu_node structure and that were blocking the current
2319 * RCU grace period. The caller must hold the specified rnp->lock with
2320 * irqs disabled, and this lock is released upon return, but irqs remain
2323 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2324 struct rcu_node *rnp, unsigned long flags)
2325 __releases(rnp->lock)
2329 struct rcu_node *rnp_p;
2331 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2332 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2333 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2334 return; /* Still need more quiescent states! */
2337 rnp_p = rnp->parent;
2338 if (rnp_p == NULL) {
2340 * Only one rcu_node structure in the tree, so don't
2341 * try to report up to its nonexistent parent!
2343 rcu_report_qs_rsp(rsp, flags);
2347 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2349 mask = rnp->grpmask;
2350 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2351 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2352 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2356 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2357 * structure. This must be called from the specified CPU.
2360 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2362 unsigned long flags;
2365 struct rcu_node *rnp;
2368 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2369 if ((rdp->cpu_no_qs.b.norm &&
2370 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2371 rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2375 * The grace period in which this quiescent state was
2376 * recorded has ended, so don't report it upwards.
2377 * We will instead need a new quiescent state that lies
2378 * within the current grace period.
2380 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2381 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2382 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2385 mask = rdp->grpmask;
2386 if ((rnp->qsmask & mask) == 0) {
2387 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2389 rdp->core_needs_qs = false;
2392 * This GP can't end until cpu checks in, so all of our
2393 * callbacks can be processed during the next GP.
2395 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2397 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2398 /* ^^^ Released rnp->lock */
2400 rcu_gp_kthread_wake(rsp);
2405 * Check to see if there is a new grace period of which this CPU
2406 * is not yet aware, and if so, set up local rcu_data state for it.
2407 * Otherwise, see if this CPU has just passed through its first
2408 * quiescent state for this grace period, and record that fact if so.
2411 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2413 /* Check for grace-period ends and beginnings. */
2414 note_gp_changes(rsp, rdp);
2417 * Does this CPU still need to do its part for current grace period?
2418 * If no, return and let the other CPUs do their part as well.
2420 if (!rdp->core_needs_qs)
2424 * Was there a quiescent state since the beginning of the grace
2425 * period? If no, then exit and wait for the next call.
2427 if (rdp->cpu_no_qs.b.norm &&
2428 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2432 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2435 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2439 * Send the specified CPU's RCU callbacks to the orphanage. The
2440 * specified CPU must be offline, and the caller must hold the
2444 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2445 struct rcu_node *rnp, struct rcu_data *rdp)
2447 /* No-CBs CPUs do not have orphanable callbacks. */
2448 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2452 * Orphan the callbacks. First adjust the counts. This is safe
2453 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2454 * cannot be running now. Thus no memory barrier is required.
2456 if (rdp->nxtlist != NULL) {
2457 rsp->qlen_lazy += rdp->qlen_lazy;
2458 rsp->qlen += rdp->qlen;
2459 rdp->n_cbs_orphaned += rdp->qlen;
2461 WRITE_ONCE(rdp->qlen, 0);
2465 * Next, move those callbacks still needing a grace period to
2466 * the orphanage, where some other CPU will pick them up.
2467 * Some of the callbacks might have gone partway through a grace
2468 * period, but that is too bad. They get to start over because we
2469 * cannot assume that grace periods are synchronized across CPUs.
2470 * We don't bother updating the ->nxttail[] array yet, instead
2471 * we just reset the whole thing later on.
2473 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2474 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2475 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2476 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2480 * Then move the ready-to-invoke callbacks to the orphanage,
2481 * where some other CPU will pick them up. These will not be
2482 * required to pass though another grace period: They are done.
2484 if (rdp->nxtlist != NULL) {
2485 *rsp->orphan_donetail = rdp->nxtlist;
2486 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2490 * Finally, initialize the rcu_data structure's list to empty and
2491 * disallow further callbacks on this CPU.
2493 init_callback_list(rdp);
2494 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2498 * Adopt the RCU callbacks from the specified rcu_state structure's
2499 * orphanage. The caller must hold the ->orphan_lock.
2501 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2504 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2506 /* No-CBs CPUs are handled specially. */
2507 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2508 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2511 /* Do the accounting first. */
2512 rdp->qlen_lazy += rsp->qlen_lazy;
2513 rdp->qlen += rsp->qlen;
2514 rdp->n_cbs_adopted += rsp->qlen;
2515 if (rsp->qlen_lazy != rsp->qlen)
2516 rcu_idle_count_callbacks_posted();
2521 * We do not need a memory barrier here because the only way we
2522 * can get here if there is an rcu_barrier() in flight is if
2523 * we are the task doing the rcu_barrier().
2526 /* First adopt the ready-to-invoke callbacks. */
2527 if (rsp->orphan_donelist != NULL) {
2528 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2529 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2530 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2531 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2532 rdp->nxttail[i] = rsp->orphan_donetail;
2533 rsp->orphan_donelist = NULL;
2534 rsp->orphan_donetail = &rsp->orphan_donelist;
2537 /* And then adopt the callbacks that still need a grace period. */
2538 if (rsp->orphan_nxtlist != NULL) {
2539 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2540 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2541 rsp->orphan_nxtlist = NULL;
2542 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2547 * Trace the fact that this CPU is going offline.
2549 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2551 RCU_TRACE(unsigned long mask);
2552 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2553 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2555 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2558 RCU_TRACE(mask = rdp->grpmask);
2559 trace_rcu_grace_period(rsp->name,
2560 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2565 * All CPUs for the specified rcu_node structure have gone offline,
2566 * and all tasks that were preempted within an RCU read-side critical
2567 * section while running on one of those CPUs have since exited their RCU
2568 * read-side critical section. Some other CPU is reporting this fact with
2569 * the specified rcu_node structure's ->lock held and interrupts disabled.
2570 * This function therefore goes up the tree of rcu_node structures,
2571 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2572 * the leaf rcu_node structure's ->qsmaskinit field has already been
2575 * This function does check that the specified rcu_node structure has
2576 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2577 * prematurely. That said, invoking it after the fact will cost you
2578 * a needless lock acquisition. So once it has done its work, don't
2581 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2584 struct rcu_node *rnp = rnp_leaf;
2586 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2587 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2590 mask = rnp->grpmask;
2594 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2595 rnp->qsmaskinit &= ~mask;
2596 rnp->qsmask &= ~mask;
2597 if (rnp->qsmaskinit) {
2598 raw_spin_unlock_rcu_node(rnp);
2599 /* irqs remain disabled. */
2602 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2607 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2608 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2611 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
2613 unsigned long flags;
2615 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2616 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2618 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2621 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2622 mask = rdp->grpmask;
2623 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
2624 rnp->qsmaskinitnext &= ~mask;
2625 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2629 * The CPU has been completely removed, and some other CPU is reporting
2630 * this fact from process context. Do the remainder of the cleanup,
2631 * including orphaning the outgoing CPU's RCU callbacks, and also
2632 * adopting them. There can only be one CPU hotplug operation at a time,
2633 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2635 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2637 unsigned long flags;
2638 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2639 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2641 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2644 /* Adjust any no-longer-needed kthreads. */
2645 rcu_boost_kthread_setaffinity(rnp, -1);
2647 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2648 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2649 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2650 rcu_adopt_orphan_cbs(rsp, flags);
2651 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2653 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2654 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2655 cpu, rdp->qlen, rdp->nxtlist);
2659 * Invoke any RCU callbacks that have made it to the end of their grace
2660 * period. Thottle as specified by rdp->blimit.
2662 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2664 unsigned long flags;
2665 struct rcu_head *next, *list, **tail;
2666 long bl, count, count_lazy;
2669 /* If no callbacks are ready, just return. */
2670 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2671 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2672 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2673 need_resched(), is_idle_task(current),
2674 rcu_is_callbacks_kthread());
2679 * Extract the list of ready callbacks, disabling to prevent
2680 * races with call_rcu() from interrupt handlers.
2682 local_irq_save(flags);
2683 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2685 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2686 list = rdp->nxtlist;
2687 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2688 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2689 tail = rdp->nxttail[RCU_DONE_TAIL];
2690 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2691 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2692 rdp->nxttail[i] = &rdp->nxtlist;
2693 local_irq_restore(flags);
2695 /* Invoke callbacks. */
2696 count = count_lazy = 0;
2700 debug_rcu_head_unqueue(list);
2701 if (__rcu_reclaim(rsp->name, list))
2704 /* Stop only if limit reached and CPU has something to do. */
2705 if (++count >= bl &&
2707 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2711 local_irq_save(flags);
2712 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2713 is_idle_task(current),
2714 rcu_is_callbacks_kthread());
2716 /* Update count, and requeue any remaining callbacks. */
2718 *tail = rdp->nxtlist;
2719 rdp->nxtlist = list;
2720 for (i = 0; i < RCU_NEXT_SIZE; i++)
2721 if (&rdp->nxtlist == rdp->nxttail[i])
2722 rdp->nxttail[i] = tail;
2726 smp_mb(); /* List handling before counting for rcu_barrier(). */
2727 rdp->qlen_lazy -= count_lazy;
2728 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2729 rdp->n_cbs_invoked += count;
2731 /* Reinstate batch limit if we have worked down the excess. */
2732 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2733 rdp->blimit = blimit;
2735 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2736 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2737 rdp->qlen_last_fqs_check = 0;
2738 rdp->n_force_qs_snap = rsp->n_force_qs;
2739 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2740 rdp->qlen_last_fqs_check = rdp->qlen;
2741 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2743 local_irq_restore(flags);
2745 /* Re-invoke RCU core processing if there are callbacks remaining. */
2746 if (cpu_has_callbacks_ready_to_invoke(rdp))
2751 * Check to see if this CPU is in a non-context-switch quiescent state
2752 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2753 * Also schedule RCU core processing.
2755 * This function must be called from hardirq context. It is normally
2756 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2757 * false, there is no point in invoking rcu_check_callbacks().
2759 void rcu_check_callbacks(int user)
2761 trace_rcu_utilization(TPS("Start scheduler-tick"));
2762 increment_cpu_stall_ticks();
2763 if (user || rcu_is_cpu_rrupt_from_idle()) {
2766 * Get here if this CPU took its interrupt from user
2767 * mode or from the idle loop, and if this is not a
2768 * nested interrupt. In this case, the CPU is in
2769 * a quiescent state, so note it.
2771 * No memory barrier is required here because both
2772 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2773 * variables that other CPUs neither access nor modify,
2774 * at least not while the corresponding CPU is online.
2780 } else if (!in_softirq()) {
2783 * Get here if this CPU did not take its interrupt from
2784 * softirq, in other words, if it is not interrupting
2785 * a rcu_bh read-side critical section. This is an _bh
2786 * critical section, so note it.
2791 rcu_preempt_check_callbacks();
2795 rcu_note_voluntary_context_switch(current);
2796 trace_rcu_utilization(TPS("End scheduler-tick"));
2800 * Scan the leaf rcu_node structures, processing dyntick state for any that
2801 * have not yet encountered a quiescent state, using the function specified.
2802 * Also initiate boosting for any threads blocked on the root rcu_node.
2804 * The caller must have suppressed start of new grace periods.
2806 static void force_qs_rnp(struct rcu_state *rsp,
2807 int (*f)(struct rcu_data *rsp, bool *isidle,
2808 unsigned long *maxj),
2809 bool *isidle, unsigned long *maxj)
2813 unsigned long flags;
2815 struct rcu_node *rnp;
2817 rcu_for_each_leaf_node(rsp, rnp) {
2818 cond_resched_rcu_qs();
2820 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2821 if (rnp->qsmask == 0) {
2822 if (rcu_state_p == &rcu_sched_state ||
2823 rsp != rcu_state_p ||
2824 rcu_preempt_blocked_readers_cgp(rnp)) {
2826 * No point in scanning bits because they
2827 * are all zero. But we might need to
2828 * priority-boost blocked readers.
2830 rcu_initiate_boost(rnp, flags);
2831 /* rcu_initiate_boost() releases rnp->lock */
2835 (rnp->parent->qsmask & rnp->grpmask)) {
2837 * Race between grace-period
2838 * initialization and task exiting RCU
2839 * read-side critical section: Report.
2841 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2842 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2848 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2849 if ((rnp->qsmask & bit) != 0) {
2850 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2855 /* Idle/offline CPUs, report (releases rnp->lock. */
2856 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2858 /* Nothing to do here, so just drop the lock. */
2859 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2865 * Force quiescent states on reluctant CPUs, and also detect which
2866 * CPUs are in dyntick-idle mode.
2868 static void force_quiescent_state(struct rcu_state *rsp)
2870 unsigned long flags;
2872 struct rcu_node *rnp;
2873 struct rcu_node *rnp_old = NULL;
2875 /* Funnel through hierarchy to reduce memory contention. */
2876 rnp = __this_cpu_read(rsp->rda->mynode);
2877 for (; rnp != NULL; rnp = rnp->parent) {
2878 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2879 !raw_spin_trylock(&rnp->fqslock);
2880 if (rnp_old != NULL)
2881 raw_spin_unlock(&rnp_old->fqslock);
2883 rsp->n_force_qs_lh++;
2888 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2890 /* Reached the root of the rcu_node tree, acquire lock. */
2891 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2892 raw_spin_unlock(&rnp_old->fqslock);
2893 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2894 rsp->n_force_qs_lh++;
2895 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2896 return; /* Someone beat us to it. */
2898 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2899 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2900 rcu_gp_kthread_wake(rsp);
2904 * This does the RCU core processing work for the specified rcu_state
2905 * and rcu_data structures. This may be called only from the CPU to
2906 * whom the rdp belongs.
2909 __rcu_process_callbacks(struct rcu_state *rsp)
2911 unsigned long flags;
2913 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2915 WARN_ON_ONCE(rdp->beenonline == 0);
2917 /* Update RCU state based on any recent quiescent states. */
2918 rcu_check_quiescent_state(rsp, rdp);
2920 /* Does this CPU require a not-yet-started grace period? */
2921 local_irq_save(flags);
2922 if (cpu_needs_another_gp(rsp, rdp)) {
2923 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2924 needwake = rcu_start_gp(rsp);
2925 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2927 rcu_gp_kthread_wake(rsp);
2929 local_irq_restore(flags);
2932 /* If there are callbacks ready, invoke them. */
2933 if (cpu_has_callbacks_ready_to_invoke(rdp))
2934 invoke_rcu_callbacks(rsp, rdp);
2936 /* Do any needed deferred wakeups of rcuo kthreads. */
2937 do_nocb_deferred_wakeup(rdp);
2941 * Do RCU core processing for the current CPU.
2943 static void rcu_process_callbacks(struct softirq_action *unused)
2945 struct rcu_state *rsp;
2947 if (cpu_is_offline(smp_processor_id()))
2949 trace_rcu_utilization(TPS("Start RCU core"));
2950 for_each_rcu_flavor(rsp)
2951 __rcu_process_callbacks(rsp);
2952 trace_rcu_utilization(TPS("End RCU core"));
2956 * Schedule RCU callback invocation. If the specified type of RCU
2957 * does not support RCU priority boosting, just do a direct call,
2958 * otherwise wake up the per-CPU kernel kthread. Note that because we
2959 * are running on the current CPU with softirqs disabled, the
2960 * rcu_cpu_kthread_task cannot disappear out from under us.
2962 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2964 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2966 if (likely(!rsp->boost)) {
2967 rcu_do_batch(rsp, rdp);
2970 invoke_rcu_callbacks_kthread();
2973 static void invoke_rcu_core(void)
2975 if (cpu_online(smp_processor_id()))
2976 raise_softirq(RCU_SOFTIRQ);
2980 * Handle any core-RCU processing required by a call_rcu() invocation.
2982 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2983 struct rcu_head *head, unsigned long flags)
2988 * If called from an extended quiescent state, invoke the RCU
2989 * core in order to force a re-evaluation of RCU's idleness.
2991 if (!rcu_is_watching())
2994 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2995 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2999 * Force the grace period if too many callbacks or too long waiting.
3000 * Enforce hysteresis, and don't invoke force_quiescent_state()
3001 * if some other CPU has recently done so. Also, don't bother
3002 * invoking force_quiescent_state() if the newly enqueued callback
3003 * is the only one waiting for a grace period to complete.
3005 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
3007 /* Are we ignoring a completed grace period? */
3008 note_gp_changes(rsp, rdp);
3010 /* Start a new grace period if one not already started. */
3011 if (!rcu_gp_in_progress(rsp)) {
3012 struct rcu_node *rnp_root = rcu_get_root(rsp);
3014 raw_spin_lock_rcu_node(rnp_root);
3015 needwake = rcu_start_gp(rsp);
3016 raw_spin_unlock_rcu_node(rnp_root);
3018 rcu_gp_kthread_wake(rsp);
3020 /* Give the grace period a kick. */
3021 rdp->blimit = LONG_MAX;
3022 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3023 *rdp->nxttail[RCU_DONE_TAIL] != head)
3024 force_quiescent_state(rsp);
3025 rdp->n_force_qs_snap = rsp->n_force_qs;
3026 rdp->qlen_last_fqs_check = rdp->qlen;
3032 * RCU callback function to leak a callback.
3034 static void rcu_leak_callback(struct rcu_head *rhp)
3039 * Helper function for call_rcu() and friends. The cpu argument will
3040 * normally be -1, indicating "currently running CPU". It may specify
3041 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3042 * is expected to specify a CPU.
3045 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3046 struct rcu_state *rsp, int cpu, bool lazy)
3048 unsigned long flags;
3049 struct rcu_data *rdp;
3051 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
3052 if (debug_rcu_head_queue(head)) {
3053 /* Probable double call_rcu(), so leak the callback. */
3054 WRITE_ONCE(head->func, rcu_leak_callback);
3055 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3062 * Opportunistically note grace-period endings and beginnings.
3063 * Note that we might see a beginning right after we see an
3064 * end, but never vice versa, since this CPU has to pass through
3065 * a quiescent state betweentimes.
3067 local_irq_save(flags);
3068 rdp = this_cpu_ptr(rsp->rda);
3070 /* Add the callback to our list. */
3071 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3075 rdp = per_cpu_ptr(rsp->rda, cpu);
3076 if (likely(rdp->mynode)) {
3077 /* Post-boot, so this should be for a no-CBs CPU. */
3078 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3079 WARN_ON_ONCE(offline);
3080 /* Offline CPU, _call_rcu() illegal, leak callback. */
3081 local_irq_restore(flags);
3085 * Very early boot, before rcu_init(). Initialize if needed
3086 * and then drop through to queue the callback.
3089 WARN_ON_ONCE(!rcu_is_watching());
3090 if (!likely(rdp->nxtlist))
3091 init_default_callback_list(rdp);
3093 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3097 rcu_idle_count_callbacks_posted();
3098 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3099 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3100 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3102 if (__is_kfree_rcu_offset((unsigned long)func))
3103 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3104 rdp->qlen_lazy, rdp->qlen);
3106 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3108 /* Go handle any RCU core processing required. */
3109 __call_rcu_core(rsp, rdp, head, flags);
3110 local_irq_restore(flags);
3114 * Queue an RCU-sched callback for invocation after a grace period.
3116 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3118 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3120 EXPORT_SYMBOL_GPL(call_rcu_sched);
3123 * Queue an RCU callback for invocation after a quicker grace period.
3125 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3127 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3129 EXPORT_SYMBOL_GPL(call_rcu_bh);
3132 * Queue an RCU callback for lazy invocation after a grace period.
3133 * This will likely be later named something like "call_rcu_lazy()",
3134 * but this change will require some way of tagging the lazy RCU
3135 * callbacks in the list of pending callbacks. Until then, this
3136 * function may only be called from __kfree_rcu().
3138 void kfree_call_rcu(struct rcu_head *head,
3139 rcu_callback_t func)
3141 __call_rcu(head, func, rcu_state_p, -1, 1);
3143 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3146 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3147 * any blocking grace-period wait automatically implies a grace period
3148 * if there is only one CPU online at any point time during execution
3149 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3150 * occasionally incorrectly indicate that there are multiple CPUs online
3151 * when there was in fact only one the whole time, as this just adds
3152 * some overhead: RCU still operates correctly.
3154 static inline int rcu_blocking_is_gp(void)
3158 might_sleep(); /* Check for RCU read-side critical section. */
3160 ret = num_online_cpus() <= 1;
3166 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3168 * Control will return to the caller some time after a full rcu-sched
3169 * grace period has elapsed, in other words after all currently executing
3170 * rcu-sched read-side critical sections have completed. These read-side
3171 * critical sections are delimited by rcu_read_lock_sched() and
3172 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3173 * local_irq_disable(), and so on may be used in place of
3174 * rcu_read_lock_sched().
3176 * This means that all preempt_disable code sequences, including NMI and
3177 * non-threaded hardware-interrupt handlers, in progress on entry will
3178 * have completed before this primitive returns. However, this does not
3179 * guarantee that softirq handlers will have completed, since in some
3180 * kernels, these handlers can run in process context, and can block.
3182 * Note that this guarantee implies further memory-ordering guarantees.
3183 * On systems with more than one CPU, when synchronize_sched() returns,
3184 * each CPU is guaranteed to have executed a full memory barrier since the
3185 * end of its last RCU-sched read-side critical section whose beginning
3186 * preceded the call to synchronize_sched(). In addition, each CPU having
3187 * an RCU read-side critical section that extends beyond the return from
3188 * synchronize_sched() is guaranteed to have executed a full memory barrier
3189 * after the beginning of synchronize_sched() and before the beginning of
3190 * that RCU read-side critical section. Note that these guarantees include
3191 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3192 * that are executing in the kernel.
3194 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3195 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3196 * to have executed a full memory barrier during the execution of
3197 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3198 * again only if the system has more than one CPU).
3200 * This primitive provides the guarantees made by the (now removed)
3201 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3202 * guarantees that rcu_read_lock() sections will have completed.
3203 * In "classic RCU", these two guarantees happen to be one and
3204 * the same, but can differ in realtime RCU implementations.
3206 void synchronize_sched(void)
3208 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3209 lock_is_held(&rcu_lock_map) ||
3210 lock_is_held(&rcu_sched_lock_map),
3211 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3212 if (rcu_blocking_is_gp())
3214 if (rcu_gp_is_expedited())
3215 synchronize_sched_expedited();
3217 wait_rcu_gp(call_rcu_sched);
3219 EXPORT_SYMBOL_GPL(synchronize_sched);
3222 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3224 * Control will return to the caller some time after a full rcu_bh grace
3225 * period has elapsed, in other words after all currently executing rcu_bh
3226 * read-side critical sections have completed. RCU read-side critical
3227 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3228 * and may be nested.
3230 * See the description of synchronize_sched() for more detailed information
3231 * on memory ordering guarantees.
3233 void synchronize_rcu_bh(void)
3235 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3236 lock_is_held(&rcu_lock_map) ||
3237 lock_is_held(&rcu_sched_lock_map),
3238 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3239 if (rcu_blocking_is_gp())
3241 if (rcu_gp_is_expedited())
3242 synchronize_rcu_bh_expedited();
3244 wait_rcu_gp(call_rcu_bh);
3246 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3249 * get_state_synchronize_rcu - Snapshot current RCU state
3251 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3252 * to determine whether or not a full grace period has elapsed in the
3255 unsigned long get_state_synchronize_rcu(void)
3258 * Any prior manipulation of RCU-protected data must happen
3259 * before the load from ->gpnum.
3264 * Make sure this load happens before the purportedly
3265 * time-consuming work between get_state_synchronize_rcu()
3266 * and cond_synchronize_rcu().
3268 return smp_load_acquire(&rcu_state_p->gpnum);
3270 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3273 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3275 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3277 * If a full RCU grace period has elapsed since the earlier call to
3278 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3279 * synchronize_rcu() to wait for a full grace period.
3281 * Yes, this function does not take counter wrap into account. But
3282 * counter wrap is harmless. If the counter wraps, we have waited for
3283 * more than 2 billion grace periods (and way more on a 64-bit system!),
3284 * so waiting for one additional grace period should be just fine.
3286 void cond_synchronize_rcu(unsigned long oldstate)
3288 unsigned long newstate;
3291 * Ensure that this load happens before any RCU-destructive
3292 * actions the caller might carry out after we return.
3294 newstate = smp_load_acquire(&rcu_state_p->completed);
3295 if (ULONG_CMP_GE(oldstate, newstate))
3298 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3301 * get_state_synchronize_sched - Snapshot current RCU-sched state
3303 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3304 * to determine whether or not a full grace period has elapsed in the
3307 unsigned long get_state_synchronize_sched(void)
3310 * Any prior manipulation of RCU-protected data must happen
3311 * before the load from ->gpnum.
3316 * Make sure this load happens before the purportedly
3317 * time-consuming work between get_state_synchronize_sched()
3318 * and cond_synchronize_sched().
3320 return smp_load_acquire(&rcu_sched_state.gpnum);
3322 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3325 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3327 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3329 * If a full RCU-sched grace period has elapsed since the earlier call to
3330 * get_state_synchronize_sched(), just return. Otherwise, invoke
3331 * synchronize_sched() to wait for a full grace period.
3333 * Yes, this function does not take counter wrap into account. But
3334 * counter wrap is harmless. If the counter wraps, we have waited for
3335 * more than 2 billion grace periods (and way more on a 64-bit system!),
3336 * so waiting for one additional grace period should be just fine.
3338 void cond_synchronize_sched(unsigned long oldstate)
3340 unsigned long newstate;
3343 * Ensure that this load happens before any RCU-destructive
3344 * actions the caller might carry out after we return.
3346 newstate = smp_load_acquire(&rcu_sched_state.completed);
3347 if (ULONG_CMP_GE(oldstate, newstate))
3348 synchronize_sched();
3350 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3352 /* Adjust sequence number for start of update-side operation. */
3353 static void rcu_seq_start(unsigned long *sp)
3355 WRITE_ONCE(*sp, *sp + 1);
3356 smp_mb(); /* Ensure update-side operation after counter increment. */
3357 WARN_ON_ONCE(!(*sp & 0x1));
3360 /* Adjust sequence number for end of update-side operation. */
3361 static void rcu_seq_end(unsigned long *sp)
3363 smp_mb(); /* Ensure update-side operation before counter increment. */
3364 WRITE_ONCE(*sp, *sp + 1);
3365 WARN_ON_ONCE(*sp & 0x1);
3368 /* Take a snapshot of the update side's sequence number. */
3369 static unsigned long rcu_seq_snap(unsigned long *sp)
3373 s = (READ_ONCE(*sp) + 3) & ~0x1;
3374 smp_mb(); /* Above access must not bleed into critical section. */
3379 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3380 * full update-side operation has occurred.
3382 static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3384 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3387 /* Wrapper functions for expedited grace periods. */
3388 static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
3390 rcu_seq_start(&rsp->expedited_sequence);
3392 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
3394 rcu_seq_end(&rsp->expedited_sequence);
3395 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3397 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
3399 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3400 return rcu_seq_snap(&rsp->expedited_sequence);
3402 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
3404 return rcu_seq_done(&rsp->expedited_sequence, s);
3408 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3409 * recent CPU-online activity. Note that these masks are not cleared
3410 * when CPUs go offline, so they reflect the union of all CPUs that have
3411 * ever been online. This means that this function normally takes its
3412 * no-work-to-do fastpath.
3414 static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
3417 unsigned long flags;
3419 unsigned long oldmask;
3420 int ncpus = READ_ONCE(rsp->ncpus);
3421 struct rcu_node *rnp;
3422 struct rcu_node *rnp_up;
3424 /* If no new CPUs onlined since last time, nothing to do. */
3425 if (likely(ncpus == rsp->ncpus_snap))
3427 rsp->ncpus_snap = ncpus;
3430 * Each pass through the following loop propagates newly onlined
3431 * CPUs for the current rcu_node structure up the rcu_node tree.
3433 rcu_for_each_leaf_node(rsp, rnp) {
3434 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3435 if (rnp->expmaskinit == rnp->expmaskinitnext) {
3436 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3437 continue; /* No new CPUs, nothing to do. */
3440 /* Update this node's mask, track old value for propagation. */
3441 oldmask = rnp->expmaskinit;
3442 rnp->expmaskinit = rnp->expmaskinitnext;
3443 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3445 /* If was already nonzero, nothing to propagate. */
3449 /* Propagate the new CPU up the tree. */
3450 mask = rnp->grpmask;
3451 rnp_up = rnp->parent;
3454 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
3455 if (rnp_up->expmaskinit)
3457 rnp_up->expmaskinit |= mask;
3458 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
3461 mask = rnp_up->grpmask;
3462 rnp_up = rnp_up->parent;
3468 * Reset the ->expmask values in the rcu_node tree in preparation for
3469 * a new expedited grace period.
3471 static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
3473 unsigned long flags;
3474 struct rcu_node *rnp;
3476 sync_exp_reset_tree_hotplug(rsp);
3477 rcu_for_each_node_breadth_first(rsp, rnp) {
3478 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3479 WARN_ON_ONCE(rnp->expmask);
3480 rnp->expmask = rnp->expmaskinit;
3481 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3486 * Return non-zero if there is no RCU expedited grace period in progress
3487 * for the specified rcu_node structure, in other words, if all CPUs and
3488 * tasks covered by the specified rcu_node structure have done their bit
3489 * for the current expedited grace period. Works only for preemptible
3490 * RCU -- other RCU implementation use other means.
3492 * Caller must hold the root rcu_node's exp_funnel_mutex.
3494 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
3496 return rnp->exp_tasks == NULL &&
3497 READ_ONCE(rnp->expmask) == 0;
3501 * Report the exit from RCU read-side critical section for the last task
3502 * that queued itself during or before the current expedited preemptible-RCU
3503 * grace period. This event is reported either to the rcu_node structure on
3504 * which the task was queued or to one of that rcu_node structure's ancestors,
3505 * recursively up the tree. (Calm down, calm down, we do the recursion
3508 * Caller must hold the root rcu_node's exp_funnel_mutex and the
3509 * specified rcu_node structure's ->lock.
3511 static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
3512 bool wake, unsigned long flags)
3513 __releases(rnp->lock)
3518 if (!sync_rcu_preempt_exp_done(rnp)) {
3520 rcu_initiate_boost(rnp, flags);
3522 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3525 if (rnp->parent == NULL) {
3526 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3528 smp_mb(); /* EGP done before wake_up(). */
3529 wake_up(&rsp->expedited_wq);
3533 mask = rnp->grpmask;
3534 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
3536 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
3537 WARN_ON_ONCE(!(rnp->expmask & mask));
3538 rnp->expmask &= ~mask;
3543 * Report expedited quiescent state for specified node. This is a
3544 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3546 * Caller must hold the root rcu_node's exp_funnel_mutex.
3548 static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
3549 struct rcu_node *rnp, bool wake)
3551 unsigned long flags;
3553 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3554 __rcu_report_exp_rnp(rsp, rnp, wake, flags);
3558 * Report expedited quiescent state for multiple CPUs, all covered by the
3559 * specified leaf rcu_node structure. Caller must hold the root
3560 * rcu_node's exp_funnel_mutex.
3562 static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
3563 unsigned long mask, bool wake)
3565 unsigned long flags;
3567 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3568 if (!(rnp->expmask & mask)) {
3569 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3572 rnp->expmask &= ~mask;
3573 __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
3577 * Report expedited quiescent state for specified rcu_data (CPU).
3578 * Caller must hold the root rcu_node's exp_funnel_mutex.
3580 static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
3583 rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
3586 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3587 static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp,
3588 struct rcu_data *rdp,
3589 atomic_long_t *stat, unsigned long s)
3591 if (rcu_exp_gp_seq_done(rsp, s)) {
3593 mutex_unlock(&rnp->exp_funnel_mutex);
3595 mutex_unlock(&rdp->exp_funnel_mutex);
3596 /* Ensure test happens before caller kfree(). */
3597 smp_mb__before_atomic(); /* ^^^ */
3598 atomic_long_inc(stat);
3605 * Funnel-lock acquisition for expedited grace periods. Returns a
3606 * pointer to the root rcu_node structure, or NULL if some other
3607 * task did the expedited grace period for us.
3609 static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
3611 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
3612 struct rcu_node *rnp0;
3613 struct rcu_node *rnp1 = NULL;
3616 * First try directly acquiring the root lock in order to reduce
3617 * latency in the common case where expedited grace periods are
3618 * rare. We check mutex_is_locked() to avoid pathological levels of
3619 * memory contention on ->exp_funnel_mutex in the heavy-load case.
3621 rnp0 = rcu_get_root(rsp);
3622 if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) {
3623 if (mutex_trylock(&rnp0->exp_funnel_mutex)) {
3624 if (sync_exp_work_done(rsp, rnp0, NULL,
3625 &rdp->expedited_workdone0, s))
3632 * Each pass through the following loop works its way
3633 * up the rcu_node tree, returning if others have done the
3634 * work or otherwise falls through holding the root rnp's
3635 * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure
3636 * can be inexact, as it is just promoting locality and is not
3637 * strictly needed for correctness.
3639 if (sync_exp_work_done(rsp, NULL, NULL, &rdp->expedited_workdone1, s))
3641 mutex_lock(&rdp->exp_funnel_mutex);
3643 for (; rnp0 != NULL; rnp0 = rnp0->parent) {
3644 if (sync_exp_work_done(rsp, rnp1, rdp,
3645 &rdp->expedited_workdone2, s))
3647 mutex_lock(&rnp0->exp_funnel_mutex);
3649 mutex_unlock(&rnp1->exp_funnel_mutex);
3651 mutex_unlock(&rdp->exp_funnel_mutex);
3654 if (sync_exp_work_done(rsp, rnp1, rdp,
3655 &rdp->expedited_workdone3, s))
3660 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3661 static void sync_sched_exp_handler(void *data)
3663 struct rcu_data *rdp;
3664 struct rcu_node *rnp;
3665 struct rcu_state *rsp = data;
3667 rdp = this_cpu_ptr(rsp->rda);
3669 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
3670 __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
3672 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
3673 resched_cpu(smp_processor_id());
3676 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
3677 static void sync_sched_exp_online_cleanup(int cpu)
3679 struct rcu_data *rdp;
3681 struct rcu_node *rnp;
3682 struct rcu_state *rsp = &rcu_sched_state;
3684 rdp = per_cpu_ptr(rsp->rda, cpu);
3686 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
3688 ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
3693 * Select the nodes that the upcoming expedited grace period needs
3696 static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
3697 smp_call_func_t func)
3700 unsigned long flags;
3702 unsigned long mask_ofl_test;
3703 unsigned long mask_ofl_ipi;
3705 struct rcu_node *rnp;
3707 sync_exp_reset_tree(rsp);
3708 rcu_for_each_leaf_node(rsp, rnp) {
3709 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3711 /* Each pass checks a CPU for identity, offline, and idle. */
3713 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
3714 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3715 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3717 if (raw_smp_processor_id() == cpu ||
3718 !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3719 mask_ofl_test |= rdp->grpmask;
3721 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
3724 * Need to wait for any blocked tasks as well. Note that
3725 * additional blocking tasks will also block the expedited
3726 * GP until such time as the ->expmask bits are cleared.
3728 if (rcu_preempt_has_tasks(rnp))
3729 rnp->exp_tasks = rnp->blkd_tasks.next;
3730 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3732 /* IPI the remaining CPUs for expedited quiescent state. */
3734 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3735 if (!(mask_ofl_ipi & mask))
3738 ret = smp_call_function_single(cpu, func, rsp, 0);
3740 mask_ofl_ipi &= ~mask;
3743 /* Failed, raced with offline. */
3744 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3745 if (cpu_online(cpu) &&
3746 (rnp->expmask & mask)) {
3747 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3748 schedule_timeout_uninterruptible(1);
3749 if (cpu_online(cpu) &&
3750 (rnp->expmask & mask))
3752 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3754 if (!(rnp->expmask & mask))
3755 mask_ofl_ipi &= ~mask;
3756 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3758 /* Report quiescent states for those that went offline. */
3759 mask_ofl_test |= mask_ofl_ipi;
3761 rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
3765 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
3768 unsigned long jiffies_stall;
3769 unsigned long jiffies_start;
3772 struct rcu_node *rnp;
3773 struct rcu_node *rnp_root = rcu_get_root(rsp);
3776 jiffies_stall = rcu_jiffies_till_stall_check();
3777 jiffies_start = jiffies;
3780 ret = wait_event_interruptible_timeout(
3782 sync_rcu_preempt_exp_done(rnp_root),
3784 if (ret > 0 || sync_rcu_preempt_exp_done(rnp_root))
3787 /* Hit a signal, disable CPU stall warnings. */
3788 wait_event(rsp->expedited_wq,
3789 sync_rcu_preempt_exp_done(rnp_root));
3792 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3795 rcu_for_each_leaf_node(rsp, rnp) {
3796 ndetected = rcu_print_task_exp_stall(rnp);
3798 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3799 struct rcu_data *rdp;
3801 if (!(rnp->expmask & mask))
3804 rdp = per_cpu_ptr(rsp->rda, cpu);
3805 pr_cont(" %d-%c%c%c", cpu,
3806 "O."[cpu_online(cpu)],
3807 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
3808 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
3812 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
3813 jiffies - jiffies_start, rsp->expedited_sequence,
3814 rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]);
3816 pr_err("blocking rcu_node structures:");
3817 rcu_for_each_node_breadth_first(rsp, rnp) {
3818 if (rnp == rnp_root)
3819 continue; /* printed unconditionally */
3820 if (sync_rcu_preempt_exp_done(rnp))
3822 pr_cont(" l=%u:%d-%d:%#lx/%c",
3823 rnp->level, rnp->grplo, rnp->grphi,
3825 ".T"[!!rnp->exp_tasks]);
3829 rcu_for_each_leaf_node(rsp, rnp) {
3831 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3832 if (!(rnp->expmask & mask))
3837 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
3842 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3844 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3845 * approach to force the grace period to end quickly. This consumes
3846 * significant time on all CPUs and is unfriendly to real-time workloads,
3847 * so is thus not recommended for any sort of common-case code. In fact,
3848 * if you are using synchronize_sched_expedited() in a loop, please
3849 * restructure your code to batch your updates, and then use a single
3850 * synchronize_sched() instead.
3852 * This implementation can be thought of as an application of sequence
3853 * locking to expedited grace periods, but using the sequence counter to
3854 * determine when someone else has already done the work instead of for
3857 void synchronize_sched_expedited(void)
3860 struct rcu_node *rnp;
3861 struct rcu_state *rsp = &rcu_sched_state;
3863 /* If only one CPU, this is automatically a grace period. */
3864 if (rcu_blocking_is_gp())
3867 /* If expedited grace periods are prohibited, fall back to normal. */
3868 if (rcu_gp_is_normal()) {
3869 wait_rcu_gp(call_rcu_sched);
3873 /* Take a snapshot of the sequence number. */
3874 s = rcu_exp_gp_seq_snap(rsp);
3876 rnp = exp_funnel_lock(rsp, s);
3878 return; /* Someone else did our work for us. */
3880 rcu_exp_gp_seq_start(rsp);
3881 sync_rcu_exp_select_cpus(rsp, sync_sched_exp_handler);
3882 synchronize_sched_expedited_wait(rsp);
3884 rcu_exp_gp_seq_end(rsp);
3885 mutex_unlock(&rnp->exp_funnel_mutex);
3887 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3890 * Check to see if there is any immediate RCU-related work to be done
3891 * by the current CPU, for the specified type of RCU, returning 1 if so.
3892 * The checks are in order of increasing expense: checks that can be
3893 * carried out against CPU-local state are performed first. However,
3894 * we must check for CPU stalls first, else we might not get a chance.
3896 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3898 struct rcu_node *rnp = rdp->mynode;
3900 rdp->n_rcu_pending++;
3902 /* Check for CPU stalls, if enabled. */
3903 check_cpu_stall(rsp, rdp);
3905 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3906 if (rcu_nohz_full_cpu(rsp))
3909 /* Is the RCU core waiting for a quiescent state from this CPU? */
3910 if (rcu_scheduler_fully_active &&
3911 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3912 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3913 rdp->n_rp_core_needs_qs++;
3914 } else if (rdp->core_needs_qs &&
3915 (!rdp->cpu_no_qs.b.norm ||
3916 rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3917 rdp->n_rp_report_qs++;
3921 /* Does this CPU have callbacks ready to invoke? */
3922 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3923 rdp->n_rp_cb_ready++;
3927 /* Has RCU gone idle with this CPU needing another grace period? */
3928 if (cpu_needs_another_gp(rsp, rdp)) {
3929 rdp->n_rp_cpu_needs_gp++;
3933 /* Has another RCU grace period completed? */
3934 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3935 rdp->n_rp_gp_completed++;
3939 /* Has a new RCU grace period started? */
3940 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3941 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3942 rdp->n_rp_gp_started++;
3946 /* Does this CPU need a deferred NOCB wakeup? */
3947 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3948 rdp->n_rp_nocb_defer_wakeup++;
3953 rdp->n_rp_need_nothing++;
3958 * Check to see if there is any immediate RCU-related work to be done
3959 * by the current CPU, returning 1 if so. This function is part of the
3960 * RCU implementation; it is -not- an exported member of the RCU API.
3962 static int rcu_pending(void)
3964 struct rcu_state *rsp;
3966 for_each_rcu_flavor(rsp)
3967 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3973 * Return true if the specified CPU has any callback. If all_lazy is
3974 * non-NULL, store an indication of whether all callbacks are lazy.
3975 * (If there are no callbacks, all of them are deemed to be lazy.)
3977 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3981 struct rcu_data *rdp;
3982 struct rcu_state *rsp;
3984 for_each_rcu_flavor(rsp) {
3985 rdp = this_cpu_ptr(rsp->rda);
3989 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
4000 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
4001 * the compiler is expected to optimize this away.
4003 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
4004 int cpu, unsigned long done)
4006 trace_rcu_barrier(rsp->name, s, cpu,
4007 atomic_read(&rsp->barrier_cpu_count), done);
4011 * RCU callback function for _rcu_barrier(). If we are last, wake
4012 * up the task executing _rcu_barrier().
4014 static void rcu_barrier_callback(struct rcu_head *rhp)
4016 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
4017 struct rcu_state *rsp = rdp->rsp;
4019 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
4020 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
4021 complete(&rsp->barrier_completion);
4023 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
4028 * Called with preemption disabled, and from cross-cpu IRQ context.
4030 static void rcu_barrier_func(void *type)
4032 struct rcu_state *rsp = type;
4033 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
4035 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
4036 atomic_inc(&rsp->barrier_cpu_count);
4037 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
4041 * Orchestrate the specified type of RCU barrier, waiting for all
4042 * RCU callbacks of the specified type to complete.
4044 static void _rcu_barrier(struct rcu_state *rsp)
4047 struct rcu_data *rdp;
4048 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
4050 _rcu_barrier_trace(rsp, "Begin", -1, s);
4052 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4053 mutex_lock(&rsp->barrier_mutex);
4055 /* Did someone else do our work for us? */
4056 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
4057 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
4058 smp_mb(); /* caller's subsequent code after above check. */
4059 mutex_unlock(&rsp->barrier_mutex);
4063 /* Mark the start of the barrier operation. */
4064 rcu_seq_start(&rsp->barrier_sequence);
4065 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
4068 * Initialize the count to one rather than to zero in order to
4069 * avoid a too-soon return to zero in case of a short grace period
4070 * (or preemption of this task). Exclude CPU-hotplug operations
4071 * to ensure that no offline CPU has callbacks queued.
4073 init_completion(&rsp->barrier_completion);
4074 atomic_set(&rsp->barrier_cpu_count, 1);
4078 * Force each CPU with callbacks to register a new callback.
4079 * When that callback is invoked, we will know that all of the
4080 * corresponding CPU's preceding callbacks have been invoked.
4082 for_each_possible_cpu(cpu) {
4083 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
4085 rdp = per_cpu_ptr(rsp->rda, cpu);
4086 if (rcu_is_nocb_cpu(cpu)) {
4087 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
4088 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
4089 rsp->barrier_sequence);
4091 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
4092 rsp->barrier_sequence);
4093 smp_mb__before_atomic();
4094 atomic_inc(&rsp->barrier_cpu_count);
4095 __call_rcu(&rdp->barrier_head,
4096 rcu_barrier_callback, rsp, cpu, 0);
4098 } else if (READ_ONCE(rdp->qlen)) {
4099 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
4100 rsp->barrier_sequence);
4101 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
4103 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
4104 rsp->barrier_sequence);
4110 * Now that we have an rcu_barrier_callback() callback on each
4111 * CPU, and thus each counted, remove the initial count.
4113 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
4114 complete(&rsp->barrier_completion);
4116 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4117 wait_for_completion(&rsp->barrier_completion);
4119 /* Mark the end of the barrier operation. */
4120 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
4121 rcu_seq_end(&rsp->barrier_sequence);
4123 /* Other rcu_barrier() invocations can now safely proceed. */
4124 mutex_unlock(&rsp->barrier_mutex);
4128 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4130 void rcu_barrier_bh(void)
4132 _rcu_barrier(&rcu_bh_state);
4134 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
4137 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4139 void rcu_barrier_sched(void)
4141 _rcu_barrier(&rcu_sched_state);
4143 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
4146 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4147 * first CPU in a given leaf rcu_node structure coming online. The caller
4148 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4151 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4154 struct rcu_node *rnp = rnp_leaf;
4157 mask = rnp->grpmask;
4161 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4162 rnp->qsmaskinit |= mask;
4163 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4168 * Do boot-time initialization of a CPU's per-CPU RCU data.
4171 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
4173 unsigned long flags;
4174 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4175 struct rcu_node *rnp = rcu_get_root(rsp);
4177 /* Set up local state, ensuring consistent view of global state. */
4178 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4179 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
4180 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
4181 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
4182 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
4185 mutex_init(&rdp->exp_funnel_mutex);
4186 rcu_boot_init_nocb_percpu_data(rdp);
4187 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4191 * Initialize a CPU's per-CPU RCU data. Note that only one online or
4192 * offline event can be happening at a given time. Note also that we
4193 * can accept some slop in the rsp->completed access due to the fact
4194 * that this CPU cannot possibly have any RCU callbacks in flight yet.
4197 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
4199 unsigned long flags;
4201 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4202 struct rcu_node *rnp = rcu_get_root(rsp);
4204 /* Set up local state, ensuring consistent view of global state. */
4205 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4206 rdp->qlen_last_fqs_check = 0;
4207 rdp->n_force_qs_snap = rsp->n_force_qs;
4208 rdp->blimit = blimit;
4210 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
4211 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
4212 rcu_sysidle_init_percpu_data(rdp->dynticks);
4213 atomic_set(&rdp->dynticks->dynticks,
4214 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
4215 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4218 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4219 * propagation up the rcu_node tree will happen at the beginning
4220 * of the next grace period.
4223 mask = rdp->grpmask;
4224 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4225 rnp->qsmaskinitnext |= mask;
4226 rnp->expmaskinitnext |= mask;
4227 if (!rdp->beenonline)
4228 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
4229 rdp->beenonline = true; /* We have now been online. */
4230 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
4231 rdp->completed = rnp->completed;
4232 rdp->cpu_no_qs.b.norm = true;
4233 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
4234 rdp->core_needs_qs = false;
4235 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
4236 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4239 static void rcu_prepare_cpu(int cpu)
4241 struct rcu_state *rsp;
4243 for_each_rcu_flavor(rsp)
4244 rcu_init_percpu_data(cpu, rsp);
4248 * Handle CPU online/offline notification events.
4250 int rcu_cpu_notify(struct notifier_block *self,
4251 unsigned long action, void *hcpu)
4253 long cpu = (long)hcpu;
4254 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
4255 struct rcu_node *rnp = rdp->mynode;
4256 struct rcu_state *rsp;
4259 case CPU_UP_PREPARE:
4260 case CPU_UP_PREPARE_FROZEN:
4261 rcu_prepare_cpu(cpu);
4262 rcu_prepare_kthreads(cpu);
4263 rcu_spawn_all_nocb_kthreads(cpu);
4266 case CPU_DOWN_FAILED:
4267 sync_sched_exp_online_cleanup(cpu);
4268 rcu_boost_kthread_setaffinity(rnp, -1);
4270 case CPU_DOWN_PREPARE:
4271 rcu_boost_kthread_setaffinity(rnp, cpu);
4274 case CPU_DYING_FROZEN:
4275 for_each_rcu_flavor(rsp)
4276 rcu_cleanup_dying_cpu(rsp);
4278 case CPU_DYING_IDLE:
4279 /* QS for any half-done expedited RCU-sched GP. */
4281 rcu_report_exp_rdp(&rcu_sched_state,
4282 this_cpu_ptr(rcu_sched_state.rda), true);
4285 for_each_rcu_flavor(rsp) {
4286 rcu_cleanup_dying_idle_cpu(cpu, rsp);
4290 case CPU_DEAD_FROZEN:
4291 case CPU_UP_CANCELED:
4292 case CPU_UP_CANCELED_FROZEN:
4293 for_each_rcu_flavor(rsp) {
4294 rcu_cleanup_dead_cpu(cpu, rsp);
4295 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
4304 static int rcu_pm_notify(struct notifier_block *self,
4305 unsigned long action, void *hcpu)
4308 case PM_HIBERNATION_PREPARE:
4309 case PM_SUSPEND_PREPARE:
4310 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4313 case PM_POST_HIBERNATION:
4314 case PM_POST_SUSPEND:
4315 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4316 rcu_unexpedite_gp();
4325 * Spawn the kthreads that handle each RCU flavor's grace periods.
4327 static int __init rcu_spawn_gp_kthread(void)
4329 unsigned long flags;
4330 int kthread_prio_in = kthread_prio;
4331 struct rcu_node *rnp;
4332 struct rcu_state *rsp;
4333 struct sched_param sp;
4334 struct task_struct *t;
4336 /* Force priority into range. */
4337 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4339 else if (kthread_prio < 0)
4341 else if (kthread_prio > 99)
4343 if (kthread_prio != kthread_prio_in)
4344 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4345 kthread_prio, kthread_prio_in);
4347 rcu_scheduler_fully_active = 1;
4348 for_each_rcu_flavor(rsp) {
4349 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4351 rnp = rcu_get_root(rsp);
4352 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4353 rsp->gp_kthread = t;
4355 sp.sched_priority = kthread_prio;
4356 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4358 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4361 rcu_spawn_nocb_kthreads();
4362 rcu_spawn_boost_kthreads();
4365 early_initcall(rcu_spawn_gp_kthread);
4368 * This function is invoked towards the end of the scheduler's initialization
4369 * process. Before this is called, the idle task might contain
4370 * RCU read-side critical sections (during which time, this idle
4371 * task is booting the system). After this function is called, the
4372 * idle tasks are prohibited from containing RCU read-side critical
4373 * sections. This function also enables RCU lockdep checking.
4375 void rcu_scheduler_starting(void)
4377 WARN_ON(num_online_cpus() != 1);
4378 WARN_ON(nr_context_switches() > 0);
4379 rcu_scheduler_active = 1;
4383 * Compute the per-level fanout, either using the exact fanout specified
4384 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4386 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4390 if (rcu_fanout_exact) {
4391 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4392 for (i = rcu_num_lvls - 2; i >= 0; i--)
4393 levelspread[i] = RCU_FANOUT;
4399 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4401 levelspread[i] = (cprv + ccur - 1) / ccur;
4408 * Helper function for rcu_init() that initializes one rcu_state structure.
4410 static void __init rcu_init_one(struct rcu_state *rsp)
4412 static const char * const buf[] = RCU_NODE_NAME_INIT;
4413 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4414 static const char * const exp[] = RCU_EXP_NAME_INIT;
4415 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4416 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4417 static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS];
4418 static u8 fl_mask = 0x1;
4420 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4421 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4425 struct rcu_node *rnp;
4427 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4429 /* Silence gcc 4.8 false positive about array index out of range. */
4430 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4431 panic("rcu_init_one: rcu_num_lvls out of range");
4433 /* Initialize the level-tracking arrays. */
4435 for (i = 0; i < rcu_num_lvls; i++)
4436 levelcnt[i] = num_rcu_lvl[i];
4437 for (i = 1; i < rcu_num_lvls; i++)
4438 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4439 rcu_init_levelspread(levelspread, levelcnt);
4440 rsp->flavor_mask = fl_mask;
4443 /* Initialize the elements themselves, starting from the leaves. */
4445 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4446 cpustride *= levelspread[i];
4447 rnp = rsp->level[i];
4448 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4449 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4450 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4451 &rcu_node_class[i], buf[i]);
4452 raw_spin_lock_init(&rnp->fqslock);
4453 lockdep_set_class_and_name(&rnp->fqslock,
4454 &rcu_fqs_class[i], fqs[i]);
4455 rnp->gpnum = rsp->gpnum;
4456 rnp->completed = rsp->completed;
4458 rnp->qsmaskinit = 0;
4459 rnp->grplo = j * cpustride;
4460 rnp->grphi = (j + 1) * cpustride - 1;
4461 if (rnp->grphi >= nr_cpu_ids)
4462 rnp->grphi = nr_cpu_ids - 1;
4468 rnp->grpnum = j % levelspread[i - 1];
4469 rnp->grpmask = 1UL << rnp->grpnum;
4470 rnp->parent = rsp->level[i - 1] +
4471 j / levelspread[i - 1];
4474 INIT_LIST_HEAD(&rnp->blkd_tasks);
4475 rcu_init_one_nocb(rnp);
4476 mutex_init(&rnp->exp_funnel_mutex);
4477 lockdep_set_class_and_name(&rnp->exp_funnel_mutex,
4478 &rcu_exp_class[i], exp[i]);
4482 init_waitqueue_head(&rsp->gp_wq);
4483 init_waitqueue_head(&rsp->expedited_wq);
4484 rnp = rsp->level[rcu_num_lvls - 1];
4485 for_each_possible_cpu(i) {
4486 while (i > rnp->grphi)
4488 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4489 rcu_boot_init_percpu_data(i, rsp);
4491 list_add(&rsp->flavors, &rcu_struct_flavors);
4495 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4496 * replace the definitions in tree.h because those are needed to size
4497 * the ->node array in the rcu_state structure.
4499 static void __init rcu_init_geometry(void)
4503 int rcu_capacity[RCU_NUM_LVLS];
4506 * Initialize any unspecified boot parameters.
4507 * The default values of jiffies_till_first_fqs and
4508 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4509 * value, which is a function of HZ, then adding one for each
4510 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4512 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4513 if (jiffies_till_first_fqs == ULONG_MAX)
4514 jiffies_till_first_fqs = d;
4515 if (jiffies_till_next_fqs == ULONG_MAX)
4516 jiffies_till_next_fqs = d;
4518 /* If the compile-time values are accurate, just leave. */
4519 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4520 nr_cpu_ids == NR_CPUS)
4522 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4523 rcu_fanout_leaf, nr_cpu_ids);
4526 * The boot-time rcu_fanout_leaf parameter must be at least two
4527 * and cannot exceed the number of bits in the rcu_node masks.
4528 * Complain and fall back to the compile-time values if this
4529 * limit is exceeded.
4531 if (rcu_fanout_leaf < 2 ||
4532 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4533 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4539 * Compute number of nodes that can be handled an rcu_node tree
4540 * with the given number of levels.
4542 rcu_capacity[0] = rcu_fanout_leaf;
4543 for (i = 1; i < RCU_NUM_LVLS; i++)
4544 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4547 * The tree must be able to accommodate the configured number of CPUs.
4548 * If this limit is exceeded, fall back to the compile-time values.
4550 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4551 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4556 /* Calculate the number of levels in the tree. */
4557 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4559 rcu_num_lvls = i + 1;
4561 /* Calculate the number of rcu_nodes at each level of the tree. */
4562 for (i = 0; i < rcu_num_lvls; i++) {
4563 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4564 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4567 /* Calculate the total number of rcu_node structures. */
4569 for (i = 0; i < rcu_num_lvls; i++)
4570 rcu_num_nodes += num_rcu_lvl[i];
4574 * Dump out the structure of the rcu_node combining tree associated
4575 * with the rcu_state structure referenced by rsp.
4577 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4580 struct rcu_node *rnp;
4582 pr_info("rcu_node tree layout dump\n");
4584 rcu_for_each_node_breadth_first(rsp, rnp) {
4585 if (rnp->level != level) {
4590 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4595 void __init rcu_init(void)
4599 rcu_early_boot_tests();
4601 rcu_bootup_announce();
4602 rcu_init_geometry();
4603 rcu_init_one(&rcu_bh_state);
4604 rcu_init_one(&rcu_sched_state);
4606 rcu_dump_rcu_node_tree(&rcu_sched_state);
4607 __rcu_init_preempt();
4608 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4611 * We don't need protection against CPU-hotplug here because
4612 * this is called early in boot, before either interrupts
4613 * or the scheduler are operational.
4615 cpu_notifier(rcu_cpu_notify, 0);
4616 pm_notifier(rcu_pm_notify, 0);
4617 for_each_online_cpu(cpu)
4618 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
4621 #include "tree_plugin.h"