2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/export.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/module.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <linux/prefetch.h>
54 #include <linux/delay.h>
55 #include <linux/stop_machine.h>
56 #include <linux/random.h>
57 #include <linux/trace_events.h>
58 #include <linux/suspend.h>
63 MODULE_ALIAS("rcutree");
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
71 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
73 static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS];
76 * In order to export the rcu_state name to the tracing tools, it
77 * needs to be added in the __tracepoint_string section.
78 * This requires defining a separate variable tp_<sname>_varname
79 * that points to the string being used, and this will allow
80 * the tracing userspace tools to be able to decipher the string
81 * address to the matching string.
84 # define DEFINE_RCU_TPS(sname) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
87 # define RCU_STATE_NAME(sname) sname##_varname
89 # define DEFINE_RCU_TPS(sname)
90 # define RCU_STATE_NAME(sname) __stringify(sname)
93 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
94 DEFINE_RCU_TPS(sname) \
95 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
96 struct rcu_state sname##_state = { \
97 .level = { &sname##_state.node[0] }, \
98 .rda = &sname##_data, \
100 .fqs_state = RCU_GP_IDLE, \
101 .gpnum = 0UL - 300UL, \
102 .completed = 0UL - 300UL, \
103 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
104 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
105 .orphan_donetail = &sname##_state.orphan_donelist, \
106 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
107 .name = RCU_STATE_NAME(sname), \
111 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
112 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
114 static struct rcu_state *const rcu_state_p;
115 static struct rcu_data __percpu *const rcu_data_p;
116 LIST_HEAD(rcu_struct_flavors);
118 /* Dump rcu_node combining tree at boot to verify correct setup. */
119 static bool dump_tree;
120 module_param(dump_tree, bool, 0444);
121 /* Control rcu_node-tree auto-balancing at boot time. */
122 static bool rcu_fanout_exact;
123 module_param(rcu_fanout_exact, bool, 0444);
124 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
125 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
126 module_param(rcu_fanout_leaf, int, 0444);
127 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
128 /* Number of rcu_nodes at specified level. */
129 static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
130 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
133 * The rcu_scheduler_active variable transitions from zero to one just
134 * before the first task is spawned. So when this variable is zero, RCU
135 * can assume that there is but one task, allowing RCU to (for example)
136 * optimize synchronize_sched() to a simple barrier(). When this variable
137 * is one, RCU must actually do all the hard work required to detect real
138 * grace periods. This variable is also used to suppress boot-time false
139 * positives from lockdep-RCU error checking.
141 int rcu_scheduler_active __read_mostly;
142 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
156 static int rcu_scheduler_fully_active __read_mostly;
158 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
159 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163 static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
166 /* rcuc/rcub kthread realtime priority */
167 #ifdef CONFIG_RCU_KTHREAD_PRIO
168 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
169 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
170 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
171 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
172 module_param(kthread_prio, int, 0644);
174 /* Delay in jiffies for grace-period initialization delays, debug only. */
176 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
177 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
178 module_param(gp_preinit_delay, int, 0644);
179 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
180 static const int gp_preinit_delay;
181 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
183 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
184 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
185 module_param(gp_init_delay, int, 0644);
186 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
187 static const int gp_init_delay;
188 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
190 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
191 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
192 module_param(gp_cleanup_delay, int, 0644);
193 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
194 static const int gp_cleanup_delay;
195 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
198 * Number of grace periods between delays, normalized by the duration of
199 * the delay. The longer the the delay, the more the grace periods between
200 * each delay. The reason for this normalization is that it means that,
201 * for non-zero delays, the overall slowdown of grace periods is constant
202 * regardless of the duration of the delay. This arrangement balances
203 * the need for long delays to increase some race probabilities with the
204 * need for fast grace periods to increase other race probabilities.
206 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
209 * Track the rcutorture test sequence number and the update version
210 * number within a given test. The rcutorture_testseq is incremented
211 * on every rcutorture module load and unload, so has an odd value
212 * when a test is running. The rcutorture_vernum is set to zero
213 * when rcutorture starts and is incremented on each rcutorture update.
214 * These variables enable correlating rcutorture output with the
215 * RCU tracing information.
217 unsigned long rcutorture_testseq;
218 unsigned long rcutorture_vernum;
221 * Compute the mask of online CPUs for the specified rcu_node structure.
222 * This will not be stable unless the rcu_node structure's ->lock is
223 * held, but the bit corresponding to the current CPU will be stable
226 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
228 return READ_ONCE(rnp->qsmaskinitnext);
232 * Return true if an RCU grace period is in progress. The READ_ONCE()s
233 * permit this function to be invoked without holding the root rcu_node
234 * structure's ->lock, but of course results can be subject to change.
236 static int rcu_gp_in_progress(struct rcu_state *rsp)
238 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
242 * Note a quiescent state. Because we do not need to know
243 * how many quiescent states passed, just if there was at least
244 * one since the start of the grace period, this just sets a flag.
245 * The caller must have disabled preemption.
247 void rcu_sched_qs(void)
249 if (__this_cpu_read(rcu_sched_data.cpu_no_qs.s)) {
250 trace_rcu_grace_period(TPS("rcu_sched"),
251 __this_cpu_read(rcu_sched_data.gpnum),
253 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
254 if (__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) {
255 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
256 rcu_report_exp_rdp(&rcu_sched_state,
257 this_cpu_ptr(&rcu_sched_data),
265 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
266 trace_rcu_grace_period(TPS("rcu_bh"),
267 __this_cpu_read(rcu_bh_data.gpnum),
269 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
273 static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
275 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
276 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
277 .dynticks = ATOMIC_INIT(1),
278 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
279 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
280 .dynticks_idle = ATOMIC_INIT(1),
281 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
284 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
285 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
288 * Let the RCU core know that this CPU has gone through the scheduler,
289 * which is a quiescent state. This is called when the need for a
290 * quiescent state is urgent, so we burn an atomic operation and full
291 * memory barriers to let the RCU core know about it, regardless of what
292 * this CPU might (or might not) do in the near future.
294 * We inform the RCU core by emulating a zero-duration dyntick-idle
295 * period, which we in turn do by incrementing the ->dynticks counter
298 static void rcu_momentary_dyntick_idle(void)
301 struct rcu_data *rdp;
302 struct rcu_dynticks *rdtp;
304 struct rcu_state *rsp;
306 local_irq_save(flags);
309 * Yes, we can lose flag-setting operations. This is OK, because
310 * the flag will be set again after some delay.
312 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
313 raw_cpu_write(rcu_sched_qs_mask, 0);
315 /* Find the flavor that needs a quiescent state. */
316 for_each_rcu_flavor(rsp) {
317 rdp = raw_cpu_ptr(rsp->rda);
318 if (!(resched_mask & rsp->flavor_mask))
320 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
321 if (READ_ONCE(rdp->mynode->completed) !=
322 READ_ONCE(rdp->cond_resched_completed))
326 * Pretend to be momentarily idle for the quiescent state.
327 * This allows the grace-period kthread to record the
328 * quiescent state, with no need for this CPU to do anything
331 rdtp = this_cpu_ptr(&rcu_dynticks);
332 smp_mb__before_atomic(); /* Earlier stuff before QS. */
333 atomic_add(2, &rdtp->dynticks); /* QS. */
334 smp_mb__after_atomic(); /* Later stuff after QS. */
337 local_irq_restore(flags);
341 * Note a context switch. This is a quiescent state for RCU-sched,
342 * and requires special handling for preemptible RCU.
343 * The caller must have disabled preemption.
345 void rcu_note_context_switch(void)
347 trace_rcu_utilization(TPS("Start context switch"));
349 rcu_preempt_note_context_switch();
350 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
351 rcu_momentary_dyntick_idle();
352 trace_rcu_utilization(TPS("End context switch"));
354 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
357 * Register a quiescent state for all RCU flavors. If there is an
358 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
359 * dyntick-idle quiescent state visible to other CPUs (but only for those
360 * RCU flavors in desperate need of a quiescent state, which will normally
361 * be none of them). Either way, do a lightweight quiescent state for
364 void rcu_all_qs(void)
366 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
367 rcu_momentary_dyntick_idle();
368 this_cpu_inc(rcu_qs_ctr);
370 EXPORT_SYMBOL_GPL(rcu_all_qs);
372 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
373 static long qhimark = 10000; /* If this many pending, ignore blimit. */
374 static long qlowmark = 100; /* Once only this many pending, use blimit. */
376 module_param(blimit, long, 0444);
377 module_param(qhimark, long, 0444);
378 module_param(qlowmark, long, 0444);
380 static ulong jiffies_till_first_fqs = ULONG_MAX;
381 static ulong jiffies_till_next_fqs = ULONG_MAX;
383 module_param(jiffies_till_first_fqs, ulong, 0644);
384 module_param(jiffies_till_next_fqs, ulong, 0644);
387 * How long the grace period must be before we start recruiting
388 * quiescent-state help from rcu_note_context_switch().
390 static ulong jiffies_till_sched_qs = HZ / 20;
391 module_param(jiffies_till_sched_qs, ulong, 0644);
393 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
394 struct rcu_data *rdp);
395 static void force_qs_rnp(struct rcu_state *rsp,
396 int (*f)(struct rcu_data *rsp, bool *isidle,
397 unsigned long *maxj),
398 bool *isidle, unsigned long *maxj);
399 static void force_quiescent_state(struct rcu_state *rsp);
400 static int rcu_pending(void);
403 * Return the number of RCU batches started thus far for debug & stats.
405 unsigned long rcu_batches_started(void)
407 return rcu_state_p->gpnum;
409 EXPORT_SYMBOL_GPL(rcu_batches_started);
412 * Return the number of RCU-sched batches started thus far for debug & stats.
414 unsigned long rcu_batches_started_sched(void)
416 return rcu_sched_state.gpnum;
418 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
421 * Return the number of RCU BH batches started thus far for debug & stats.
423 unsigned long rcu_batches_started_bh(void)
425 return rcu_bh_state.gpnum;
427 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
430 * Return the number of RCU batches completed thus far for debug & stats.
432 unsigned long rcu_batches_completed(void)
434 return rcu_state_p->completed;
436 EXPORT_SYMBOL_GPL(rcu_batches_completed);
439 * Return the number of RCU-sched batches completed thus far for debug & stats.
441 unsigned long rcu_batches_completed_sched(void)
443 return rcu_sched_state.completed;
445 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
448 * Return the number of RCU BH batches completed thus far for debug & stats.
450 unsigned long rcu_batches_completed_bh(void)
452 return rcu_bh_state.completed;
454 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
457 * Force a quiescent state.
459 void rcu_force_quiescent_state(void)
461 force_quiescent_state(rcu_state_p);
463 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
466 * Force a quiescent state for RCU BH.
468 void rcu_bh_force_quiescent_state(void)
470 force_quiescent_state(&rcu_bh_state);
472 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
475 * Force a quiescent state for RCU-sched.
477 void rcu_sched_force_quiescent_state(void)
479 force_quiescent_state(&rcu_sched_state);
481 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
484 * Show the state of the grace-period kthreads.
486 void show_rcu_gp_kthreads(void)
488 struct rcu_state *rsp;
490 for_each_rcu_flavor(rsp) {
491 pr_info("%s: wait state: %d ->state: %#lx\n",
492 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
493 /* sched_show_task(rsp->gp_kthread); */
496 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
499 * Record the number of times rcutorture tests have been initiated and
500 * terminated. This information allows the debugfs tracing stats to be
501 * correlated to the rcutorture messages, even when the rcutorture module
502 * is being repeatedly loaded and unloaded. In other words, we cannot
503 * store this state in rcutorture itself.
505 void rcutorture_record_test_transition(void)
507 rcutorture_testseq++;
508 rcutorture_vernum = 0;
510 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
513 * Send along grace-period-related data for rcutorture diagnostics.
515 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
516 unsigned long *gpnum, unsigned long *completed)
518 struct rcu_state *rsp = NULL;
527 case RCU_SCHED_FLAVOR:
528 rsp = &rcu_sched_state;
534 *flags = READ_ONCE(rsp->gp_flags);
535 *gpnum = READ_ONCE(rsp->gpnum);
536 *completed = READ_ONCE(rsp->completed);
543 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
546 * Record the number of writer passes through the current rcutorture test.
547 * This is also used to correlate debugfs tracing stats with the rcutorture
550 void rcutorture_record_progress(unsigned long vernum)
554 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
557 * Does the CPU have callbacks ready to be invoked?
560 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
562 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
563 rdp->nxttail[RCU_DONE_TAIL] != NULL;
567 * Return the root node of the specified rcu_state structure.
569 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
571 return &rsp->node[0];
575 * Is there any need for future grace periods?
576 * Interrupts must be disabled. If the caller does not hold the root
577 * rnp_node structure's ->lock, the results are advisory only.
579 static int rcu_future_needs_gp(struct rcu_state *rsp)
581 struct rcu_node *rnp = rcu_get_root(rsp);
582 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
583 int *fp = &rnp->need_future_gp[idx];
585 return READ_ONCE(*fp);
589 * Does the current CPU require a not-yet-started grace period?
590 * The caller must have disabled interrupts to prevent races with
591 * normal callback registry.
594 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
598 if (rcu_gp_in_progress(rsp))
599 return 0; /* No, a grace period is already in progress. */
600 if (rcu_future_needs_gp(rsp))
601 return 1; /* Yes, a no-CBs CPU needs one. */
602 if (!rdp->nxttail[RCU_NEXT_TAIL])
603 return 0; /* No, this is a no-CBs (or offline) CPU. */
604 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
605 return 1; /* Yes, this CPU has newly registered callbacks. */
606 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
607 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
608 ULONG_CMP_LT(READ_ONCE(rsp->completed),
609 rdp->nxtcompleted[i]))
610 return 1; /* Yes, CBs for future grace period. */
611 return 0; /* No grace period needed. */
615 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
617 * If the new value of the ->dynticks_nesting counter now is zero,
618 * we really have entered idle, and must do the appropriate accounting.
619 * The caller must have disabled interrupts.
621 static void rcu_eqs_enter_common(long long oldval, bool user)
623 struct rcu_state *rsp;
624 struct rcu_data *rdp;
625 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
627 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
628 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
629 !user && !is_idle_task(current)) {
630 struct task_struct *idle __maybe_unused =
631 idle_task(smp_processor_id());
633 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
634 ftrace_dump(DUMP_ORIG);
635 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
636 current->pid, current->comm,
637 idle->pid, idle->comm); /* must be idle task! */
639 for_each_rcu_flavor(rsp) {
640 rdp = this_cpu_ptr(rsp->rda);
641 do_nocb_deferred_wakeup(rdp);
643 rcu_prepare_for_idle();
644 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
645 smp_mb__before_atomic(); /* See above. */
646 atomic_inc(&rdtp->dynticks);
647 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
648 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
649 atomic_read(&rdtp->dynticks) & 0x1);
650 rcu_dynticks_task_enter();
653 * It is illegal to enter an extended quiescent state while
654 * in an RCU read-side critical section.
656 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
657 "Illegal idle entry in RCU read-side critical section.");
658 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
659 "Illegal idle entry in RCU-bh read-side critical section.");
660 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
661 "Illegal idle entry in RCU-sched read-side critical section.");
665 * Enter an RCU extended quiescent state, which can be either the
666 * idle loop or adaptive-tickless usermode execution.
668 static void rcu_eqs_enter(bool user)
671 struct rcu_dynticks *rdtp;
673 rdtp = this_cpu_ptr(&rcu_dynticks);
674 oldval = rdtp->dynticks_nesting;
675 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
676 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
677 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
678 rdtp->dynticks_nesting = 0;
679 rcu_eqs_enter_common(oldval, user);
681 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
686 * rcu_idle_enter - inform RCU that current CPU is entering idle
688 * Enter idle mode, in other words, -leave- the mode in which RCU
689 * read-side critical sections can occur. (Though RCU read-side
690 * critical sections can occur in irq handlers in idle, a possibility
691 * handled by irq_enter() and irq_exit().)
693 * We crowbar the ->dynticks_nesting field to zero to allow for
694 * the possibility of usermode upcalls having messed up our count
695 * of interrupt nesting level during the prior busy period.
697 void rcu_idle_enter(void)
701 local_irq_save(flags);
702 rcu_eqs_enter(false);
703 rcu_sysidle_enter(0);
704 local_irq_restore(flags);
706 EXPORT_SYMBOL_GPL(rcu_idle_enter);
708 #ifdef CONFIG_NO_HZ_FULL
710 * rcu_user_enter - inform RCU that we are resuming userspace.
712 * Enter RCU idle mode right before resuming userspace. No use of RCU
713 * is permitted between this call and rcu_user_exit(). This way the
714 * CPU doesn't need to maintain the tick for RCU maintenance purposes
715 * when the CPU runs in userspace.
717 void rcu_user_enter(void)
721 #endif /* CONFIG_NO_HZ_FULL */
724 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
726 * Exit from an interrupt handler, which might possibly result in entering
727 * idle mode, in other words, leaving the mode in which read-side critical
728 * sections can occur.
730 * This code assumes that the idle loop never does anything that might
731 * result in unbalanced calls to irq_enter() and irq_exit(). If your
732 * architecture violates this assumption, RCU will give you what you
733 * deserve, good and hard. But very infrequently and irreproducibly.
735 * Use things like work queues to work around this limitation.
737 * You have been warned.
739 void rcu_irq_exit(void)
743 struct rcu_dynticks *rdtp;
745 local_irq_save(flags);
746 rdtp = this_cpu_ptr(&rcu_dynticks);
747 oldval = rdtp->dynticks_nesting;
748 rdtp->dynticks_nesting--;
749 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
750 rdtp->dynticks_nesting < 0);
751 if (rdtp->dynticks_nesting)
752 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
754 rcu_eqs_enter_common(oldval, true);
755 rcu_sysidle_enter(1);
756 local_irq_restore(flags);
760 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
762 * If the new value of the ->dynticks_nesting counter was previously zero,
763 * we really have exited idle, and must do the appropriate accounting.
764 * The caller must have disabled interrupts.
766 static void rcu_eqs_exit_common(long long oldval, int user)
768 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
770 rcu_dynticks_task_exit();
771 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
772 atomic_inc(&rdtp->dynticks);
773 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
774 smp_mb__after_atomic(); /* See above. */
775 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
776 !(atomic_read(&rdtp->dynticks) & 0x1));
777 rcu_cleanup_after_idle();
778 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
779 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
780 !user && !is_idle_task(current)) {
781 struct task_struct *idle __maybe_unused =
782 idle_task(smp_processor_id());
784 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
785 oldval, rdtp->dynticks_nesting);
786 ftrace_dump(DUMP_ORIG);
787 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
788 current->pid, current->comm,
789 idle->pid, idle->comm); /* must be idle task! */
794 * Exit an RCU extended quiescent state, which can be either the
795 * idle loop or adaptive-tickless usermode execution.
797 static void rcu_eqs_exit(bool user)
799 struct rcu_dynticks *rdtp;
802 rdtp = this_cpu_ptr(&rcu_dynticks);
803 oldval = rdtp->dynticks_nesting;
804 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
805 if (oldval & DYNTICK_TASK_NEST_MASK) {
806 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
808 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
809 rcu_eqs_exit_common(oldval, user);
814 * rcu_idle_exit - inform RCU that current CPU is leaving idle
816 * Exit idle mode, in other words, -enter- the mode in which RCU
817 * read-side critical sections can occur.
819 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
820 * allow for the possibility of usermode upcalls messing up our count
821 * of interrupt nesting level during the busy period that is just
824 void rcu_idle_exit(void)
828 local_irq_save(flags);
831 local_irq_restore(flags);
833 EXPORT_SYMBOL_GPL(rcu_idle_exit);
835 #ifdef CONFIG_NO_HZ_FULL
837 * rcu_user_exit - inform RCU that we are exiting userspace.
839 * Exit RCU idle mode while entering the kernel because it can
840 * run a RCU read side critical section anytime.
842 void rcu_user_exit(void)
846 #endif /* CONFIG_NO_HZ_FULL */
849 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
851 * Enter an interrupt handler, which might possibly result in exiting
852 * idle mode, in other words, entering the mode in which read-side critical
853 * sections can occur.
855 * Note that the Linux kernel is fully capable of entering an interrupt
856 * handler that it never exits, for example when doing upcalls to
857 * user mode! This code assumes that the idle loop never does upcalls to
858 * user mode. If your architecture does do upcalls from the idle loop (or
859 * does anything else that results in unbalanced calls to the irq_enter()
860 * and irq_exit() functions), RCU will give you what you deserve, good
861 * and hard. But very infrequently and irreproducibly.
863 * Use things like work queues to work around this limitation.
865 * You have been warned.
867 void rcu_irq_enter(void)
870 struct rcu_dynticks *rdtp;
873 local_irq_save(flags);
874 rdtp = this_cpu_ptr(&rcu_dynticks);
875 oldval = rdtp->dynticks_nesting;
876 rdtp->dynticks_nesting++;
877 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
878 rdtp->dynticks_nesting == 0);
880 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
882 rcu_eqs_exit_common(oldval, true);
884 local_irq_restore(flags);
888 * rcu_nmi_enter - inform RCU of entry to NMI context
890 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
891 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
892 * that the CPU is active. This implementation permits nested NMIs, as
893 * long as the nesting level does not overflow an int. (You will probably
894 * run out of stack space first.)
896 void rcu_nmi_enter(void)
898 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
901 /* Complain about underflow. */
902 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
905 * If idle from RCU viewpoint, atomically increment ->dynticks
906 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
907 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
908 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
909 * to be in the outermost NMI handler that interrupted an RCU-idle
910 * period (observation due to Andy Lutomirski).
912 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
913 smp_mb__before_atomic(); /* Force delay from prior write. */
914 atomic_inc(&rdtp->dynticks);
915 /* atomic_inc() before later RCU read-side crit sects */
916 smp_mb__after_atomic(); /* See above. */
917 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
920 rdtp->dynticks_nmi_nesting += incby;
925 * rcu_nmi_exit - inform RCU of exit from NMI context
927 * If we are returning from the outermost NMI handler that interrupted an
928 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
929 * to let the RCU grace-period handling know that the CPU is back to
932 void rcu_nmi_exit(void)
934 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
937 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
938 * (We are exiting an NMI handler, so RCU better be paying attention
941 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
942 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
945 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
946 * leave it in non-RCU-idle state.
948 if (rdtp->dynticks_nmi_nesting != 1) {
949 rdtp->dynticks_nmi_nesting -= 2;
953 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
954 rdtp->dynticks_nmi_nesting = 0;
955 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
956 smp_mb__before_atomic(); /* See above. */
957 atomic_inc(&rdtp->dynticks);
958 smp_mb__after_atomic(); /* Force delay to next write. */
959 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
963 * __rcu_is_watching - are RCU read-side critical sections safe?
965 * Return true if RCU is watching the running CPU, which means that
966 * this CPU can safely enter RCU read-side critical sections. Unlike
967 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
968 * least disabled preemption.
970 bool notrace __rcu_is_watching(void)
972 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
976 * rcu_is_watching - see if RCU thinks that the current CPU is idle
978 * If the current CPU is in its idle loop and is neither in an interrupt
979 * or NMI handler, return true.
981 bool notrace rcu_is_watching(void)
985 preempt_disable_notrace();
986 ret = __rcu_is_watching();
987 preempt_enable_notrace();
990 EXPORT_SYMBOL_GPL(rcu_is_watching);
992 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
995 * Is the current CPU online? Disable preemption to avoid false positives
996 * that could otherwise happen due to the current CPU number being sampled,
997 * this task being preempted, its old CPU being taken offline, resuming
998 * on some other CPU, then determining that its old CPU is now offline.
999 * It is OK to use RCU on an offline processor during initial boot, hence
1000 * the check for rcu_scheduler_fully_active. Note also that it is OK
1001 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1002 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1003 * offline to continue to use RCU for one jiffy after marking itself
1004 * offline in the cpu_online_mask. This leniency is necessary given the
1005 * non-atomic nature of the online and offline processing, for example,
1006 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1009 * This is also why RCU internally marks CPUs online during the
1010 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1012 * Disable checking if in an NMI handler because we cannot safely report
1013 * errors from NMI handlers anyway.
1015 bool rcu_lockdep_current_cpu_online(void)
1017 struct rcu_data *rdp;
1018 struct rcu_node *rnp;
1024 rdp = this_cpu_ptr(&rcu_sched_data);
1026 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1027 !rcu_scheduler_fully_active;
1031 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1033 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1036 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1038 * If the current CPU is idle or running at a first-level (not nested)
1039 * interrupt from idle, return true. The caller must have at least
1040 * disabled preemption.
1042 static int rcu_is_cpu_rrupt_from_idle(void)
1044 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1048 * Snapshot the specified CPU's dynticks counter so that we can later
1049 * credit them with an implicit quiescent state. Return 1 if this CPU
1050 * is in dynticks idle mode, which is an extended quiescent state.
1052 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1053 bool *isidle, unsigned long *maxj)
1055 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1056 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1057 if ((rdp->dynticks_snap & 0x1) == 0) {
1058 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1061 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1062 rdp->mynode->gpnum))
1063 WRITE_ONCE(rdp->gpwrap, true);
1069 * Return true if the specified CPU has passed through a quiescent
1070 * state by virtue of being in or having passed through an dynticks
1071 * idle state since the last call to dyntick_save_progress_counter()
1072 * for this same CPU, or by virtue of having been offline.
1074 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1075 bool *isidle, unsigned long *maxj)
1081 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1082 snap = (unsigned int)rdp->dynticks_snap;
1085 * If the CPU passed through or entered a dynticks idle phase with
1086 * no active irq/NMI handlers, then we can safely pretend that the CPU
1087 * already acknowledged the request to pass through a quiescent
1088 * state. Either way, that CPU cannot possibly be in an RCU
1089 * read-side critical section that started before the beginning
1090 * of the current RCU grace period.
1092 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1093 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1094 rdp->dynticks_fqs++;
1099 * Check for the CPU being offline, but only if the grace period
1100 * is old enough. We don't need to worry about the CPU changing
1101 * state: If we see it offline even once, it has been through a
1104 * The reason for insisting that the grace period be at least
1105 * one jiffy old is that CPUs that are not quite online and that
1106 * have just gone offline can still execute RCU read-side critical
1109 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1110 return 0; /* Grace period is not old enough. */
1112 if (cpu_is_offline(rdp->cpu)) {
1113 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1119 * A CPU running for an extended time within the kernel can
1120 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1121 * even context-switching back and forth between a pair of
1122 * in-kernel CPU-bound tasks cannot advance grace periods.
1123 * So if the grace period is old enough, make the CPU pay attention.
1124 * Note that the unsynchronized assignments to the per-CPU
1125 * rcu_sched_qs_mask variable are safe. Yes, setting of
1126 * bits can be lost, but they will be set again on the next
1127 * force-quiescent-state pass. So lost bit sets do not result
1128 * in incorrect behavior, merely in a grace period lasting
1129 * a few jiffies longer than it might otherwise. Because
1130 * there are at most four threads involved, and because the
1131 * updates are only once every few jiffies, the probability of
1132 * lossage (and thus of slight grace-period extension) is
1135 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1136 * is set too high, we override with half of the RCU CPU stall
1139 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1140 if (ULONG_CMP_GE(jiffies,
1141 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1142 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1143 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1144 WRITE_ONCE(rdp->cond_resched_completed,
1145 READ_ONCE(rdp->mynode->completed));
1146 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1148 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1149 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1150 rdp->rsp->jiffies_resched += 5; /* Enable beating. */
1151 } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1152 /* Time to beat on that CPU again! */
1153 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1154 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1161 static void record_gp_stall_check_time(struct rcu_state *rsp)
1163 unsigned long j = jiffies;
1167 smp_wmb(); /* Record start time before stall time. */
1168 j1 = rcu_jiffies_till_stall_check();
1169 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1170 rsp->jiffies_resched = j + j1 / 2;
1171 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1175 * Complain about starvation of grace-period kthread.
1177 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1183 gpa = READ_ONCE(rsp->gp_activity);
1184 if (j - gpa > 2 * HZ)
1185 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n",
1187 rsp->gpnum, rsp->completed,
1188 rsp->gp_flags, rsp->gp_state,
1189 rsp->gp_kthread ? rsp->gp_kthread->state : 0);
1193 * Dump stacks of all tasks running on stalled CPUs.
1195 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1198 unsigned long flags;
1199 struct rcu_node *rnp;
1201 rcu_for_each_leaf_node(rsp, rnp) {
1202 raw_spin_lock_irqsave(&rnp->lock, flags);
1203 if (rnp->qsmask != 0) {
1204 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1205 if (rnp->qsmask & (1UL << cpu))
1206 dump_cpu_task(rnp->grplo + cpu);
1208 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1212 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1216 unsigned long flags;
1220 struct rcu_node *rnp = rcu_get_root(rsp);
1223 /* Only let one CPU complain about others per time interval. */
1225 raw_spin_lock_irqsave(&rnp->lock, flags);
1226 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1227 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1228 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1231 WRITE_ONCE(rsp->jiffies_stall,
1232 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1233 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1236 * OK, time to rat on our buddy...
1237 * See Documentation/RCU/stallwarn.txt for info on how to debug
1238 * RCU CPU stall warnings.
1240 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1242 print_cpu_stall_info_begin();
1243 rcu_for_each_leaf_node(rsp, rnp) {
1244 raw_spin_lock_irqsave(&rnp->lock, flags);
1245 ndetected += rcu_print_task_stall(rnp);
1246 if (rnp->qsmask != 0) {
1247 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1248 if (rnp->qsmask & (1UL << cpu)) {
1249 print_cpu_stall_info(rsp,
1254 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1257 print_cpu_stall_info_end();
1258 for_each_possible_cpu(cpu)
1259 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1260 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1261 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1262 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1264 rcu_dump_cpu_stacks(rsp);
1266 if (READ_ONCE(rsp->gpnum) != gpnum ||
1267 READ_ONCE(rsp->completed) == gpnum) {
1268 pr_err("INFO: Stall ended before state dump start\n");
1271 gpa = READ_ONCE(rsp->gp_activity);
1272 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1273 rsp->name, j - gpa, j, gpa,
1274 jiffies_till_next_fqs,
1275 rcu_get_root(rsp)->qsmask);
1276 /* In this case, the current CPU might be at fault. */
1277 sched_show_task(current);
1281 /* Complain about tasks blocking the grace period. */
1282 rcu_print_detail_task_stall(rsp);
1284 rcu_check_gp_kthread_starvation(rsp);
1286 force_quiescent_state(rsp); /* Kick them all. */
1289 static void print_cpu_stall(struct rcu_state *rsp)
1292 unsigned long flags;
1293 struct rcu_node *rnp = rcu_get_root(rsp);
1297 * OK, time to rat on ourselves...
1298 * See Documentation/RCU/stallwarn.txt for info on how to debug
1299 * RCU CPU stall warnings.
1301 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1302 print_cpu_stall_info_begin();
1303 print_cpu_stall_info(rsp, smp_processor_id());
1304 print_cpu_stall_info_end();
1305 for_each_possible_cpu(cpu)
1306 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1307 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1308 jiffies - rsp->gp_start,
1309 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1311 rcu_check_gp_kthread_starvation(rsp);
1313 rcu_dump_cpu_stacks(rsp);
1315 raw_spin_lock_irqsave(&rnp->lock, flags);
1316 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1317 WRITE_ONCE(rsp->jiffies_stall,
1318 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1319 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1322 * Attempt to revive the RCU machinery by forcing a context switch.
1324 * A context switch would normally allow the RCU state machine to make
1325 * progress and it could be we're stuck in kernel space without context
1326 * switches for an entirely unreasonable amount of time.
1328 resched_cpu(smp_processor_id());
1331 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1333 unsigned long completed;
1334 unsigned long gpnum;
1338 struct rcu_node *rnp;
1340 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1345 * Lots of memory barriers to reject false positives.
1347 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1348 * then rsp->gp_start, and finally rsp->completed. These values
1349 * are updated in the opposite order with memory barriers (or
1350 * equivalent) during grace-period initialization and cleanup.
1351 * Now, a false positive can occur if we get an new value of
1352 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1353 * the memory barriers, the only way that this can happen is if one
1354 * grace period ends and another starts between these two fetches.
1355 * Detect this by comparing rsp->completed with the previous fetch
1358 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1359 * and rsp->gp_start suffice to forestall false positives.
1361 gpnum = READ_ONCE(rsp->gpnum);
1362 smp_rmb(); /* Pick up ->gpnum first... */
1363 js = READ_ONCE(rsp->jiffies_stall);
1364 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1365 gps = READ_ONCE(rsp->gp_start);
1366 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1367 completed = READ_ONCE(rsp->completed);
1368 if (ULONG_CMP_GE(completed, gpnum) ||
1369 ULONG_CMP_LT(j, js) ||
1370 ULONG_CMP_GE(gps, js))
1371 return; /* No stall or GP completed since entering function. */
1373 if (rcu_gp_in_progress(rsp) &&
1374 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1376 /* We haven't checked in, so go dump stack. */
1377 print_cpu_stall(rsp);
1379 } else if (rcu_gp_in_progress(rsp) &&
1380 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1382 /* They had a few time units to dump stack, so complain. */
1383 print_other_cpu_stall(rsp, gpnum);
1388 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1390 * Set the stall-warning timeout way off into the future, thus preventing
1391 * any RCU CPU stall-warning messages from appearing in the current set of
1392 * RCU grace periods.
1394 * The caller must disable hard irqs.
1396 void rcu_cpu_stall_reset(void)
1398 struct rcu_state *rsp;
1400 for_each_rcu_flavor(rsp)
1401 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1405 * Initialize the specified rcu_data structure's default callback list
1406 * to empty. The default callback list is the one that is not used by
1407 * no-callbacks CPUs.
1409 static void init_default_callback_list(struct rcu_data *rdp)
1413 rdp->nxtlist = NULL;
1414 for (i = 0; i < RCU_NEXT_SIZE; i++)
1415 rdp->nxttail[i] = &rdp->nxtlist;
1419 * Initialize the specified rcu_data structure's callback list to empty.
1421 static void init_callback_list(struct rcu_data *rdp)
1423 if (init_nocb_callback_list(rdp))
1425 init_default_callback_list(rdp);
1429 * Determine the value that ->completed will have at the end of the
1430 * next subsequent grace period. This is used to tag callbacks so that
1431 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1432 * been dyntick-idle for an extended period with callbacks under the
1433 * influence of RCU_FAST_NO_HZ.
1435 * The caller must hold rnp->lock with interrupts disabled.
1437 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1438 struct rcu_node *rnp)
1441 * If RCU is idle, we just wait for the next grace period.
1442 * But we can only be sure that RCU is idle if we are looking
1443 * at the root rcu_node structure -- otherwise, a new grace
1444 * period might have started, but just not yet gotten around
1445 * to initializing the current non-root rcu_node structure.
1447 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1448 return rnp->completed + 1;
1451 * Otherwise, wait for a possible partial grace period and
1452 * then the subsequent full grace period.
1454 return rnp->completed + 2;
1458 * Trace-event helper function for rcu_start_future_gp() and
1459 * rcu_nocb_wait_gp().
1461 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1462 unsigned long c, const char *s)
1464 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1465 rnp->completed, c, rnp->level,
1466 rnp->grplo, rnp->grphi, s);
1470 * Start some future grace period, as needed to handle newly arrived
1471 * callbacks. The required future grace periods are recorded in each
1472 * rcu_node structure's ->need_future_gp field. Returns true if there
1473 * is reason to awaken the grace-period kthread.
1475 * The caller must hold the specified rcu_node structure's ->lock.
1477 static bool __maybe_unused
1478 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1479 unsigned long *c_out)
1484 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1487 * Pick up grace-period number for new callbacks. If this
1488 * grace period is already marked as needed, return to the caller.
1490 c = rcu_cbs_completed(rdp->rsp, rnp);
1491 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1492 if (rnp->need_future_gp[c & 0x1]) {
1493 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1498 * If either this rcu_node structure or the root rcu_node structure
1499 * believe that a grace period is in progress, then we must wait
1500 * for the one following, which is in "c". Because our request
1501 * will be noticed at the end of the current grace period, we don't
1502 * need to explicitly start one. We only do the lockless check
1503 * of rnp_root's fields if the current rcu_node structure thinks
1504 * there is no grace period in flight, and because we hold rnp->lock,
1505 * the only possible change is when rnp_root's two fields are
1506 * equal, in which case rnp_root->gpnum might be concurrently
1507 * incremented. But that is OK, as it will just result in our
1508 * doing some extra useless work.
1510 if (rnp->gpnum != rnp->completed ||
1511 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1512 rnp->need_future_gp[c & 0x1]++;
1513 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1518 * There might be no grace period in progress. If we don't already
1519 * hold it, acquire the root rcu_node structure's lock in order to
1520 * start one (if needed).
1522 if (rnp != rnp_root) {
1523 raw_spin_lock(&rnp_root->lock);
1524 smp_mb__after_unlock_lock();
1528 * Get a new grace-period number. If there really is no grace
1529 * period in progress, it will be smaller than the one we obtained
1530 * earlier. Adjust callbacks as needed. Note that even no-CBs
1531 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1533 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1534 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1535 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1536 rdp->nxtcompleted[i] = c;
1539 * If the needed for the required grace period is already
1540 * recorded, trace and leave.
1542 if (rnp_root->need_future_gp[c & 0x1]) {
1543 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1547 /* Record the need for the future grace period. */
1548 rnp_root->need_future_gp[c & 0x1]++;
1550 /* If a grace period is not already in progress, start one. */
1551 if (rnp_root->gpnum != rnp_root->completed) {
1552 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1554 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1555 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1558 if (rnp != rnp_root)
1559 raw_spin_unlock(&rnp_root->lock);
1567 * Clean up any old requests for the just-ended grace period. Also return
1568 * whether any additional grace periods have been requested. Also invoke
1569 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1570 * waiting for this grace period to complete.
1572 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1574 int c = rnp->completed;
1576 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1578 rcu_nocb_gp_cleanup(rsp, rnp);
1579 rnp->need_future_gp[c & 0x1] = 0;
1580 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1581 trace_rcu_future_gp(rnp, rdp, c,
1582 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1587 * Awaken the grace-period kthread for the specified flavor of RCU.
1588 * Don't do a self-awaken, and don't bother awakening when there is
1589 * nothing for the grace-period kthread to do (as in several CPUs
1590 * raced to awaken, and we lost), and finally don't try to awaken
1591 * a kthread that has not yet been created.
1593 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1595 if (current == rsp->gp_kthread ||
1596 !READ_ONCE(rsp->gp_flags) ||
1599 wake_up(&rsp->gp_wq);
1603 * If there is room, assign a ->completed number to any callbacks on
1604 * this CPU that have not already been assigned. Also accelerate any
1605 * callbacks that were previously assigned a ->completed number that has
1606 * since proven to be too conservative, which can happen if callbacks get
1607 * assigned a ->completed number while RCU is idle, but with reference to
1608 * a non-root rcu_node structure. This function is idempotent, so it does
1609 * not hurt to call it repeatedly. Returns an flag saying that we should
1610 * awaken the RCU grace-period kthread.
1612 * The caller must hold rnp->lock with interrupts disabled.
1614 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1615 struct rcu_data *rdp)
1621 /* If the CPU has no callbacks, nothing to do. */
1622 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1626 * Starting from the sublist containing the callbacks most
1627 * recently assigned a ->completed number and working down, find the
1628 * first sublist that is not assignable to an upcoming grace period.
1629 * Such a sublist has something in it (first two tests) and has
1630 * a ->completed number assigned that will complete sooner than
1631 * the ->completed number for newly arrived callbacks (last test).
1633 * The key point is that any later sublist can be assigned the
1634 * same ->completed number as the newly arrived callbacks, which
1635 * means that the callbacks in any of these later sublist can be
1636 * grouped into a single sublist, whether or not they have already
1637 * been assigned a ->completed number.
1639 c = rcu_cbs_completed(rsp, rnp);
1640 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1641 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1642 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1646 * If there are no sublist for unassigned callbacks, leave.
1647 * At the same time, advance "i" one sublist, so that "i" will
1648 * index into the sublist where all the remaining callbacks should
1651 if (++i >= RCU_NEXT_TAIL)
1655 * Assign all subsequent callbacks' ->completed number to the next
1656 * full grace period and group them all in the sublist initially
1659 for (; i <= RCU_NEXT_TAIL; i++) {
1660 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1661 rdp->nxtcompleted[i] = c;
1663 /* Record any needed additional grace periods. */
1664 ret = rcu_start_future_gp(rnp, rdp, NULL);
1666 /* Trace depending on how much we were able to accelerate. */
1667 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1668 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1670 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1675 * Move any callbacks whose grace period has completed to the
1676 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1677 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1678 * sublist. This function is idempotent, so it does not hurt to
1679 * invoke it repeatedly. As long as it is not invoked -too- often...
1680 * Returns true if the RCU grace-period kthread needs to be awakened.
1682 * The caller must hold rnp->lock with interrupts disabled.
1684 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1685 struct rcu_data *rdp)
1689 /* If the CPU has no callbacks, nothing to do. */
1690 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1694 * Find all callbacks whose ->completed numbers indicate that they
1695 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1697 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1698 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1700 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1702 /* Clean up any sublist tail pointers that were misordered above. */
1703 for (j = RCU_WAIT_TAIL; j < i; j++)
1704 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1706 /* Copy down callbacks to fill in empty sublists. */
1707 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1708 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1710 rdp->nxttail[j] = rdp->nxttail[i];
1711 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1714 /* Classify any remaining callbacks. */
1715 return rcu_accelerate_cbs(rsp, rnp, rdp);
1719 * Update CPU-local rcu_data state to record the beginnings and ends of
1720 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1721 * structure corresponding to the current CPU, and must have irqs disabled.
1722 * Returns true if the grace-period kthread needs to be awakened.
1724 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1725 struct rcu_data *rdp)
1729 /* Handle the ends of any preceding grace periods first. */
1730 if (rdp->completed == rnp->completed &&
1731 !unlikely(READ_ONCE(rdp->gpwrap))) {
1733 /* No grace period end, so just accelerate recent callbacks. */
1734 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1738 /* Advance callbacks. */
1739 ret = rcu_advance_cbs(rsp, rnp, rdp);
1741 /* Remember that we saw this grace-period completion. */
1742 rdp->completed = rnp->completed;
1743 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1746 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1748 * If the current grace period is waiting for this CPU,
1749 * set up to detect a quiescent state, otherwise don't
1750 * go looking for one.
1752 rdp->gpnum = rnp->gpnum;
1753 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1754 rdp->cpu_no_qs.b.norm = true;
1755 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1756 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1757 zero_cpu_stall_ticks(rdp);
1758 WRITE_ONCE(rdp->gpwrap, false);
1763 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1765 unsigned long flags;
1767 struct rcu_node *rnp;
1769 local_irq_save(flags);
1771 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1772 rdp->completed == READ_ONCE(rnp->completed) &&
1773 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1774 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1775 local_irq_restore(flags);
1778 smp_mb__after_unlock_lock();
1779 needwake = __note_gp_changes(rsp, rnp, rdp);
1780 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1782 rcu_gp_kthread_wake(rsp);
1785 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1788 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1789 schedule_timeout_uninterruptible(delay);
1793 * Initialize a new grace period. Return 0 if no grace period required.
1795 static int rcu_gp_init(struct rcu_state *rsp)
1797 unsigned long oldmask;
1798 struct rcu_data *rdp;
1799 struct rcu_node *rnp = rcu_get_root(rsp);
1801 WRITE_ONCE(rsp->gp_activity, jiffies);
1802 raw_spin_lock_irq(&rnp->lock);
1803 smp_mb__after_unlock_lock();
1804 if (!READ_ONCE(rsp->gp_flags)) {
1805 /* Spurious wakeup, tell caller to go back to sleep. */
1806 raw_spin_unlock_irq(&rnp->lock);
1809 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1811 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1813 * Grace period already in progress, don't start another.
1814 * Not supposed to be able to happen.
1816 raw_spin_unlock_irq(&rnp->lock);
1820 /* Advance to a new grace period and initialize state. */
1821 record_gp_stall_check_time(rsp);
1822 /* Record GP times before starting GP, hence smp_store_release(). */
1823 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1824 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1825 raw_spin_unlock_irq(&rnp->lock);
1828 * Apply per-leaf buffered online and offline operations to the
1829 * rcu_node tree. Note that this new grace period need not wait
1830 * for subsequent online CPUs, and that quiescent-state forcing
1831 * will handle subsequent offline CPUs.
1833 rcu_for_each_leaf_node(rsp, rnp) {
1834 rcu_gp_slow(rsp, gp_preinit_delay);
1835 raw_spin_lock_irq(&rnp->lock);
1836 smp_mb__after_unlock_lock();
1837 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1838 !rnp->wait_blkd_tasks) {
1839 /* Nothing to do on this leaf rcu_node structure. */
1840 raw_spin_unlock_irq(&rnp->lock);
1844 /* Record old state, apply changes to ->qsmaskinit field. */
1845 oldmask = rnp->qsmaskinit;
1846 rnp->qsmaskinit = rnp->qsmaskinitnext;
1848 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1849 if (!oldmask != !rnp->qsmaskinit) {
1850 if (!oldmask) /* First online CPU for this rcu_node. */
1851 rcu_init_new_rnp(rnp);
1852 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1853 rnp->wait_blkd_tasks = true;
1854 else /* Last offline CPU and can propagate. */
1855 rcu_cleanup_dead_rnp(rnp);
1859 * If all waited-on tasks from prior grace period are
1860 * done, and if all this rcu_node structure's CPUs are
1861 * still offline, propagate up the rcu_node tree and
1862 * clear ->wait_blkd_tasks. Otherwise, if one of this
1863 * rcu_node structure's CPUs has since come back online,
1864 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1865 * checks for this, so just call it unconditionally).
1867 if (rnp->wait_blkd_tasks &&
1868 (!rcu_preempt_has_tasks(rnp) ||
1870 rnp->wait_blkd_tasks = false;
1871 rcu_cleanup_dead_rnp(rnp);
1874 raw_spin_unlock_irq(&rnp->lock);
1878 * Set the quiescent-state-needed bits in all the rcu_node
1879 * structures for all currently online CPUs in breadth-first order,
1880 * starting from the root rcu_node structure, relying on the layout
1881 * of the tree within the rsp->node[] array. Note that other CPUs
1882 * will access only the leaves of the hierarchy, thus seeing that no
1883 * grace period is in progress, at least until the corresponding
1884 * leaf node has been initialized. In addition, we have excluded
1885 * CPU-hotplug operations.
1887 * The grace period cannot complete until the initialization
1888 * process finishes, because this kthread handles both.
1890 rcu_for_each_node_breadth_first(rsp, rnp) {
1891 rcu_gp_slow(rsp, gp_init_delay);
1892 raw_spin_lock_irq(&rnp->lock);
1893 smp_mb__after_unlock_lock();
1894 rdp = this_cpu_ptr(rsp->rda);
1895 rcu_preempt_check_blocked_tasks(rnp);
1896 rnp->qsmask = rnp->qsmaskinit;
1897 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1898 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1899 WRITE_ONCE(rnp->completed, rsp->completed);
1900 if (rnp == rdp->mynode)
1901 (void)__note_gp_changes(rsp, rnp, rdp);
1902 rcu_preempt_boost_start_gp(rnp);
1903 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1904 rnp->level, rnp->grplo,
1905 rnp->grphi, rnp->qsmask);
1906 raw_spin_unlock_irq(&rnp->lock);
1907 cond_resched_rcu_qs();
1908 WRITE_ONCE(rsp->gp_activity, jiffies);
1915 * Helper function for wait_event_interruptible_timeout() wakeup
1916 * at force-quiescent-state time.
1918 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1920 struct rcu_node *rnp = rcu_get_root(rsp);
1922 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1923 *gfp = READ_ONCE(rsp->gp_flags);
1924 if (*gfp & RCU_GP_FLAG_FQS)
1927 /* The current grace period has completed. */
1928 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1935 * Do one round of quiescent-state forcing.
1937 static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1939 int fqs_state = fqs_state_in;
1940 bool isidle = false;
1942 struct rcu_node *rnp = rcu_get_root(rsp);
1944 WRITE_ONCE(rsp->gp_activity, jiffies);
1946 if (fqs_state == RCU_SAVE_DYNTICK) {
1947 /* Collect dyntick-idle snapshots. */
1948 if (is_sysidle_rcu_state(rsp)) {
1950 maxj = jiffies - ULONG_MAX / 4;
1952 force_qs_rnp(rsp, dyntick_save_progress_counter,
1954 rcu_sysidle_report_gp(rsp, isidle, maxj);
1955 fqs_state = RCU_FORCE_QS;
1957 /* Handle dyntick-idle and offline CPUs. */
1959 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1961 /* Clear flag to prevent immediate re-entry. */
1962 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1963 raw_spin_lock_irq(&rnp->lock);
1964 smp_mb__after_unlock_lock();
1965 WRITE_ONCE(rsp->gp_flags,
1966 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
1967 raw_spin_unlock_irq(&rnp->lock);
1973 * Clean up after the old grace period.
1975 static void rcu_gp_cleanup(struct rcu_state *rsp)
1977 unsigned long gp_duration;
1978 bool needgp = false;
1980 struct rcu_data *rdp;
1981 struct rcu_node *rnp = rcu_get_root(rsp);
1983 WRITE_ONCE(rsp->gp_activity, jiffies);
1984 raw_spin_lock_irq(&rnp->lock);
1985 smp_mb__after_unlock_lock();
1986 gp_duration = jiffies - rsp->gp_start;
1987 if (gp_duration > rsp->gp_max)
1988 rsp->gp_max = gp_duration;
1991 * We know the grace period is complete, but to everyone else
1992 * it appears to still be ongoing. But it is also the case
1993 * that to everyone else it looks like there is nothing that
1994 * they can do to advance the grace period. It is therefore
1995 * safe for us to drop the lock in order to mark the grace
1996 * period as completed in all of the rcu_node structures.
1998 raw_spin_unlock_irq(&rnp->lock);
2001 * Propagate new ->completed value to rcu_node structures so
2002 * that other CPUs don't have to wait until the start of the next
2003 * grace period to process their callbacks. This also avoids
2004 * some nasty RCU grace-period initialization races by forcing
2005 * the end of the current grace period to be completely recorded in
2006 * all of the rcu_node structures before the beginning of the next
2007 * grace period is recorded in any of the rcu_node structures.
2009 rcu_for_each_node_breadth_first(rsp, rnp) {
2010 raw_spin_lock_irq(&rnp->lock);
2011 smp_mb__after_unlock_lock();
2012 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2013 WARN_ON_ONCE(rnp->qsmask);
2014 WRITE_ONCE(rnp->completed, rsp->gpnum);
2015 rdp = this_cpu_ptr(rsp->rda);
2016 if (rnp == rdp->mynode)
2017 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2018 /* smp_mb() provided by prior unlock-lock pair. */
2019 nocb += rcu_future_gp_cleanup(rsp, rnp);
2020 raw_spin_unlock_irq(&rnp->lock);
2021 cond_resched_rcu_qs();
2022 WRITE_ONCE(rsp->gp_activity, jiffies);
2023 rcu_gp_slow(rsp, gp_cleanup_delay);
2025 rnp = rcu_get_root(rsp);
2026 raw_spin_lock_irq(&rnp->lock);
2027 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
2028 rcu_nocb_gp_set(rnp, nocb);
2030 /* Declare grace period done. */
2031 WRITE_ONCE(rsp->completed, rsp->gpnum);
2032 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2033 rsp->fqs_state = RCU_GP_IDLE;
2034 rdp = this_cpu_ptr(rsp->rda);
2035 /* Advance CBs to reduce false positives below. */
2036 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2037 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2038 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2039 trace_rcu_grace_period(rsp->name,
2040 READ_ONCE(rsp->gpnum),
2043 raw_spin_unlock_irq(&rnp->lock);
2047 * Body of kthread that handles grace periods.
2049 static int __noreturn rcu_gp_kthread(void *arg)
2055 struct rcu_state *rsp = arg;
2056 struct rcu_node *rnp = rcu_get_root(rsp);
2058 rcu_bind_gp_kthread();
2061 /* Handle grace-period start. */
2063 trace_rcu_grace_period(rsp->name,
2064 READ_ONCE(rsp->gpnum),
2066 rsp->gp_state = RCU_GP_WAIT_GPS;
2067 wait_event_interruptible(rsp->gp_wq,
2068 READ_ONCE(rsp->gp_flags) &
2070 rsp->gp_state = RCU_GP_DONE_GPS;
2071 /* Locking provides needed memory barrier. */
2072 if (rcu_gp_init(rsp))
2074 cond_resched_rcu_qs();
2075 WRITE_ONCE(rsp->gp_activity, jiffies);
2076 WARN_ON(signal_pending(current));
2077 trace_rcu_grace_period(rsp->name,
2078 READ_ONCE(rsp->gpnum),
2082 /* Handle quiescent-state forcing. */
2083 fqs_state = RCU_SAVE_DYNTICK;
2084 j = jiffies_till_first_fqs;
2087 jiffies_till_first_fqs = HZ;
2092 rsp->jiffies_force_qs = jiffies + j;
2093 trace_rcu_grace_period(rsp->name,
2094 READ_ONCE(rsp->gpnum),
2096 rsp->gp_state = RCU_GP_WAIT_FQS;
2097 ret = wait_event_interruptible_timeout(rsp->gp_wq,
2098 rcu_gp_fqs_check_wake(rsp, &gf), j);
2099 rsp->gp_state = RCU_GP_DOING_FQS;
2100 /* Locking provides needed memory barriers. */
2101 /* If grace period done, leave loop. */
2102 if (!READ_ONCE(rnp->qsmask) &&
2103 !rcu_preempt_blocked_readers_cgp(rnp))
2105 /* If time for quiescent-state forcing, do it. */
2106 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2107 (gf & RCU_GP_FLAG_FQS)) {
2108 trace_rcu_grace_period(rsp->name,
2109 READ_ONCE(rsp->gpnum),
2111 fqs_state = rcu_gp_fqs(rsp, fqs_state);
2112 trace_rcu_grace_period(rsp->name,
2113 READ_ONCE(rsp->gpnum),
2115 cond_resched_rcu_qs();
2116 WRITE_ONCE(rsp->gp_activity, jiffies);
2118 /* Deal with stray signal. */
2119 cond_resched_rcu_qs();
2120 WRITE_ONCE(rsp->gp_activity, jiffies);
2121 WARN_ON(signal_pending(current));
2122 trace_rcu_grace_period(rsp->name,
2123 READ_ONCE(rsp->gpnum),
2126 j = jiffies_till_next_fqs;
2129 jiffies_till_next_fqs = HZ;
2132 jiffies_till_next_fqs = 1;
2136 /* Handle grace-period end. */
2137 rsp->gp_state = RCU_GP_CLEANUP;
2138 rcu_gp_cleanup(rsp);
2139 rsp->gp_state = RCU_GP_CLEANED;
2144 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2145 * in preparation for detecting the next grace period. The caller must hold
2146 * the root node's ->lock and hard irqs must be disabled.
2148 * Note that it is legal for a dying CPU (which is marked as offline) to
2149 * invoke this function. This can happen when the dying CPU reports its
2152 * Returns true if the grace-period kthread must be awakened.
2155 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2156 struct rcu_data *rdp)
2158 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2160 * Either we have not yet spawned the grace-period
2161 * task, this CPU does not need another grace period,
2162 * or a grace period is already in progress.
2163 * Either way, don't start a new grace period.
2167 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2168 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2172 * We can't do wakeups while holding the rnp->lock, as that
2173 * could cause possible deadlocks with the rq->lock. Defer
2174 * the wakeup to our caller.
2180 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2181 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2182 * is invoked indirectly from rcu_advance_cbs(), which would result in
2183 * endless recursion -- or would do so if it wasn't for the self-deadlock
2184 * that is encountered beforehand.
2186 * Returns true if the grace-period kthread needs to be awakened.
2188 static bool rcu_start_gp(struct rcu_state *rsp)
2190 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2191 struct rcu_node *rnp = rcu_get_root(rsp);
2195 * If there is no grace period in progress right now, any
2196 * callbacks we have up to this point will be satisfied by the
2197 * next grace period. Also, advancing the callbacks reduces the
2198 * probability of false positives from cpu_needs_another_gp()
2199 * resulting in pointless grace periods. So, advance callbacks
2200 * then start the grace period!
2202 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2203 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2208 * Report a full set of quiescent states to the specified rcu_state
2209 * data structure. This involves cleaning up after the prior grace
2210 * period and letting rcu_start_gp() start up the next grace period
2211 * if one is needed. Note that the caller must hold rnp->lock, which
2212 * is released before return.
2214 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2215 __releases(rcu_get_root(rsp)->lock)
2217 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2218 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2219 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2220 rcu_gp_kthread_wake(rsp);
2224 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2225 * Allows quiescent states for a group of CPUs to be reported at one go
2226 * to the specified rcu_node structure, though all the CPUs in the group
2227 * must be represented by the same rcu_node structure (which need not be a
2228 * leaf rcu_node structure, though it often will be). The gps parameter
2229 * is the grace-period snapshot, which means that the quiescent states
2230 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2231 * must be held upon entry, and it is released before return.
2234 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2235 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2236 __releases(rnp->lock)
2238 unsigned long oldmask = 0;
2239 struct rcu_node *rnp_c;
2241 /* Walk up the rcu_node hierarchy. */
2243 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2246 * Our bit has already been cleared, or the
2247 * relevant grace period is already over, so done.
2249 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2252 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2253 rnp->qsmask &= ~mask;
2254 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2255 mask, rnp->qsmask, rnp->level,
2256 rnp->grplo, rnp->grphi,
2258 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2260 /* Other bits still set at this level, so done. */
2261 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2264 mask = rnp->grpmask;
2265 if (rnp->parent == NULL) {
2267 /* No more levels. Exit loop holding root lock. */
2271 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2274 raw_spin_lock_irqsave(&rnp->lock, flags);
2275 smp_mb__after_unlock_lock();
2276 oldmask = rnp_c->qsmask;
2280 * Get here if we are the last CPU to pass through a quiescent
2281 * state for this grace period. Invoke rcu_report_qs_rsp()
2282 * to clean up and start the next grace period if one is needed.
2284 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2288 * Record a quiescent state for all tasks that were previously queued
2289 * on the specified rcu_node structure and that were blocking the current
2290 * RCU grace period. The caller must hold the specified rnp->lock with
2291 * irqs disabled, and this lock is released upon return, but irqs remain
2294 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2295 struct rcu_node *rnp, unsigned long flags)
2296 __releases(rnp->lock)
2300 struct rcu_node *rnp_p;
2302 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2303 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2304 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2305 return; /* Still need more quiescent states! */
2308 rnp_p = rnp->parent;
2309 if (rnp_p == NULL) {
2311 * Only one rcu_node structure in the tree, so don't
2312 * try to report up to its nonexistent parent!
2314 rcu_report_qs_rsp(rsp, flags);
2318 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2320 mask = rnp->grpmask;
2321 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2322 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
2323 smp_mb__after_unlock_lock();
2324 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2328 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2329 * structure. This must be either called from the specified CPU, or
2330 * called when the specified CPU is known to be offline (and when it is
2331 * also known that no other CPU is concurrently trying to help the offline
2332 * CPU). The lastcomp argument is used to make sure we are still in the
2333 * grace period of interest. We don't want to end the current grace period
2334 * based on quiescent states detected in an earlier grace period!
2337 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2339 unsigned long flags;
2342 struct rcu_node *rnp;
2345 raw_spin_lock_irqsave(&rnp->lock, flags);
2346 smp_mb__after_unlock_lock();
2347 if ((rdp->cpu_no_qs.b.norm &&
2348 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2349 rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2353 * The grace period in which this quiescent state was
2354 * recorded has ended, so don't report it upwards.
2355 * We will instead need a new quiescent state that lies
2356 * within the current grace period.
2358 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2359 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2360 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2363 mask = rdp->grpmask;
2364 if ((rnp->qsmask & mask) == 0) {
2365 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2367 rdp->core_needs_qs = 0;
2370 * This GP can't end until cpu checks in, so all of our
2371 * callbacks can be processed during the next GP.
2373 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2375 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2376 /* ^^^ Released rnp->lock */
2378 rcu_gp_kthread_wake(rsp);
2383 * Check to see if there is a new grace period of which this CPU
2384 * is not yet aware, and if so, set up local rcu_data state for it.
2385 * Otherwise, see if this CPU has just passed through its first
2386 * quiescent state for this grace period, and record that fact if so.
2389 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2391 /* Check for grace-period ends and beginnings. */
2392 note_gp_changes(rsp, rdp);
2395 * Does this CPU still need to do its part for current grace period?
2396 * If no, return and let the other CPUs do their part as well.
2398 if (!rdp->core_needs_qs)
2402 * Was there a quiescent state since the beginning of the grace
2403 * period? If no, then exit and wait for the next call.
2405 if (rdp->cpu_no_qs.b.norm &&
2406 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2410 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2413 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2417 * Send the specified CPU's RCU callbacks to the orphanage. The
2418 * specified CPU must be offline, and the caller must hold the
2422 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2423 struct rcu_node *rnp, struct rcu_data *rdp)
2425 /* No-CBs CPUs do not have orphanable callbacks. */
2426 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2430 * Orphan the callbacks. First adjust the counts. This is safe
2431 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2432 * cannot be running now. Thus no memory barrier is required.
2434 if (rdp->nxtlist != NULL) {
2435 rsp->qlen_lazy += rdp->qlen_lazy;
2436 rsp->qlen += rdp->qlen;
2437 rdp->n_cbs_orphaned += rdp->qlen;
2439 WRITE_ONCE(rdp->qlen, 0);
2443 * Next, move those callbacks still needing a grace period to
2444 * the orphanage, where some other CPU will pick them up.
2445 * Some of the callbacks might have gone partway through a grace
2446 * period, but that is too bad. They get to start over because we
2447 * cannot assume that grace periods are synchronized across CPUs.
2448 * We don't bother updating the ->nxttail[] array yet, instead
2449 * we just reset the whole thing later on.
2451 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2452 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2453 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2454 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2458 * Then move the ready-to-invoke callbacks to the orphanage,
2459 * where some other CPU will pick them up. These will not be
2460 * required to pass though another grace period: They are done.
2462 if (rdp->nxtlist != NULL) {
2463 *rsp->orphan_donetail = rdp->nxtlist;
2464 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2468 * Finally, initialize the rcu_data structure's list to empty and
2469 * disallow further callbacks on this CPU.
2471 init_callback_list(rdp);
2472 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2476 * Adopt the RCU callbacks from the specified rcu_state structure's
2477 * orphanage. The caller must hold the ->orphan_lock.
2479 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2482 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2484 /* No-CBs CPUs are handled specially. */
2485 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2486 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2489 /* Do the accounting first. */
2490 rdp->qlen_lazy += rsp->qlen_lazy;
2491 rdp->qlen += rsp->qlen;
2492 rdp->n_cbs_adopted += rsp->qlen;
2493 if (rsp->qlen_lazy != rsp->qlen)
2494 rcu_idle_count_callbacks_posted();
2499 * We do not need a memory barrier here because the only way we
2500 * can get here if there is an rcu_barrier() in flight is if
2501 * we are the task doing the rcu_barrier().
2504 /* First adopt the ready-to-invoke callbacks. */
2505 if (rsp->orphan_donelist != NULL) {
2506 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2507 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2508 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2509 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2510 rdp->nxttail[i] = rsp->orphan_donetail;
2511 rsp->orphan_donelist = NULL;
2512 rsp->orphan_donetail = &rsp->orphan_donelist;
2515 /* And then adopt the callbacks that still need a grace period. */
2516 if (rsp->orphan_nxtlist != NULL) {
2517 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2518 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2519 rsp->orphan_nxtlist = NULL;
2520 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2525 * Trace the fact that this CPU is going offline.
2527 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2529 RCU_TRACE(unsigned long mask);
2530 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2531 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2533 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2536 RCU_TRACE(mask = rdp->grpmask);
2537 trace_rcu_grace_period(rsp->name,
2538 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2543 * All CPUs for the specified rcu_node structure have gone offline,
2544 * and all tasks that were preempted within an RCU read-side critical
2545 * section while running on one of those CPUs have since exited their RCU
2546 * read-side critical section. Some other CPU is reporting this fact with
2547 * the specified rcu_node structure's ->lock held and interrupts disabled.
2548 * This function therefore goes up the tree of rcu_node structures,
2549 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2550 * the leaf rcu_node structure's ->qsmaskinit field has already been
2553 * This function does check that the specified rcu_node structure has
2554 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2555 * prematurely. That said, invoking it after the fact will cost you
2556 * a needless lock acquisition. So once it has done its work, don't
2559 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2562 struct rcu_node *rnp = rnp_leaf;
2564 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2565 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2568 mask = rnp->grpmask;
2572 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2573 smp_mb__after_unlock_lock(); /* GP memory ordering. */
2574 rnp->qsmaskinit &= ~mask;
2575 rnp->qsmask &= ~mask;
2576 if (rnp->qsmaskinit) {
2577 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2580 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2585 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
2586 * function. We now remove it from the rcu_node tree's ->qsmaskinit
2589 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
2591 unsigned long flags;
2593 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2594 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2596 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2599 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
2600 mask = rdp->grpmask;
2601 raw_spin_lock_irqsave(&rnp->lock, flags);
2602 smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */
2603 rnp->qsmaskinitnext &= ~mask;
2604 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2608 * The CPU has been completely removed, and some other CPU is reporting
2609 * this fact from process context. Do the remainder of the cleanup,
2610 * including orphaning the outgoing CPU's RCU callbacks, and also
2611 * adopting them. There can only be one CPU hotplug operation at a time,
2612 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2614 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2616 unsigned long flags;
2617 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2618 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2620 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2623 /* Adjust any no-longer-needed kthreads. */
2624 rcu_boost_kthread_setaffinity(rnp, -1);
2626 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2627 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2628 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2629 rcu_adopt_orphan_cbs(rsp, flags);
2630 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2632 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2633 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2634 cpu, rdp->qlen, rdp->nxtlist);
2638 * Invoke any RCU callbacks that have made it to the end of their grace
2639 * period. Thottle as specified by rdp->blimit.
2641 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2643 unsigned long flags;
2644 struct rcu_head *next, *list, **tail;
2645 long bl, count, count_lazy;
2648 /* If no callbacks are ready, just return. */
2649 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2650 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2651 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2652 need_resched(), is_idle_task(current),
2653 rcu_is_callbacks_kthread());
2658 * Extract the list of ready callbacks, disabling to prevent
2659 * races with call_rcu() from interrupt handlers.
2661 local_irq_save(flags);
2662 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2664 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2665 list = rdp->nxtlist;
2666 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2667 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2668 tail = rdp->nxttail[RCU_DONE_TAIL];
2669 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2670 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2671 rdp->nxttail[i] = &rdp->nxtlist;
2672 local_irq_restore(flags);
2674 /* Invoke callbacks. */
2675 count = count_lazy = 0;
2679 debug_rcu_head_unqueue(list);
2680 if (__rcu_reclaim(rsp->name, list))
2683 /* Stop only if limit reached and CPU has something to do. */
2684 if (++count >= bl &&
2686 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2690 local_irq_save(flags);
2691 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2692 is_idle_task(current),
2693 rcu_is_callbacks_kthread());
2695 /* Update count, and requeue any remaining callbacks. */
2697 *tail = rdp->nxtlist;
2698 rdp->nxtlist = list;
2699 for (i = 0; i < RCU_NEXT_SIZE; i++)
2700 if (&rdp->nxtlist == rdp->nxttail[i])
2701 rdp->nxttail[i] = tail;
2705 smp_mb(); /* List handling before counting for rcu_barrier(). */
2706 rdp->qlen_lazy -= count_lazy;
2707 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2708 rdp->n_cbs_invoked += count;
2710 /* Reinstate batch limit if we have worked down the excess. */
2711 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2712 rdp->blimit = blimit;
2714 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2715 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2716 rdp->qlen_last_fqs_check = 0;
2717 rdp->n_force_qs_snap = rsp->n_force_qs;
2718 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2719 rdp->qlen_last_fqs_check = rdp->qlen;
2720 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2722 local_irq_restore(flags);
2724 /* Re-invoke RCU core processing if there are callbacks remaining. */
2725 if (cpu_has_callbacks_ready_to_invoke(rdp))
2730 * Check to see if this CPU is in a non-context-switch quiescent state
2731 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2732 * Also schedule RCU core processing.
2734 * This function must be called from hardirq context. It is normally
2735 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2736 * false, there is no point in invoking rcu_check_callbacks().
2738 void rcu_check_callbacks(int user)
2740 trace_rcu_utilization(TPS("Start scheduler-tick"));
2741 increment_cpu_stall_ticks();
2742 if (user || rcu_is_cpu_rrupt_from_idle()) {
2745 * Get here if this CPU took its interrupt from user
2746 * mode or from the idle loop, and if this is not a
2747 * nested interrupt. In this case, the CPU is in
2748 * a quiescent state, so note it.
2750 * No memory barrier is required here because both
2751 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2752 * variables that other CPUs neither access nor modify,
2753 * at least not while the corresponding CPU is online.
2759 } else if (!in_softirq()) {
2762 * Get here if this CPU did not take its interrupt from
2763 * softirq, in other words, if it is not interrupting
2764 * a rcu_bh read-side critical section. This is an _bh
2765 * critical section, so note it.
2770 rcu_preempt_check_callbacks();
2774 rcu_note_voluntary_context_switch(current);
2775 trace_rcu_utilization(TPS("End scheduler-tick"));
2779 * Scan the leaf rcu_node structures, processing dyntick state for any that
2780 * have not yet encountered a quiescent state, using the function specified.
2781 * Also initiate boosting for any threads blocked on the root rcu_node.
2783 * The caller must have suppressed start of new grace periods.
2785 static void force_qs_rnp(struct rcu_state *rsp,
2786 int (*f)(struct rcu_data *rsp, bool *isidle,
2787 unsigned long *maxj),
2788 bool *isidle, unsigned long *maxj)
2792 unsigned long flags;
2794 struct rcu_node *rnp;
2796 rcu_for_each_leaf_node(rsp, rnp) {
2797 cond_resched_rcu_qs();
2799 raw_spin_lock_irqsave(&rnp->lock, flags);
2800 smp_mb__after_unlock_lock();
2801 if (rnp->qsmask == 0) {
2802 if (rcu_state_p == &rcu_sched_state ||
2803 rsp != rcu_state_p ||
2804 rcu_preempt_blocked_readers_cgp(rnp)) {
2806 * No point in scanning bits because they
2807 * are all zero. But we might need to
2808 * priority-boost blocked readers.
2810 rcu_initiate_boost(rnp, flags);
2811 /* rcu_initiate_boost() releases rnp->lock */
2815 (rnp->parent->qsmask & rnp->grpmask)) {
2817 * Race between grace-period
2818 * initialization and task exiting RCU
2819 * read-side critical section: Report.
2821 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2822 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2828 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2829 if ((rnp->qsmask & bit) != 0) {
2830 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2835 /* Idle/offline CPUs, report (releases rnp->lock. */
2836 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2838 /* Nothing to do here, so just drop the lock. */
2839 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2845 * Force quiescent states on reluctant CPUs, and also detect which
2846 * CPUs are in dyntick-idle mode.
2848 static void force_quiescent_state(struct rcu_state *rsp)
2850 unsigned long flags;
2852 struct rcu_node *rnp;
2853 struct rcu_node *rnp_old = NULL;
2855 /* Funnel through hierarchy to reduce memory contention. */
2856 rnp = __this_cpu_read(rsp->rda->mynode);
2857 for (; rnp != NULL; rnp = rnp->parent) {
2858 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2859 !raw_spin_trylock(&rnp->fqslock);
2860 if (rnp_old != NULL)
2861 raw_spin_unlock(&rnp_old->fqslock);
2863 rsp->n_force_qs_lh++;
2868 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2870 /* Reached the root of the rcu_node tree, acquire lock. */
2871 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2872 smp_mb__after_unlock_lock();
2873 raw_spin_unlock(&rnp_old->fqslock);
2874 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2875 rsp->n_force_qs_lh++;
2876 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2877 return; /* Someone beat us to it. */
2879 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2880 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2881 rcu_gp_kthread_wake(rsp);
2885 * This does the RCU core processing work for the specified rcu_state
2886 * and rcu_data structures. This may be called only from the CPU to
2887 * whom the rdp belongs.
2890 __rcu_process_callbacks(struct rcu_state *rsp)
2892 unsigned long flags;
2894 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2896 WARN_ON_ONCE(rdp->beenonline == 0);
2898 /* Update RCU state based on any recent quiescent states. */
2899 rcu_check_quiescent_state(rsp, rdp);
2901 /* Does this CPU require a not-yet-started grace period? */
2902 local_irq_save(flags);
2903 if (cpu_needs_another_gp(rsp, rdp)) {
2904 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2905 needwake = rcu_start_gp(rsp);
2906 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2908 rcu_gp_kthread_wake(rsp);
2910 local_irq_restore(flags);
2913 /* If there are callbacks ready, invoke them. */
2914 if (cpu_has_callbacks_ready_to_invoke(rdp))
2915 invoke_rcu_callbacks(rsp, rdp);
2917 /* Do any needed deferred wakeups of rcuo kthreads. */
2918 do_nocb_deferred_wakeup(rdp);
2922 * Do RCU core processing for the current CPU.
2924 static void rcu_process_callbacks(struct softirq_action *unused)
2926 struct rcu_state *rsp;
2928 if (cpu_is_offline(smp_processor_id()))
2930 trace_rcu_utilization(TPS("Start RCU core"));
2931 for_each_rcu_flavor(rsp)
2932 __rcu_process_callbacks(rsp);
2933 trace_rcu_utilization(TPS("End RCU core"));
2937 * Schedule RCU callback invocation. If the specified type of RCU
2938 * does not support RCU priority boosting, just do a direct call,
2939 * otherwise wake up the per-CPU kernel kthread. Note that because we
2940 * are running on the current CPU with softirqs disabled, the
2941 * rcu_cpu_kthread_task cannot disappear out from under us.
2943 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2945 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2947 if (likely(!rsp->boost)) {
2948 rcu_do_batch(rsp, rdp);
2951 invoke_rcu_callbacks_kthread();
2954 static void invoke_rcu_core(void)
2956 if (cpu_online(smp_processor_id()))
2957 raise_softirq(RCU_SOFTIRQ);
2961 * Handle any core-RCU processing required by a call_rcu() invocation.
2963 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2964 struct rcu_head *head, unsigned long flags)
2969 * If called from an extended quiescent state, invoke the RCU
2970 * core in order to force a re-evaluation of RCU's idleness.
2972 if (!rcu_is_watching())
2975 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2976 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2980 * Force the grace period if too many callbacks or too long waiting.
2981 * Enforce hysteresis, and don't invoke force_quiescent_state()
2982 * if some other CPU has recently done so. Also, don't bother
2983 * invoking force_quiescent_state() if the newly enqueued callback
2984 * is the only one waiting for a grace period to complete.
2986 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2988 /* Are we ignoring a completed grace period? */
2989 note_gp_changes(rsp, rdp);
2991 /* Start a new grace period if one not already started. */
2992 if (!rcu_gp_in_progress(rsp)) {
2993 struct rcu_node *rnp_root = rcu_get_root(rsp);
2995 raw_spin_lock(&rnp_root->lock);
2996 smp_mb__after_unlock_lock();
2997 needwake = rcu_start_gp(rsp);
2998 raw_spin_unlock(&rnp_root->lock);
3000 rcu_gp_kthread_wake(rsp);
3002 /* Give the grace period a kick. */
3003 rdp->blimit = LONG_MAX;
3004 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3005 *rdp->nxttail[RCU_DONE_TAIL] != head)
3006 force_quiescent_state(rsp);
3007 rdp->n_force_qs_snap = rsp->n_force_qs;
3008 rdp->qlen_last_fqs_check = rdp->qlen;
3014 * RCU callback function to leak a callback.
3016 static void rcu_leak_callback(struct rcu_head *rhp)
3021 * Helper function for call_rcu() and friends. The cpu argument will
3022 * normally be -1, indicating "currently running CPU". It may specify
3023 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3024 * is expected to specify a CPU.
3027 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
3028 struct rcu_state *rsp, int cpu, bool lazy)
3030 unsigned long flags;
3031 struct rcu_data *rdp;
3033 WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
3034 if (debug_rcu_head_queue(head)) {
3035 /* Probable double call_rcu(), so leak the callback. */
3036 WRITE_ONCE(head->func, rcu_leak_callback);
3037 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3044 * Opportunistically note grace-period endings and beginnings.
3045 * Note that we might see a beginning right after we see an
3046 * end, but never vice versa, since this CPU has to pass through
3047 * a quiescent state betweentimes.
3049 local_irq_save(flags);
3050 rdp = this_cpu_ptr(rsp->rda);
3052 /* Add the callback to our list. */
3053 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3057 rdp = per_cpu_ptr(rsp->rda, cpu);
3058 if (likely(rdp->mynode)) {
3059 /* Post-boot, so this should be for a no-CBs CPU. */
3060 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3061 WARN_ON_ONCE(offline);
3062 /* Offline CPU, _call_rcu() illegal, leak callback. */
3063 local_irq_restore(flags);
3067 * Very early boot, before rcu_init(). Initialize if needed
3068 * and then drop through to queue the callback.
3071 WARN_ON_ONCE(!rcu_is_watching());
3072 if (!likely(rdp->nxtlist))
3073 init_default_callback_list(rdp);
3075 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3079 rcu_idle_count_callbacks_posted();
3080 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3081 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3082 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3084 if (__is_kfree_rcu_offset((unsigned long)func))
3085 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3086 rdp->qlen_lazy, rdp->qlen);
3088 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3090 /* Go handle any RCU core processing required. */
3091 __call_rcu_core(rsp, rdp, head, flags);
3092 local_irq_restore(flags);
3096 * Queue an RCU-sched callback for invocation after a grace period.
3098 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
3100 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3102 EXPORT_SYMBOL_GPL(call_rcu_sched);
3105 * Queue an RCU callback for invocation after a quicker grace period.
3107 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
3109 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3111 EXPORT_SYMBOL_GPL(call_rcu_bh);
3114 * Queue an RCU callback for lazy invocation after a grace period.
3115 * This will likely be later named something like "call_rcu_lazy()",
3116 * but this change will require some way of tagging the lazy RCU
3117 * callbacks in the list of pending callbacks. Until then, this
3118 * function may only be called from __kfree_rcu().
3120 void kfree_call_rcu(struct rcu_head *head,
3121 void (*func)(struct rcu_head *rcu))
3123 __call_rcu(head, func, rcu_state_p, -1, 1);
3125 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3128 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3129 * any blocking grace-period wait automatically implies a grace period
3130 * if there is only one CPU online at any point time during execution
3131 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3132 * occasionally incorrectly indicate that there are multiple CPUs online
3133 * when there was in fact only one the whole time, as this just adds
3134 * some overhead: RCU still operates correctly.
3136 static inline int rcu_blocking_is_gp(void)
3140 might_sleep(); /* Check for RCU read-side critical section. */
3142 ret = num_online_cpus() <= 1;
3148 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3150 * Control will return to the caller some time after a full rcu-sched
3151 * grace period has elapsed, in other words after all currently executing
3152 * rcu-sched read-side critical sections have completed. These read-side
3153 * critical sections are delimited by rcu_read_lock_sched() and
3154 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3155 * local_irq_disable(), and so on may be used in place of
3156 * rcu_read_lock_sched().
3158 * This means that all preempt_disable code sequences, including NMI and
3159 * non-threaded hardware-interrupt handlers, in progress on entry will
3160 * have completed before this primitive returns. However, this does not
3161 * guarantee that softirq handlers will have completed, since in some
3162 * kernels, these handlers can run in process context, and can block.
3164 * Note that this guarantee implies further memory-ordering guarantees.
3165 * On systems with more than one CPU, when synchronize_sched() returns,
3166 * each CPU is guaranteed to have executed a full memory barrier since the
3167 * end of its last RCU-sched read-side critical section whose beginning
3168 * preceded the call to synchronize_sched(). In addition, each CPU having
3169 * an RCU read-side critical section that extends beyond the return from
3170 * synchronize_sched() is guaranteed to have executed a full memory barrier
3171 * after the beginning of synchronize_sched() and before the beginning of
3172 * that RCU read-side critical section. Note that these guarantees include
3173 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3174 * that are executing in the kernel.
3176 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3177 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3178 * to have executed a full memory barrier during the execution of
3179 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3180 * again only if the system has more than one CPU).
3182 * This primitive provides the guarantees made by the (now removed)
3183 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3184 * guarantees that rcu_read_lock() sections will have completed.
3185 * In "classic RCU", these two guarantees happen to be one and
3186 * the same, but can differ in realtime RCU implementations.
3188 void synchronize_sched(void)
3190 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3191 lock_is_held(&rcu_lock_map) ||
3192 lock_is_held(&rcu_sched_lock_map),
3193 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3194 if (rcu_blocking_is_gp())
3196 if (rcu_gp_is_expedited())
3197 synchronize_sched_expedited();
3199 wait_rcu_gp(call_rcu_sched);
3201 EXPORT_SYMBOL_GPL(synchronize_sched);
3204 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3206 * Control will return to the caller some time after a full rcu_bh grace
3207 * period has elapsed, in other words after all currently executing rcu_bh
3208 * read-side critical sections have completed. RCU read-side critical
3209 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3210 * and may be nested.
3212 * See the description of synchronize_sched() for more detailed information
3213 * on memory ordering guarantees.
3215 void synchronize_rcu_bh(void)
3217 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3218 lock_is_held(&rcu_lock_map) ||
3219 lock_is_held(&rcu_sched_lock_map),
3220 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3221 if (rcu_blocking_is_gp())
3223 if (rcu_gp_is_expedited())
3224 synchronize_rcu_bh_expedited();
3226 wait_rcu_gp(call_rcu_bh);
3228 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3231 * get_state_synchronize_rcu - Snapshot current RCU state
3233 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3234 * to determine whether or not a full grace period has elapsed in the
3237 unsigned long get_state_synchronize_rcu(void)
3240 * Any prior manipulation of RCU-protected data must happen
3241 * before the load from ->gpnum.
3246 * Make sure this load happens before the purportedly
3247 * time-consuming work between get_state_synchronize_rcu()
3248 * and cond_synchronize_rcu().
3250 return smp_load_acquire(&rcu_state_p->gpnum);
3252 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3255 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3257 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3259 * If a full RCU grace period has elapsed since the earlier call to
3260 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3261 * synchronize_rcu() to wait for a full grace period.
3263 * Yes, this function does not take counter wrap into account. But
3264 * counter wrap is harmless. If the counter wraps, we have waited for
3265 * more than 2 billion grace periods (and way more on a 64-bit system!),
3266 * so waiting for one additional grace period should be just fine.
3268 void cond_synchronize_rcu(unsigned long oldstate)
3270 unsigned long newstate;
3273 * Ensure that this load happens before any RCU-destructive
3274 * actions the caller might carry out after we return.
3276 newstate = smp_load_acquire(&rcu_state_p->completed);
3277 if (ULONG_CMP_GE(oldstate, newstate))
3280 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3283 * get_state_synchronize_sched - Snapshot current RCU-sched state
3285 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3286 * to determine whether or not a full grace period has elapsed in the
3289 unsigned long get_state_synchronize_sched(void)
3292 * Any prior manipulation of RCU-protected data must happen
3293 * before the load from ->gpnum.
3298 * Make sure this load happens before the purportedly
3299 * time-consuming work between get_state_synchronize_sched()
3300 * and cond_synchronize_sched().
3302 return smp_load_acquire(&rcu_sched_state.gpnum);
3304 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3307 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3309 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3311 * If a full RCU-sched grace period has elapsed since the earlier call to
3312 * get_state_synchronize_sched(), just return. Otherwise, invoke
3313 * synchronize_sched() to wait for a full grace period.
3315 * Yes, this function does not take counter wrap into account. But
3316 * counter wrap is harmless. If the counter wraps, we have waited for
3317 * more than 2 billion grace periods (and way more on a 64-bit system!),
3318 * so waiting for one additional grace period should be just fine.
3320 void cond_synchronize_sched(unsigned long oldstate)
3322 unsigned long newstate;
3325 * Ensure that this load happens before any RCU-destructive
3326 * actions the caller might carry out after we return.
3328 newstate = smp_load_acquire(&rcu_sched_state.completed);
3329 if (ULONG_CMP_GE(oldstate, newstate))
3330 synchronize_sched();
3332 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3334 /* Adjust sequence number for start of update-side operation. */
3335 static void rcu_seq_start(unsigned long *sp)
3337 WRITE_ONCE(*sp, *sp + 1);
3338 smp_mb(); /* Ensure update-side operation after counter increment. */
3339 WARN_ON_ONCE(!(*sp & 0x1));
3342 /* Adjust sequence number for end of update-side operation. */
3343 static void rcu_seq_end(unsigned long *sp)
3345 smp_mb(); /* Ensure update-side operation before counter increment. */
3346 WRITE_ONCE(*sp, *sp + 1);
3347 WARN_ON_ONCE(*sp & 0x1);
3350 /* Take a snapshot of the update side's sequence number. */
3351 static unsigned long rcu_seq_snap(unsigned long *sp)
3355 smp_mb(); /* Caller's modifications seen first by other CPUs. */
3356 s = (READ_ONCE(*sp) + 3) & ~0x1;
3357 smp_mb(); /* Above access must not bleed into critical section. */
3362 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3363 * full update-side operation has occurred.
3365 static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3367 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3370 /* Wrapper functions for expedited grace periods. */
3371 static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
3373 rcu_seq_start(&rsp->expedited_sequence);
3375 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
3377 rcu_seq_end(&rsp->expedited_sequence);
3378 smp_mb(); /* Ensure that consecutive grace periods serialize. */
3380 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
3382 return rcu_seq_snap(&rsp->expedited_sequence);
3384 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
3386 return rcu_seq_done(&rsp->expedited_sequence, s);
3390 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3391 * recent CPU-online activity. Note that these masks are not cleared
3392 * when CPUs go offline, so they reflect the union of all CPUs that have
3393 * ever been online. This means that this function normally takes its
3394 * no-work-to-do fastpath.
3396 static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
3399 unsigned long flags;
3401 unsigned long oldmask;
3402 int ncpus = READ_ONCE(rsp->ncpus);
3403 struct rcu_node *rnp;
3404 struct rcu_node *rnp_up;
3406 /* If no new CPUs onlined since last time, nothing to do. */
3407 if (likely(ncpus == rsp->ncpus_snap))
3409 rsp->ncpus_snap = ncpus;
3412 * Each pass through the following loop propagates newly onlined
3413 * CPUs for the current rcu_node structure up the rcu_node tree.
3415 rcu_for_each_leaf_node(rsp, rnp) {
3416 raw_spin_lock_irqsave(&rnp->lock, flags);
3417 smp_mb__after_unlock_lock();
3418 if (rnp->expmaskinit == rnp->expmaskinitnext) {
3419 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3420 continue; /* No new CPUs, nothing to do. */
3423 /* Update this node's mask, track old value for propagation. */
3424 oldmask = rnp->expmaskinit;
3425 rnp->expmaskinit = rnp->expmaskinitnext;
3426 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3428 /* If was already nonzero, nothing to propagate. */
3432 /* Propagate the new CPU up the tree. */
3433 mask = rnp->grpmask;
3434 rnp_up = rnp->parent;
3437 raw_spin_lock_irqsave(&rnp_up->lock, flags);
3438 smp_mb__after_unlock_lock();
3439 if (rnp_up->expmaskinit)
3441 rnp_up->expmaskinit |= mask;
3442 raw_spin_unlock_irqrestore(&rnp_up->lock, flags);
3445 mask = rnp_up->grpmask;
3446 rnp_up = rnp_up->parent;
3452 * Reset the ->expmask values in the rcu_node tree in preparation for
3453 * a new expedited grace period.
3455 static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
3457 unsigned long flags;
3458 struct rcu_node *rnp;
3460 sync_exp_reset_tree_hotplug(rsp);
3461 rcu_for_each_node_breadth_first(rsp, rnp) {
3462 raw_spin_lock_irqsave(&rnp->lock, flags);
3463 smp_mb__after_unlock_lock();
3464 WARN_ON_ONCE(rnp->expmask);
3465 rnp->expmask = rnp->expmaskinit;
3466 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3471 * Return non-zero if there is no RCU expedited grace period in progress
3472 * for the specified rcu_node structure, in other words, if all CPUs and
3473 * tasks covered by the specified rcu_node structure have done their bit
3474 * for the current expedited grace period. Works only for preemptible
3475 * RCU -- other RCU implementation use other means.
3477 * Caller must hold the root rcu_node's exp_funnel_mutex.
3479 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
3481 return rnp->exp_tasks == NULL &&
3482 READ_ONCE(rnp->expmask) == 0;
3486 * Report the exit from RCU read-side critical section for the last task
3487 * that queued itself during or before the current expedited preemptible-RCU
3488 * grace period. This event is reported either to the rcu_node structure on
3489 * which the task was queued or to one of that rcu_node structure's ancestors,
3490 * recursively up the tree. (Calm down, calm down, we do the recursion
3493 * Caller must hold the root rcu_node's exp_funnel_mutex and the
3494 * specified rcu_node structure's ->lock.
3496 static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
3497 bool wake, unsigned long flags)
3498 __releases(rnp->lock)
3503 if (!sync_rcu_preempt_exp_done(rnp)) {
3505 rcu_initiate_boost(rnp, flags);
3507 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3510 if (rnp->parent == NULL) {
3511 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3513 smp_mb(); /* EGP done before wake_up(). */
3514 wake_up(&rsp->expedited_wq);
3518 mask = rnp->grpmask;
3519 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
3521 raw_spin_lock(&rnp->lock); /* irqs already disabled */
3522 smp_mb__after_unlock_lock();
3523 WARN_ON_ONCE(!(rnp->expmask & mask));
3524 rnp->expmask &= ~mask;
3529 * Report expedited quiescent state for specified node. This is a
3530 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3532 * Caller must hold the root rcu_node's exp_funnel_mutex.
3534 static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
3535 struct rcu_node *rnp, bool wake)
3537 unsigned long flags;
3539 raw_spin_lock_irqsave(&rnp->lock, flags);
3540 smp_mb__after_unlock_lock();
3541 __rcu_report_exp_rnp(rsp, rnp, wake, flags);
3545 * Report expedited quiescent state for multiple CPUs, all covered by the
3546 * specified leaf rcu_node structure. Caller must hold the root
3547 * rcu_node's exp_funnel_mutex.
3549 static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
3550 unsigned long mask, bool wake)
3552 unsigned long flags;
3554 raw_spin_lock_irqsave(&rnp->lock, flags);
3555 smp_mb__after_unlock_lock();
3556 WARN_ON_ONCE((rnp->expmask & mask) != mask);
3557 rnp->expmask &= ~mask;
3558 __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
3562 * Report expedited quiescent state for specified rcu_data (CPU).
3563 * Caller must hold the root rcu_node's exp_funnel_mutex.
3565 static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
3568 rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
3571 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3572 static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp,
3573 struct rcu_data *rdp,
3574 atomic_long_t *stat, unsigned long s)
3576 if (rcu_exp_gp_seq_done(rsp, s)) {
3578 mutex_unlock(&rnp->exp_funnel_mutex);
3580 mutex_unlock(&rdp->exp_funnel_mutex);
3581 /* Ensure test happens before caller kfree(). */
3582 smp_mb__before_atomic(); /* ^^^ */
3583 atomic_long_inc(stat);
3590 * Funnel-lock acquisition for expedited grace periods. Returns a
3591 * pointer to the root rcu_node structure, or NULL if some other
3592 * task did the expedited grace period for us.
3594 static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
3596 struct rcu_data *rdp;
3597 struct rcu_node *rnp0;
3598 struct rcu_node *rnp1 = NULL;
3601 * First try directly acquiring the root lock in order to reduce
3602 * latency in the common case where expedited grace periods are
3603 * rare. We check mutex_is_locked() to avoid pathological levels of
3604 * memory contention on ->exp_funnel_mutex in the heavy-load case.
3606 rnp0 = rcu_get_root(rsp);
3607 if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) {
3608 if (mutex_trylock(&rnp0->exp_funnel_mutex)) {
3609 if (sync_exp_work_done(rsp, rnp0, NULL,
3610 &rsp->expedited_workdone0, s))
3617 * Each pass through the following loop works its way
3618 * up the rcu_node tree, returning if others have done the
3619 * work or otherwise falls through holding the root rnp's
3620 * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure
3621 * can be inexact, as it is just promoting locality and is not
3622 * strictly needed for correctness.
3624 rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
3625 if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s))
3627 mutex_lock(&rdp->exp_funnel_mutex);
3629 for (; rnp0 != NULL; rnp0 = rnp0->parent) {
3630 if (sync_exp_work_done(rsp, rnp1, rdp,
3631 &rsp->expedited_workdone2, s))
3633 mutex_lock(&rnp0->exp_funnel_mutex);
3635 mutex_unlock(&rnp1->exp_funnel_mutex);
3637 mutex_unlock(&rdp->exp_funnel_mutex);
3640 if (sync_exp_work_done(rsp, rnp1, rdp,
3641 &rsp->expedited_workdone3, s))
3646 /* Invoked on each online non-idle CPU for expedited quiescent state. */
3647 static void synchronize_sched_expedited_cpu_stop(void *data)
3649 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
3650 resched_cpu(smp_processor_id());
3654 * Select the nodes that the upcoming expedited grace period needs
3657 static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
3658 smp_call_func_t func)
3661 unsigned long flags;
3663 unsigned long mask_ofl_test;
3664 unsigned long mask_ofl_ipi;
3666 struct rcu_node *rnp;
3668 sync_exp_reset_tree(rsp);
3669 rcu_for_each_leaf_node(rsp, rnp) {
3670 raw_spin_lock_irqsave(&rnp->lock, flags);
3671 smp_mb__after_unlock_lock();
3673 /* Each pass checks a CPU for identity, offline, and idle. */
3675 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
3676 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3677 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3679 if (raw_smp_processor_id() == cpu ||
3680 cpu_is_offline(cpu) ||
3681 !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3682 mask_ofl_test |= rdp->grpmask;
3684 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
3687 * Need to wait for any blocked tasks as well. Note that
3688 * additional blocking tasks will also block the expedited
3689 * GP until such time as the ->expmask bits are cleared.
3691 if (rcu_preempt_has_tasks(rnp))
3692 rnp->exp_tasks = rnp->blkd_tasks.next;
3693 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3695 /* IPI the remaining CPUs for expedited quiescent state. */
3697 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3698 if (!(mask_ofl_ipi & mask))
3700 ret = smp_call_function_single(cpu, func, rsp, 0);
3702 mask_ofl_ipi &= ~mask;
3704 /* Report quiescent states for those that went offline. */
3705 mask_ofl_test |= mask_ofl_ipi;
3707 rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
3711 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
3714 unsigned long jiffies_stall;
3715 unsigned long jiffies_start;
3717 struct rcu_node *rnp;
3718 struct rcu_node *rnp_root = rcu_get_root(rsp);
3721 jiffies_stall = rcu_jiffies_till_stall_check();
3722 jiffies_start = jiffies;
3725 ret = wait_event_interruptible_timeout(
3727 sync_rcu_preempt_exp_done(rnp_root),
3732 /* Hit a signal, disable CPU stall warnings. */
3733 wait_event(rsp->expedited_wq,
3734 sync_rcu_preempt_exp_done(rnp_root));
3737 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3739 rcu_for_each_leaf_node(rsp, rnp) {
3740 (void)rcu_print_task_exp_stall(rnp);
3742 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3743 struct rcu_data *rdp;
3745 if (!(rnp->expmask & mask))
3747 rdp = per_cpu_ptr(rsp->rda, cpu);
3748 pr_cont(" %d-%c%c%c", cpu,
3749 "O."[cpu_online(cpu)],
3750 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
3751 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
3755 pr_cont(" } %lu jiffies s: %lu\n",
3756 jiffies - jiffies_start, rsp->expedited_sequence);
3757 rcu_for_each_leaf_node(rsp, rnp) {
3759 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3760 if (!(rnp->expmask & mask))
3765 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
3770 * synchronize_sched_expedited - Brute-force RCU-sched grace period
3772 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3773 * approach to force the grace period to end quickly. This consumes
3774 * significant time on all CPUs and is unfriendly to real-time workloads,
3775 * so is thus not recommended for any sort of common-case code. In fact,
3776 * if you are using synchronize_sched_expedited() in a loop, please
3777 * restructure your code to batch your updates, and then use a single
3778 * synchronize_sched() instead.
3780 * This implementation can be thought of as an application of sequence
3781 * locking to expedited grace periods, but using the sequence counter to
3782 * determine when someone else has already done the work instead of for
3785 void synchronize_sched_expedited(void)
3788 struct rcu_node *rnp;
3789 struct rcu_state *rsp = &rcu_sched_state;
3791 /* Take a snapshot of the sequence number. */
3792 s = rcu_exp_gp_seq_snap(rsp);
3794 rnp = exp_funnel_lock(rsp, s);
3796 return; /* Someone else did our work for us. */
3798 rcu_exp_gp_seq_start(rsp);
3799 sync_rcu_exp_select_cpus(rsp, synchronize_sched_expedited_cpu_stop);
3800 synchronize_sched_expedited_wait(rsp);
3802 rcu_exp_gp_seq_end(rsp);
3803 mutex_unlock(&rnp->exp_funnel_mutex);
3805 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3808 * Check to see if there is any immediate RCU-related work to be done
3809 * by the current CPU, for the specified type of RCU, returning 1 if so.
3810 * The checks are in order of increasing expense: checks that can be
3811 * carried out against CPU-local state are performed first. However,
3812 * we must check for CPU stalls first, else we might not get a chance.
3814 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3816 struct rcu_node *rnp = rdp->mynode;
3818 rdp->n_rcu_pending++;
3820 /* Check for CPU stalls, if enabled. */
3821 check_cpu_stall(rsp, rdp);
3823 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3824 if (rcu_nohz_full_cpu(rsp))
3827 /* Is the RCU core waiting for a quiescent state from this CPU? */
3828 if (rcu_scheduler_fully_active &&
3829 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3830 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3831 rdp->n_rp_core_needs_qs++;
3832 } else if (rdp->core_needs_qs &&
3833 (!rdp->cpu_no_qs.b.norm ||
3834 rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3835 rdp->n_rp_report_qs++;
3839 /* Does this CPU have callbacks ready to invoke? */
3840 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3841 rdp->n_rp_cb_ready++;
3845 /* Has RCU gone idle with this CPU needing another grace period? */
3846 if (cpu_needs_another_gp(rsp, rdp)) {
3847 rdp->n_rp_cpu_needs_gp++;
3851 /* Has another RCU grace period completed? */
3852 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3853 rdp->n_rp_gp_completed++;
3857 /* Has a new RCU grace period started? */
3858 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3859 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3860 rdp->n_rp_gp_started++;
3864 /* Does this CPU need a deferred NOCB wakeup? */
3865 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3866 rdp->n_rp_nocb_defer_wakeup++;
3871 rdp->n_rp_need_nothing++;
3876 * Check to see if there is any immediate RCU-related work to be done
3877 * by the current CPU, returning 1 if so. This function is part of the
3878 * RCU implementation; it is -not- an exported member of the RCU API.
3880 static int rcu_pending(void)
3882 struct rcu_state *rsp;
3884 for_each_rcu_flavor(rsp)
3885 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3891 * Return true if the specified CPU has any callback. If all_lazy is
3892 * non-NULL, store an indication of whether all callbacks are lazy.
3893 * (If there are no callbacks, all of them are deemed to be lazy.)
3895 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3899 struct rcu_data *rdp;
3900 struct rcu_state *rsp;
3902 for_each_rcu_flavor(rsp) {
3903 rdp = this_cpu_ptr(rsp->rda);
3907 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3918 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3919 * the compiler is expected to optimize this away.
3921 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3922 int cpu, unsigned long done)
3924 trace_rcu_barrier(rsp->name, s, cpu,
3925 atomic_read(&rsp->barrier_cpu_count), done);
3929 * RCU callback function for _rcu_barrier(). If we are last, wake
3930 * up the task executing _rcu_barrier().
3932 static void rcu_barrier_callback(struct rcu_head *rhp)
3934 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3935 struct rcu_state *rsp = rdp->rsp;
3937 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3938 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
3939 complete(&rsp->barrier_completion);
3941 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
3946 * Called with preemption disabled, and from cross-cpu IRQ context.
3948 static void rcu_barrier_func(void *type)
3950 struct rcu_state *rsp = type;
3951 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3953 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
3954 atomic_inc(&rsp->barrier_cpu_count);
3955 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3959 * Orchestrate the specified type of RCU barrier, waiting for all
3960 * RCU callbacks of the specified type to complete.
3962 static void _rcu_barrier(struct rcu_state *rsp)
3965 struct rcu_data *rdp;
3966 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3968 _rcu_barrier_trace(rsp, "Begin", -1, s);
3970 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3971 mutex_lock(&rsp->barrier_mutex);
3973 /* Did someone else do our work for us? */
3974 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3975 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
3976 smp_mb(); /* caller's subsequent code after above check. */
3977 mutex_unlock(&rsp->barrier_mutex);
3981 /* Mark the start of the barrier operation. */
3982 rcu_seq_start(&rsp->barrier_sequence);
3983 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
3986 * Initialize the count to one rather than to zero in order to
3987 * avoid a too-soon return to zero in case of a short grace period
3988 * (or preemption of this task). Exclude CPU-hotplug operations
3989 * to ensure that no offline CPU has callbacks queued.
3991 init_completion(&rsp->barrier_completion);
3992 atomic_set(&rsp->barrier_cpu_count, 1);
3996 * Force each CPU with callbacks to register a new callback.
3997 * When that callback is invoked, we will know that all of the
3998 * corresponding CPU's preceding callbacks have been invoked.
4000 for_each_possible_cpu(cpu) {
4001 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
4003 rdp = per_cpu_ptr(rsp->rda, cpu);
4004 if (rcu_is_nocb_cpu(cpu)) {
4005 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
4006 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
4007 rsp->barrier_sequence);
4009 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
4010 rsp->barrier_sequence);
4011 smp_mb__before_atomic();
4012 atomic_inc(&rsp->barrier_cpu_count);
4013 __call_rcu(&rdp->barrier_head,
4014 rcu_barrier_callback, rsp, cpu, 0);
4016 } else if (READ_ONCE(rdp->qlen)) {
4017 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
4018 rsp->barrier_sequence);
4019 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
4021 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
4022 rsp->barrier_sequence);
4028 * Now that we have an rcu_barrier_callback() callback on each
4029 * CPU, and thus each counted, remove the initial count.
4031 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
4032 complete(&rsp->barrier_completion);
4034 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4035 wait_for_completion(&rsp->barrier_completion);
4037 /* Mark the end of the barrier operation. */
4038 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
4039 rcu_seq_end(&rsp->barrier_sequence);
4041 /* Other rcu_barrier() invocations can now safely proceed. */
4042 mutex_unlock(&rsp->barrier_mutex);
4046 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4048 void rcu_barrier_bh(void)
4050 _rcu_barrier(&rcu_bh_state);
4052 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
4055 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4057 void rcu_barrier_sched(void)
4059 _rcu_barrier(&rcu_sched_state);
4061 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
4064 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4065 * first CPU in a given leaf rcu_node structure coming online. The caller
4066 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4069 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4072 struct rcu_node *rnp = rnp_leaf;
4075 mask = rnp->grpmask;
4079 raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */
4080 rnp->qsmaskinit |= mask;
4081 raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
4086 * Do boot-time initialization of a CPU's per-CPU RCU data.
4089 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
4091 unsigned long flags;
4092 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4093 struct rcu_node *rnp = rcu_get_root(rsp);
4095 /* Set up local state, ensuring consistent view of global state. */
4096 raw_spin_lock_irqsave(&rnp->lock, flags);
4097 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
4098 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
4099 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
4100 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
4103 mutex_init(&rdp->exp_funnel_mutex);
4104 rcu_boot_init_nocb_percpu_data(rdp);
4105 raw_spin_unlock_irqrestore(&rnp->lock, flags);
4109 * Initialize a CPU's per-CPU RCU data. Note that only one online or
4110 * offline event can be happening at a given time. Note also that we
4111 * can accept some slop in the rsp->completed access due to the fact
4112 * that this CPU cannot possibly have any RCU callbacks in flight yet.
4115 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
4117 unsigned long flags;
4119 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4120 struct rcu_node *rnp = rcu_get_root(rsp);
4122 /* Set up local state, ensuring consistent view of global state. */
4123 raw_spin_lock_irqsave(&rnp->lock, flags);
4124 rdp->qlen_last_fqs_check = 0;
4125 rdp->n_force_qs_snap = rsp->n_force_qs;
4126 rdp->blimit = blimit;
4128 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
4129 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
4130 rcu_sysidle_init_percpu_data(rdp->dynticks);
4131 atomic_set(&rdp->dynticks->dynticks,
4132 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
4133 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
4136 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4137 * propagation up the rcu_node tree will happen at the beginning
4138 * of the next grace period.
4141 mask = rdp->grpmask;
4142 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
4143 smp_mb__after_unlock_lock();
4144 rnp->qsmaskinitnext |= mask;
4145 rnp->expmaskinitnext |= mask;
4146 if (!rdp->beenonline)
4147 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
4148 rdp->beenonline = true; /* We have now been online. */
4149 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
4150 rdp->completed = rnp->completed;
4151 rdp->cpu_no_qs.b.norm = true;
4152 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
4153 rdp->core_needs_qs = false;
4154 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
4155 raw_spin_unlock_irqrestore(&rnp->lock, flags);
4158 static void rcu_prepare_cpu(int cpu)
4160 struct rcu_state *rsp;
4162 for_each_rcu_flavor(rsp)
4163 rcu_init_percpu_data(cpu, rsp);
4167 * Handle CPU online/offline notification events.
4169 int rcu_cpu_notify(struct notifier_block *self,
4170 unsigned long action, void *hcpu)
4172 long cpu = (long)hcpu;
4173 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
4174 struct rcu_node *rnp = rdp->mynode;
4175 struct rcu_state *rsp;
4178 case CPU_UP_PREPARE:
4179 case CPU_UP_PREPARE_FROZEN:
4180 rcu_prepare_cpu(cpu);
4181 rcu_prepare_kthreads(cpu);
4182 rcu_spawn_all_nocb_kthreads(cpu);
4185 case CPU_DOWN_FAILED:
4186 rcu_boost_kthread_setaffinity(rnp, -1);
4188 case CPU_DOWN_PREPARE:
4189 rcu_boost_kthread_setaffinity(rnp, cpu);
4192 case CPU_DYING_FROZEN:
4193 for_each_rcu_flavor(rsp)
4194 rcu_cleanup_dying_cpu(rsp);
4196 case CPU_DYING_IDLE:
4197 /* QS for any half-done expedited RCU-sched GP. */
4200 for_each_rcu_flavor(rsp) {
4201 rcu_cleanup_dying_idle_cpu(cpu, rsp);
4205 case CPU_DEAD_FROZEN:
4206 case CPU_UP_CANCELED:
4207 case CPU_UP_CANCELED_FROZEN:
4208 for_each_rcu_flavor(rsp) {
4209 rcu_cleanup_dead_cpu(cpu, rsp);
4210 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
4219 static int rcu_pm_notify(struct notifier_block *self,
4220 unsigned long action, void *hcpu)
4223 case PM_HIBERNATION_PREPARE:
4224 case PM_SUSPEND_PREPARE:
4225 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4228 case PM_POST_HIBERNATION:
4229 case PM_POST_SUSPEND:
4230 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4231 rcu_unexpedite_gp();
4240 * Spawn the kthreads that handle each RCU flavor's grace periods.
4242 static int __init rcu_spawn_gp_kthread(void)
4244 unsigned long flags;
4245 int kthread_prio_in = kthread_prio;
4246 struct rcu_node *rnp;
4247 struct rcu_state *rsp;
4248 struct sched_param sp;
4249 struct task_struct *t;
4251 /* Force priority into range. */
4252 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4254 else if (kthread_prio < 0)
4256 else if (kthread_prio > 99)
4258 if (kthread_prio != kthread_prio_in)
4259 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4260 kthread_prio, kthread_prio_in);
4262 rcu_scheduler_fully_active = 1;
4263 for_each_rcu_flavor(rsp) {
4264 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4266 rnp = rcu_get_root(rsp);
4267 raw_spin_lock_irqsave(&rnp->lock, flags);
4268 rsp->gp_kthread = t;
4270 sp.sched_priority = kthread_prio;
4271 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4274 raw_spin_unlock_irqrestore(&rnp->lock, flags);
4276 rcu_spawn_nocb_kthreads();
4277 rcu_spawn_boost_kthreads();
4280 early_initcall(rcu_spawn_gp_kthread);
4283 * This function is invoked towards the end of the scheduler's initialization
4284 * process. Before this is called, the idle task might contain
4285 * RCU read-side critical sections (during which time, this idle
4286 * task is booting the system). After this function is called, the
4287 * idle tasks are prohibited from containing RCU read-side critical
4288 * sections. This function also enables RCU lockdep checking.
4290 void rcu_scheduler_starting(void)
4292 WARN_ON(num_online_cpus() != 1);
4293 WARN_ON(nr_context_switches() > 0);
4294 rcu_scheduler_active = 1;
4298 * Compute the per-level fanout, either using the exact fanout specified
4299 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4301 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4305 if (rcu_fanout_exact) {
4306 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4307 for (i = rcu_num_lvls - 2; i >= 0; i--)
4308 levelspread[i] = RCU_FANOUT;
4314 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4316 levelspread[i] = (cprv + ccur - 1) / ccur;
4323 * Helper function for rcu_init() that initializes one rcu_state structure.
4325 static void __init rcu_init_one(struct rcu_state *rsp,
4326 struct rcu_data __percpu *rda)
4328 static const char * const buf[] = RCU_NODE_NAME_INIT;
4329 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4330 static const char * const exp[] = RCU_EXP_NAME_INIT;
4331 static u8 fl_mask = 0x1;
4333 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4334 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4338 struct rcu_node *rnp;
4340 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4342 /* Silence gcc 4.8 false positive about array index out of range. */
4343 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4344 panic("rcu_init_one: rcu_num_lvls out of range");
4346 /* Initialize the level-tracking arrays. */
4348 for (i = 0; i < rcu_num_lvls; i++)
4349 levelcnt[i] = num_rcu_lvl[i];
4350 for (i = 1; i < rcu_num_lvls; i++)
4351 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4352 rcu_init_levelspread(levelspread, levelcnt);
4353 rsp->flavor_mask = fl_mask;
4356 /* Initialize the elements themselves, starting from the leaves. */
4358 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4359 cpustride *= levelspread[i];
4360 rnp = rsp->level[i];
4361 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4362 raw_spin_lock_init(&rnp->lock);
4363 lockdep_set_class_and_name(&rnp->lock,
4364 &rcu_node_class[i], buf[i]);
4365 raw_spin_lock_init(&rnp->fqslock);
4366 lockdep_set_class_and_name(&rnp->fqslock,
4367 &rcu_fqs_class[i], fqs[i]);
4368 rnp->gpnum = rsp->gpnum;
4369 rnp->completed = rsp->completed;
4371 rnp->qsmaskinit = 0;
4372 rnp->grplo = j * cpustride;
4373 rnp->grphi = (j + 1) * cpustride - 1;
4374 if (rnp->grphi >= nr_cpu_ids)
4375 rnp->grphi = nr_cpu_ids - 1;
4381 rnp->grpnum = j % levelspread[i - 1];
4382 rnp->grpmask = 1UL << rnp->grpnum;
4383 rnp->parent = rsp->level[i - 1] +
4384 j / levelspread[i - 1];
4387 INIT_LIST_HEAD(&rnp->blkd_tasks);
4388 rcu_init_one_nocb(rnp);
4389 mutex_init(&rnp->exp_funnel_mutex);
4390 lockdep_set_class_and_name(&rnp->exp_funnel_mutex,
4391 &rcu_exp_class[i], exp[i]);
4395 init_waitqueue_head(&rsp->gp_wq);
4396 init_waitqueue_head(&rsp->expedited_wq);
4397 rnp = rsp->level[rcu_num_lvls - 1];
4398 for_each_possible_cpu(i) {
4399 while (i > rnp->grphi)
4401 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4402 rcu_boot_init_percpu_data(i, rsp);
4404 list_add(&rsp->flavors, &rcu_struct_flavors);
4408 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4409 * replace the definitions in tree.h because those are needed to size
4410 * the ->node array in the rcu_state structure.
4412 static void __init rcu_init_geometry(void)
4416 int rcu_capacity[RCU_NUM_LVLS];
4419 * Initialize any unspecified boot parameters.
4420 * The default values of jiffies_till_first_fqs and
4421 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4422 * value, which is a function of HZ, then adding one for each
4423 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4425 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4426 if (jiffies_till_first_fqs == ULONG_MAX)
4427 jiffies_till_first_fqs = d;
4428 if (jiffies_till_next_fqs == ULONG_MAX)
4429 jiffies_till_next_fqs = d;
4431 /* If the compile-time values are accurate, just leave. */
4432 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4433 nr_cpu_ids == NR_CPUS)
4435 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4436 rcu_fanout_leaf, nr_cpu_ids);
4439 * The boot-time rcu_fanout_leaf parameter is only permitted
4440 * to increase the leaf-level fanout, not decrease it. Of course,
4441 * the leaf-level fanout cannot exceed the number of bits in
4442 * the rcu_node masks. Complain and fall back to the compile-
4443 * time values if these limits are exceeded.
4445 if (rcu_fanout_leaf < RCU_FANOUT_LEAF ||
4446 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4447 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4453 * Compute number of nodes that can be handled an rcu_node tree
4454 * with the given number of levels.
4456 rcu_capacity[0] = rcu_fanout_leaf;
4457 for (i = 1; i < RCU_NUM_LVLS; i++)
4458 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4461 * The tree must be able to accommodate the configured number of CPUs.
4462 * If this limit is exceeded than we have a serious problem elsewhere.
4464 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1])
4465 panic("rcu_init_geometry: rcu_capacity[] is too small");
4467 /* Calculate the number of levels in the tree. */
4468 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4470 rcu_num_lvls = i + 1;
4472 /* Calculate the number of rcu_nodes at each level of the tree. */
4473 for (i = 0; i < rcu_num_lvls; i++) {
4474 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4475 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4478 /* Calculate the total number of rcu_node structures. */
4480 for (i = 0; i < rcu_num_lvls; i++)
4481 rcu_num_nodes += num_rcu_lvl[i];
4485 * Dump out the structure of the rcu_node combining tree associated
4486 * with the rcu_state structure referenced by rsp.
4488 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4491 struct rcu_node *rnp;
4493 pr_info("rcu_node tree layout dump\n");
4495 rcu_for_each_node_breadth_first(rsp, rnp) {
4496 if (rnp->level != level) {
4501 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4506 void __init rcu_init(void)
4510 rcu_early_boot_tests();
4512 rcu_bootup_announce();
4513 rcu_init_geometry();
4514 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
4515 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
4517 rcu_dump_rcu_node_tree(&rcu_sched_state);
4518 __rcu_init_preempt();
4519 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4522 * We don't need protection against CPU-hotplug here because
4523 * this is called early in boot, before either interrupts
4524 * or the scheduler are operational.
4526 cpu_notifier(rcu_cpu_notify, 0);
4527 pm_notifier(rcu_pm_notify, 0);
4528 for_each_online_cpu(cpu)
4529 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
4532 #include "tree_plugin.h"