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/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
53 #include <linux/delay.h>
54 #include <linux/stop_machine.h>
55 #include <linux/random.h>
56 #include <linux/trace_events.h>
57 #include <linux/suspend.h>
62 #ifdef MODULE_PARAM_PREFIX
63 #undef MODULE_PARAM_PREFIX
65 #define MODULE_PARAM_PREFIX "rcutree."
67 /* Data structures. */
70 * In order to export the rcu_state name to the tracing tools, it
71 * needs to be added in the __tracepoint_string section.
72 * This requires defining a separate variable tp_<sname>_varname
73 * that points to the string being used, and this will allow
74 * the tracing userspace tools to be able to decipher the string
75 * address to the matching string.
78 # define DEFINE_RCU_TPS(sname) \
79 static char sname##_varname[] = #sname; \
80 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
81 # define RCU_STATE_NAME(sname) sname##_varname
83 # define DEFINE_RCU_TPS(sname)
84 # define RCU_STATE_NAME(sname) __stringify(sname)
87 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
88 DEFINE_RCU_TPS(sname) \
89 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
90 struct rcu_state sname##_state = { \
91 .level = { &sname##_state.node[0] }, \
92 .rda = &sname##_data, \
94 .gp_state = RCU_GP_IDLE, \
95 .gpnum = 0UL - 300UL, \
96 .completed = 0UL - 300UL, \
97 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
98 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
99 .orphan_donetail = &sname##_state.orphan_donelist, \
100 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
101 .name = RCU_STATE_NAME(sname), \
103 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
104 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
107 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
110 static struct rcu_state *const rcu_state_p;
111 LIST_HEAD(rcu_struct_flavors);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree;
115 module_param(dump_tree, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact;
118 module_param(rcu_fanout_exact, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121 module_param(rcu_fanout_leaf, int, 0444);
122 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123 /* Number of rcu_nodes at specified level. */
124 static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126 /* panic() on RCU Stall sysctl. */
127 int sysctl_panic_on_rcu_stall __read_mostly;
130 * The rcu_scheduler_active variable is initialized to the value
131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
132 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
133 * RCU can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_rcu() to a simple barrier(). When this variable
135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
136 * to detect real grace periods. This variable is also used to suppress
137 * boot-time false positives from lockdep-RCU error checking. Finally, it
138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
139 * is fully initialized, including all of its kthreads having been spawned.
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);
165 static void sync_sched_exp_online_cleanup(int cpu);
167 /* rcuc/rcub kthread realtime priority */
168 #ifdef CONFIG_RCU_KTHREAD_PRIO
169 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
170 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
171 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
172 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
173 module_param(kthread_prio, int, 0644);
175 /* Delay in jiffies for grace-period initialization delays, debug only. */
177 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
178 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
179 module_param(gp_preinit_delay, int, 0644);
180 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
181 static const int gp_preinit_delay;
182 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
184 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
185 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
186 module_param(gp_init_delay, int, 0644);
187 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
188 static const int gp_init_delay;
189 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
191 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
192 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
193 module_param(gp_cleanup_delay, int, 0644);
194 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
195 static const int gp_cleanup_delay;
196 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
199 * Number of grace periods between delays, normalized by the duration of
200 * the delay. The longer the the delay, the more the grace periods between
201 * each delay. The reason for this normalization is that it means that,
202 * for non-zero delays, the overall slowdown of grace periods is constant
203 * regardless of the duration of the delay. This arrangement balances
204 * the need for long delays to increase some race probabilities with the
205 * need for fast grace periods to increase other race probabilities.
207 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
210 * Track the rcutorture test sequence number and the update version
211 * number within a given test. The rcutorture_testseq is incremented
212 * on every rcutorture module load and unload, so has an odd value
213 * when a test is running. The rcutorture_vernum is set to zero
214 * when rcutorture starts and is incremented on each rcutorture update.
215 * These variables enable correlating rcutorture output with the
216 * RCU tracing information.
218 unsigned long rcutorture_testseq;
219 unsigned long rcutorture_vernum;
222 * Compute the mask of online CPUs for the specified rcu_node structure.
223 * This will not be stable unless the rcu_node structure's ->lock is
224 * held, but the bit corresponding to the current CPU will be stable
227 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
229 return READ_ONCE(rnp->qsmaskinitnext);
233 * Return true if an RCU grace period is in progress. The READ_ONCE()s
234 * permit this function to be invoked without holding the root rcu_node
235 * structure's ->lock, but of course results can be subject to change.
237 static int rcu_gp_in_progress(struct rcu_state *rsp)
239 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
243 * Note a quiescent state. Because we do not need to know
244 * how many quiescent states passed, just if there was at least
245 * one since the start of the grace period, this just sets a flag.
246 * The caller must have disabled preemption.
248 void rcu_sched_qs(void)
250 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
252 trace_rcu_grace_period(TPS("rcu_sched"),
253 __this_cpu_read(rcu_sched_data.gpnum),
255 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
256 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
258 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
259 rcu_report_exp_rdp(&rcu_sched_state,
260 this_cpu_ptr(&rcu_sched_data), true);
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 */
285 * Record entry into an extended quiescent state. This is only to be
286 * called when not already in an extended quiescent state.
288 static void rcu_dynticks_eqs_enter(void)
290 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
294 * CPUs seeing atomic_inc_return() must see prior RCU read-side
295 * critical sections, and we also must force ordering with the
298 special = atomic_inc_return(&rdtp->dynticks);
299 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && special & 0x1);
303 * Record exit from an extended quiescent state. This is only to be
304 * called from an extended quiescent state.
306 static void rcu_dynticks_eqs_exit(void)
308 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
312 * CPUs seeing atomic_inc_return() must see prior idle sojourns,
313 * and we also must force ordering with the next RCU read-side
316 special = atomic_inc_return(&rdtp->dynticks);
317 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !(special & 0x1));
321 * Reset the current CPU's ->dynticks counter to indicate that the
322 * newly onlined CPU is no longer in an extended quiescent state.
323 * This will either leave the counter unchanged, or increment it
324 * to the next non-quiescent value.
326 * The non-atomic test/increment sequence works because the upper bits
327 * of the ->dynticks counter are manipulated only by the corresponding CPU,
328 * or when the corresponding CPU is offline.
330 static void rcu_dynticks_eqs_online(void)
332 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
334 if (atomic_read(&rdtp->dynticks) & 0x1)
336 atomic_add(0x1, &rdtp->dynticks);
340 * Is the current CPU in an extended quiescent state?
342 * No ordering, as we are sampling CPU-local information.
344 bool rcu_dynticks_curr_cpu_in_eqs(void)
346 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
348 return !(atomic_read(&rdtp->dynticks) & 0x1);
352 * Snapshot the ->dynticks counter with full ordering so as to allow
353 * stable comparison of this counter with past and future snapshots.
355 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
357 int snap = atomic_add_return(0, &rdtp->dynticks);
363 * Return true if the snapshot returned from rcu_dynticks_snap()
364 * indicates that RCU is in an extended quiescent state.
366 static bool rcu_dynticks_in_eqs(int snap)
368 return !(snap & 0x1);
372 * Return true if the CPU corresponding to the specified rcu_dynticks
373 * structure has spent some time in an extended quiescent state since
374 * rcu_dynticks_snap() returned the specified snapshot.
376 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
378 return snap != rcu_dynticks_snap(rdtp);
382 * Do a double-increment of the ->dynticks counter to emulate a
383 * momentary idle-CPU quiescent state.
385 static void rcu_dynticks_momentary_idle(void)
387 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
388 int special = atomic_add_return(2, &rdtp->dynticks);
390 /* It is illegal to call this from idle state. */
391 WARN_ON_ONCE(!(special & 0x1));
394 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
395 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
398 * Let the RCU core know that this CPU has gone through the scheduler,
399 * which is a quiescent state. This is called when the need for a
400 * quiescent state is urgent, so we burn an atomic operation and full
401 * memory barriers to let the RCU core know about it, regardless of what
402 * this CPU might (or might not) do in the near future.
404 * We inform the RCU core by emulating a zero-duration dyntick-idle
405 * period, which we in turn do by incrementing the ->dynticks counter
408 * The caller must have disabled interrupts.
410 static void rcu_momentary_dyntick_idle(void)
412 struct rcu_data *rdp;
414 struct rcu_state *rsp;
417 * Yes, we can lose flag-setting operations. This is OK, because
418 * the flag will be set again after some delay.
420 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
421 raw_cpu_write(rcu_sched_qs_mask, 0);
423 /* Find the flavor that needs a quiescent state. */
424 for_each_rcu_flavor(rsp) {
425 rdp = raw_cpu_ptr(rsp->rda);
426 if (!(resched_mask & rsp->flavor_mask))
428 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
429 if (READ_ONCE(rdp->mynode->completed) !=
430 READ_ONCE(rdp->cond_resched_completed))
434 * Pretend to be momentarily idle for the quiescent state.
435 * This allows the grace-period kthread to record the
436 * quiescent state, with no need for this CPU to do anything
439 rcu_dynticks_momentary_idle();
445 * Note a context switch. This is a quiescent state for RCU-sched,
446 * and requires special handling for preemptible RCU.
447 * The caller must have disabled interrupts.
449 void rcu_note_context_switch(void)
451 barrier(); /* Avoid RCU read-side critical sections leaking down. */
452 trace_rcu_utilization(TPS("Start context switch"));
454 rcu_preempt_note_context_switch();
455 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
456 rcu_momentary_dyntick_idle();
457 trace_rcu_utilization(TPS("End context switch"));
458 barrier(); /* Avoid RCU read-side critical sections leaking up. */
460 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
463 * Register a quiescent state for all RCU flavors. If there is an
464 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
465 * dyntick-idle quiescent state visible to other CPUs (but only for those
466 * RCU flavors in desperate need of a quiescent state, which will normally
467 * be none of them). Either way, do a lightweight quiescent state for
470 * The barrier() calls are redundant in the common case when this is
471 * called externally, but just in case this is called from within this
475 void rcu_all_qs(void)
479 barrier(); /* Avoid RCU read-side critical sections leaking down. */
480 if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) {
481 local_irq_save(flags);
482 rcu_momentary_dyntick_idle();
483 local_irq_restore(flags);
485 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))) {
487 * Yes, we just checked a per-CPU variable with preemption
488 * enabled, so we might be migrated to some other CPU at
489 * this point. That is OK because in that case, the
490 * migration will supply the needed quiescent state.
491 * We might end up needlessly disabling preemption and
492 * invoking rcu_sched_qs() on the destination CPU, but
493 * the probability and cost are both quite low, so this
494 * should not be a problem in practice.
500 this_cpu_inc(rcu_qs_ctr);
501 barrier(); /* Avoid RCU read-side critical sections leaking up. */
503 EXPORT_SYMBOL_GPL(rcu_all_qs);
505 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
506 static long qhimark = 10000; /* If this many pending, ignore blimit. */
507 static long qlowmark = 100; /* Once only this many pending, use blimit. */
509 module_param(blimit, long, 0444);
510 module_param(qhimark, long, 0444);
511 module_param(qlowmark, long, 0444);
513 static ulong jiffies_till_first_fqs = ULONG_MAX;
514 static ulong jiffies_till_next_fqs = ULONG_MAX;
515 static bool rcu_kick_kthreads;
517 module_param(jiffies_till_first_fqs, ulong, 0644);
518 module_param(jiffies_till_next_fqs, ulong, 0644);
519 module_param(rcu_kick_kthreads, bool, 0644);
522 * How long the grace period must be before we start recruiting
523 * quiescent-state help from rcu_note_context_switch().
525 static ulong jiffies_till_sched_qs = HZ / 20;
526 module_param(jiffies_till_sched_qs, ulong, 0644);
528 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
529 struct rcu_data *rdp);
530 static void force_qs_rnp(struct rcu_state *rsp,
531 int (*f)(struct rcu_data *rsp, bool *isidle,
532 unsigned long *maxj),
533 bool *isidle, unsigned long *maxj);
534 static void force_quiescent_state(struct rcu_state *rsp);
535 static int rcu_pending(void);
538 * Return the number of RCU batches started thus far for debug & stats.
540 unsigned long rcu_batches_started(void)
542 return rcu_state_p->gpnum;
544 EXPORT_SYMBOL_GPL(rcu_batches_started);
547 * Return the number of RCU-sched batches started thus far for debug & stats.
549 unsigned long rcu_batches_started_sched(void)
551 return rcu_sched_state.gpnum;
553 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
556 * Return the number of RCU BH batches started thus far for debug & stats.
558 unsigned long rcu_batches_started_bh(void)
560 return rcu_bh_state.gpnum;
562 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
565 * Return the number of RCU batches completed thus far for debug & stats.
567 unsigned long rcu_batches_completed(void)
569 return rcu_state_p->completed;
571 EXPORT_SYMBOL_GPL(rcu_batches_completed);
574 * Return the number of RCU-sched batches completed thus far for debug & stats.
576 unsigned long rcu_batches_completed_sched(void)
578 return rcu_sched_state.completed;
580 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
583 * Return the number of RCU BH batches completed thus far for debug & stats.
585 unsigned long rcu_batches_completed_bh(void)
587 return rcu_bh_state.completed;
589 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
592 * Return the number of RCU expedited batches completed thus far for
593 * debug & stats. Odd numbers mean that a batch is in progress, even
594 * numbers mean idle. The value returned will thus be roughly double
595 * the cumulative batches since boot.
597 unsigned long rcu_exp_batches_completed(void)
599 return rcu_state_p->expedited_sequence;
601 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
604 * Return the number of RCU-sched expedited batches completed thus far
605 * for debug & stats. Similar to rcu_exp_batches_completed().
607 unsigned long rcu_exp_batches_completed_sched(void)
609 return rcu_sched_state.expedited_sequence;
611 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
614 * Force a quiescent state.
616 void rcu_force_quiescent_state(void)
618 force_quiescent_state(rcu_state_p);
620 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
623 * Force a quiescent state for RCU BH.
625 void rcu_bh_force_quiescent_state(void)
627 force_quiescent_state(&rcu_bh_state);
629 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
632 * Force a quiescent state for RCU-sched.
634 void rcu_sched_force_quiescent_state(void)
636 force_quiescent_state(&rcu_sched_state);
638 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
641 * Show the state of the grace-period kthreads.
643 void show_rcu_gp_kthreads(void)
645 struct rcu_state *rsp;
647 for_each_rcu_flavor(rsp) {
648 pr_info("%s: wait state: %d ->state: %#lx\n",
649 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
650 /* sched_show_task(rsp->gp_kthread); */
653 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
656 * Record the number of times rcutorture tests have been initiated and
657 * terminated. This information allows the debugfs tracing stats to be
658 * correlated to the rcutorture messages, even when the rcutorture module
659 * is being repeatedly loaded and unloaded. In other words, we cannot
660 * store this state in rcutorture itself.
662 void rcutorture_record_test_transition(void)
664 rcutorture_testseq++;
665 rcutorture_vernum = 0;
667 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
670 * Send along grace-period-related data for rcutorture diagnostics.
672 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
673 unsigned long *gpnum, unsigned long *completed)
675 struct rcu_state *rsp = NULL;
684 case RCU_SCHED_FLAVOR:
685 rsp = &rcu_sched_state;
691 *flags = READ_ONCE(rsp->gp_flags);
692 *gpnum = READ_ONCE(rsp->gpnum);
693 *completed = READ_ONCE(rsp->completed);
700 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
703 * Record the number of writer passes through the current rcutorture test.
704 * This is also used to correlate debugfs tracing stats with the rcutorture
707 void rcutorture_record_progress(unsigned long vernum)
711 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
714 * Does the CPU have callbacks ready to be invoked?
717 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
719 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
720 rdp->nxttail[RCU_NEXT_TAIL] != NULL;
724 * Return the root node of the specified rcu_state structure.
726 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
728 return &rsp->node[0];
732 * Is there any need for future grace periods?
733 * Interrupts must be disabled. If the caller does not hold the root
734 * rnp_node structure's ->lock, the results are advisory only.
736 static int rcu_future_needs_gp(struct rcu_state *rsp)
738 struct rcu_node *rnp = rcu_get_root(rsp);
739 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
740 int *fp = &rnp->need_future_gp[idx];
742 return READ_ONCE(*fp);
746 * Does the current CPU require a not-yet-started grace period?
747 * The caller must have disabled interrupts to prevent races with
748 * normal callback registry.
751 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
755 if (rcu_gp_in_progress(rsp))
756 return false; /* No, a grace period is already in progress. */
757 if (rcu_future_needs_gp(rsp))
758 return true; /* Yes, a no-CBs CPU needs one. */
759 if (!rdp->nxttail[RCU_NEXT_TAIL])
760 return false; /* No, this is a no-CBs (or offline) CPU. */
761 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
762 return true; /* Yes, CPU has newly registered callbacks. */
763 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
764 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
765 ULONG_CMP_LT(READ_ONCE(rsp->completed),
766 rdp->nxtcompleted[i]))
767 return true; /* Yes, CBs for future grace period. */
768 return false; /* No grace period needed. */
772 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
774 * If the new value of the ->dynticks_nesting counter now is zero,
775 * we really have entered idle, and must do the appropriate accounting.
776 * The caller must have disabled interrupts.
778 static void rcu_eqs_enter_common(long long oldval, bool user)
780 struct rcu_state *rsp;
781 struct rcu_data *rdp;
782 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
784 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
785 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
786 !user && !is_idle_task(current)) {
787 struct task_struct *idle __maybe_unused =
788 idle_task(smp_processor_id());
790 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
791 rcu_ftrace_dump(DUMP_ORIG);
792 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
793 current->pid, current->comm,
794 idle->pid, idle->comm); /* must be idle task! */
796 for_each_rcu_flavor(rsp) {
797 rdp = this_cpu_ptr(rsp->rda);
798 do_nocb_deferred_wakeup(rdp);
800 rcu_prepare_for_idle();
801 rcu_dynticks_eqs_enter();
802 rcu_dynticks_task_enter();
805 * It is illegal to enter an extended quiescent state while
806 * in an RCU read-side critical section.
808 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
809 "Illegal idle entry in RCU read-side critical section.");
810 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
811 "Illegal idle entry in RCU-bh read-side critical section.");
812 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
813 "Illegal idle entry in RCU-sched read-side critical section.");
817 * Enter an RCU extended quiescent state, which can be either the
818 * idle loop or adaptive-tickless usermode execution.
820 static void rcu_eqs_enter(bool user)
823 struct rcu_dynticks *rdtp;
825 rdtp = this_cpu_ptr(&rcu_dynticks);
826 oldval = rdtp->dynticks_nesting;
827 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
828 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
829 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
830 rdtp->dynticks_nesting = 0;
831 rcu_eqs_enter_common(oldval, user);
833 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
838 * rcu_idle_enter - inform RCU that current CPU is entering idle
840 * Enter idle mode, in other words, -leave- the mode in which RCU
841 * read-side critical sections can occur. (Though RCU read-side
842 * critical sections can occur in irq handlers in idle, a possibility
843 * handled by irq_enter() and irq_exit().)
845 * We crowbar the ->dynticks_nesting field to zero to allow for
846 * the possibility of usermode upcalls having messed up our count
847 * of interrupt nesting level during the prior busy period.
849 void rcu_idle_enter(void)
853 local_irq_save(flags);
854 rcu_eqs_enter(false);
855 rcu_sysidle_enter(0);
856 local_irq_restore(flags);
858 EXPORT_SYMBOL_GPL(rcu_idle_enter);
860 #ifdef CONFIG_NO_HZ_FULL
862 * rcu_user_enter - inform RCU that we are resuming userspace.
864 * Enter RCU idle mode right before resuming userspace. No use of RCU
865 * is permitted between this call and rcu_user_exit(). This way the
866 * CPU doesn't need to maintain the tick for RCU maintenance purposes
867 * when the CPU runs in userspace.
869 void rcu_user_enter(void)
873 #endif /* CONFIG_NO_HZ_FULL */
876 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
878 * Exit from an interrupt handler, which might possibly result in entering
879 * idle mode, in other words, leaving the mode in which read-side critical
880 * sections can occur. The caller must have disabled interrupts.
882 * This code assumes that the idle loop never does anything that might
883 * result in unbalanced calls to irq_enter() and irq_exit(). If your
884 * architecture violates this assumption, RCU will give you what you
885 * deserve, good and hard. But very infrequently and irreproducibly.
887 * Use things like work queues to work around this limitation.
889 * You have been warned.
891 void rcu_irq_exit(void)
894 struct rcu_dynticks *rdtp;
896 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
897 rdtp = this_cpu_ptr(&rcu_dynticks);
898 oldval = rdtp->dynticks_nesting;
899 rdtp->dynticks_nesting--;
900 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
901 rdtp->dynticks_nesting < 0);
902 if (rdtp->dynticks_nesting)
903 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
905 rcu_eqs_enter_common(oldval, true);
906 rcu_sysidle_enter(1);
910 * Wrapper for rcu_irq_exit() where interrupts are enabled.
912 void rcu_irq_exit_irqson(void)
916 local_irq_save(flags);
918 local_irq_restore(flags);
922 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
924 * If the new value of the ->dynticks_nesting counter was previously zero,
925 * we really have exited idle, and must do the appropriate accounting.
926 * The caller must have disabled interrupts.
928 static void rcu_eqs_exit_common(long long oldval, int user)
930 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
932 rcu_dynticks_task_exit();
933 rcu_dynticks_eqs_exit();
934 rcu_cleanup_after_idle();
935 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
936 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
937 !user && !is_idle_task(current)) {
938 struct task_struct *idle __maybe_unused =
939 idle_task(smp_processor_id());
941 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
942 oldval, rdtp->dynticks_nesting);
943 rcu_ftrace_dump(DUMP_ORIG);
944 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
945 current->pid, current->comm,
946 idle->pid, idle->comm); /* must be idle task! */
951 * Exit an RCU extended quiescent state, which can be either the
952 * idle loop or adaptive-tickless usermode execution.
954 static void rcu_eqs_exit(bool user)
956 struct rcu_dynticks *rdtp;
959 rdtp = this_cpu_ptr(&rcu_dynticks);
960 oldval = rdtp->dynticks_nesting;
961 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
962 if (oldval & DYNTICK_TASK_NEST_MASK) {
963 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
965 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
966 rcu_eqs_exit_common(oldval, user);
971 * rcu_idle_exit - inform RCU that current CPU is leaving idle
973 * Exit idle mode, in other words, -enter- the mode in which RCU
974 * read-side critical sections can occur.
976 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
977 * allow for the possibility of usermode upcalls messing up our count
978 * of interrupt nesting level during the busy period that is just
981 void rcu_idle_exit(void)
985 local_irq_save(flags);
988 local_irq_restore(flags);
990 EXPORT_SYMBOL_GPL(rcu_idle_exit);
992 #ifdef CONFIG_NO_HZ_FULL
994 * rcu_user_exit - inform RCU that we are exiting userspace.
996 * Exit RCU idle mode while entering the kernel because it can
997 * run a RCU read side critical section anytime.
999 void rcu_user_exit(void)
1003 #endif /* CONFIG_NO_HZ_FULL */
1006 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1008 * Enter an interrupt handler, which might possibly result in exiting
1009 * idle mode, in other words, entering the mode in which read-side critical
1010 * sections can occur. The caller must have disabled interrupts.
1012 * Note that the Linux kernel is fully capable of entering an interrupt
1013 * handler that it never exits, for example when doing upcalls to
1014 * user mode! This code assumes that the idle loop never does upcalls to
1015 * user mode. If your architecture does do upcalls from the idle loop (or
1016 * does anything else that results in unbalanced calls to the irq_enter()
1017 * and irq_exit() functions), RCU will give you what you deserve, good
1018 * and hard. But very infrequently and irreproducibly.
1020 * Use things like work queues to work around this limitation.
1022 * You have been warned.
1024 void rcu_irq_enter(void)
1026 struct rcu_dynticks *rdtp;
1029 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
1030 rdtp = this_cpu_ptr(&rcu_dynticks);
1031 oldval = rdtp->dynticks_nesting;
1032 rdtp->dynticks_nesting++;
1033 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
1034 rdtp->dynticks_nesting == 0);
1036 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
1038 rcu_eqs_exit_common(oldval, true);
1039 rcu_sysidle_exit(1);
1043 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1045 void rcu_irq_enter_irqson(void)
1047 unsigned long flags;
1049 local_irq_save(flags);
1051 local_irq_restore(flags);
1055 * rcu_nmi_enter - inform RCU of entry to NMI context
1057 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1058 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1059 * that the CPU is active. This implementation permits nested NMIs, as
1060 * long as the nesting level does not overflow an int. (You will probably
1061 * run out of stack space first.)
1063 void rcu_nmi_enter(void)
1065 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1068 /* Complain about underflow. */
1069 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
1072 * If idle from RCU viewpoint, atomically increment ->dynticks
1073 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1074 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1075 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1076 * to be in the outermost NMI handler that interrupted an RCU-idle
1077 * period (observation due to Andy Lutomirski).
1079 if (rcu_dynticks_curr_cpu_in_eqs()) {
1080 rcu_dynticks_eqs_exit();
1083 rdtp->dynticks_nmi_nesting += incby;
1088 * rcu_nmi_exit - inform RCU of exit from NMI context
1090 * If we are returning from the outermost NMI handler that interrupted an
1091 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
1092 * to let the RCU grace-period handling know that the CPU is back to
1095 void rcu_nmi_exit(void)
1097 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1100 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1101 * (We are exiting an NMI handler, so RCU better be paying attention
1104 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
1105 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
1108 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1109 * leave it in non-RCU-idle state.
1111 if (rdtp->dynticks_nmi_nesting != 1) {
1112 rdtp->dynticks_nmi_nesting -= 2;
1116 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1117 rdtp->dynticks_nmi_nesting = 0;
1118 rcu_dynticks_eqs_enter();
1122 * __rcu_is_watching - are RCU read-side critical sections safe?
1124 * Return true if RCU is watching the running CPU, which means that
1125 * this CPU can safely enter RCU read-side critical sections. Unlike
1126 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1127 * least disabled preemption.
1129 bool notrace __rcu_is_watching(void)
1131 return !rcu_dynticks_curr_cpu_in_eqs();
1135 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1137 * If the current CPU is in its idle loop and is neither in an interrupt
1138 * or NMI handler, return true.
1140 bool notrace rcu_is_watching(void)
1144 preempt_disable_notrace();
1145 ret = __rcu_is_watching();
1146 preempt_enable_notrace();
1149 EXPORT_SYMBOL_GPL(rcu_is_watching);
1151 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1154 * Is the current CPU online? Disable preemption to avoid false positives
1155 * that could otherwise happen due to the current CPU number being sampled,
1156 * this task being preempted, its old CPU being taken offline, resuming
1157 * on some other CPU, then determining that its old CPU is now offline.
1158 * It is OK to use RCU on an offline processor during initial boot, hence
1159 * the check for rcu_scheduler_fully_active. Note also that it is OK
1160 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1161 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1162 * offline to continue to use RCU for one jiffy after marking itself
1163 * offline in the cpu_online_mask. This leniency is necessary given the
1164 * non-atomic nature of the online and offline processing, for example,
1165 * the fact that a CPU enters the scheduler after completing the teardown
1168 * This is also why RCU internally marks CPUs online during in the
1169 * preparation phase and offline after the CPU has been taken down.
1171 * Disable checking if in an NMI handler because we cannot safely report
1172 * errors from NMI handlers anyway.
1174 bool rcu_lockdep_current_cpu_online(void)
1176 struct rcu_data *rdp;
1177 struct rcu_node *rnp;
1183 rdp = this_cpu_ptr(&rcu_sched_data);
1185 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1186 !rcu_scheduler_fully_active;
1190 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1192 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1195 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1197 * If the current CPU is idle or running at a first-level (not nested)
1198 * interrupt from idle, return true. The caller must have at least
1199 * disabled preemption.
1201 static int rcu_is_cpu_rrupt_from_idle(void)
1203 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1207 * Snapshot the specified CPU's dynticks counter so that we can later
1208 * credit them with an implicit quiescent state. Return 1 if this CPU
1209 * is in dynticks idle mode, which is an extended quiescent state.
1211 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1212 bool *isidle, unsigned long *maxj)
1214 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1215 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1216 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1217 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1218 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1219 rdp->mynode->gpnum))
1220 WRITE_ONCE(rdp->gpwrap, true);
1227 * Return true if the specified CPU has passed through a quiescent
1228 * state by virtue of being in or having passed through an dynticks
1229 * idle state since the last call to dyntick_save_progress_counter()
1230 * for this same CPU, or by virtue of having been offline.
1232 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1233 bool *isidle, unsigned long *maxj)
1237 unsigned long rjtsc;
1238 struct rcu_node *rnp;
1241 * If the CPU passed through or entered a dynticks idle phase with
1242 * no active irq/NMI handlers, then we can safely pretend that the CPU
1243 * already acknowledged the request to pass through a quiescent
1244 * state. Either way, that CPU cannot possibly be in an RCU
1245 * read-side critical section that started before the beginning
1246 * of the current RCU grace period.
1248 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1249 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1250 rdp->dynticks_fqs++;
1254 /* Compute and saturate jiffies_till_sched_qs. */
1255 jtsq = jiffies_till_sched_qs;
1256 rjtsc = rcu_jiffies_till_stall_check();
1257 if (jtsq > rjtsc / 2) {
1258 WRITE_ONCE(jiffies_till_sched_qs, rjtsc);
1260 } else if (jtsq < 1) {
1261 WRITE_ONCE(jiffies_till_sched_qs, 1);
1266 * Has this CPU encountered a cond_resched_rcu_qs() since the
1267 * beginning of the grace period? For this to be the case,
1268 * the CPU has to have noticed the current grace period. This
1269 * might not be the case for nohz_full CPUs looping in the kernel.
1272 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1273 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_qs_ctr, rdp->cpu) &&
1274 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1275 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1279 /* Check for the CPU being offline. */
1280 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1281 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1287 * A CPU running for an extended time within the kernel can
1288 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1289 * even context-switching back and forth between a pair of
1290 * in-kernel CPU-bound tasks cannot advance grace periods.
1291 * So if the grace period is old enough, make the CPU pay attention.
1292 * Note that the unsynchronized assignments to the per-CPU
1293 * rcu_sched_qs_mask variable are safe. Yes, setting of
1294 * bits can be lost, but they will be set again on the next
1295 * force-quiescent-state pass. So lost bit sets do not result
1296 * in incorrect behavior, merely in a grace period lasting
1297 * a few jiffies longer than it might otherwise. Because
1298 * there are at most four threads involved, and because the
1299 * updates are only once every few jiffies, the probability of
1300 * lossage (and thus of slight grace-period extension) is
1303 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1304 * is set too high, we override with half of the RCU CPU stall
1307 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1308 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1309 time_after(jiffies, rdp->rsp->jiffies_resched)) {
1310 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1311 WRITE_ONCE(rdp->cond_resched_completed,
1312 READ_ONCE(rdp->mynode->completed));
1313 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1315 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1317 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1321 * If more than halfway to RCU CPU stall-warning time, do
1322 * a resched_cpu() to try to loosen things up a bit.
1324 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2)
1325 resched_cpu(rdp->cpu);
1330 static void record_gp_stall_check_time(struct rcu_state *rsp)
1332 unsigned long j = jiffies;
1336 smp_wmb(); /* Record start time before stall time. */
1337 j1 = rcu_jiffies_till_stall_check();
1338 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1339 rsp->jiffies_resched = j + j1 / 2;
1340 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1344 * Convert a ->gp_state value to a character string.
1346 static const char *gp_state_getname(short gs)
1348 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1350 return gp_state_names[gs];
1354 * Complain about starvation of grace-period kthread.
1356 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1362 gpa = READ_ONCE(rsp->gp_activity);
1363 if (j - gpa > 2 * HZ) {
1364 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1366 rsp->gpnum, rsp->completed,
1368 gp_state_getname(rsp->gp_state), rsp->gp_state,
1369 rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
1370 if (rsp->gp_kthread) {
1371 sched_show_task(rsp->gp_kthread);
1372 wake_up_process(rsp->gp_kthread);
1378 * Dump stacks of all tasks running on stalled CPUs. First try using
1379 * NMIs, but fall back to manual remote stack tracing on architectures
1380 * that don't support NMI-based stack dumps. The NMI-triggered stack
1381 * traces are more accurate because they are printed by the target CPU.
1383 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1386 unsigned long flags;
1387 struct rcu_node *rnp;
1389 rcu_for_each_leaf_node(rsp, rnp) {
1390 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1391 for_each_leaf_node_possible_cpu(rnp, cpu)
1392 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1393 if (!trigger_single_cpu_backtrace(cpu))
1395 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1400 * If too much time has passed in the current grace period, and if
1401 * so configured, go kick the relevant kthreads.
1403 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1407 if (!rcu_kick_kthreads)
1409 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1410 if (time_after(jiffies, j) && rsp->gp_kthread &&
1411 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1412 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1413 rcu_ftrace_dump(DUMP_ALL);
1414 wake_up_process(rsp->gp_kthread);
1415 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1419 static inline void panic_on_rcu_stall(void)
1421 if (sysctl_panic_on_rcu_stall)
1422 panic("RCU Stall\n");
1425 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1429 unsigned long flags;
1433 struct rcu_node *rnp = rcu_get_root(rsp);
1436 /* Kick and suppress, if so configured. */
1437 rcu_stall_kick_kthreads(rsp);
1438 if (rcu_cpu_stall_suppress)
1441 /* Only let one CPU complain about others per time interval. */
1443 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1444 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1445 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1446 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1449 WRITE_ONCE(rsp->jiffies_stall,
1450 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1451 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1454 * OK, time to rat on our buddy...
1455 * See Documentation/RCU/stallwarn.txt for info on how to debug
1456 * RCU CPU stall warnings.
1458 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1460 print_cpu_stall_info_begin();
1461 rcu_for_each_leaf_node(rsp, rnp) {
1462 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1463 ndetected += rcu_print_task_stall(rnp);
1464 if (rnp->qsmask != 0) {
1465 for_each_leaf_node_possible_cpu(rnp, cpu)
1466 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1467 print_cpu_stall_info(rsp, cpu);
1471 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1474 print_cpu_stall_info_end();
1475 for_each_possible_cpu(cpu)
1476 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1477 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1478 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1479 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1481 rcu_dump_cpu_stacks(rsp);
1483 /* Complain about tasks blocking the grace period. */
1484 rcu_print_detail_task_stall(rsp);
1486 if (READ_ONCE(rsp->gpnum) != gpnum ||
1487 READ_ONCE(rsp->completed) == gpnum) {
1488 pr_err("INFO: Stall ended before state dump start\n");
1491 gpa = READ_ONCE(rsp->gp_activity);
1492 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1493 rsp->name, j - gpa, j, gpa,
1494 jiffies_till_next_fqs,
1495 rcu_get_root(rsp)->qsmask);
1496 /* In this case, the current CPU might be at fault. */
1497 sched_show_task(current);
1501 rcu_check_gp_kthread_starvation(rsp);
1503 panic_on_rcu_stall();
1505 force_quiescent_state(rsp); /* Kick them all. */
1508 static void print_cpu_stall(struct rcu_state *rsp)
1511 unsigned long flags;
1512 struct rcu_node *rnp = rcu_get_root(rsp);
1515 /* Kick and suppress, if so configured. */
1516 rcu_stall_kick_kthreads(rsp);
1517 if (rcu_cpu_stall_suppress)
1521 * OK, time to rat on ourselves...
1522 * See Documentation/RCU/stallwarn.txt for info on how to debug
1523 * RCU CPU stall warnings.
1525 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1526 print_cpu_stall_info_begin();
1527 print_cpu_stall_info(rsp, smp_processor_id());
1528 print_cpu_stall_info_end();
1529 for_each_possible_cpu(cpu)
1530 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1531 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1532 jiffies - rsp->gp_start,
1533 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1535 rcu_check_gp_kthread_starvation(rsp);
1537 rcu_dump_cpu_stacks(rsp);
1539 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1540 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1541 WRITE_ONCE(rsp->jiffies_stall,
1542 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1543 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1545 panic_on_rcu_stall();
1548 * Attempt to revive the RCU machinery by forcing a context switch.
1550 * A context switch would normally allow the RCU state machine to make
1551 * progress and it could be we're stuck in kernel space without context
1552 * switches for an entirely unreasonable amount of time.
1554 resched_cpu(smp_processor_id());
1557 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1559 unsigned long completed;
1560 unsigned long gpnum;
1564 struct rcu_node *rnp;
1566 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1567 !rcu_gp_in_progress(rsp))
1569 rcu_stall_kick_kthreads(rsp);
1573 * Lots of memory barriers to reject false positives.
1575 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1576 * then rsp->gp_start, and finally rsp->completed. These values
1577 * are updated in the opposite order with memory barriers (or
1578 * equivalent) during grace-period initialization and cleanup.
1579 * Now, a false positive can occur if we get an new value of
1580 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1581 * the memory barriers, the only way that this can happen is if one
1582 * grace period ends and another starts between these two fetches.
1583 * Detect this by comparing rsp->completed with the previous fetch
1586 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1587 * and rsp->gp_start suffice to forestall false positives.
1589 gpnum = READ_ONCE(rsp->gpnum);
1590 smp_rmb(); /* Pick up ->gpnum first... */
1591 js = READ_ONCE(rsp->jiffies_stall);
1592 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1593 gps = READ_ONCE(rsp->gp_start);
1594 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1595 completed = READ_ONCE(rsp->completed);
1596 if (ULONG_CMP_GE(completed, gpnum) ||
1597 ULONG_CMP_LT(j, js) ||
1598 ULONG_CMP_GE(gps, js))
1599 return; /* No stall or GP completed since entering function. */
1601 if (rcu_gp_in_progress(rsp) &&
1602 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1604 /* We haven't checked in, so go dump stack. */
1605 print_cpu_stall(rsp);
1607 } else if (rcu_gp_in_progress(rsp) &&
1608 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1610 /* They had a few time units to dump stack, so complain. */
1611 print_other_cpu_stall(rsp, gpnum);
1616 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1618 * Set the stall-warning timeout way off into the future, thus preventing
1619 * any RCU CPU stall-warning messages from appearing in the current set of
1620 * RCU grace periods.
1622 * The caller must disable hard irqs.
1624 void rcu_cpu_stall_reset(void)
1626 struct rcu_state *rsp;
1628 for_each_rcu_flavor(rsp)
1629 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1633 * Initialize the specified rcu_data structure's default callback list
1634 * to empty. The default callback list is the one that is not used by
1635 * no-callbacks CPUs.
1637 static void init_default_callback_list(struct rcu_data *rdp)
1641 rdp->nxtlist = NULL;
1642 for (i = 0; i < RCU_NEXT_SIZE; i++)
1643 rdp->nxttail[i] = &rdp->nxtlist;
1647 * Initialize the specified rcu_data structure's callback list to empty.
1649 static void init_callback_list(struct rcu_data *rdp)
1651 if (init_nocb_callback_list(rdp))
1653 init_default_callback_list(rdp);
1657 * Determine the value that ->completed will have at the end of the
1658 * next subsequent grace period. This is used to tag callbacks so that
1659 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1660 * been dyntick-idle for an extended period with callbacks under the
1661 * influence of RCU_FAST_NO_HZ.
1663 * The caller must hold rnp->lock with interrupts disabled.
1665 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1666 struct rcu_node *rnp)
1669 * If RCU is idle, we just wait for the next grace period.
1670 * But we can only be sure that RCU is idle if we are looking
1671 * at the root rcu_node structure -- otherwise, a new grace
1672 * period might have started, but just not yet gotten around
1673 * to initializing the current non-root rcu_node structure.
1675 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1676 return rnp->completed + 1;
1679 * Otherwise, wait for a possible partial grace period and
1680 * then the subsequent full grace period.
1682 return rnp->completed + 2;
1686 * Trace-event helper function for rcu_start_future_gp() and
1687 * rcu_nocb_wait_gp().
1689 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1690 unsigned long c, const char *s)
1692 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1693 rnp->completed, c, rnp->level,
1694 rnp->grplo, rnp->grphi, s);
1698 * Start some future grace period, as needed to handle newly arrived
1699 * callbacks. The required future grace periods are recorded in each
1700 * rcu_node structure's ->need_future_gp field. Returns true if there
1701 * is reason to awaken the grace-period kthread.
1703 * The caller must hold the specified rcu_node structure's ->lock.
1705 static bool __maybe_unused
1706 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1707 unsigned long *c_out)
1712 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1715 * Pick up grace-period number for new callbacks. If this
1716 * grace period is already marked as needed, return to the caller.
1718 c = rcu_cbs_completed(rdp->rsp, rnp);
1719 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1720 if (rnp->need_future_gp[c & 0x1]) {
1721 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1726 * If either this rcu_node structure or the root rcu_node structure
1727 * believe that a grace period is in progress, then we must wait
1728 * for the one following, which is in "c". Because our request
1729 * will be noticed at the end of the current grace period, we don't
1730 * need to explicitly start one. We only do the lockless check
1731 * of rnp_root's fields if the current rcu_node structure thinks
1732 * there is no grace period in flight, and because we hold rnp->lock,
1733 * the only possible change is when rnp_root's two fields are
1734 * equal, in which case rnp_root->gpnum might be concurrently
1735 * incremented. But that is OK, as it will just result in our
1736 * doing some extra useless work.
1738 if (rnp->gpnum != rnp->completed ||
1739 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1740 rnp->need_future_gp[c & 0x1]++;
1741 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1746 * There might be no grace period in progress. If we don't already
1747 * hold it, acquire the root rcu_node structure's lock in order to
1748 * start one (if needed).
1750 if (rnp != rnp_root)
1751 raw_spin_lock_rcu_node(rnp_root);
1754 * Get a new grace-period number. If there really is no grace
1755 * period in progress, it will be smaller than the one we obtained
1756 * earlier. Adjust callbacks as needed. Note that even no-CBs
1757 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1759 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1760 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1761 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1762 rdp->nxtcompleted[i] = c;
1765 * If the needed for the required grace period is already
1766 * recorded, trace and leave.
1768 if (rnp_root->need_future_gp[c & 0x1]) {
1769 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1773 /* Record the need for the future grace period. */
1774 rnp_root->need_future_gp[c & 0x1]++;
1776 /* If a grace period is not already in progress, start one. */
1777 if (rnp_root->gpnum != rnp_root->completed) {
1778 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1780 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1781 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1784 if (rnp != rnp_root)
1785 raw_spin_unlock_rcu_node(rnp_root);
1793 * Clean up any old requests for the just-ended grace period. Also return
1794 * whether any additional grace periods have been requested. Also invoke
1795 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1796 * waiting for this grace period to complete.
1798 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1800 int c = rnp->completed;
1802 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1804 rnp->need_future_gp[c & 0x1] = 0;
1805 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1806 trace_rcu_future_gp(rnp, rdp, c,
1807 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1812 * Awaken the grace-period kthread for the specified flavor of RCU.
1813 * Don't do a self-awaken, and don't bother awakening when there is
1814 * nothing for the grace-period kthread to do (as in several CPUs
1815 * raced to awaken, and we lost), and finally don't try to awaken
1816 * a kthread that has not yet been created.
1818 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1820 if (current == rsp->gp_kthread ||
1821 !READ_ONCE(rsp->gp_flags) ||
1824 swake_up(&rsp->gp_wq);
1828 * If there is room, assign a ->completed number to any callbacks on
1829 * this CPU that have not already been assigned. Also accelerate any
1830 * callbacks that were previously assigned a ->completed number that has
1831 * since proven to be too conservative, which can happen if callbacks get
1832 * assigned a ->completed number while RCU is idle, but with reference to
1833 * a non-root rcu_node structure. This function is idempotent, so it does
1834 * not hurt to call it repeatedly. Returns an flag saying that we should
1835 * awaken the RCU grace-period kthread.
1837 * The caller must hold rnp->lock with interrupts disabled.
1839 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1840 struct rcu_data *rdp)
1846 /* If the CPU has no callbacks, nothing to do. */
1847 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1851 * Starting from the sublist containing the callbacks most
1852 * recently assigned a ->completed number and working down, find the
1853 * first sublist that is not assignable to an upcoming grace period.
1854 * Such a sublist has something in it (first two tests) and has
1855 * a ->completed number assigned that will complete sooner than
1856 * the ->completed number for newly arrived callbacks (last test).
1858 * The key point is that any later sublist can be assigned the
1859 * same ->completed number as the newly arrived callbacks, which
1860 * means that the callbacks in any of these later sublist can be
1861 * grouped into a single sublist, whether or not they have already
1862 * been assigned a ->completed number.
1864 c = rcu_cbs_completed(rsp, rnp);
1865 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1866 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1867 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1871 * If there are no sublist for unassigned callbacks, leave.
1872 * At the same time, advance "i" one sublist, so that "i" will
1873 * index into the sublist where all the remaining callbacks should
1876 if (++i >= RCU_NEXT_TAIL)
1880 * Assign all subsequent callbacks' ->completed number to the next
1881 * full grace period and group them all in the sublist initially
1884 for (; i <= RCU_NEXT_TAIL; i++) {
1885 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1886 rdp->nxtcompleted[i] = c;
1888 /* Record any needed additional grace periods. */
1889 ret = rcu_start_future_gp(rnp, rdp, NULL);
1891 /* Trace depending on how much we were able to accelerate. */
1892 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1893 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1895 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1900 * Move any callbacks whose grace period has completed to the
1901 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1902 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1903 * sublist. This function is idempotent, so it does not hurt to
1904 * invoke it repeatedly. As long as it is not invoked -too- often...
1905 * Returns true if the RCU grace-period kthread needs to be awakened.
1907 * The caller must hold rnp->lock with interrupts disabled.
1909 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1910 struct rcu_data *rdp)
1914 /* If the CPU has no callbacks, nothing to do. */
1915 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1919 * Find all callbacks whose ->completed numbers indicate that they
1920 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1922 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1923 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1925 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1927 /* Clean up any sublist tail pointers that were misordered above. */
1928 for (j = RCU_WAIT_TAIL; j < i; j++)
1929 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1931 /* Copy down callbacks to fill in empty sublists. */
1932 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1933 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1935 rdp->nxttail[j] = rdp->nxttail[i];
1936 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1939 /* Classify any remaining callbacks. */
1940 return rcu_accelerate_cbs(rsp, rnp, rdp);
1944 * Update CPU-local rcu_data state to record the beginnings and ends of
1945 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1946 * structure corresponding to the current CPU, and must have irqs disabled.
1947 * Returns true if the grace-period kthread needs to be awakened.
1949 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1950 struct rcu_data *rdp)
1955 /* Handle the ends of any preceding grace periods first. */
1956 if (rdp->completed == rnp->completed &&
1957 !unlikely(READ_ONCE(rdp->gpwrap))) {
1959 /* No grace period end, so just accelerate recent callbacks. */
1960 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1964 /* Advance callbacks. */
1965 ret = rcu_advance_cbs(rsp, rnp, rdp);
1967 /* Remember that we saw this grace-period completion. */
1968 rdp->completed = rnp->completed;
1969 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1972 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1974 * If the current grace period is waiting for this CPU,
1975 * set up to detect a quiescent state, otherwise don't
1976 * go looking for one.
1978 rdp->gpnum = rnp->gpnum;
1979 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1980 need_gp = !!(rnp->qsmask & rdp->grpmask);
1981 rdp->cpu_no_qs.b.norm = need_gp;
1982 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1983 rdp->core_needs_qs = need_gp;
1984 zero_cpu_stall_ticks(rdp);
1985 WRITE_ONCE(rdp->gpwrap, false);
1990 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1992 unsigned long flags;
1994 struct rcu_node *rnp;
1996 local_irq_save(flags);
1998 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1999 rdp->completed == READ_ONCE(rnp->completed) &&
2000 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
2001 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
2002 local_irq_restore(flags);
2005 needwake = __note_gp_changes(rsp, rnp, rdp);
2006 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2008 rcu_gp_kthread_wake(rsp);
2011 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
2014 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
2015 schedule_timeout_uninterruptible(delay);
2019 * Initialize a new grace period. Return false if no grace period required.
2021 static bool rcu_gp_init(struct rcu_state *rsp)
2023 unsigned long oldmask;
2024 struct rcu_data *rdp;
2025 struct rcu_node *rnp = rcu_get_root(rsp);
2027 WRITE_ONCE(rsp->gp_activity, jiffies);
2028 raw_spin_lock_irq_rcu_node(rnp);
2029 if (!READ_ONCE(rsp->gp_flags)) {
2030 /* Spurious wakeup, tell caller to go back to sleep. */
2031 raw_spin_unlock_irq_rcu_node(rnp);
2034 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
2036 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
2038 * Grace period already in progress, don't start another.
2039 * Not supposed to be able to happen.
2041 raw_spin_unlock_irq_rcu_node(rnp);
2045 /* Advance to a new grace period and initialize state. */
2046 record_gp_stall_check_time(rsp);
2047 /* Record GP times before starting GP, hence smp_store_release(). */
2048 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
2049 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
2050 raw_spin_unlock_irq_rcu_node(rnp);
2053 * Apply per-leaf buffered online and offline operations to the
2054 * rcu_node tree. Note that this new grace period need not wait
2055 * for subsequent online CPUs, and that quiescent-state forcing
2056 * will handle subsequent offline CPUs.
2058 rcu_for_each_leaf_node(rsp, rnp) {
2059 rcu_gp_slow(rsp, gp_preinit_delay);
2060 raw_spin_lock_irq_rcu_node(rnp);
2061 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2062 !rnp->wait_blkd_tasks) {
2063 /* Nothing to do on this leaf rcu_node structure. */
2064 raw_spin_unlock_irq_rcu_node(rnp);
2068 /* Record old state, apply changes to ->qsmaskinit field. */
2069 oldmask = rnp->qsmaskinit;
2070 rnp->qsmaskinit = rnp->qsmaskinitnext;
2072 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2073 if (!oldmask != !rnp->qsmaskinit) {
2074 if (!oldmask) /* First online CPU for this rcu_node. */
2075 rcu_init_new_rnp(rnp);
2076 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2077 rnp->wait_blkd_tasks = true;
2078 else /* Last offline CPU and can propagate. */
2079 rcu_cleanup_dead_rnp(rnp);
2083 * If all waited-on tasks from prior grace period are
2084 * done, and if all this rcu_node structure's CPUs are
2085 * still offline, propagate up the rcu_node tree and
2086 * clear ->wait_blkd_tasks. Otherwise, if one of this
2087 * rcu_node structure's CPUs has since come back online,
2088 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2089 * checks for this, so just call it unconditionally).
2091 if (rnp->wait_blkd_tasks &&
2092 (!rcu_preempt_has_tasks(rnp) ||
2094 rnp->wait_blkd_tasks = false;
2095 rcu_cleanup_dead_rnp(rnp);
2098 raw_spin_unlock_irq_rcu_node(rnp);
2102 * Set the quiescent-state-needed bits in all the rcu_node
2103 * structures for all currently online CPUs in breadth-first order,
2104 * starting from the root rcu_node structure, relying on the layout
2105 * of the tree within the rsp->node[] array. Note that other CPUs
2106 * will access only the leaves of the hierarchy, thus seeing that no
2107 * grace period is in progress, at least until the corresponding
2108 * leaf node has been initialized.
2110 * The grace period cannot complete until the initialization
2111 * process finishes, because this kthread handles both.
2113 rcu_for_each_node_breadth_first(rsp, rnp) {
2114 rcu_gp_slow(rsp, gp_init_delay);
2115 raw_spin_lock_irq_rcu_node(rnp);
2116 rdp = this_cpu_ptr(rsp->rda);
2117 rcu_preempt_check_blocked_tasks(rnp);
2118 rnp->qsmask = rnp->qsmaskinit;
2119 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2120 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2121 WRITE_ONCE(rnp->completed, rsp->completed);
2122 if (rnp == rdp->mynode)
2123 (void)__note_gp_changes(rsp, rnp, rdp);
2124 rcu_preempt_boost_start_gp(rnp);
2125 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2126 rnp->level, rnp->grplo,
2127 rnp->grphi, rnp->qsmask);
2128 raw_spin_unlock_irq_rcu_node(rnp);
2129 cond_resched_rcu_qs();
2130 WRITE_ONCE(rsp->gp_activity, jiffies);
2137 * Helper function for wait_event_interruptible_timeout() wakeup
2138 * at force-quiescent-state time.
2140 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2142 struct rcu_node *rnp = rcu_get_root(rsp);
2144 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2145 *gfp = READ_ONCE(rsp->gp_flags);
2146 if (*gfp & RCU_GP_FLAG_FQS)
2149 /* The current grace period has completed. */
2150 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2157 * Do one round of quiescent-state forcing.
2159 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2161 bool isidle = false;
2163 struct rcu_node *rnp = rcu_get_root(rsp);
2165 WRITE_ONCE(rsp->gp_activity, jiffies);
2168 /* Collect dyntick-idle snapshots. */
2169 if (is_sysidle_rcu_state(rsp)) {
2171 maxj = jiffies - ULONG_MAX / 4;
2173 force_qs_rnp(rsp, dyntick_save_progress_counter,
2175 rcu_sysidle_report_gp(rsp, isidle, maxj);
2177 /* Handle dyntick-idle and offline CPUs. */
2179 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
2181 /* Clear flag to prevent immediate re-entry. */
2182 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2183 raw_spin_lock_irq_rcu_node(rnp);
2184 WRITE_ONCE(rsp->gp_flags,
2185 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2186 raw_spin_unlock_irq_rcu_node(rnp);
2191 * Clean up after the old grace period.
2193 static void rcu_gp_cleanup(struct rcu_state *rsp)
2195 unsigned long gp_duration;
2196 bool needgp = false;
2198 struct rcu_data *rdp;
2199 struct rcu_node *rnp = rcu_get_root(rsp);
2200 struct swait_queue_head *sq;
2202 WRITE_ONCE(rsp->gp_activity, jiffies);
2203 raw_spin_lock_irq_rcu_node(rnp);
2204 gp_duration = jiffies - rsp->gp_start;
2205 if (gp_duration > rsp->gp_max)
2206 rsp->gp_max = gp_duration;
2209 * We know the grace period is complete, but to everyone else
2210 * it appears to still be ongoing. But it is also the case
2211 * that to everyone else it looks like there is nothing that
2212 * they can do to advance the grace period. It is therefore
2213 * safe for us to drop the lock in order to mark the grace
2214 * period as completed in all of the rcu_node structures.
2216 raw_spin_unlock_irq_rcu_node(rnp);
2219 * Propagate new ->completed value to rcu_node structures so
2220 * that other CPUs don't have to wait until the start of the next
2221 * grace period to process their callbacks. This also avoids
2222 * some nasty RCU grace-period initialization races by forcing
2223 * the end of the current grace period to be completely recorded in
2224 * all of the rcu_node structures before the beginning of the next
2225 * grace period is recorded in any of the rcu_node structures.
2227 rcu_for_each_node_breadth_first(rsp, rnp) {
2228 raw_spin_lock_irq_rcu_node(rnp);
2229 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2230 WARN_ON_ONCE(rnp->qsmask);
2231 WRITE_ONCE(rnp->completed, rsp->gpnum);
2232 rdp = this_cpu_ptr(rsp->rda);
2233 if (rnp == rdp->mynode)
2234 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2235 /* smp_mb() provided by prior unlock-lock pair. */
2236 nocb += rcu_future_gp_cleanup(rsp, rnp);
2237 sq = rcu_nocb_gp_get(rnp);
2238 raw_spin_unlock_irq_rcu_node(rnp);
2239 rcu_nocb_gp_cleanup(sq);
2240 cond_resched_rcu_qs();
2241 WRITE_ONCE(rsp->gp_activity, jiffies);
2242 rcu_gp_slow(rsp, gp_cleanup_delay);
2244 rnp = rcu_get_root(rsp);
2245 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2246 rcu_nocb_gp_set(rnp, nocb);
2248 /* Declare grace period done. */
2249 WRITE_ONCE(rsp->completed, rsp->gpnum);
2250 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2251 rsp->gp_state = RCU_GP_IDLE;
2252 rdp = this_cpu_ptr(rsp->rda);
2253 /* Advance CBs to reduce false positives below. */
2254 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2255 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2256 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2257 trace_rcu_grace_period(rsp->name,
2258 READ_ONCE(rsp->gpnum),
2261 raw_spin_unlock_irq_rcu_node(rnp);
2265 * Body of kthread that handles grace periods.
2267 static int __noreturn rcu_gp_kthread(void *arg)
2273 struct rcu_state *rsp = arg;
2274 struct rcu_node *rnp = rcu_get_root(rsp);
2276 rcu_bind_gp_kthread();
2279 /* Handle grace-period start. */
2281 trace_rcu_grace_period(rsp->name,
2282 READ_ONCE(rsp->gpnum),
2284 rsp->gp_state = RCU_GP_WAIT_GPS;
2285 swait_event_interruptible(rsp->gp_wq,
2286 READ_ONCE(rsp->gp_flags) &
2288 rsp->gp_state = RCU_GP_DONE_GPS;
2289 /* Locking provides needed memory barrier. */
2290 if (rcu_gp_init(rsp))
2292 cond_resched_rcu_qs();
2293 WRITE_ONCE(rsp->gp_activity, jiffies);
2294 WARN_ON(signal_pending(current));
2295 trace_rcu_grace_period(rsp->name,
2296 READ_ONCE(rsp->gpnum),
2300 /* Handle quiescent-state forcing. */
2301 first_gp_fqs = true;
2302 j = jiffies_till_first_fqs;
2305 jiffies_till_first_fqs = HZ;
2310 rsp->jiffies_force_qs = jiffies + j;
2311 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2314 trace_rcu_grace_period(rsp->name,
2315 READ_ONCE(rsp->gpnum),
2317 rsp->gp_state = RCU_GP_WAIT_FQS;
2318 ret = swait_event_interruptible_timeout(rsp->gp_wq,
2319 rcu_gp_fqs_check_wake(rsp, &gf), j);
2320 rsp->gp_state = RCU_GP_DOING_FQS;
2321 /* Locking provides needed memory barriers. */
2322 /* If grace period done, leave loop. */
2323 if (!READ_ONCE(rnp->qsmask) &&
2324 !rcu_preempt_blocked_readers_cgp(rnp))
2326 /* If time for quiescent-state forcing, do it. */
2327 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2328 (gf & RCU_GP_FLAG_FQS)) {
2329 trace_rcu_grace_period(rsp->name,
2330 READ_ONCE(rsp->gpnum),
2332 rcu_gp_fqs(rsp, first_gp_fqs);
2333 first_gp_fqs = false;
2334 trace_rcu_grace_period(rsp->name,
2335 READ_ONCE(rsp->gpnum),
2337 cond_resched_rcu_qs();
2338 WRITE_ONCE(rsp->gp_activity, jiffies);
2339 ret = 0; /* Force full wait till next FQS. */
2340 j = jiffies_till_next_fqs;
2343 jiffies_till_next_fqs = HZ;
2346 jiffies_till_next_fqs = 1;
2349 /* Deal with stray signal. */
2350 cond_resched_rcu_qs();
2351 WRITE_ONCE(rsp->gp_activity, jiffies);
2352 WARN_ON(signal_pending(current));
2353 trace_rcu_grace_period(rsp->name,
2354 READ_ONCE(rsp->gpnum),
2356 ret = 1; /* Keep old FQS timing. */
2358 if (time_after(jiffies, rsp->jiffies_force_qs))
2361 j = rsp->jiffies_force_qs - j;
2365 /* Handle grace-period end. */
2366 rsp->gp_state = RCU_GP_CLEANUP;
2367 rcu_gp_cleanup(rsp);
2368 rsp->gp_state = RCU_GP_CLEANED;
2373 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2374 * in preparation for detecting the next grace period. The caller must hold
2375 * the root node's ->lock and hard irqs must be disabled.
2377 * Note that it is legal for a dying CPU (which is marked as offline) to
2378 * invoke this function. This can happen when the dying CPU reports its
2381 * Returns true if the grace-period kthread must be awakened.
2384 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2385 struct rcu_data *rdp)
2387 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2389 * Either we have not yet spawned the grace-period
2390 * task, this CPU does not need another grace period,
2391 * or a grace period is already in progress.
2392 * Either way, don't start a new grace period.
2396 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2397 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2401 * We can't do wakeups while holding the rnp->lock, as that
2402 * could cause possible deadlocks with the rq->lock. Defer
2403 * the wakeup to our caller.
2409 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2410 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2411 * is invoked indirectly from rcu_advance_cbs(), which would result in
2412 * endless recursion -- or would do so if it wasn't for the self-deadlock
2413 * that is encountered beforehand.
2415 * Returns true if the grace-period kthread needs to be awakened.
2417 static bool rcu_start_gp(struct rcu_state *rsp)
2419 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2420 struct rcu_node *rnp = rcu_get_root(rsp);
2424 * If there is no grace period in progress right now, any
2425 * callbacks we have up to this point will be satisfied by the
2426 * next grace period. Also, advancing the callbacks reduces the
2427 * probability of false positives from cpu_needs_another_gp()
2428 * resulting in pointless grace periods. So, advance callbacks
2429 * then start the grace period!
2431 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2432 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2437 * Report a full set of quiescent states to the specified rcu_state data
2438 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2439 * kthread if another grace period is required. Whether we wake
2440 * the grace-period kthread or it awakens itself for the next round
2441 * of quiescent-state forcing, that kthread will clean up after the
2442 * just-completed grace period. Note that the caller must hold rnp->lock,
2443 * which is released before return.
2445 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2446 __releases(rcu_get_root(rsp)->lock)
2448 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2449 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2450 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2451 rcu_gp_kthread_wake(rsp);
2455 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2456 * Allows quiescent states for a group of CPUs to be reported at one go
2457 * to the specified rcu_node structure, though all the CPUs in the group
2458 * must be represented by the same rcu_node structure (which need not be a
2459 * leaf rcu_node structure, though it often will be). The gps parameter
2460 * is the grace-period snapshot, which means that the quiescent states
2461 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2462 * must be held upon entry, and it is released before return.
2465 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2466 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2467 __releases(rnp->lock)
2469 unsigned long oldmask = 0;
2470 struct rcu_node *rnp_c;
2472 /* Walk up the rcu_node hierarchy. */
2474 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2477 * Our bit has already been cleared, or the
2478 * relevant grace period is already over, so done.
2480 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2483 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2484 rnp->qsmask &= ~mask;
2485 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2486 mask, rnp->qsmask, rnp->level,
2487 rnp->grplo, rnp->grphi,
2489 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2491 /* Other bits still set at this level, so done. */
2492 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2495 mask = rnp->grpmask;
2496 if (rnp->parent == NULL) {
2498 /* No more levels. Exit loop holding root lock. */
2502 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2505 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2506 oldmask = rnp_c->qsmask;
2510 * Get here if we are the last CPU to pass through a quiescent
2511 * state for this grace period. Invoke rcu_report_qs_rsp()
2512 * to clean up and start the next grace period if one is needed.
2514 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2518 * Record a quiescent state for all tasks that were previously queued
2519 * on the specified rcu_node structure and that were blocking the current
2520 * RCU grace period. The caller must hold the specified rnp->lock with
2521 * irqs disabled, and this lock is released upon return, but irqs remain
2524 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2525 struct rcu_node *rnp, unsigned long flags)
2526 __releases(rnp->lock)
2530 struct rcu_node *rnp_p;
2532 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2533 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2534 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2535 return; /* Still need more quiescent states! */
2538 rnp_p = rnp->parent;
2539 if (rnp_p == NULL) {
2541 * Only one rcu_node structure in the tree, so don't
2542 * try to report up to its nonexistent parent!
2544 rcu_report_qs_rsp(rsp, flags);
2548 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2550 mask = rnp->grpmask;
2551 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2552 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2553 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2557 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2558 * structure. This must be called from the specified CPU.
2561 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2563 unsigned long flags;
2566 struct rcu_node *rnp;
2569 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2570 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2571 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2574 * The grace period in which this quiescent state was
2575 * recorded has ended, so don't report it upwards.
2576 * We will instead need a new quiescent state that lies
2577 * within the current grace period.
2579 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2580 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2581 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2584 mask = rdp->grpmask;
2585 if ((rnp->qsmask & mask) == 0) {
2586 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2588 rdp->core_needs_qs = false;
2591 * This GP can't end until cpu checks in, so all of our
2592 * callbacks can be processed during the next GP.
2594 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2596 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2597 /* ^^^ Released rnp->lock */
2599 rcu_gp_kthread_wake(rsp);
2604 * Check to see if there is a new grace period of which this CPU
2605 * is not yet aware, and if so, set up local rcu_data state for it.
2606 * Otherwise, see if this CPU has just passed through its first
2607 * quiescent state for this grace period, and record that fact if so.
2610 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2612 /* Check for grace-period ends and beginnings. */
2613 note_gp_changes(rsp, rdp);
2616 * Does this CPU still need to do its part for current grace period?
2617 * If no, return and let the other CPUs do their part as well.
2619 if (!rdp->core_needs_qs)
2623 * Was there a quiescent state since the beginning of the grace
2624 * period? If no, then exit and wait for the next call.
2626 if (rdp->cpu_no_qs.b.norm)
2630 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2633 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2637 * Send the specified CPU's RCU callbacks to the orphanage. The
2638 * specified CPU must be offline, and the caller must hold the
2642 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2643 struct rcu_node *rnp, struct rcu_data *rdp)
2645 /* No-CBs CPUs do not have orphanable callbacks. */
2646 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2650 * Orphan the callbacks. First adjust the counts. This is safe
2651 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2652 * cannot be running now. Thus no memory barrier is required.
2654 if (rdp->nxtlist != NULL) {
2655 rsp->qlen_lazy += rdp->qlen_lazy;
2656 rsp->qlen += rdp->qlen;
2657 rdp->n_cbs_orphaned += rdp->qlen;
2659 WRITE_ONCE(rdp->qlen, 0);
2663 * Next, move those callbacks still needing a grace period to
2664 * the orphanage, where some other CPU will pick them up.
2665 * Some of the callbacks might have gone partway through a grace
2666 * period, but that is too bad. They get to start over because we
2667 * cannot assume that grace periods are synchronized across CPUs.
2668 * We don't bother updating the ->nxttail[] array yet, instead
2669 * we just reset the whole thing later on.
2671 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2672 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2673 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2674 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2678 * Then move the ready-to-invoke callbacks to the orphanage,
2679 * where some other CPU will pick them up. These will not be
2680 * required to pass though another grace period: They are done.
2682 if (rdp->nxtlist != NULL) {
2683 *rsp->orphan_donetail = rdp->nxtlist;
2684 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2688 * Finally, initialize the rcu_data structure's list to empty and
2689 * disallow further callbacks on this CPU.
2691 init_callback_list(rdp);
2692 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2696 * Adopt the RCU callbacks from the specified rcu_state structure's
2697 * orphanage. The caller must hold the ->orphan_lock.
2699 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2702 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2704 /* No-CBs CPUs are handled specially. */
2705 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2706 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2709 /* Do the accounting first. */
2710 rdp->qlen_lazy += rsp->qlen_lazy;
2711 rdp->qlen += rsp->qlen;
2712 rdp->n_cbs_adopted += rsp->qlen;
2713 if (rsp->qlen_lazy != rsp->qlen)
2714 rcu_idle_count_callbacks_posted();
2719 * We do not need a memory barrier here because the only way we
2720 * can get here if there is an rcu_barrier() in flight is if
2721 * we are the task doing the rcu_barrier().
2724 /* First adopt the ready-to-invoke callbacks. */
2725 if (rsp->orphan_donelist != NULL) {
2726 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2727 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2728 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2729 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2730 rdp->nxttail[i] = rsp->orphan_donetail;
2731 rsp->orphan_donelist = NULL;
2732 rsp->orphan_donetail = &rsp->orphan_donelist;
2735 /* And then adopt the callbacks that still need a grace period. */
2736 if (rsp->orphan_nxtlist != NULL) {
2737 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2738 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2739 rsp->orphan_nxtlist = NULL;
2740 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2745 * Trace the fact that this CPU is going offline.
2747 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2749 RCU_TRACE(unsigned long mask);
2750 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2751 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2753 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2756 RCU_TRACE(mask = rdp->grpmask);
2757 trace_rcu_grace_period(rsp->name,
2758 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2763 * All CPUs for the specified rcu_node structure have gone offline,
2764 * and all tasks that were preempted within an RCU read-side critical
2765 * section while running on one of those CPUs have since exited their RCU
2766 * read-side critical section. Some other CPU is reporting this fact with
2767 * the specified rcu_node structure's ->lock held and interrupts disabled.
2768 * This function therefore goes up the tree of rcu_node structures,
2769 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2770 * the leaf rcu_node structure's ->qsmaskinit field has already been
2773 * This function does check that the specified rcu_node structure has
2774 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2775 * prematurely. That said, invoking it after the fact will cost you
2776 * a needless lock acquisition. So once it has done its work, don't
2779 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2782 struct rcu_node *rnp = rnp_leaf;
2784 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2785 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2788 mask = rnp->grpmask;
2792 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2793 rnp->qsmaskinit &= ~mask;
2794 rnp->qsmask &= ~mask;
2795 if (rnp->qsmaskinit) {
2796 raw_spin_unlock_rcu_node(rnp);
2797 /* irqs remain disabled. */
2800 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2805 * The CPU has been completely removed, and some other CPU is reporting
2806 * this fact from process context. Do the remainder of the cleanup,
2807 * including orphaning the outgoing CPU's RCU callbacks, and also
2808 * adopting them. There can only be one CPU hotplug operation at a time,
2809 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2811 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2813 unsigned long flags;
2814 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2815 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2817 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2820 /* Adjust any no-longer-needed kthreads. */
2821 rcu_boost_kthread_setaffinity(rnp, -1);
2823 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2824 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2825 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2826 rcu_adopt_orphan_cbs(rsp, flags);
2827 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2829 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2830 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2831 cpu, rdp->qlen, rdp->nxtlist);
2835 * Invoke any RCU callbacks that have made it to the end of their grace
2836 * period. Thottle as specified by rdp->blimit.
2838 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2840 unsigned long flags;
2841 struct rcu_head *next, *list, **tail;
2842 long bl, count, count_lazy;
2845 /* If no callbacks are ready, just return. */
2846 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2847 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2848 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2849 need_resched(), is_idle_task(current),
2850 rcu_is_callbacks_kthread());
2855 * Extract the list of ready callbacks, disabling to prevent
2856 * races with call_rcu() from interrupt handlers.
2858 local_irq_save(flags);
2859 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2861 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2862 list = rdp->nxtlist;
2863 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2864 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2865 tail = rdp->nxttail[RCU_DONE_TAIL];
2866 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2867 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2868 rdp->nxttail[i] = &rdp->nxtlist;
2869 local_irq_restore(flags);
2871 /* Invoke callbacks. */
2872 count = count_lazy = 0;
2876 debug_rcu_head_unqueue(list);
2877 if (__rcu_reclaim(rsp->name, list))
2880 /* Stop only if limit reached and CPU has something to do. */
2881 if (++count >= bl &&
2883 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2887 local_irq_save(flags);
2888 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2889 is_idle_task(current),
2890 rcu_is_callbacks_kthread());
2892 /* Update count, and requeue any remaining callbacks. */
2894 *tail = rdp->nxtlist;
2895 rdp->nxtlist = list;
2896 for (i = 0; i < RCU_NEXT_SIZE; i++)
2897 if (&rdp->nxtlist == rdp->nxttail[i])
2898 rdp->nxttail[i] = tail;
2902 smp_mb(); /* List handling before counting for rcu_barrier(). */
2903 rdp->qlen_lazy -= count_lazy;
2904 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2905 rdp->n_cbs_invoked += count;
2907 /* Reinstate batch limit if we have worked down the excess. */
2908 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2909 rdp->blimit = blimit;
2911 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2912 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2913 rdp->qlen_last_fqs_check = 0;
2914 rdp->n_force_qs_snap = rsp->n_force_qs;
2915 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2916 rdp->qlen_last_fqs_check = rdp->qlen;
2917 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2919 local_irq_restore(flags);
2921 /* Re-invoke RCU core processing if there are callbacks remaining. */
2922 if (cpu_has_callbacks_ready_to_invoke(rdp))
2927 * Check to see if this CPU is in a non-context-switch quiescent state
2928 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2929 * Also schedule RCU core processing.
2931 * This function must be called from hardirq context. It is normally
2932 * invoked from the scheduling-clock interrupt.
2934 void rcu_check_callbacks(int user)
2936 trace_rcu_utilization(TPS("Start scheduler-tick"));
2937 increment_cpu_stall_ticks();
2938 if (user || rcu_is_cpu_rrupt_from_idle()) {
2941 * Get here if this CPU took its interrupt from user
2942 * mode or from the idle loop, and if this is not a
2943 * nested interrupt. In this case, the CPU is in
2944 * a quiescent state, so note it.
2946 * No memory barrier is required here because both
2947 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2948 * variables that other CPUs neither access nor modify,
2949 * at least not while the corresponding CPU is online.
2955 } else if (!in_softirq()) {
2958 * Get here if this CPU did not take its interrupt from
2959 * softirq, in other words, if it is not interrupting
2960 * a rcu_bh read-side critical section. This is an _bh
2961 * critical section, so note it.
2966 rcu_preempt_check_callbacks();
2970 rcu_note_voluntary_context_switch(current);
2971 trace_rcu_utilization(TPS("End scheduler-tick"));
2975 * Scan the leaf rcu_node structures, processing dyntick state for any that
2976 * have not yet encountered a quiescent state, using the function specified.
2977 * Also initiate boosting for any threads blocked on the root rcu_node.
2979 * The caller must have suppressed start of new grace periods.
2981 static void force_qs_rnp(struct rcu_state *rsp,
2982 int (*f)(struct rcu_data *rsp, bool *isidle,
2983 unsigned long *maxj),
2984 bool *isidle, unsigned long *maxj)
2987 unsigned long flags;
2989 struct rcu_node *rnp;
2991 rcu_for_each_leaf_node(rsp, rnp) {
2992 cond_resched_rcu_qs();
2994 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2995 if (rnp->qsmask == 0) {
2996 if (rcu_state_p == &rcu_sched_state ||
2997 rsp != rcu_state_p ||
2998 rcu_preempt_blocked_readers_cgp(rnp)) {
3000 * No point in scanning bits because they
3001 * are all zero. But we might need to
3002 * priority-boost blocked readers.
3004 rcu_initiate_boost(rnp, flags);
3005 /* rcu_initiate_boost() releases rnp->lock */
3009 (rnp->parent->qsmask & rnp->grpmask)) {
3011 * Race between grace-period
3012 * initialization and task exiting RCU
3013 * read-side critical section: Report.
3015 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
3016 /* rcu_report_unblock_qs_rnp() rlses ->lock */
3020 for_each_leaf_node_possible_cpu(rnp, cpu) {
3021 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
3022 if ((rnp->qsmask & bit) != 0) {
3023 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
3028 /* Idle/offline CPUs, report (releases rnp->lock. */
3029 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
3031 /* Nothing to do here, so just drop the lock. */
3032 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3038 * Force quiescent states on reluctant CPUs, and also detect which
3039 * CPUs are in dyntick-idle mode.
3041 static void force_quiescent_state(struct rcu_state *rsp)
3043 unsigned long flags;
3045 struct rcu_node *rnp;
3046 struct rcu_node *rnp_old = NULL;
3048 /* Funnel through hierarchy to reduce memory contention. */
3049 rnp = __this_cpu_read(rsp->rda->mynode);
3050 for (; rnp != NULL; rnp = rnp->parent) {
3051 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
3052 !raw_spin_trylock(&rnp->fqslock);
3053 if (rnp_old != NULL)
3054 raw_spin_unlock(&rnp_old->fqslock);
3056 rsp->n_force_qs_lh++;
3061 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
3063 /* Reached the root of the rcu_node tree, acquire lock. */
3064 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
3065 raw_spin_unlock(&rnp_old->fqslock);
3066 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
3067 rsp->n_force_qs_lh++;
3068 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
3069 return; /* Someone beat us to it. */
3071 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
3072 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
3073 rcu_gp_kthread_wake(rsp);
3077 * This does the RCU core processing work for the specified rcu_state
3078 * and rcu_data structures. This may be called only from the CPU to
3079 * whom the rdp belongs.
3082 __rcu_process_callbacks(struct rcu_state *rsp)
3084 unsigned long flags;
3086 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3088 WARN_ON_ONCE(rdp->beenonline == 0);
3090 /* Update RCU state based on any recent quiescent states. */
3091 rcu_check_quiescent_state(rsp, rdp);
3093 /* Does this CPU require a not-yet-started grace period? */
3094 local_irq_save(flags);
3095 if (cpu_needs_another_gp(rsp, rdp)) {
3096 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
3097 needwake = rcu_start_gp(rsp);
3098 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
3100 rcu_gp_kthread_wake(rsp);
3102 local_irq_restore(flags);
3105 /* If there are callbacks ready, invoke them. */
3106 if (cpu_has_callbacks_ready_to_invoke(rdp))
3107 invoke_rcu_callbacks(rsp, rdp);
3109 /* Do any needed deferred wakeups of rcuo kthreads. */
3110 do_nocb_deferred_wakeup(rdp);
3114 * Do RCU core processing for the current CPU.
3116 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
3118 struct rcu_state *rsp;
3120 if (cpu_is_offline(smp_processor_id()))
3122 trace_rcu_utilization(TPS("Start RCU core"));
3123 for_each_rcu_flavor(rsp)
3124 __rcu_process_callbacks(rsp);
3125 trace_rcu_utilization(TPS("End RCU core"));
3129 * Schedule RCU callback invocation. If the specified type of RCU
3130 * does not support RCU priority boosting, just do a direct call,
3131 * otherwise wake up the per-CPU kernel kthread. Note that because we
3132 * are running on the current CPU with softirqs disabled, the
3133 * rcu_cpu_kthread_task cannot disappear out from under us.
3135 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
3137 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
3139 if (likely(!rsp->boost)) {
3140 rcu_do_batch(rsp, rdp);
3143 invoke_rcu_callbacks_kthread();
3146 static void invoke_rcu_core(void)
3148 if (cpu_online(smp_processor_id()))
3149 raise_softirq(RCU_SOFTIRQ);
3153 * Handle any core-RCU processing required by a call_rcu() invocation.
3155 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
3156 struct rcu_head *head, unsigned long flags)
3161 * If called from an extended quiescent state, invoke the RCU
3162 * core in order to force a re-evaluation of RCU's idleness.
3164 if (!rcu_is_watching())
3167 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3168 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3172 * Force the grace period if too many callbacks or too long waiting.
3173 * Enforce hysteresis, and don't invoke force_quiescent_state()
3174 * if some other CPU has recently done so. Also, don't bother
3175 * invoking force_quiescent_state() if the newly enqueued callback
3176 * is the only one waiting for a grace period to complete.
3178 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
3180 /* Are we ignoring a completed grace period? */
3181 note_gp_changes(rsp, rdp);
3183 /* Start a new grace period if one not already started. */
3184 if (!rcu_gp_in_progress(rsp)) {
3185 struct rcu_node *rnp_root = rcu_get_root(rsp);
3187 raw_spin_lock_rcu_node(rnp_root);
3188 needwake = rcu_start_gp(rsp);
3189 raw_spin_unlock_rcu_node(rnp_root);
3191 rcu_gp_kthread_wake(rsp);
3193 /* Give the grace period a kick. */
3194 rdp->blimit = LONG_MAX;
3195 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3196 *rdp->nxttail[RCU_DONE_TAIL] != head)
3197 force_quiescent_state(rsp);
3198 rdp->n_force_qs_snap = rsp->n_force_qs;
3199 rdp->qlen_last_fqs_check = rdp->qlen;
3205 * RCU callback function to leak a callback.
3207 static void rcu_leak_callback(struct rcu_head *rhp)
3212 * Helper function for call_rcu() and friends. The cpu argument will
3213 * normally be -1, indicating "currently running CPU". It may specify
3214 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3215 * is expected to specify a CPU.
3218 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3219 struct rcu_state *rsp, int cpu, bool lazy)
3221 unsigned long flags;
3222 struct rcu_data *rdp;
3224 /* Misaligned rcu_head! */
3225 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3227 if (debug_rcu_head_queue(head)) {
3228 /* Probable double call_rcu(), so leak the callback. */
3229 WRITE_ONCE(head->func, rcu_leak_callback);
3230 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3235 local_irq_save(flags);
3236 rdp = this_cpu_ptr(rsp->rda);
3238 /* Add the callback to our list. */
3239 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3243 rdp = per_cpu_ptr(rsp->rda, cpu);
3244 if (likely(rdp->mynode)) {
3245 /* Post-boot, so this should be for a no-CBs CPU. */
3246 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3247 WARN_ON_ONCE(offline);
3248 /* Offline CPU, _call_rcu() illegal, leak callback. */
3249 local_irq_restore(flags);
3253 * Very early boot, before rcu_init(). Initialize if needed
3254 * and then drop through to queue the callback.
3257 WARN_ON_ONCE(!rcu_is_watching());
3258 if (!likely(rdp->nxtlist))
3259 init_default_callback_list(rdp);
3261 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3265 rcu_idle_count_callbacks_posted();
3266 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3267 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3268 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3270 if (__is_kfree_rcu_offset((unsigned long)func))
3271 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3272 rdp->qlen_lazy, rdp->qlen);
3274 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3276 /* Go handle any RCU core processing required. */
3277 __call_rcu_core(rsp, rdp, head, flags);
3278 local_irq_restore(flags);
3282 * Queue an RCU-sched callback for invocation after a grace period.
3284 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3286 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3288 EXPORT_SYMBOL_GPL(call_rcu_sched);
3291 * Queue an RCU callback for invocation after a quicker grace period.
3293 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3295 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3297 EXPORT_SYMBOL_GPL(call_rcu_bh);
3300 * Queue an RCU callback for lazy invocation after a grace period.
3301 * This will likely be later named something like "call_rcu_lazy()",
3302 * but this change will require some way of tagging the lazy RCU
3303 * callbacks in the list of pending callbacks. Until then, this
3304 * function may only be called from __kfree_rcu().
3306 void kfree_call_rcu(struct rcu_head *head,
3307 rcu_callback_t func)
3309 __call_rcu(head, func, rcu_state_p, -1, 1);
3311 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3314 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3315 * any blocking grace-period wait automatically implies a grace period
3316 * if there is only one CPU online at any point time during execution
3317 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3318 * occasionally incorrectly indicate that there are multiple CPUs online
3319 * when there was in fact only one the whole time, as this just adds
3320 * some overhead: RCU still operates correctly.
3322 static inline int rcu_blocking_is_gp(void)
3326 might_sleep(); /* Check for RCU read-side critical section. */
3328 ret = num_online_cpus() <= 1;
3334 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3336 * Control will return to the caller some time after a full rcu-sched
3337 * grace period has elapsed, in other words after all currently executing
3338 * rcu-sched read-side critical sections have completed. These read-side
3339 * critical sections are delimited by rcu_read_lock_sched() and
3340 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3341 * local_irq_disable(), and so on may be used in place of
3342 * rcu_read_lock_sched().
3344 * This means that all preempt_disable code sequences, including NMI and
3345 * non-threaded hardware-interrupt handlers, in progress on entry will
3346 * have completed before this primitive returns. However, this does not
3347 * guarantee that softirq handlers will have completed, since in some
3348 * kernels, these handlers can run in process context, and can block.
3350 * Note that this guarantee implies further memory-ordering guarantees.
3351 * On systems with more than one CPU, when synchronize_sched() returns,
3352 * each CPU is guaranteed to have executed a full memory barrier since the
3353 * end of its last RCU-sched read-side critical section whose beginning
3354 * preceded the call to synchronize_sched(). In addition, each CPU having
3355 * an RCU read-side critical section that extends beyond the return from
3356 * synchronize_sched() is guaranteed to have executed a full memory barrier
3357 * after the beginning of synchronize_sched() and before the beginning of
3358 * that RCU read-side critical section. Note that these guarantees include
3359 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3360 * that are executing in the kernel.
3362 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3363 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3364 * to have executed a full memory barrier during the execution of
3365 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3366 * again only if the system has more than one CPU).
3368 * This primitive provides the guarantees made by the (now removed)
3369 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3370 * guarantees that rcu_read_lock() sections will have completed.
3371 * In "classic RCU", these two guarantees happen to be one and
3372 * the same, but can differ in realtime RCU implementations.
3374 void synchronize_sched(void)
3376 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3377 lock_is_held(&rcu_lock_map) ||
3378 lock_is_held(&rcu_sched_lock_map),
3379 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3380 if (rcu_blocking_is_gp())
3382 if (rcu_gp_is_expedited())
3383 synchronize_sched_expedited();
3385 wait_rcu_gp(call_rcu_sched);
3387 EXPORT_SYMBOL_GPL(synchronize_sched);
3390 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3392 * Control will return to the caller some time after a full rcu_bh grace
3393 * period has elapsed, in other words after all currently executing rcu_bh
3394 * read-side critical sections have completed. RCU read-side critical
3395 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3396 * and may be nested.
3398 * See the description of synchronize_sched() for more detailed information
3399 * on memory ordering guarantees.
3401 void synchronize_rcu_bh(void)
3403 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3404 lock_is_held(&rcu_lock_map) ||
3405 lock_is_held(&rcu_sched_lock_map),
3406 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3407 if (rcu_blocking_is_gp())
3409 if (rcu_gp_is_expedited())
3410 synchronize_rcu_bh_expedited();
3412 wait_rcu_gp(call_rcu_bh);
3414 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3417 * get_state_synchronize_rcu - Snapshot current RCU state
3419 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3420 * to determine whether or not a full grace period has elapsed in the
3423 unsigned long get_state_synchronize_rcu(void)
3426 * Any prior manipulation of RCU-protected data must happen
3427 * before the load from ->gpnum.
3432 * Make sure this load happens before the purportedly
3433 * time-consuming work between get_state_synchronize_rcu()
3434 * and cond_synchronize_rcu().
3436 return smp_load_acquire(&rcu_state_p->gpnum);
3438 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3441 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3443 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3445 * If a full RCU grace period has elapsed since the earlier call to
3446 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3447 * synchronize_rcu() to wait for a full grace period.
3449 * Yes, this function does not take counter wrap into account. But
3450 * counter wrap is harmless. If the counter wraps, we have waited for
3451 * more than 2 billion grace periods (and way more on a 64-bit system!),
3452 * so waiting for one additional grace period should be just fine.
3454 void cond_synchronize_rcu(unsigned long oldstate)
3456 unsigned long newstate;
3459 * Ensure that this load happens before any RCU-destructive
3460 * actions the caller might carry out after we return.
3462 newstate = smp_load_acquire(&rcu_state_p->completed);
3463 if (ULONG_CMP_GE(oldstate, newstate))
3466 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3469 * get_state_synchronize_sched - Snapshot current RCU-sched state
3471 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3472 * to determine whether or not a full grace period has elapsed in the
3475 unsigned long get_state_synchronize_sched(void)
3478 * Any prior manipulation of RCU-protected data must happen
3479 * before the load from ->gpnum.
3484 * Make sure this load happens before the purportedly
3485 * time-consuming work between get_state_synchronize_sched()
3486 * and cond_synchronize_sched().
3488 return smp_load_acquire(&rcu_sched_state.gpnum);
3490 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3493 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3495 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3497 * If a full RCU-sched grace period has elapsed since the earlier call to
3498 * get_state_synchronize_sched(), just return. Otherwise, invoke
3499 * synchronize_sched() to wait for a full grace period.
3501 * Yes, this function does not take counter wrap into account. But
3502 * counter wrap is harmless. If the counter wraps, we have waited for
3503 * more than 2 billion grace periods (and way more on a 64-bit system!),
3504 * so waiting for one additional grace period should be just fine.
3506 void cond_synchronize_sched(unsigned long oldstate)
3508 unsigned long newstate;
3511 * Ensure that this load happens before any RCU-destructive
3512 * actions the caller might carry out after we return.
3514 newstate = smp_load_acquire(&rcu_sched_state.completed);
3515 if (ULONG_CMP_GE(oldstate, newstate))
3516 synchronize_sched();
3518 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3520 /* Adjust sequence number for start of update-side operation. */
3521 static void rcu_seq_start(unsigned long *sp)
3523 WRITE_ONCE(*sp, *sp + 1);
3524 smp_mb(); /* Ensure update-side operation after counter increment. */
3525 WARN_ON_ONCE(!(*sp & 0x1));
3528 /* Adjust sequence number for end of update-side operation. */
3529 static void rcu_seq_end(unsigned long *sp)
3531 smp_mb(); /* Ensure update-side operation before counter increment. */
3532 WRITE_ONCE(*sp, *sp + 1);
3533 WARN_ON_ONCE(*sp & 0x1);
3536 /* Take a snapshot of the update side's sequence number. */
3537 static unsigned long rcu_seq_snap(unsigned long *sp)
3541 s = (READ_ONCE(*sp) + 3) & ~0x1;
3542 smp_mb(); /* Above access must not bleed into critical section. */
3547 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3548 * full update-side operation has occurred.
3550 static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3552 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3556 * Check to see if there is any immediate RCU-related work to be done
3557 * by the current CPU, for the specified type of RCU, returning 1 if so.
3558 * The checks are in order of increasing expense: checks that can be
3559 * carried out against CPU-local state are performed first. However,
3560 * we must check for CPU stalls first, else we might not get a chance.
3562 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3564 struct rcu_node *rnp = rdp->mynode;
3566 rdp->n_rcu_pending++;
3568 /* Check for CPU stalls, if enabled. */
3569 check_cpu_stall(rsp, rdp);
3571 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3572 if (rcu_nohz_full_cpu(rsp))
3575 /* Is the RCU core waiting for a quiescent state from this CPU? */
3576 if (rcu_scheduler_fully_active &&
3577 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3578 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3579 rdp->n_rp_core_needs_qs++;
3580 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3581 rdp->n_rp_report_qs++;
3585 /* Does this CPU have callbacks ready to invoke? */
3586 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3587 rdp->n_rp_cb_ready++;
3591 /* Has RCU gone idle with this CPU needing another grace period? */
3592 if (cpu_needs_another_gp(rsp, rdp)) {
3593 rdp->n_rp_cpu_needs_gp++;
3597 /* Has another RCU grace period completed? */
3598 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3599 rdp->n_rp_gp_completed++;
3603 /* Has a new RCU grace period started? */
3604 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3605 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3606 rdp->n_rp_gp_started++;
3610 /* Does this CPU need a deferred NOCB wakeup? */
3611 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3612 rdp->n_rp_nocb_defer_wakeup++;
3617 rdp->n_rp_need_nothing++;
3622 * Check to see if there is any immediate RCU-related work to be done
3623 * by the current CPU, returning 1 if so. This function is part of the
3624 * RCU implementation; it is -not- an exported member of the RCU API.
3626 static int rcu_pending(void)
3628 struct rcu_state *rsp;
3630 for_each_rcu_flavor(rsp)
3631 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3637 * Return true if the specified CPU has any callback. If all_lazy is
3638 * non-NULL, store an indication of whether all callbacks are lazy.
3639 * (If there are no callbacks, all of them are deemed to be lazy.)
3641 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3645 struct rcu_data *rdp;
3646 struct rcu_state *rsp;
3648 for_each_rcu_flavor(rsp) {
3649 rdp = this_cpu_ptr(rsp->rda);
3653 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3664 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3665 * the compiler is expected to optimize this away.
3667 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3668 int cpu, unsigned long done)
3670 trace_rcu_barrier(rsp->name, s, cpu,
3671 atomic_read(&rsp->barrier_cpu_count), done);
3675 * RCU callback function for _rcu_barrier(). If we are last, wake
3676 * up the task executing _rcu_barrier().
3678 static void rcu_barrier_callback(struct rcu_head *rhp)
3680 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3681 struct rcu_state *rsp = rdp->rsp;
3683 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3684 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
3685 complete(&rsp->barrier_completion);
3687 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
3692 * Called with preemption disabled, and from cross-cpu IRQ context.
3694 static void rcu_barrier_func(void *type)
3696 struct rcu_state *rsp = type;
3697 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3699 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
3700 atomic_inc(&rsp->barrier_cpu_count);
3701 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3705 * Orchestrate the specified type of RCU barrier, waiting for all
3706 * RCU callbacks of the specified type to complete.
3708 static void _rcu_barrier(struct rcu_state *rsp)
3711 struct rcu_data *rdp;
3712 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3714 _rcu_barrier_trace(rsp, "Begin", -1, s);
3716 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3717 mutex_lock(&rsp->barrier_mutex);
3719 /* Did someone else do our work for us? */
3720 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3721 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
3722 smp_mb(); /* caller's subsequent code after above check. */
3723 mutex_unlock(&rsp->barrier_mutex);
3727 /* Mark the start of the barrier operation. */
3728 rcu_seq_start(&rsp->barrier_sequence);
3729 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
3732 * Initialize the count to one rather than to zero in order to
3733 * avoid a too-soon return to zero in case of a short grace period
3734 * (or preemption of this task). Exclude CPU-hotplug operations
3735 * to ensure that no offline CPU has callbacks queued.
3737 init_completion(&rsp->barrier_completion);
3738 atomic_set(&rsp->barrier_cpu_count, 1);
3742 * Force each CPU with callbacks to register a new callback.
3743 * When that callback is invoked, we will know that all of the
3744 * corresponding CPU's preceding callbacks have been invoked.
3746 for_each_possible_cpu(cpu) {
3747 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3749 rdp = per_cpu_ptr(rsp->rda, cpu);
3750 if (rcu_is_nocb_cpu(cpu)) {
3751 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3752 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3753 rsp->barrier_sequence);
3755 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3756 rsp->barrier_sequence);
3757 smp_mb__before_atomic();
3758 atomic_inc(&rsp->barrier_cpu_count);
3759 __call_rcu(&rdp->barrier_head,
3760 rcu_barrier_callback, rsp, cpu, 0);
3762 } else if (READ_ONCE(rdp->qlen)) {
3763 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3764 rsp->barrier_sequence);
3765 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3767 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3768 rsp->barrier_sequence);
3774 * Now that we have an rcu_barrier_callback() callback on each
3775 * CPU, and thus each counted, remove the initial count.
3777 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3778 complete(&rsp->barrier_completion);
3780 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3781 wait_for_completion(&rsp->barrier_completion);
3783 /* Mark the end of the barrier operation. */
3784 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
3785 rcu_seq_end(&rsp->barrier_sequence);
3787 /* Other rcu_barrier() invocations can now safely proceed. */
3788 mutex_unlock(&rsp->barrier_mutex);
3792 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3794 void rcu_barrier_bh(void)
3796 _rcu_barrier(&rcu_bh_state);
3798 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3801 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3803 void rcu_barrier_sched(void)
3805 _rcu_barrier(&rcu_sched_state);
3807 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3810 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3811 * first CPU in a given leaf rcu_node structure coming online. The caller
3812 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3815 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3818 struct rcu_node *rnp = rnp_leaf;
3821 mask = rnp->grpmask;
3825 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3826 rnp->qsmaskinit |= mask;
3827 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3832 * Do boot-time initialization of a CPU's per-CPU RCU data.
3835 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3837 unsigned long flags;
3838 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3839 struct rcu_node *rnp = rcu_get_root(rsp);
3841 /* Set up local state, ensuring consistent view of global state. */
3842 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3843 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3844 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3845 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3846 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3849 rcu_boot_init_nocb_percpu_data(rdp);
3850 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3854 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3855 * offline event can be happening at a given time. Note also that we
3856 * can accept some slop in the rsp->completed access due to the fact
3857 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3860 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3862 unsigned long flags;
3863 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3864 struct rcu_node *rnp = rcu_get_root(rsp);
3866 /* Set up local state, ensuring consistent view of global state. */
3867 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3868 rdp->qlen_last_fqs_check = 0;
3869 rdp->n_force_qs_snap = rsp->n_force_qs;
3870 rdp->blimit = blimit;
3872 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3873 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3874 rcu_sysidle_init_percpu_data(rdp->dynticks);
3875 rcu_dynticks_eqs_online();
3876 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3879 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3880 * propagation up the rcu_node tree will happen at the beginning
3881 * of the next grace period.
3884 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3885 if (!rdp->beenonline)
3886 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
3887 rdp->beenonline = true; /* We have now been online. */
3888 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3889 rdp->completed = rnp->completed;
3890 rdp->cpu_no_qs.b.norm = true;
3891 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
3892 rdp->core_needs_qs = false;
3893 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3894 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3897 int rcutree_prepare_cpu(unsigned int cpu)
3899 struct rcu_state *rsp;
3901 for_each_rcu_flavor(rsp)
3902 rcu_init_percpu_data(cpu, rsp);
3904 rcu_prepare_kthreads(cpu);
3905 rcu_spawn_all_nocb_kthreads(cpu);
3910 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3912 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3914 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3917 int rcutree_online_cpu(unsigned int cpu)
3919 sync_sched_exp_online_cleanup(cpu);
3920 rcutree_affinity_setting(cpu, -1);
3924 int rcutree_offline_cpu(unsigned int cpu)
3926 rcutree_affinity_setting(cpu, cpu);
3931 int rcutree_dying_cpu(unsigned int cpu)
3933 struct rcu_state *rsp;
3935 for_each_rcu_flavor(rsp)
3936 rcu_cleanup_dying_cpu(rsp);
3940 int rcutree_dead_cpu(unsigned int cpu)
3942 struct rcu_state *rsp;
3944 for_each_rcu_flavor(rsp) {
3945 rcu_cleanup_dead_cpu(cpu, rsp);
3946 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3952 * Mark the specified CPU as being online so that subsequent grace periods
3953 * (both expedited and normal) will wait on it. Note that this means that
3954 * incoming CPUs are not allowed to use RCU read-side critical sections
3955 * until this function is called. Failing to observe this restriction
3956 * will result in lockdep splats.
3958 void rcu_cpu_starting(unsigned int cpu)
3960 unsigned long flags;
3962 struct rcu_data *rdp;
3963 struct rcu_node *rnp;
3964 struct rcu_state *rsp;
3966 for_each_rcu_flavor(rsp) {
3967 rdp = per_cpu_ptr(rsp->rda, cpu);
3969 mask = rdp->grpmask;
3970 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3971 rnp->qsmaskinitnext |= mask;
3972 rnp->expmaskinitnext |= mask;
3973 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3977 #ifdef CONFIG_HOTPLUG_CPU
3979 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3980 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3982 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3983 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3986 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3988 unsigned long flags;
3990 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3991 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3993 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3994 mask = rdp->grpmask;
3995 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3996 rnp->qsmaskinitnext &= ~mask;
3997 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4000 void rcu_report_dead(unsigned int cpu)
4002 struct rcu_state *rsp;
4004 /* QS for any half-done expedited RCU-sched GP. */
4006 rcu_report_exp_rdp(&rcu_sched_state,
4007 this_cpu_ptr(rcu_sched_state.rda), true);
4009 for_each_rcu_flavor(rsp)
4010 rcu_cleanup_dying_idle_cpu(cpu, rsp);
4014 static int rcu_pm_notify(struct notifier_block *self,
4015 unsigned long action, void *hcpu)
4018 case PM_HIBERNATION_PREPARE:
4019 case PM_SUSPEND_PREPARE:
4020 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4023 case PM_POST_HIBERNATION:
4024 case PM_POST_SUSPEND:
4025 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4026 rcu_unexpedite_gp();
4035 * Spawn the kthreads that handle each RCU flavor's grace periods.
4037 static int __init rcu_spawn_gp_kthread(void)
4039 unsigned long flags;
4040 int kthread_prio_in = kthread_prio;
4041 struct rcu_node *rnp;
4042 struct rcu_state *rsp;
4043 struct sched_param sp;
4044 struct task_struct *t;
4046 /* Force priority into range. */
4047 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4049 else if (kthread_prio < 0)
4051 else if (kthread_prio > 99)
4053 if (kthread_prio != kthread_prio_in)
4054 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4055 kthread_prio, kthread_prio_in);
4057 rcu_scheduler_fully_active = 1;
4058 for_each_rcu_flavor(rsp) {
4059 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4061 rnp = rcu_get_root(rsp);
4062 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4063 rsp->gp_kthread = t;
4065 sp.sched_priority = kthread_prio;
4066 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4068 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4071 rcu_spawn_nocb_kthreads();
4072 rcu_spawn_boost_kthreads();
4075 early_initcall(rcu_spawn_gp_kthread);
4078 * This function is invoked towards the end of the scheduler's
4079 * initialization process. Before this is called, the idle task might
4080 * contain synchronous grace-period primitives (during which time, this idle
4081 * task is booting the system, and such primitives are no-ops). After this
4082 * function is called, any synchronous grace-period primitives are run as
4083 * expedited, with the requesting task driving the grace period forward.
4084 * A later core_initcall() rcu_exp_runtime_mode() will switch to full
4085 * runtime RCU functionality.
4087 void rcu_scheduler_starting(void)
4089 WARN_ON(num_online_cpus() != 1);
4090 WARN_ON(nr_context_switches() > 0);
4091 rcu_test_sync_prims();
4092 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4093 rcu_test_sync_prims();
4097 * Compute the per-level fanout, either using the exact fanout specified
4098 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4100 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4104 if (rcu_fanout_exact) {
4105 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4106 for (i = rcu_num_lvls - 2; i >= 0; i--)
4107 levelspread[i] = RCU_FANOUT;
4113 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4115 levelspread[i] = (cprv + ccur - 1) / ccur;
4122 * Helper function for rcu_init() that initializes one rcu_state structure.
4124 static void __init rcu_init_one(struct rcu_state *rsp)
4126 static const char * const buf[] = RCU_NODE_NAME_INIT;
4127 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4128 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4129 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4130 static u8 fl_mask = 0x1;
4132 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4133 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4137 struct rcu_node *rnp;
4139 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4141 /* Silence gcc 4.8 false positive about array index out of range. */
4142 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4143 panic("rcu_init_one: rcu_num_lvls out of range");
4145 /* Initialize the level-tracking arrays. */
4147 for (i = 0; i < rcu_num_lvls; i++)
4148 levelcnt[i] = num_rcu_lvl[i];
4149 for (i = 1; i < rcu_num_lvls; i++)
4150 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4151 rcu_init_levelspread(levelspread, levelcnt);
4152 rsp->flavor_mask = fl_mask;
4155 /* Initialize the elements themselves, starting from the leaves. */
4157 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4158 cpustride *= levelspread[i];
4159 rnp = rsp->level[i];
4160 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4161 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4162 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4163 &rcu_node_class[i], buf[i]);
4164 raw_spin_lock_init(&rnp->fqslock);
4165 lockdep_set_class_and_name(&rnp->fqslock,
4166 &rcu_fqs_class[i], fqs[i]);
4167 rnp->gpnum = rsp->gpnum;
4168 rnp->completed = rsp->completed;
4170 rnp->qsmaskinit = 0;
4171 rnp->grplo = j * cpustride;
4172 rnp->grphi = (j + 1) * cpustride - 1;
4173 if (rnp->grphi >= nr_cpu_ids)
4174 rnp->grphi = nr_cpu_ids - 1;
4180 rnp->grpnum = j % levelspread[i - 1];
4181 rnp->grpmask = 1UL << rnp->grpnum;
4182 rnp->parent = rsp->level[i - 1] +
4183 j / levelspread[i - 1];
4186 INIT_LIST_HEAD(&rnp->blkd_tasks);
4187 rcu_init_one_nocb(rnp);
4188 init_waitqueue_head(&rnp->exp_wq[0]);
4189 init_waitqueue_head(&rnp->exp_wq[1]);
4190 init_waitqueue_head(&rnp->exp_wq[2]);
4191 init_waitqueue_head(&rnp->exp_wq[3]);
4192 spin_lock_init(&rnp->exp_lock);
4196 init_swait_queue_head(&rsp->gp_wq);
4197 init_swait_queue_head(&rsp->expedited_wq);
4198 rnp = rsp->level[rcu_num_lvls - 1];
4199 for_each_possible_cpu(i) {
4200 while (i > rnp->grphi)
4202 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4203 rcu_boot_init_percpu_data(i, rsp);
4205 list_add(&rsp->flavors, &rcu_struct_flavors);
4209 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4210 * replace the definitions in tree.h because those are needed to size
4211 * the ->node array in the rcu_state structure.
4213 static void __init rcu_init_geometry(void)
4217 int rcu_capacity[RCU_NUM_LVLS];
4220 * Initialize any unspecified boot parameters.
4221 * The default values of jiffies_till_first_fqs and
4222 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4223 * value, which is a function of HZ, then adding one for each
4224 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4226 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4227 if (jiffies_till_first_fqs == ULONG_MAX)
4228 jiffies_till_first_fqs = d;
4229 if (jiffies_till_next_fqs == ULONG_MAX)
4230 jiffies_till_next_fqs = d;
4232 /* If the compile-time values are accurate, just leave. */
4233 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4234 nr_cpu_ids == NR_CPUS)
4236 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4237 rcu_fanout_leaf, nr_cpu_ids);
4240 * The boot-time rcu_fanout_leaf parameter must be at least two
4241 * and cannot exceed the number of bits in the rcu_node masks.
4242 * Complain and fall back to the compile-time values if this
4243 * limit is exceeded.
4245 if (rcu_fanout_leaf < 2 ||
4246 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4247 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4253 * Compute number of nodes that can be handled an rcu_node tree
4254 * with the given number of levels.
4256 rcu_capacity[0] = rcu_fanout_leaf;
4257 for (i = 1; i < RCU_NUM_LVLS; i++)
4258 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4261 * The tree must be able to accommodate the configured number of CPUs.
4262 * If this limit is exceeded, fall back to the compile-time values.
4264 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4265 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4270 /* Calculate the number of levels in the tree. */
4271 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4273 rcu_num_lvls = i + 1;
4275 /* Calculate the number of rcu_nodes at each level of the tree. */
4276 for (i = 0; i < rcu_num_lvls; i++) {
4277 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4278 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4281 /* Calculate the total number of rcu_node structures. */
4283 for (i = 0; i < rcu_num_lvls; i++)
4284 rcu_num_nodes += num_rcu_lvl[i];
4288 * Dump out the structure of the rcu_node combining tree associated
4289 * with the rcu_state structure referenced by rsp.
4291 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4294 struct rcu_node *rnp;
4296 pr_info("rcu_node tree layout dump\n");
4298 rcu_for_each_node_breadth_first(rsp, rnp) {
4299 if (rnp->level != level) {
4304 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4309 void __init rcu_init(void)
4313 rcu_early_boot_tests();
4315 rcu_bootup_announce();
4316 rcu_init_geometry();
4317 rcu_init_one(&rcu_bh_state);
4318 rcu_init_one(&rcu_sched_state);
4320 rcu_dump_rcu_node_tree(&rcu_sched_state);
4321 __rcu_init_preempt();
4322 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4325 * We don't need protection against CPU-hotplug here because
4326 * this is called early in boot, before either interrupts
4327 * or the scheduler are operational.
4329 pm_notifier(rcu_pm_notify, 0);
4330 for_each_online_cpu(cpu) {
4331 rcutree_prepare_cpu(cpu);
4332 rcu_cpu_starting(cpu);
4336 #include "tree_exp.h"
4337 #include "tree_plugin.h"