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 -
31 #define pr_fmt(fmt) "rcu: " fmt
33 #include <linux/types.h>
34 #include <linux/kernel.h>
35 #include <linux/init.h>
36 #include <linux/spinlock.h>
37 #include <linux/smp.h>
38 #include <linux/rcupdate_wait.h>
39 #include <linux/interrupt.h>
40 #include <linux/sched.h>
41 #include <linux/sched/debug.h>
42 #include <linux/nmi.h>
43 #include <linux/atomic.h>
44 #include <linux/bitops.h>
45 #include <linux/export.h>
46 #include <linux/completion.h>
47 #include <linux/moduleparam.h>
48 #include <linux/percpu.h>
49 #include <linux/notifier.h>
50 #include <linux/cpu.h>
51 #include <linux/mutex.h>
52 #include <linux/time.h>
53 #include <linux/kernel_stat.h>
54 #include <linux/wait.h>
55 #include <linux/kthread.h>
56 #include <uapi/linux/sched/types.h>
57 #include <linux/prefetch.h>
58 #include <linux/delay.h>
59 #include <linux/stop_machine.h>
60 #include <linux/random.h>
61 #include <linux/trace_events.h>
62 #include <linux/suspend.h>
63 #include <linux/ftrace.h>
68 #ifdef MODULE_PARAM_PREFIX
69 #undef MODULE_PARAM_PREFIX
71 #define MODULE_PARAM_PREFIX "rcutree."
73 /* Data structures. */
76 * In order to export the rcu_state name to the tracing tools, it
77 * needs to be added in the __tracepoint_string section.
78 * This requires defining a separate variable tp_<sname>_varname
79 * that points to the string being used, and this will allow
80 * the tracing userspace tools to be able to decipher the string
81 * address to the matching string.
84 # define DEFINE_RCU_TPS(sname) \
85 static char sname##_varname[] = #sname; \
86 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
87 # define RCU_STATE_NAME(sname) sname##_varname
89 # define DEFINE_RCU_TPS(sname)
90 # define RCU_STATE_NAME(sname) __stringify(sname)
93 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
94 DEFINE_RCU_TPS(sname) \
95 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
96 struct rcu_state sname##_state = { \
97 .level = { &sname##_state.node[0] }, \
98 .rda = &sname##_data, \
100 .gp_state = RCU_GP_IDLE, \
101 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT, \
102 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
103 .name = RCU_STATE_NAME(sname), \
105 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
106 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
107 .ofl_lock = __SPIN_LOCK_UNLOCKED(sname##_state.ofl_lock), \
110 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
111 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
113 static struct rcu_state *const rcu_state_p;
114 LIST_HEAD(rcu_struct_flavors);
116 /* Dump rcu_node combining tree at boot to verify correct setup. */
117 static bool dump_tree;
118 module_param(dump_tree, bool, 0444);
119 /* Control rcu_node-tree auto-balancing at boot time. */
120 static bool rcu_fanout_exact;
121 module_param(rcu_fanout_exact, bool, 0444);
122 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
123 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
124 module_param(rcu_fanout_leaf, int, 0444);
125 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
126 /* Number of rcu_nodes at specified level. */
127 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
128 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
129 /* panic() on RCU Stall sysctl. */
130 int sysctl_panic_on_rcu_stall __read_mostly;
133 * The rcu_scheduler_active variable is initialized to the value
134 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
135 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
136 * RCU can assume that there is but one task, allowing RCU to (for example)
137 * optimize synchronize_rcu() to a simple barrier(). When this variable
138 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
139 * to detect real grace periods. This variable is also used to suppress
140 * boot-time false positives from lockdep-RCU error checking. Finally, it
141 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
142 * is fully initialized, including all of its kthreads having been spawned.
144 int rcu_scheduler_active __read_mostly;
145 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
148 * The rcu_scheduler_fully_active variable transitions from zero to one
149 * during the early_initcall() processing, which is after the scheduler
150 * is capable of creating new tasks. So RCU processing (for example,
151 * creating tasks for RCU priority boosting) must be delayed until after
152 * rcu_scheduler_fully_active transitions from zero to one. We also
153 * currently delay invocation of any RCU callbacks until after this point.
155 * It might later prove better for people registering RCU callbacks during
156 * early boot to take responsibility for these callbacks, but one step at
159 static int rcu_scheduler_fully_active __read_mostly;
162 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
163 struct rcu_node *rnp, unsigned long gps, unsigned long flags);
164 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
165 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
166 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
167 static void invoke_rcu_core(void);
168 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
169 static void rcu_report_exp_rdp(struct rcu_state *rsp,
170 struct rcu_data *rdp, bool wake);
171 static void sync_sched_exp_online_cleanup(int cpu);
173 /* rcuc/rcub kthread realtime priority */
174 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
175 module_param(kthread_prio, int, 0644);
177 /* Delay in jiffies for grace-period initialization delays, debug only. */
179 static int gp_preinit_delay;
180 module_param(gp_preinit_delay, int, 0444);
181 static int gp_init_delay;
182 module_param(gp_init_delay, int, 0444);
183 static int gp_cleanup_delay;
184 module_param(gp_cleanup_delay, int, 0444);
187 * Number of grace periods between delays, normalized by the duration of
188 * the delay. The longer the delay, the more the grace periods between
189 * each delay. The reason for this normalization is that it means that,
190 * for non-zero delays, the overall slowdown of grace periods is constant
191 * regardless of the duration of the delay. This arrangement balances
192 * the need for long delays to increase some race probabilities with the
193 * need for fast grace periods to increase other race probabilities.
195 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
198 * Track the rcutorture test sequence number and the update version
199 * number within a given test. The rcutorture_testseq is incremented
200 * on every rcutorture module load and unload, so has an odd value
201 * when a test is running. The rcutorture_vernum is set to zero
202 * when rcutorture starts and is incremented on each rcutorture update.
203 * These variables enable correlating rcutorture output with the
204 * RCU tracing information.
206 unsigned long rcutorture_testseq;
207 unsigned long rcutorture_vernum;
210 * Compute the mask of online CPUs for the specified rcu_node structure.
211 * This will not be stable unless the rcu_node structure's ->lock is
212 * held, but the bit corresponding to the current CPU will be stable
215 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
217 return READ_ONCE(rnp->qsmaskinitnext);
221 * Return true if an RCU grace period is in progress. The READ_ONCE()s
222 * permit this function to be invoked without holding the root rcu_node
223 * structure's ->lock, but of course results can be subject to change.
225 static int rcu_gp_in_progress(struct rcu_state *rsp)
227 return rcu_seq_state(rcu_seq_current(&rsp->gp_seq));
231 * Note a quiescent state. Because we do not need to know
232 * how many quiescent states passed, just if there was at least
233 * one since the start of the grace period, this just sets a flag.
234 * The caller must have disabled preemption.
236 void rcu_sched_qs(void)
238 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
239 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
241 trace_rcu_grace_period(TPS("rcu_sched"),
242 __this_cpu_read(rcu_sched_data.gp_seq),
244 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
245 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
247 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
248 rcu_report_exp_rdp(&rcu_sched_state,
249 this_cpu_ptr(&rcu_sched_data), true);
254 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
255 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
256 trace_rcu_grace_period(TPS("rcu_bh"),
257 __this_cpu_read(rcu_bh_data.gp_seq),
259 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
264 * Steal a bit from the bottom of ->dynticks for idle entry/exit
265 * control. Initially this is for TLB flushing.
267 #define RCU_DYNTICK_CTRL_MASK 0x1
268 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
269 #ifndef rcu_eqs_special_exit
270 #define rcu_eqs_special_exit() do { } while (0)
273 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
274 .dynticks_nesting = 1,
275 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
276 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
280 * Record entry into an extended quiescent state. This is only to be
281 * called when not already in an extended quiescent state.
283 static void rcu_dynticks_eqs_enter(void)
285 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
289 * CPUs seeing atomic_add_return() must see prior RCU read-side
290 * critical sections, and we also must force ordering with the
293 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
294 /* Better be in an extended quiescent state! */
295 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
296 (seq & RCU_DYNTICK_CTRL_CTR));
297 /* Better not have special action (TLB flush) pending! */
298 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
299 (seq & RCU_DYNTICK_CTRL_MASK));
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_add_return() must see prior idle sojourns,
313 * and we also must force ordering with the next RCU read-side
316 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
317 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
318 !(seq & RCU_DYNTICK_CTRL_CTR));
319 if (seq & RCU_DYNTICK_CTRL_MASK) {
320 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
321 smp_mb__after_atomic(); /* _exit after clearing mask. */
322 /* Prefer duplicate flushes to losing a flush. */
323 rcu_eqs_special_exit();
328 * Reset the current CPU's ->dynticks counter to indicate that the
329 * newly onlined CPU is no longer in an extended quiescent state.
330 * This will either leave the counter unchanged, or increment it
331 * to the next non-quiescent value.
333 * The non-atomic test/increment sequence works because the upper bits
334 * of the ->dynticks counter are manipulated only by the corresponding CPU,
335 * or when the corresponding CPU is offline.
337 static void rcu_dynticks_eqs_online(void)
339 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
341 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
343 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
347 * Is the current CPU in an extended quiescent state?
349 * No ordering, as we are sampling CPU-local information.
351 bool rcu_dynticks_curr_cpu_in_eqs(void)
353 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
355 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
359 * Snapshot the ->dynticks counter with full ordering so as to allow
360 * stable comparison of this counter with past and future snapshots.
362 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
364 int snap = atomic_add_return(0, &rdtp->dynticks);
366 return snap & ~RCU_DYNTICK_CTRL_MASK;
370 * Return true if the snapshot returned from rcu_dynticks_snap()
371 * indicates that RCU is in an extended quiescent state.
373 static bool rcu_dynticks_in_eqs(int snap)
375 return !(snap & RCU_DYNTICK_CTRL_CTR);
379 * Return true if the CPU corresponding to the specified rcu_dynticks
380 * structure has spent some time in an extended quiescent state since
381 * rcu_dynticks_snap() returned the specified snapshot.
383 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
385 return snap != rcu_dynticks_snap(rdtp);
389 * Set the special (bottom) bit of the specified CPU so that it
390 * will take special action (such as flushing its TLB) on the
391 * next exit from an extended quiescent state. Returns true if
392 * the bit was successfully set, or false if the CPU was not in
393 * an extended quiescent state.
395 bool rcu_eqs_special_set(int cpu)
399 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
402 old = atomic_read(&rdtp->dynticks);
403 if (old & RCU_DYNTICK_CTRL_CTR)
405 new = old | RCU_DYNTICK_CTRL_MASK;
406 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
411 * Let the RCU core know that this CPU has gone through the scheduler,
412 * which is a quiescent state. This is called when the need for a
413 * quiescent state is urgent, so we burn an atomic operation and full
414 * memory barriers to let the RCU core know about it, regardless of what
415 * this CPU might (or might not) do in the near future.
417 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
419 * The caller must have disabled interrupts and must not be idle.
421 static void rcu_momentary_dyntick_idle(void)
423 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
426 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
427 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
428 /* It is illegal to call this from idle state. */
429 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
433 * Note a context switch. This is a quiescent state for RCU-sched,
434 * and requires special handling for preemptible RCU.
435 * The caller must have disabled interrupts.
437 void rcu_note_context_switch(bool preempt)
439 barrier(); /* Avoid RCU read-side critical sections leaking down. */
440 trace_rcu_utilization(TPS("Start context switch"));
442 rcu_preempt_note_context_switch(preempt);
443 /* Load rcu_urgent_qs before other flags. */
444 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
446 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
447 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
448 rcu_momentary_dyntick_idle();
449 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
451 rcu_tasks_qs(current);
453 trace_rcu_utilization(TPS("End context switch"));
454 barrier(); /* Avoid RCU read-side critical sections leaking up. */
456 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
459 * Register a quiescent state for all RCU flavors. If there is an
460 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
461 * dyntick-idle quiescent state visible to other CPUs (but only for those
462 * RCU flavors in desperate need of a quiescent state, which will normally
463 * be none of them). Either way, do a lightweight quiescent state for
466 * The barrier() calls are redundant in the common case when this is
467 * called externally, but just in case this is called from within this
471 void rcu_all_qs(void)
475 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
478 /* Load rcu_urgent_qs before other flags. */
479 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
483 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
484 barrier(); /* Avoid RCU read-side critical sections leaking down. */
485 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
486 local_irq_save(flags);
487 rcu_momentary_dyntick_idle();
488 local_irq_restore(flags);
490 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
492 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
493 barrier(); /* Avoid RCU read-side critical sections leaking up. */
496 EXPORT_SYMBOL_GPL(rcu_all_qs);
498 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
499 static long blimit = DEFAULT_RCU_BLIMIT;
500 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
501 static long qhimark = DEFAULT_RCU_QHIMARK;
502 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
503 static long qlowmark = DEFAULT_RCU_QLOMARK;
505 module_param(blimit, long, 0444);
506 module_param(qhimark, long, 0444);
507 module_param(qlowmark, long, 0444);
509 static ulong jiffies_till_first_fqs = ULONG_MAX;
510 static ulong jiffies_till_next_fqs = ULONG_MAX;
511 static bool rcu_kick_kthreads;
513 module_param(jiffies_till_first_fqs, ulong, 0644);
514 module_param(jiffies_till_next_fqs, ulong, 0644);
515 module_param(rcu_kick_kthreads, bool, 0644);
518 * How long the grace period must be before we start recruiting
519 * quiescent-state help from rcu_note_context_switch().
521 static ulong jiffies_till_sched_qs = HZ / 10;
522 module_param(jiffies_till_sched_qs, ulong, 0444);
524 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
525 static void force_quiescent_state(struct rcu_state *rsp);
526 static int rcu_pending(void);
529 * Return the number of RCU GPs completed thus far for debug & stats.
531 unsigned long rcu_get_gp_seq(void)
533 return READ_ONCE(rcu_state_p->gp_seq);
535 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
538 * Return the number of RCU-sched GPs completed thus far for debug & stats.
540 unsigned long rcu_sched_get_gp_seq(void)
542 return READ_ONCE(rcu_sched_state.gp_seq);
544 EXPORT_SYMBOL_GPL(rcu_sched_get_gp_seq);
547 * Return the number of RCU-bh GPs completed thus far for debug & stats.
549 unsigned long rcu_bh_get_gp_seq(void)
551 return READ_ONCE(rcu_bh_state.gp_seq);
553 EXPORT_SYMBOL_GPL(rcu_bh_get_gp_seq);
556 * Return the number of RCU expedited batches completed thus far for
557 * debug & stats. Odd numbers mean that a batch is in progress, even
558 * numbers mean idle. The value returned will thus be roughly double
559 * the cumulative batches since boot.
561 unsigned long rcu_exp_batches_completed(void)
563 return rcu_state_p->expedited_sequence;
565 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
568 * Return the number of RCU-sched expedited batches completed thus far
569 * for debug & stats. Similar to rcu_exp_batches_completed().
571 unsigned long rcu_exp_batches_completed_sched(void)
573 return rcu_sched_state.expedited_sequence;
575 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
578 * Force a quiescent state.
580 void rcu_force_quiescent_state(void)
582 force_quiescent_state(rcu_state_p);
584 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
587 * Force a quiescent state for RCU BH.
589 void rcu_bh_force_quiescent_state(void)
591 force_quiescent_state(&rcu_bh_state);
593 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
596 * Force a quiescent state for RCU-sched.
598 void rcu_sched_force_quiescent_state(void)
600 force_quiescent_state(&rcu_sched_state);
602 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
605 * Show the state of the grace-period kthreads.
607 void show_rcu_gp_kthreads(void)
610 struct rcu_data *rdp;
611 struct rcu_node *rnp;
612 struct rcu_state *rsp;
614 for_each_rcu_flavor(rsp) {
615 pr_info("%s: wait state: %d ->state: %#lx\n",
616 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
617 rcu_for_each_node_breadth_first(rsp, rnp) {
618 if (ULONG_CMP_GE(rsp->gp_seq, rnp->gp_seq_needed))
620 pr_info("\trcu_node %d:%d ->gp_seq %lu ->gp_seq_needed %lu\n",
621 rnp->grplo, rnp->grphi, rnp->gp_seq,
623 if (!rcu_is_leaf_node(rnp))
625 for_each_leaf_node_possible_cpu(rnp, cpu) {
626 rdp = per_cpu_ptr(rsp->rda, cpu);
628 ULONG_CMP_GE(rsp->gp_seq,
631 pr_info("\tcpu %d ->gp_seq_needed %lu\n",
632 cpu, rdp->gp_seq_needed);
635 /* sched_show_task(rsp->gp_kthread); */
638 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
641 * Record the number of times rcutorture tests have been initiated and
642 * terminated. This information allows the debugfs tracing stats to be
643 * correlated to the rcutorture messages, even when the rcutorture module
644 * is being repeatedly loaded and unloaded. In other words, we cannot
645 * store this state in rcutorture itself.
647 void rcutorture_record_test_transition(void)
649 rcutorture_testseq++;
650 rcutorture_vernum = 0;
652 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
655 * Send along grace-period-related data for rcutorture diagnostics.
657 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
658 unsigned long *gp_seq)
660 struct rcu_state *rsp = NULL;
669 case RCU_SCHED_FLAVOR:
670 rsp = &rcu_sched_state;
677 *flags = READ_ONCE(rsp->gp_flags);
678 *gp_seq = rcu_seq_current(&rsp->gp_seq);
680 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
683 * Record the number of writer passes through the current rcutorture test.
684 * This is also used to correlate debugfs tracing stats with the rcutorture
687 void rcutorture_record_progress(unsigned long vernum)
691 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
694 * Return the root node of the specified rcu_state structure.
696 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
698 return &rsp->node[0];
702 * Enter an RCU extended quiescent state, which can be either the
703 * idle loop or adaptive-tickless usermode execution.
705 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
706 * the possibility of usermode upcalls having messed up our count
707 * of interrupt nesting level during the prior busy period.
709 static void rcu_eqs_enter(bool user)
711 struct rcu_state *rsp;
712 struct rcu_data *rdp;
713 struct rcu_dynticks *rdtp;
715 rdtp = this_cpu_ptr(&rcu_dynticks);
716 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
717 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
718 rdtp->dynticks_nesting == 0);
719 if (rdtp->dynticks_nesting != 1) {
720 rdtp->dynticks_nesting--;
724 lockdep_assert_irqs_disabled();
725 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
726 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
727 for_each_rcu_flavor(rsp) {
728 rdp = this_cpu_ptr(rsp->rda);
729 do_nocb_deferred_wakeup(rdp);
731 rcu_prepare_for_idle();
732 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
733 rcu_dynticks_eqs_enter();
734 rcu_dynticks_task_enter();
738 * rcu_idle_enter - inform RCU that current CPU is entering idle
740 * Enter idle mode, in other words, -leave- the mode in which RCU
741 * read-side critical sections can occur. (Though RCU read-side
742 * critical sections can occur in irq handlers in idle, a possibility
743 * handled by irq_enter() and irq_exit().)
745 * If you add or remove a call to rcu_idle_enter(), be sure to test with
746 * CONFIG_RCU_EQS_DEBUG=y.
748 void rcu_idle_enter(void)
750 lockdep_assert_irqs_disabled();
751 rcu_eqs_enter(false);
754 #ifdef CONFIG_NO_HZ_FULL
756 * rcu_user_enter - inform RCU that we are resuming userspace.
758 * Enter RCU idle mode right before resuming userspace. No use of RCU
759 * is permitted between this call and rcu_user_exit(). This way the
760 * CPU doesn't need to maintain the tick for RCU maintenance purposes
761 * when the CPU runs in userspace.
763 * If you add or remove a call to rcu_user_enter(), be sure to test with
764 * CONFIG_RCU_EQS_DEBUG=y.
766 void rcu_user_enter(void)
768 lockdep_assert_irqs_disabled();
771 #endif /* CONFIG_NO_HZ_FULL */
774 * rcu_nmi_exit - inform RCU of exit from NMI context
776 * If we are returning from the outermost NMI handler that interrupted an
777 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
778 * to let the RCU grace-period handling know that the CPU is back to
781 * If you add or remove a call to rcu_nmi_exit(), be sure to test
782 * with CONFIG_RCU_EQS_DEBUG=y.
784 void rcu_nmi_exit(void)
786 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
789 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
790 * (We are exiting an NMI handler, so RCU better be paying attention
793 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
794 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
797 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
798 * leave it in non-RCU-idle state.
800 if (rdtp->dynticks_nmi_nesting != 1) {
801 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
802 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
803 rdtp->dynticks_nmi_nesting - 2);
807 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
808 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
809 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
810 rcu_dynticks_eqs_enter();
814 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
816 * Exit from an interrupt handler, which might possibly result in entering
817 * idle mode, in other words, leaving the mode in which read-side critical
818 * sections can occur. The caller must have disabled interrupts.
820 * This code assumes that the idle loop never does anything that might
821 * result in unbalanced calls to irq_enter() and irq_exit(). If your
822 * architecture's idle loop violates this assumption, RCU will give you what
823 * you deserve, good and hard. But very infrequently and irreproducibly.
825 * Use things like work queues to work around this limitation.
827 * You have been warned.
829 * If you add or remove a call to rcu_irq_exit(), be sure to test with
830 * CONFIG_RCU_EQS_DEBUG=y.
832 void rcu_irq_exit(void)
834 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
836 lockdep_assert_irqs_disabled();
837 if (rdtp->dynticks_nmi_nesting == 1)
838 rcu_prepare_for_idle();
840 if (rdtp->dynticks_nmi_nesting == 0)
841 rcu_dynticks_task_enter();
845 * Wrapper for rcu_irq_exit() where interrupts are enabled.
847 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
848 * with CONFIG_RCU_EQS_DEBUG=y.
850 void rcu_irq_exit_irqson(void)
854 local_irq_save(flags);
856 local_irq_restore(flags);
860 * Exit an RCU extended quiescent state, which can be either the
861 * idle loop or adaptive-tickless usermode execution.
863 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
864 * allow for the possibility of usermode upcalls messing up our count of
865 * interrupt nesting level during the busy period that is just now starting.
867 static void rcu_eqs_exit(bool user)
869 struct rcu_dynticks *rdtp;
872 lockdep_assert_irqs_disabled();
873 rdtp = this_cpu_ptr(&rcu_dynticks);
874 oldval = rdtp->dynticks_nesting;
875 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
877 rdtp->dynticks_nesting++;
880 rcu_dynticks_task_exit();
881 rcu_dynticks_eqs_exit();
882 rcu_cleanup_after_idle();
883 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, 1, rdtp->dynticks);
884 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
885 WRITE_ONCE(rdtp->dynticks_nesting, 1);
886 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
890 * rcu_idle_exit - inform RCU that current CPU is leaving idle
892 * Exit idle mode, in other words, -enter- the mode in which RCU
893 * read-side critical sections can occur.
895 * If you add or remove a call to rcu_idle_exit(), be sure to test with
896 * CONFIG_RCU_EQS_DEBUG=y.
898 void rcu_idle_exit(void)
902 local_irq_save(flags);
904 local_irq_restore(flags);
907 #ifdef CONFIG_NO_HZ_FULL
909 * rcu_user_exit - inform RCU that we are exiting userspace.
911 * Exit RCU idle mode while entering the kernel because it can
912 * run a RCU read side critical section anytime.
914 * If you add or remove a call to rcu_user_exit(), be sure to test with
915 * CONFIG_RCU_EQS_DEBUG=y.
917 void rcu_user_exit(void)
921 #endif /* CONFIG_NO_HZ_FULL */
924 * rcu_nmi_enter - inform RCU of entry to NMI context
926 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
927 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
928 * that the CPU is active. This implementation permits nested NMIs, as
929 * long as the nesting level does not overflow an int. (You will probably
930 * run out of stack space first.)
932 * If you add or remove a call to rcu_nmi_enter(), be sure to test
933 * with CONFIG_RCU_EQS_DEBUG=y.
935 void rcu_nmi_enter(void)
937 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
940 /* Complain about underflow. */
941 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
944 * If idle from RCU viewpoint, atomically increment ->dynticks
945 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
946 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
947 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
948 * to be in the outermost NMI handler that interrupted an RCU-idle
949 * period (observation due to Andy Lutomirski).
951 if (rcu_dynticks_curr_cpu_in_eqs()) {
952 rcu_dynticks_eqs_exit();
955 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
956 rdtp->dynticks_nmi_nesting,
957 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
958 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
959 rdtp->dynticks_nmi_nesting + incby);
964 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
966 * Enter an interrupt handler, which might possibly result in exiting
967 * idle mode, in other words, entering the mode in which read-side critical
968 * sections can occur. The caller must have disabled interrupts.
970 * Note that the Linux kernel is fully capable of entering an interrupt
971 * handler that it never exits, for example when doing upcalls to user mode!
972 * This code assumes that the idle loop never does upcalls to user mode.
973 * If your architecture's idle loop does do upcalls to user mode (or does
974 * anything else that results in unbalanced calls to the irq_enter() and
975 * irq_exit() functions), RCU will give you what you deserve, good and hard.
976 * But very infrequently and irreproducibly.
978 * Use things like work queues to work around this limitation.
980 * You have been warned.
982 * If you add or remove a call to rcu_irq_enter(), be sure to test with
983 * CONFIG_RCU_EQS_DEBUG=y.
985 void rcu_irq_enter(void)
987 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
989 lockdep_assert_irqs_disabled();
990 if (rdtp->dynticks_nmi_nesting == 0)
991 rcu_dynticks_task_exit();
993 if (rdtp->dynticks_nmi_nesting == 1)
994 rcu_cleanup_after_idle();
998 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1000 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1001 * with CONFIG_RCU_EQS_DEBUG=y.
1003 void rcu_irq_enter_irqson(void)
1005 unsigned long flags;
1007 local_irq_save(flags);
1009 local_irq_restore(flags);
1013 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1015 * Return true if RCU is watching the running CPU, which means that this
1016 * CPU can safely enter RCU read-side critical sections. In other words,
1017 * if the current CPU is in its idle loop and is neither in an interrupt
1018 * or NMI handler, return true.
1020 bool notrace rcu_is_watching(void)
1024 preempt_disable_notrace();
1025 ret = !rcu_dynticks_curr_cpu_in_eqs();
1026 preempt_enable_notrace();
1029 EXPORT_SYMBOL_GPL(rcu_is_watching);
1032 * If a holdout task is actually running, request an urgent quiescent
1033 * state from its CPU. This is unsynchronized, so migrations can cause
1034 * the request to go to the wrong CPU. Which is OK, all that will happen
1035 * is that the CPU's next context switch will be a bit slower and next
1036 * time around this task will generate another request.
1038 void rcu_request_urgent_qs_task(struct task_struct *t)
1045 return; /* This task is not running on that CPU. */
1046 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1049 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1052 * Is the current CPU online as far as RCU is concerned?
1054 * Disable preemption to avoid false positives that could otherwise
1055 * happen due to the current CPU number being sampled, this task being
1056 * preempted, its old CPU being taken offline, resuming on some other CPU,
1057 * then determining that its old CPU is now offline. Because there are
1058 * multiple flavors of RCU, and because this function can be called in the
1059 * midst of updating the flavors while a given CPU coming online or going
1060 * offline, it is necessary to check all flavors. If any of the flavors
1061 * believe that given CPU is online, it is considered to be online.
1063 * Disable checking if in an NMI handler because we cannot safely
1064 * report errors from NMI handlers anyway. In addition, it is OK to use
1065 * RCU on an offline processor during initial boot, hence the check for
1066 * rcu_scheduler_fully_active.
1068 bool rcu_lockdep_current_cpu_online(void)
1070 struct rcu_data *rdp;
1071 struct rcu_node *rnp;
1072 struct rcu_state *rsp;
1074 if (in_nmi() || !rcu_scheduler_fully_active)
1077 for_each_rcu_flavor(rsp) {
1078 rdp = this_cpu_ptr(rsp->rda);
1080 if (rdp->grpmask & rcu_rnp_online_cpus(rnp)) {
1088 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1090 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1093 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1095 * If the current CPU is idle or running at a first-level (not nested)
1096 * interrupt from idle, return true. The caller must have at least
1097 * disabled preemption.
1099 static int rcu_is_cpu_rrupt_from_idle(void)
1101 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
1102 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
1106 * We are reporting a quiescent state on behalf of some other CPU, so
1107 * it is our responsibility to check for and handle potential overflow
1108 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1109 * After all, the CPU might be in deep idle state, and thus executing no
1112 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1114 raw_lockdep_assert_held_rcu_node(rnp);
1115 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1117 WRITE_ONCE(rdp->gpwrap, true);
1118 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1119 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1123 * Snapshot the specified CPU's dynticks counter so that we can later
1124 * credit them with an implicit quiescent state. Return 1 if this CPU
1125 * is in dynticks idle mode, which is an extended quiescent state.
1127 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1129 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1130 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1131 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1132 rcu_gpnum_ovf(rdp->mynode, rdp);
1139 * Handler for the irq_work request posted when a grace period has
1140 * gone on for too long, but not yet long enough for an RCU CPU
1141 * stall warning. Set state appropriately, but just complain if
1142 * there is unexpected state on entry.
1144 static void rcu_iw_handler(struct irq_work *iwp)
1146 struct rcu_data *rdp;
1147 struct rcu_node *rnp;
1149 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1151 raw_spin_lock_rcu_node(rnp);
1152 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1153 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1154 rdp->rcu_iw_pending = false;
1156 raw_spin_unlock_rcu_node(rnp);
1160 * Return true if the specified CPU has passed through a quiescent
1161 * state by virtue of being in or having passed through an dynticks
1162 * idle state since the last call to dyntick_save_progress_counter()
1163 * for this same CPU, or by virtue of having been offline.
1165 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1170 struct rcu_node *rnp = rdp->mynode;
1173 * If the CPU passed through or entered a dynticks idle phase with
1174 * no active irq/NMI handlers, then we can safely pretend that the CPU
1175 * already acknowledged the request to pass through a quiescent
1176 * state. Either way, that CPU cannot possibly be in an RCU
1177 * read-side critical section that started before the beginning
1178 * of the current RCU grace period.
1180 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1181 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1182 rdp->dynticks_fqs++;
1183 rcu_gpnum_ovf(rnp, rdp);
1188 * Has this CPU encountered a cond_resched() since the beginning
1189 * of the grace period? For this to be the case, the CPU has to
1190 * have noticed the current grace period. This might not be the
1191 * case for nohz_full CPUs looping in the kernel.
1193 jtsq = jiffies_till_sched_qs;
1194 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1195 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1196 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1197 rcu_seq_current(&rdp->gp_seq) == rnp->gp_seq && !rdp->gpwrap) {
1198 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("rqc"));
1199 rcu_gpnum_ovf(rnp, rdp);
1201 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1202 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1203 smp_store_release(ruqp, true);
1206 /* If waiting too long on an offline CPU, complain. */
1207 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1208 time_after(jiffies, rdp->rsp->gp_start + HZ)) {
1210 struct rcu_node *rnp1;
1212 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1213 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1214 __func__, rnp->grplo, rnp->grphi, rnp->level,
1215 (long)rnp->gp_seq, (long)rnp->completedqs);
1216 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1217 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1218 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1219 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1220 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1221 __func__, rdp->cpu, ".o"[onl],
1222 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1223 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1224 return 1; /* Break things loose after complaining. */
1228 * A CPU running for an extended time within the kernel can
1229 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1230 * even context-switching back and forth between a pair of
1231 * in-kernel CPU-bound tasks cannot advance grace periods.
1232 * So if the grace period is old enough, make the CPU pay attention.
1233 * Note that the unsynchronized assignments to the per-CPU
1234 * rcu_need_heavy_qs variable are safe. Yes, setting of
1235 * bits can be lost, but they will be set again on the next
1236 * force-quiescent-state pass. So lost bit sets do not result
1237 * in incorrect behavior, merely in a grace period lasting
1238 * a few jiffies longer than it might otherwise. Because
1239 * there are at most four threads involved, and because the
1240 * updates are only once every few jiffies, the probability of
1241 * lossage (and thus of slight grace-period extension) is
1244 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1245 if (!READ_ONCE(*rnhqp) &&
1246 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1247 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1248 WRITE_ONCE(*rnhqp, true);
1249 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1250 smp_store_release(ruqp, true);
1251 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1255 * If more than halfway to RCU CPU stall-warning time, do a
1256 * resched_cpu() to try to loosen things up a bit. Also check to
1257 * see if the CPU is getting hammered with interrupts, but only
1258 * once per grace period, just to keep the IPIs down to a dull roar.
1260 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1261 resched_cpu(rdp->cpu);
1262 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1263 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1264 (rnp->ffmask & rdp->grpmask)) {
1265 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1266 rdp->rcu_iw_pending = true;
1267 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1268 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1275 static void record_gp_stall_check_time(struct rcu_state *rsp)
1277 unsigned long j = jiffies;
1281 j1 = rcu_jiffies_till_stall_check();
1282 /* Record ->gp_start before ->jiffies_stall. */
1283 smp_store_release(&rsp->jiffies_stall, j + j1); /* ^^^ */
1284 rsp->jiffies_resched = j + j1 / 2;
1285 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1289 * Convert a ->gp_state value to a character string.
1291 static const char *gp_state_getname(short gs)
1293 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1295 return gp_state_names[gs];
1299 * Complain about starvation of grace-period kthread.
1301 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1307 gpa = READ_ONCE(rsp->gp_activity);
1308 if (j - gpa > 2 * HZ) {
1309 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1311 (long)rcu_seq_current(&rsp->gp_seq),
1313 gp_state_getname(rsp->gp_state), rsp->gp_state,
1314 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1315 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1316 if (rsp->gp_kthread) {
1317 pr_err("RCU grace-period kthread stack dump:\n");
1318 sched_show_task(rsp->gp_kthread);
1319 wake_up_process(rsp->gp_kthread);
1325 * Dump stacks of all tasks running on stalled CPUs. First try using
1326 * NMIs, but fall back to manual remote stack tracing on architectures
1327 * that don't support NMI-based stack dumps. The NMI-triggered stack
1328 * traces are more accurate because they are printed by the target CPU.
1330 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1333 unsigned long flags;
1334 struct rcu_node *rnp;
1336 rcu_for_each_leaf_node(rsp, rnp) {
1337 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1338 for_each_leaf_node_possible_cpu(rnp, cpu)
1339 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1340 if (!trigger_single_cpu_backtrace(cpu))
1342 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1347 * If too much time has passed in the current grace period, and if
1348 * so configured, go kick the relevant kthreads.
1350 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1354 if (!rcu_kick_kthreads)
1356 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1357 if (time_after(jiffies, j) && rsp->gp_kthread &&
1358 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1359 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1360 rcu_ftrace_dump(DUMP_ALL);
1361 wake_up_process(rsp->gp_kthread);
1362 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1366 static void panic_on_rcu_stall(void)
1368 if (sysctl_panic_on_rcu_stall)
1369 panic("RCU Stall\n");
1372 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gp_seq)
1375 unsigned long flags;
1379 struct rcu_node *rnp = rcu_get_root(rsp);
1382 /* Kick and suppress, if so configured. */
1383 rcu_stall_kick_kthreads(rsp);
1384 if (rcu_cpu_stall_suppress)
1388 * OK, time to rat on our buddy...
1389 * See Documentation/RCU/stallwarn.txt for info on how to debug
1390 * RCU CPU stall warnings.
1392 pr_err("INFO: %s detected stalls on CPUs/tasks:", rsp->name);
1393 print_cpu_stall_info_begin();
1394 rcu_for_each_leaf_node(rsp, rnp) {
1395 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1396 ndetected += rcu_print_task_stall(rnp);
1397 if (rnp->qsmask != 0) {
1398 for_each_leaf_node_possible_cpu(rnp, cpu)
1399 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1400 print_cpu_stall_info(rsp, cpu);
1404 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1407 print_cpu_stall_info_end();
1408 for_each_possible_cpu(cpu)
1409 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1411 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
1412 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1413 (long)rcu_seq_current(&rsp->gp_seq), totqlen);
1415 rcu_dump_cpu_stacks(rsp);
1417 /* Complain about tasks blocking the grace period. */
1418 rcu_print_detail_task_stall(rsp);
1420 if (rcu_seq_current(&rsp->gp_seq) != gp_seq) {
1421 pr_err("INFO: Stall ended before state dump start\n");
1424 gpa = READ_ONCE(rsp->gp_activity);
1425 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1426 rsp->name, j - gpa, j, gpa,
1427 jiffies_till_next_fqs,
1428 rcu_get_root(rsp)->qsmask);
1429 /* In this case, the current CPU might be at fault. */
1430 sched_show_task(current);
1433 /* Rewrite if needed in case of slow consoles. */
1434 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1435 WRITE_ONCE(rsp->jiffies_stall,
1436 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1438 rcu_check_gp_kthread_starvation(rsp);
1440 panic_on_rcu_stall();
1442 force_quiescent_state(rsp); /* Kick them all. */
1445 static void print_cpu_stall(struct rcu_state *rsp)
1448 unsigned long flags;
1449 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1450 struct rcu_node *rnp = rcu_get_root(rsp);
1453 /* Kick and suppress, if so configured. */
1454 rcu_stall_kick_kthreads(rsp);
1455 if (rcu_cpu_stall_suppress)
1459 * OK, time to rat on ourselves...
1460 * See Documentation/RCU/stallwarn.txt for info on how to debug
1461 * RCU CPU stall warnings.
1463 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1464 print_cpu_stall_info_begin();
1465 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1466 print_cpu_stall_info(rsp, smp_processor_id());
1467 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1468 print_cpu_stall_info_end();
1469 for_each_possible_cpu(cpu)
1470 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1472 pr_cont(" (t=%lu jiffies g=%ld q=%lu)\n",
1473 jiffies - rsp->gp_start,
1474 (long)rcu_seq_current(&rsp->gp_seq), totqlen);
1476 rcu_check_gp_kthread_starvation(rsp);
1478 rcu_dump_cpu_stacks(rsp);
1480 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1481 /* Rewrite if needed in case of slow consoles. */
1482 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1483 WRITE_ONCE(rsp->jiffies_stall,
1484 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1485 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1487 panic_on_rcu_stall();
1490 * Attempt to revive the RCU machinery by forcing a context switch.
1492 * A context switch would normally allow the RCU state machine to make
1493 * progress and it could be we're stuck in kernel space without context
1494 * switches for an entirely unreasonable amount of time.
1496 resched_cpu(smp_processor_id());
1499 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1507 struct rcu_node *rnp;
1509 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1510 !rcu_gp_in_progress(rsp))
1512 rcu_stall_kick_kthreads(rsp);
1516 * Lots of memory barriers to reject false positives.
1518 * The idea is to pick up rsp->gp_seq, then rsp->jiffies_stall,
1519 * then rsp->gp_start, and finally another copy of rsp->gp_seq.
1520 * These values are updated in the opposite order with memory
1521 * barriers (or equivalent) during grace-period initialization
1522 * and cleanup. Now, a false positive can occur if we get an new
1523 * value of rsp->gp_start and a old value of rsp->jiffies_stall.
1524 * But given the memory barriers, the only way that this can happen
1525 * is if one grace period ends and another starts between these
1526 * two fetches. This is detected by comparing the second fetch
1527 * of rsp->gp_seq with the previous fetch from rsp->gp_seq.
1529 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1530 * and rsp->gp_start suffice to forestall false positives.
1532 gs1 = READ_ONCE(rsp->gp_seq);
1533 smp_rmb(); /* Pick up ->gp_seq first... */
1534 js = READ_ONCE(rsp->jiffies_stall);
1535 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1536 gps = READ_ONCE(rsp->gp_start);
1537 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
1538 gs2 = READ_ONCE(rsp->gp_seq);
1540 ULONG_CMP_LT(j, js) ||
1541 ULONG_CMP_GE(gps, js))
1542 return; /* No stall or GP completed since entering function. */
1544 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1545 if (rcu_gp_in_progress(rsp) &&
1546 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
1547 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1549 /* We haven't checked in, so go dump stack. */
1550 print_cpu_stall(rsp);
1552 } else if (rcu_gp_in_progress(rsp) &&
1553 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
1554 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1556 /* They had a few time units to dump stack, so complain. */
1557 print_other_cpu_stall(rsp, gs2);
1562 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1564 * Set the stall-warning timeout way off into the future, thus preventing
1565 * any RCU CPU stall-warning messages from appearing in the current set of
1566 * RCU grace periods.
1568 * The caller must disable hard irqs.
1570 void rcu_cpu_stall_reset(void)
1572 struct rcu_state *rsp;
1574 for_each_rcu_flavor(rsp)
1575 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1578 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1579 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1580 unsigned long gp_seq_req, const char *s)
1582 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gp_seq, gp_seq_req,
1583 rnp->level, rnp->grplo, rnp->grphi, s);
1587 * rcu_start_this_gp - Request the start of a particular grace period
1588 * @rnp_start: The leaf node of the CPU from which to start.
1589 * @rdp: The rcu_data corresponding to the CPU from which to start.
1590 * @gp_seq_req: The gp_seq of the grace period to start.
1592 * Start the specified grace period, as needed to handle newly arrived
1593 * callbacks. The required future grace periods are recorded in each
1594 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1595 * is reason to awaken the grace-period kthread.
1597 * The caller must hold the specified rcu_node structure's ->lock, which
1598 * is why the caller is responsible for waking the grace-period kthread.
1600 * Returns true if the GP thread needs to be awakened else false.
1602 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1603 unsigned long gp_seq_req)
1606 struct rcu_state *rsp = rdp->rsp;
1607 struct rcu_node *rnp;
1610 * Use funnel locking to either acquire the root rcu_node
1611 * structure's lock or bail out if the need for this grace period
1612 * has already been recorded -- or if that grace period has in
1613 * fact already started. If there is already a grace period in
1614 * progress in a non-leaf node, no recording is needed because the
1615 * end of the grace period will scan the leaf rcu_node structures.
1616 * Note that rnp_start->lock must not be released.
1618 raw_lockdep_assert_held_rcu_node(rnp_start);
1619 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1620 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1621 if (rnp != rnp_start)
1622 raw_spin_lock_rcu_node(rnp);
1623 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1624 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1625 (rnp != rnp_start &&
1626 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1627 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1631 rnp->gp_seq_needed = gp_seq_req;
1632 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1634 * We just marked the leaf or internal node, and a
1635 * grace period is in progress, which means that
1636 * rcu_gp_cleanup() will see the marking. Bail to
1637 * reduce contention.
1639 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1640 TPS("Startedleaf"));
1643 if (rnp != rnp_start && rnp->parent != NULL)
1644 raw_spin_unlock_rcu_node(rnp);
1646 break; /* At root, and perhaps also leaf. */
1649 /* If GP already in progress, just leave, otherwise start one. */
1650 if (rcu_gp_in_progress(rsp)) {
1651 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1654 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1655 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags | RCU_GP_FLAG_INIT);
1656 rsp->gp_req_activity = jiffies;
1657 if (!rsp->gp_kthread) {
1658 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1661 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gp_seq), TPS("newreq"));
1662 ret = true; /* Caller must wake GP kthread. */
1664 /* Push furthest requested GP to leaf node and rcu_data structure. */
1665 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1666 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1667 rdp->gp_seq_needed = rnp->gp_seq_needed;
1669 if (rnp != rnp_start)
1670 raw_spin_unlock_rcu_node(rnp);
1675 * Clean up any old requests for the just-ended grace period. Also return
1676 * whether any additional grace periods have been requested.
1678 static bool rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1681 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1683 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1685 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1686 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1687 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1692 * Awaken the grace-period kthread for the specified flavor of RCU.
1693 * Don't do a self-awaken, and don't bother awakening when there is
1694 * nothing for the grace-period kthread to do (as in several CPUs
1695 * raced to awaken, and we lost), and finally don't try to awaken
1696 * a kthread that has not yet been created.
1698 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1700 if (current == rsp->gp_kthread ||
1701 !READ_ONCE(rsp->gp_flags) ||
1704 swake_up(&rsp->gp_wq);
1708 * If there is room, assign a ->gp_seq number to any callbacks on this
1709 * CPU that have not already been assigned. Also accelerate any callbacks
1710 * that were previously assigned a ->gp_seq number that has since proven
1711 * to be too conservative, which can happen if callbacks get assigned a
1712 * ->gp_seq number while RCU is idle, but with reference to a non-root
1713 * rcu_node structure. This function is idempotent, so it does not hurt
1714 * to call it repeatedly. Returns an flag saying that we should awaken
1715 * the RCU grace-period kthread.
1717 * The caller must hold rnp->lock with interrupts disabled.
1719 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1720 struct rcu_data *rdp)
1722 unsigned long gp_seq_req;
1725 raw_lockdep_assert_held_rcu_node(rnp);
1727 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1728 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1732 * Callbacks are often registered with incomplete grace-period
1733 * information. Something about the fact that getting exact
1734 * information requires acquiring a global lock... RCU therefore
1735 * makes a conservative estimate of the grace period number at which
1736 * a given callback will become ready to invoke. The following
1737 * code checks this estimate and improves it when possible, thus
1738 * accelerating callback invocation to an earlier grace-period
1741 gp_seq_req = rcu_seq_snap(&rsp->gp_seq);
1742 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1743 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1745 /* Trace depending on how much we were able to accelerate. */
1746 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1747 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("AccWaitCB"));
1749 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("AccReadyCB"));
1754 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1755 * rcu_node structure's ->lock be held. It consults the cached value
1756 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1757 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1758 * while holding the leaf rcu_node structure's ->lock.
1760 static void rcu_accelerate_cbs_unlocked(struct rcu_state *rsp,
1761 struct rcu_node *rnp,
1762 struct rcu_data *rdp)
1767 lockdep_assert_irqs_disabled();
1768 c = rcu_seq_snap(&rsp->gp_seq);
1769 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1770 /* Old request still live, so mark recent callbacks. */
1771 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1774 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1775 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1776 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1778 rcu_gp_kthread_wake(rsp);
1782 * Move any callbacks whose grace period has completed to the
1783 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1784 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1785 * sublist. This function is idempotent, so it does not hurt to
1786 * invoke it repeatedly. As long as it is not invoked -too- often...
1787 * Returns true if the RCU grace-period kthread needs to be awakened.
1789 * The caller must hold rnp->lock with interrupts disabled.
1791 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1792 struct rcu_data *rdp)
1794 raw_lockdep_assert_held_rcu_node(rnp);
1796 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1797 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1801 * Find all callbacks whose ->gp_seq numbers indicate that they
1802 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1804 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1806 /* Classify any remaining callbacks. */
1807 return rcu_accelerate_cbs(rsp, rnp, rdp);
1811 * Update CPU-local rcu_data state to record the beginnings and ends of
1812 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1813 * structure corresponding to the current CPU, and must have irqs disabled.
1814 * Returns true if the grace-period kthread needs to be awakened.
1816 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1817 struct rcu_data *rdp)
1822 raw_lockdep_assert_held_rcu_node(rnp);
1824 if (rdp->gp_seq == rnp->gp_seq)
1825 return false; /* Nothing to do. */
1827 /* Handle the ends of any preceding grace periods first. */
1828 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1829 unlikely(READ_ONCE(rdp->gpwrap))) {
1830 ret = rcu_advance_cbs(rsp, rnp, rdp); /* Advance callbacks. */
1831 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpuend"));
1833 ret = rcu_accelerate_cbs(rsp, rnp, rdp); /* Recent callbacks. */
1836 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1837 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1838 unlikely(READ_ONCE(rdp->gpwrap))) {
1840 * If the current grace period is waiting for this CPU,
1841 * set up to detect a quiescent state, otherwise don't
1842 * go looking for one.
1844 trace_rcu_grace_period(rsp->name, rnp->gp_seq, TPS("cpustart"));
1845 need_gp = !!(rnp->qsmask & rdp->grpmask);
1846 rdp->cpu_no_qs.b.norm = need_gp;
1847 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1848 rdp->core_needs_qs = need_gp;
1849 zero_cpu_stall_ticks(rdp);
1851 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1852 if (ULONG_CMP_GE(rnp->gp_seq_needed, rdp->gp_seq_needed) || rdp->gpwrap)
1853 rdp->gp_seq_needed = rnp->gp_seq_needed;
1854 WRITE_ONCE(rdp->gpwrap, false);
1855 rcu_gpnum_ovf(rnp, rdp);
1859 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1861 unsigned long flags;
1863 struct rcu_node *rnp;
1865 local_irq_save(flags);
1867 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1868 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1869 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1870 local_irq_restore(flags);
1873 needwake = __note_gp_changes(rsp, rnp, rdp);
1874 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1876 rcu_gp_kthread_wake(rsp);
1879 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1882 !(rcu_seq_ctr(rsp->gp_seq) %
1883 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1884 schedule_timeout_uninterruptible(delay);
1888 * Initialize a new grace period. Return false if no grace period required.
1890 static bool rcu_gp_init(struct rcu_state *rsp)
1892 unsigned long flags;
1893 unsigned long oldmask;
1895 struct rcu_data *rdp;
1896 struct rcu_node *rnp = rcu_get_root(rsp);
1898 WRITE_ONCE(rsp->gp_activity, jiffies);
1899 raw_spin_lock_irq_rcu_node(rnp);
1900 if (!READ_ONCE(rsp->gp_flags)) {
1901 /* Spurious wakeup, tell caller to go back to sleep. */
1902 raw_spin_unlock_irq_rcu_node(rnp);
1905 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1907 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1909 * Grace period already in progress, don't start another.
1910 * Not supposed to be able to happen.
1912 raw_spin_unlock_irq_rcu_node(rnp);
1916 /* Advance to a new grace period and initialize state. */
1917 record_gp_stall_check_time(rsp);
1918 /* Record GP times before starting GP, hence rcu_seq_start(). */
1919 rcu_seq_start(&rsp->gp_seq);
1920 trace_rcu_grace_period(rsp->name, rsp->gp_seq, TPS("start"));
1921 raw_spin_unlock_irq_rcu_node(rnp);
1924 * Apply per-leaf buffered online and offline operations to the
1925 * rcu_node tree. Note that this new grace period need not wait
1926 * for subsequent online CPUs, and that quiescent-state forcing
1927 * will handle subsequent offline CPUs.
1929 rsp->gp_state = RCU_GP_ONOFF;
1930 rcu_for_each_leaf_node(rsp, rnp) {
1931 spin_lock(&rsp->ofl_lock);
1932 raw_spin_lock_irq_rcu_node(rnp);
1933 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1934 !rnp->wait_blkd_tasks) {
1935 /* Nothing to do on this leaf rcu_node structure. */
1936 raw_spin_unlock_irq_rcu_node(rnp);
1937 spin_unlock(&rsp->ofl_lock);
1941 /* Record old state, apply changes to ->qsmaskinit field. */
1942 oldmask = rnp->qsmaskinit;
1943 rnp->qsmaskinit = rnp->qsmaskinitnext;
1945 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1946 if (!oldmask != !rnp->qsmaskinit) {
1947 if (!oldmask) { /* First online CPU for rcu_node. */
1948 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1949 rcu_init_new_rnp(rnp);
1950 } else if (rcu_preempt_has_tasks(rnp)) {
1951 rnp->wait_blkd_tasks = true; /* blocked tasks */
1952 } else { /* Last offline CPU and can propagate. */
1953 rcu_cleanup_dead_rnp(rnp);
1958 * If all waited-on tasks from prior grace period are
1959 * done, and if all this rcu_node structure's CPUs are
1960 * still offline, propagate up the rcu_node tree and
1961 * clear ->wait_blkd_tasks. Otherwise, if one of this
1962 * rcu_node structure's CPUs has since come back online,
1963 * simply clear ->wait_blkd_tasks.
1965 if (rnp->wait_blkd_tasks &&
1966 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1967 rnp->wait_blkd_tasks = false;
1968 if (!rnp->qsmaskinit)
1969 rcu_cleanup_dead_rnp(rnp);
1972 raw_spin_unlock_irq_rcu_node(rnp);
1973 spin_unlock(&rsp->ofl_lock);
1975 rcu_gp_slow(rsp, gp_preinit_delay); /* Races with CPU hotplug. */
1978 * Set the quiescent-state-needed bits in all the rcu_node
1979 * structures for all currently online CPUs in breadth-first order,
1980 * starting from the root rcu_node structure, relying on the layout
1981 * of the tree within the rsp->node[] array. Note that other CPUs
1982 * will access only the leaves of the hierarchy, thus seeing that no
1983 * grace period is in progress, at least until the corresponding
1984 * leaf node has been initialized.
1986 * The grace period cannot complete until the initialization
1987 * process finishes, because this kthread handles both.
1989 rsp->gp_state = RCU_GP_INIT;
1990 rcu_for_each_node_breadth_first(rsp, rnp) {
1991 rcu_gp_slow(rsp, gp_init_delay);
1992 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1993 rdp = this_cpu_ptr(rsp->rda);
1994 rcu_preempt_check_blocked_tasks(rsp, rnp);
1995 rnp->qsmask = rnp->qsmaskinit;
1996 WRITE_ONCE(rnp->gp_seq, rsp->gp_seq);
1997 if (rnp == rdp->mynode)
1998 (void)__note_gp_changes(rsp, rnp, rdp);
1999 rcu_preempt_boost_start_gp(rnp);
2000 trace_rcu_grace_period_init(rsp->name, rnp->gp_seq,
2001 rnp->level, rnp->grplo,
2002 rnp->grphi, rnp->qsmask);
2003 /* Quiescent states for tasks on any now-offline CPUs. */
2004 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
2005 rnp->rcu_gp_init_mask = mask;
2006 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
2007 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2009 raw_spin_unlock_irq_rcu_node(rnp);
2010 cond_resched_tasks_rcu_qs();
2011 WRITE_ONCE(rsp->gp_activity, jiffies);
2018 * Helper function for swait_event_idle() wakeup at force-quiescent-state
2021 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2023 struct rcu_node *rnp = rcu_get_root(rsp);
2025 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2026 *gfp = READ_ONCE(rsp->gp_flags);
2027 if (*gfp & RCU_GP_FLAG_FQS)
2030 /* The current grace period has completed. */
2031 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2038 * Do one round of quiescent-state forcing.
2040 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2042 struct rcu_node *rnp = rcu_get_root(rsp);
2044 WRITE_ONCE(rsp->gp_activity, jiffies);
2047 /* Collect dyntick-idle snapshots. */
2048 force_qs_rnp(rsp, dyntick_save_progress_counter);
2050 /* Handle dyntick-idle and offline CPUs. */
2051 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2053 /* Clear flag to prevent immediate re-entry. */
2054 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2055 raw_spin_lock_irq_rcu_node(rnp);
2056 WRITE_ONCE(rsp->gp_flags,
2057 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2058 raw_spin_unlock_irq_rcu_node(rnp);
2063 * Clean up after the old grace period.
2065 static void rcu_gp_cleanup(struct rcu_state *rsp)
2067 unsigned long gp_duration;
2068 bool needgp = false;
2069 unsigned long new_gp_seq;
2070 struct rcu_data *rdp;
2071 struct rcu_node *rnp = rcu_get_root(rsp);
2072 struct swait_queue_head *sq;
2074 WRITE_ONCE(rsp->gp_activity, jiffies);
2075 raw_spin_lock_irq_rcu_node(rnp);
2076 gp_duration = jiffies - rsp->gp_start;
2077 if (gp_duration > rsp->gp_max)
2078 rsp->gp_max = gp_duration;
2081 * We know the grace period is complete, but to everyone else
2082 * it appears to still be ongoing. But it is also the case
2083 * that to everyone else it looks like there is nothing that
2084 * they can do to advance the grace period. It is therefore
2085 * safe for us to drop the lock in order to mark the grace
2086 * period as completed in all of the rcu_node structures.
2088 raw_spin_unlock_irq_rcu_node(rnp);
2091 * Propagate new ->gp_seq value to rcu_node structures so that
2092 * other CPUs don't have to wait until the start of the next grace
2093 * period to process their callbacks. This also avoids some nasty
2094 * RCU grace-period initialization races by forcing the end of
2095 * the current grace period to be completely recorded in all of
2096 * the rcu_node structures before the beginning of the next grace
2097 * period is recorded in any of the rcu_node structures.
2099 new_gp_seq = rsp->gp_seq;
2100 rcu_seq_end(&new_gp_seq);
2101 rcu_for_each_node_breadth_first(rsp, rnp) {
2102 raw_spin_lock_irq_rcu_node(rnp);
2103 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2104 dump_blkd_tasks(rsp, rnp, 10);
2105 WARN_ON_ONCE(rnp->qsmask);
2106 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2107 rdp = this_cpu_ptr(rsp->rda);
2108 if (rnp == rdp->mynode)
2109 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2110 /* smp_mb() provided by prior unlock-lock pair. */
2111 needgp = rcu_future_gp_cleanup(rsp, rnp) || needgp;
2112 sq = rcu_nocb_gp_get(rnp);
2113 raw_spin_unlock_irq_rcu_node(rnp);
2114 rcu_nocb_gp_cleanup(sq);
2115 cond_resched_tasks_rcu_qs();
2116 WRITE_ONCE(rsp->gp_activity, jiffies);
2117 rcu_gp_slow(rsp, gp_cleanup_delay);
2119 rnp = rcu_get_root(rsp);
2120 raw_spin_lock_irq_rcu_node(rnp); /* GP before rsp->gp_seq update. */
2122 /* Declare grace period done. */
2123 rcu_seq_end(&rsp->gp_seq);
2124 trace_rcu_grace_period(rsp->name, rsp->gp_seq, TPS("end"));
2125 rsp->gp_state = RCU_GP_IDLE;
2126 /* Check for GP requests since above loop. */
2127 rdp = this_cpu_ptr(rsp->rda);
2128 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2129 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2130 TPS("CleanupMore"));
2133 /* Advance CBs to reduce false positives below. */
2134 if (!rcu_accelerate_cbs(rsp, rnp, rdp) && needgp) {
2135 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2136 rsp->gp_req_activity = jiffies;
2137 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gp_seq),
2140 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags & RCU_GP_FLAG_INIT);
2142 raw_spin_unlock_irq_rcu_node(rnp);
2146 * Body of kthread that handles grace periods.
2148 static int __noreturn rcu_gp_kthread(void *arg)
2154 struct rcu_state *rsp = arg;
2155 struct rcu_node *rnp = rcu_get_root(rsp);
2157 rcu_bind_gp_kthread();
2160 /* Handle grace-period start. */
2162 trace_rcu_grace_period(rsp->name,
2163 READ_ONCE(rsp->gp_seq),
2165 rsp->gp_state = RCU_GP_WAIT_GPS;
2166 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2168 rsp->gp_state = RCU_GP_DONE_GPS;
2169 /* Locking provides needed memory barrier. */
2170 if (rcu_gp_init(rsp))
2172 cond_resched_tasks_rcu_qs();
2173 WRITE_ONCE(rsp->gp_activity, jiffies);
2174 WARN_ON(signal_pending(current));
2175 trace_rcu_grace_period(rsp->name,
2176 READ_ONCE(rsp->gp_seq),
2180 /* Handle quiescent-state forcing. */
2181 first_gp_fqs = true;
2182 j = jiffies_till_first_fqs;
2185 jiffies_till_first_fqs = HZ;
2190 rsp->jiffies_force_qs = jiffies + j;
2191 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2194 trace_rcu_grace_period(rsp->name,
2195 READ_ONCE(rsp->gp_seq),
2197 rsp->gp_state = RCU_GP_WAIT_FQS;
2198 ret = swait_event_idle_timeout(rsp->gp_wq,
2199 rcu_gp_fqs_check_wake(rsp, &gf), j);
2200 rsp->gp_state = RCU_GP_DOING_FQS;
2201 /* Locking provides needed memory barriers. */
2202 /* If grace period done, leave loop. */
2203 if (!READ_ONCE(rnp->qsmask) &&
2204 !rcu_preempt_blocked_readers_cgp(rnp))
2206 /* If time for quiescent-state forcing, do it. */
2207 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2208 (gf & RCU_GP_FLAG_FQS)) {
2209 trace_rcu_grace_period(rsp->name,
2210 READ_ONCE(rsp->gp_seq),
2212 rcu_gp_fqs(rsp, first_gp_fqs);
2213 first_gp_fqs = false;
2214 trace_rcu_grace_period(rsp->name,
2215 READ_ONCE(rsp->gp_seq),
2217 cond_resched_tasks_rcu_qs();
2218 WRITE_ONCE(rsp->gp_activity, jiffies);
2219 ret = 0; /* Force full wait till next FQS. */
2220 j = jiffies_till_next_fqs;
2223 jiffies_till_next_fqs = HZ;
2226 jiffies_till_next_fqs = 1;
2229 /* Deal with stray signal. */
2230 cond_resched_tasks_rcu_qs();
2231 WRITE_ONCE(rsp->gp_activity, jiffies);
2232 WARN_ON(signal_pending(current));
2233 trace_rcu_grace_period(rsp->name,
2234 READ_ONCE(rsp->gp_seq),
2236 ret = 1; /* Keep old FQS timing. */
2238 if (time_after(jiffies, rsp->jiffies_force_qs))
2241 j = rsp->jiffies_force_qs - j;
2245 /* Handle grace-period end. */
2246 rsp->gp_state = RCU_GP_CLEANUP;
2247 rcu_gp_cleanup(rsp);
2248 rsp->gp_state = RCU_GP_CLEANED;
2253 * Report a full set of quiescent states to the specified rcu_state data
2254 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2255 * kthread if another grace period is required. Whether we wake
2256 * the grace-period kthread or it awakens itself for the next round
2257 * of quiescent-state forcing, that kthread will clean up after the
2258 * just-completed grace period. Note that the caller must hold rnp->lock,
2259 * which is released before return.
2261 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2262 __releases(rcu_get_root(rsp)->lock)
2264 raw_lockdep_assert_held_rcu_node(rcu_get_root(rsp));
2265 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2266 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2267 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2268 rcu_gp_kthread_wake(rsp);
2272 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2273 * Allows quiescent states for a group of CPUs to be reported at one go
2274 * to the specified rcu_node structure, though all the CPUs in the group
2275 * must be represented by the same rcu_node structure (which need not be a
2276 * leaf rcu_node structure, though it often will be). The gps parameter
2277 * is the grace-period snapshot, which means that the quiescent states
2278 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2279 * must be held upon entry, and it is released before return.
2281 * As a special case, if mask is zero, the bit-already-cleared check is
2282 * disabled. This allows propagating quiescent state due to resumed tasks
2283 * during grace-period initialization.
2286 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2287 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2288 __releases(rnp->lock)
2290 unsigned long oldmask = 0;
2291 struct rcu_node *rnp_c;
2293 raw_lockdep_assert_held_rcu_node(rnp);
2295 /* Walk up the rcu_node hierarchy. */
2297 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2300 * Our bit has already been cleared, or the
2301 * relevant grace period is already over, so done.
2303 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2306 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2307 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2308 rcu_preempt_blocked_readers_cgp(rnp));
2309 rnp->qsmask &= ~mask;
2310 trace_rcu_quiescent_state_report(rsp->name, rnp->gp_seq,
2311 mask, rnp->qsmask, rnp->level,
2312 rnp->grplo, rnp->grphi,
2314 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2316 /* Other bits still set at this level, so done. */
2317 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2320 rnp->completedqs = rnp->gp_seq;
2321 mask = rnp->grpmask;
2322 if (rnp->parent == NULL) {
2324 /* No more levels. Exit loop holding root lock. */
2328 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2331 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2332 oldmask = rnp_c->qsmask;
2336 * Get here if we are the last CPU to pass through a quiescent
2337 * state for this grace period. Invoke rcu_report_qs_rsp()
2338 * to clean up and start the next grace period if one is needed.
2340 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2344 * Record a quiescent state for all tasks that were previously queued
2345 * on the specified rcu_node structure and that were blocking the current
2346 * RCU grace period. The caller must hold the specified rnp->lock with
2347 * irqs disabled, and this lock is released upon return, but irqs remain
2350 static void __maybe_unused
2351 rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2352 struct rcu_node *rnp, unsigned long flags)
2353 __releases(rnp->lock)
2357 struct rcu_node *rnp_p;
2359 raw_lockdep_assert_held_rcu_node(rnp);
2360 if (WARN_ON_ONCE(rcu_state_p == &rcu_sched_state) ||
2361 WARN_ON_ONCE(rsp != rcu_state_p) ||
2362 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2364 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2365 return; /* Still need more quiescent states! */
2368 rnp->completedqs = rnp->gp_seq;
2369 rnp_p = rnp->parent;
2370 if (rnp_p == NULL) {
2372 * Only one rcu_node structure in the tree, so don't
2373 * try to report up to its nonexistent parent!
2375 rcu_report_qs_rsp(rsp, flags);
2379 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2381 mask = rnp->grpmask;
2382 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2383 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2384 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2388 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2389 * structure. This must be called from the specified CPU.
2392 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2394 unsigned long flags;
2397 struct rcu_node *rnp;
2400 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2401 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2405 * The grace period in which this quiescent state was
2406 * recorded has ended, so don't report it upwards.
2407 * We will instead need a new quiescent state that lies
2408 * within the current grace period.
2410 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2411 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2412 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2415 mask = rdp->grpmask;
2416 if ((rnp->qsmask & mask) == 0) {
2417 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2419 rdp->core_needs_qs = false;
2422 * This GP can't end until cpu checks in, so all of our
2423 * callbacks can be processed during the next GP.
2425 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2427 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2428 /* ^^^ Released rnp->lock */
2430 rcu_gp_kthread_wake(rsp);
2435 * Check to see if there is a new grace period of which this CPU
2436 * is not yet aware, and if so, set up local rcu_data state for it.
2437 * Otherwise, see if this CPU has just passed through its first
2438 * quiescent state for this grace period, and record that fact if so.
2441 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2443 /* Check for grace-period ends and beginnings. */
2444 note_gp_changes(rsp, rdp);
2447 * Does this CPU still need to do its part for current grace period?
2448 * If no, return and let the other CPUs do their part as well.
2450 if (!rdp->core_needs_qs)
2454 * Was there a quiescent state since the beginning of the grace
2455 * period? If no, then exit and wait for the next call.
2457 if (rdp->cpu_no_qs.b.norm)
2461 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2464 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2468 * Trace the fact that this CPU is going offline.
2470 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2472 RCU_TRACE(bool blkd;)
2473 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2474 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2476 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2479 RCU_TRACE(blkd = !!(rnp->qsmask & rdp->grpmask);)
2480 trace_rcu_grace_period(rsp->name, rnp->gp_seq,
2481 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2485 * All CPUs for the specified rcu_node structure have gone offline,
2486 * and all tasks that were preempted within an RCU read-side critical
2487 * section while running on one of those CPUs have since exited their RCU
2488 * read-side critical section. Some other CPU is reporting this fact with
2489 * the specified rcu_node structure's ->lock held and interrupts disabled.
2490 * This function therefore goes up the tree of rcu_node structures,
2491 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2492 * the leaf rcu_node structure's ->qsmaskinit field has already been
2495 * This function does check that the specified rcu_node structure has
2496 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2497 * prematurely. That said, invoking it after the fact will cost you
2498 * a needless lock acquisition. So once it has done its work, don't
2501 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2504 struct rcu_node *rnp = rnp_leaf;
2506 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2507 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2508 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2509 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2512 mask = rnp->grpmask;
2516 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2517 rnp->qsmaskinit &= ~mask;
2518 /* Between grace periods, so better already be zero! */
2519 WARN_ON_ONCE(rnp->qsmask);
2520 if (rnp->qsmaskinit) {
2521 raw_spin_unlock_rcu_node(rnp);
2522 /* irqs remain disabled. */
2525 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2530 * The CPU has been completely removed, and some other CPU is reporting
2531 * this fact from process context. Do the remainder of the cleanup.
2532 * There can only be one CPU hotplug operation at a time, so no need for
2535 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2537 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2538 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2540 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2543 /* Adjust any no-longer-needed kthreads. */
2544 rcu_boost_kthread_setaffinity(rnp, -1);
2548 * Invoke any RCU callbacks that have made it to the end of their grace
2549 * period. Thottle as specified by rdp->blimit.
2551 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2553 unsigned long flags;
2554 struct rcu_head *rhp;
2555 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2558 /* If no callbacks are ready, just return. */
2559 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2560 trace_rcu_batch_start(rsp->name,
2561 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2562 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2563 trace_rcu_batch_end(rsp->name, 0,
2564 !rcu_segcblist_empty(&rdp->cblist),
2565 need_resched(), is_idle_task(current),
2566 rcu_is_callbacks_kthread());
2571 * Extract the list of ready callbacks, disabling to prevent
2572 * races with call_rcu() from interrupt handlers. Leave the
2573 * callback counts, as rcu_barrier() needs to be conservative.
2575 local_irq_save(flags);
2576 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2578 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2579 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2580 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2581 local_irq_restore(flags);
2583 /* Invoke callbacks. */
2584 rhp = rcu_cblist_dequeue(&rcl);
2585 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2586 debug_rcu_head_unqueue(rhp);
2587 if (__rcu_reclaim(rsp->name, rhp))
2588 rcu_cblist_dequeued_lazy(&rcl);
2590 * Stop only if limit reached and CPU has something to do.
2591 * Note: The rcl structure counts down from zero.
2593 if (-rcl.len >= bl &&
2595 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2599 local_irq_save(flags);
2601 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2602 is_idle_task(current), rcu_is_callbacks_kthread());
2604 /* Update counts and requeue any remaining callbacks. */
2605 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2606 smp_mb(); /* List handling before counting for rcu_barrier(). */
2607 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2609 /* Reinstate batch limit if we have worked down the excess. */
2610 count = rcu_segcblist_n_cbs(&rdp->cblist);
2611 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2612 rdp->blimit = blimit;
2614 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2615 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2616 rdp->qlen_last_fqs_check = 0;
2617 rdp->n_force_qs_snap = rsp->n_force_qs;
2618 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2619 rdp->qlen_last_fqs_check = count;
2622 * The following usually indicates a double call_rcu(). To track
2623 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2625 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2627 local_irq_restore(flags);
2629 /* Re-invoke RCU core processing if there are callbacks remaining. */
2630 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2635 * Check to see if this CPU is in a non-context-switch quiescent state
2636 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2637 * Also schedule RCU core processing.
2639 * This function must be called from hardirq context. It is normally
2640 * invoked from the scheduling-clock interrupt.
2642 void rcu_check_callbacks(int user)
2644 trace_rcu_utilization(TPS("Start scheduler-tick"));
2645 increment_cpu_stall_ticks();
2646 if (user || rcu_is_cpu_rrupt_from_idle()) {
2649 * Get here if this CPU took its interrupt from user
2650 * mode or from the idle loop, and if this is not a
2651 * nested interrupt. In this case, the CPU is in
2652 * a quiescent state, so note it.
2654 * No memory barrier is required here because both
2655 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2656 * variables that other CPUs neither access nor modify,
2657 * at least not while the corresponding CPU is online.
2662 rcu_note_voluntary_context_switch(current);
2664 } else if (!in_softirq()) {
2667 * Get here if this CPU did not take its interrupt from
2668 * softirq, in other words, if it is not interrupting
2669 * a rcu_bh read-side critical section. This is an _bh
2670 * critical section, so note it.
2675 rcu_preempt_check_callbacks();
2679 trace_rcu_utilization(TPS("End scheduler-tick"));
2683 * Scan the leaf rcu_node structures, processing dyntick state for any that
2684 * have not yet encountered a quiescent state, using the function specified.
2685 * Also initiate boosting for any threads blocked on the root rcu_node.
2687 * The caller must have suppressed start of new grace periods.
2689 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2692 unsigned long flags;
2694 struct rcu_node *rnp;
2696 rcu_for_each_leaf_node(rsp, rnp) {
2697 cond_resched_tasks_rcu_qs();
2699 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2700 if (rnp->qsmask == 0) {
2701 if (rcu_state_p == &rcu_sched_state ||
2702 rsp != rcu_state_p ||
2703 rcu_preempt_blocked_readers_cgp(rnp)) {
2705 * No point in scanning bits because they
2706 * are all zero. But we might need to
2707 * priority-boost blocked readers.
2709 rcu_initiate_boost(rnp, flags);
2710 /* rcu_initiate_boost() releases rnp->lock */
2713 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2716 for_each_leaf_node_possible_cpu(rnp, cpu) {
2717 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2718 if ((rnp->qsmask & bit) != 0) {
2719 if (f(per_cpu_ptr(rsp->rda, cpu)))
2724 /* Idle/offline CPUs, report (releases rnp->lock). */
2725 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2727 /* Nothing to do here, so just drop the lock. */
2728 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2734 * Force quiescent states on reluctant CPUs, and also detect which
2735 * CPUs are in dyntick-idle mode.
2737 static void force_quiescent_state(struct rcu_state *rsp)
2739 unsigned long flags;
2741 struct rcu_node *rnp;
2742 struct rcu_node *rnp_old = NULL;
2744 /* Funnel through hierarchy to reduce memory contention. */
2745 rnp = __this_cpu_read(rsp->rda->mynode);
2746 for (; rnp != NULL; rnp = rnp->parent) {
2747 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2748 !raw_spin_trylock(&rnp->fqslock);
2749 if (rnp_old != NULL)
2750 raw_spin_unlock(&rnp_old->fqslock);
2755 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2757 /* Reached the root of the rcu_node tree, acquire lock. */
2758 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2759 raw_spin_unlock(&rnp_old->fqslock);
2760 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2761 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2762 return; /* Someone beat us to it. */
2764 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2765 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2766 rcu_gp_kthread_wake(rsp);
2770 * This function checks for grace-period requests that fail to motivate
2771 * RCU to come out of its idle mode.
2774 rcu_check_gp_start_stall(struct rcu_state *rsp, struct rcu_node *rnp,
2775 struct rcu_data *rdp)
2777 const unsigned long gpssdelay = rcu_jiffies_till_stall_check() * HZ;
2778 unsigned long flags;
2780 struct rcu_node *rnp_root = rcu_get_root(rsp);
2781 static atomic_t warned = ATOMIC_INIT(0);
2783 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress(rsp) ||
2784 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed))
2786 j = jiffies; /* Expensive access, and in common case don't get here. */
2787 if (time_before(j, READ_ONCE(rsp->gp_req_activity) + gpssdelay) ||
2788 time_before(j, READ_ONCE(rsp->gp_activity) + gpssdelay) ||
2789 atomic_read(&warned))
2792 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2794 if (rcu_gp_in_progress(rsp) ||
2795 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2796 time_before(j, READ_ONCE(rsp->gp_req_activity) + gpssdelay) ||
2797 time_before(j, READ_ONCE(rsp->gp_activity) + gpssdelay) ||
2798 atomic_read(&warned)) {
2799 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2802 /* Hold onto the leaf lock to make others see warned==1. */
2804 if (rnp_root != rnp)
2805 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
2807 if (rcu_gp_in_progress(rsp) ||
2808 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2809 time_before(j, rsp->gp_req_activity + gpssdelay) ||
2810 time_before(j, rsp->gp_activity + gpssdelay) ||
2811 atomic_xchg(&warned, 1)) {
2812 raw_spin_unlock_rcu_node(rnp_root); /* irqs remain disabled. */
2813 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2816 pr_alert("%s: g%ld->%ld gar:%lu ga:%lu f%#x gs:%d %s->state:%#lx\n",
2817 __func__, (long)READ_ONCE(rsp->gp_seq),
2818 (long)READ_ONCE(rnp_root->gp_seq_needed),
2819 j - rsp->gp_req_activity, j - rsp->gp_activity,
2820 rsp->gp_flags, rsp->gp_state, rsp->name,
2821 rsp->gp_kthread ? rsp->gp_kthread->state : 0x1ffffL);
2823 if (rnp_root != rnp)
2824 raw_spin_unlock_rcu_node(rnp_root);
2825 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2829 * This does the RCU core processing work for the specified rcu_state
2830 * and rcu_data structures. This may be called only from the CPU to
2831 * whom the rdp belongs.
2834 __rcu_process_callbacks(struct rcu_state *rsp)
2836 unsigned long flags;
2837 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2838 struct rcu_node *rnp = rdp->mynode;
2840 WARN_ON_ONCE(!rdp->beenonline);
2842 /* Update RCU state based on any recent quiescent states. */
2843 rcu_check_quiescent_state(rsp, rdp);
2845 /* No grace period and unregistered callbacks? */
2846 if (!rcu_gp_in_progress(rsp) &&
2847 rcu_segcblist_is_enabled(&rdp->cblist)) {
2848 local_irq_save(flags);
2849 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2850 rcu_accelerate_cbs_unlocked(rsp, rnp, rdp);
2851 local_irq_restore(flags);
2854 rcu_check_gp_start_stall(rsp, rnp, rdp);
2856 /* If there are callbacks ready, invoke them. */
2857 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2858 invoke_rcu_callbacks(rsp, rdp);
2860 /* Do any needed deferred wakeups of rcuo kthreads. */
2861 do_nocb_deferred_wakeup(rdp);
2865 * Do RCU core processing for the current CPU.
2867 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2869 struct rcu_state *rsp;
2871 if (cpu_is_offline(smp_processor_id()))
2873 trace_rcu_utilization(TPS("Start RCU core"));
2874 for_each_rcu_flavor(rsp)
2875 __rcu_process_callbacks(rsp);
2876 trace_rcu_utilization(TPS("End RCU core"));
2880 * Schedule RCU callback invocation. If the specified type of RCU
2881 * does not support RCU priority boosting, just do a direct call,
2882 * otherwise wake up the per-CPU kernel kthread. Note that because we
2883 * are running on the current CPU with softirqs disabled, the
2884 * rcu_cpu_kthread_task cannot disappear out from under us.
2886 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2888 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2890 if (likely(!rsp->boost)) {
2891 rcu_do_batch(rsp, rdp);
2894 invoke_rcu_callbacks_kthread();
2897 static void invoke_rcu_core(void)
2899 if (cpu_online(smp_processor_id()))
2900 raise_softirq(RCU_SOFTIRQ);
2904 * Handle any core-RCU processing required by a call_rcu() invocation.
2906 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2907 struct rcu_head *head, unsigned long flags)
2910 * If called from an extended quiescent state, invoke the RCU
2911 * core in order to force a re-evaluation of RCU's idleness.
2913 if (!rcu_is_watching())
2916 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2917 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2921 * Force the grace period if too many callbacks or too long waiting.
2922 * Enforce hysteresis, and don't invoke force_quiescent_state()
2923 * if some other CPU has recently done so. Also, don't bother
2924 * invoking force_quiescent_state() if the newly enqueued callback
2925 * is the only one waiting for a grace period to complete.
2927 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2928 rdp->qlen_last_fqs_check + qhimark)) {
2930 /* Are we ignoring a completed grace period? */
2931 note_gp_changes(rsp, rdp);
2933 /* Start a new grace period if one not already started. */
2934 if (!rcu_gp_in_progress(rsp)) {
2935 rcu_accelerate_cbs_unlocked(rsp, rdp->mynode, rdp);
2937 /* Give the grace period a kick. */
2938 rdp->blimit = LONG_MAX;
2939 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2940 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2941 force_quiescent_state(rsp);
2942 rdp->n_force_qs_snap = rsp->n_force_qs;
2943 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2949 * RCU callback function to leak a callback.
2951 static void rcu_leak_callback(struct rcu_head *rhp)
2956 * Helper function for call_rcu() and friends. The cpu argument will
2957 * normally be -1, indicating "currently running CPU". It may specify
2958 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2959 * is expected to specify a CPU.
2962 __call_rcu(struct rcu_head *head, rcu_callback_t func,
2963 struct rcu_state *rsp, int cpu, bool lazy)
2965 unsigned long flags;
2966 struct rcu_data *rdp;
2968 /* Misaligned rcu_head! */
2969 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2971 if (debug_rcu_head_queue(head)) {
2973 * Probable double call_rcu(), so leak the callback.
2974 * Use rcu:rcu_callback trace event to find the previous
2975 * time callback was passed to __call_rcu().
2977 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
2979 WRITE_ONCE(head->func, rcu_leak_callback);
2984 local_irq_save(flags);
2985 rdp = this_cpu_ptr(rsp->rda);
2987 /* Add the callback to our list. */
2988 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
2992 rdp = per_cpu_ptr(rsp->rda, cpu);
2993 if (likely(rdp->mynode)) {
2994 /* Post-boot, so this should be for a no-CBs CPU. */
2995 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2996 WARN_ON_ONCE(offline);
2997 /* Offline CPU, _call_rcu() illegal, leak callback. */
2998 local_irq_restore(flags);
3002 * Very early boot, before rcu_init(). Initialize if needed
3003 * and then drop through to queue the callback.
3006 WARN_ON_ONCE(!rcu_is_watching());
3007 if (rcu_segcblist_empty(&rdp->cblist))
3008 rcu_segcblist_init(&rdp->cblist);
3010 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3012 rcu_idle_count_callbacks_posted();
3014 if (__is_kfree_rcu_offset((unsigned long)func))
3015 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3016 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3017 rcu_segcblist_n_cbs(&rdp->cblist));
3019 trace_rcu_callback(rsp->name, head,
3020 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3021 rcu_segcblist_n_cbs(&rdp->cblist));
3023 /* Go handle any RCU core processing required. */
3024 __call_rcu_core(rsp, rdp, head, flags);
3025 local_irq_restore(flags);
3029 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3030 * @head: structure to be used for queueing the RCU updates.
3031 * @func: actual callback function to be invoked after the grace period
3033 * The callback function will be invoked some time after a full grace
3034 * period elapses, in other words after all currently executing RCU
3035 * read-side critical sections have completed. call_rcu_sched() assumes
3036 * that the read-side critical sections end on enabling of preemption
3037 * or on voluntary preemption.
3038 * RCU read-side critical sections are delimited by:
3040 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3041 * - anything that disables preemption.
3043 * These may be nested.
3045 * See the description of call_rcu() for more detailed information on
3046 * memory ordering guarantees.
3048 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3050 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3052 EXPORT_SYMBOL_GPL(call_rcu_sched);
3055 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3056 * @head: structure to be used for queueing the RCU updates.
3057 * @func: actual callback function to be invoked after the grace period
3059 * The callback function will be invoked some time after a full grace
3060 * period elapses, in other words after all currently executing RCU
3061 * read-side critical sections have completed. call_rcu_bh() assumes
3062 * that the read-side critical sections end on completion of a softirq
3063 * handler. This means that read-side critical sections in process
3064 * context must not be interrupted by softirqs. This interface is to be
3065 * used when most of the read-side critical sections are in softirq context.
3066 * RCU read-side critical sections are delimited by:
3068 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3069 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3071 * These may be nested.
3073 * See the description of call_rcu() for more detailed information on
3074 * memory ordering guarantees.
3076 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3078 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3080 EXPORT_SYMBOL_GPL(call_rcu_bh);
3083 * Queue an RCU callback for lazy invocation after a grace period.
3084 * This will likely be later named something like "call_rcu_lazy()",
3085 * but this change will require some way of tagging the lazy RCU
3086 * callbacks in the list of pending callbacks. Until then, this
3087 * function may only be called from __kfree_rcu().
3089 void kfree_call_rcu(struct rcu_head *head,
3090 rcu_callback_t func)
3092 __call_rcu(head, func, rcu_state_p, -1, 1);
3094 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3097 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3098 * any blocking grace-period wait automatically implies a grace period
3099 * if there is only one CPU online at any point time during execution
3100 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3101 * occasionally incorrectly indicate that there are multiple CPUs online
3102 * when there was in fact only one the whole time, as this just adds
3103 * some overhead: RCU still operates correctly.
3105 static int rcu_blocking_is_gp(void)
3109 might_sleep(); /* Check for RCU read-side critical section. */
3111 ret = num_online_cpus() <= 1;
3117 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3119 * Control will return to the caller some time after a full rcu-sched
3120 * grace period has elapsed, in other words after all currently executing
3121 * rcu-sched read-side critical sections have completed. These read-side
3122 * critical sections are delimited by rcu_read_lock_sched() and
3123 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3124 * local_irq_disable(), and so on may be used in place of
3125 * rcu_read_lock_sched().
3127 * This means that all preempt_disable code sequences, including NMI and
3128 * non-threaded hardware-interrupt handlers, in progress on entry will
3129 * have completed before this primitive returns. However, this does not
3130 * guarantee that softirq handlers will have completed, since in some
3131 * kernels, these handlers can run in process context, and can block.
3133 * Note that this guarantee implies further memory-ordering guarantees.
3134 * On systems with more than one CPU, when synchronize_sched() returns,
3135 * each CPU is guaranteed to have executed a full memory barrier since the
3136 * end of its last RCU-sched read-side critical section whose beginning
3137 * preceded the call to synchronize_sched(). In addition, each CPU having
3138 * an RCU read-side critical section that extends beyond the return from
3139 * synchronize_sched() is guaranteed to have executed a full memory barrier
3140 * after the beginning of synchronize_sched() and before the beginning of
3141 * that RCU read-side critical section. Note that these guarantees include
3142 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3143 * that are executing in the kernel.
3145 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3146 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3147 * to have executed a full memory barrier during the execution of
3148 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3149 * again only if the system has more than one CPU).
3151 void synchronize_sched(void)
3153 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3154 lock_is_held(&rcu_lock_map) ||
3155 lock_is_held(&rcu_sched_lock_map),
3156 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3157 if (rcu_blocking_is_gp())
3159 if (rcu_gp_is_expedited())
3160 synchronize_sched_expedited();
3162 wait_rcu_gp(call_rcu_sched);
3164 EXPORT_SYMBOL_GPL(synchronize_sched);
3167 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3169 * Control will return to the caller some time after a full rcu_bh grace
3170 * period has elapsed, in other words after all currently executing rcu_bh
3171 * read-side critical sections have completed. RCU read-side critical
3172 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3173 * and may be nested.
3175 * See the description of synchronize_sched() for more detailed information
3176 * on memory ordering guarantees.
3178 void synchronize_rcu_bh(void)
3180 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3181 lock_is_held(&rcu_lock_map) ||
3182 lock_is_held(&rcu_sched_lock_map),
3183 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3184 if (rcu_blocking_is_gp())
3186 if (rcu_gp_is_expedited())
3187 synchronize_rcu_bh_expedited();
3189 wait_rcu_gp(call_rcu_bh);
3191 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3194 * get_state_synchronize_rcu - Snapshot current RCU state
3196 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3197 * to determine whether or not a full grace period has elapsed in the
3200 unsigned long get_state_synchronize_rcu(void)
3203 * Any prior manipulation of RCU-protected data must happen
3204 * before the load from ->gp_seq.
3207 return rcu_seq_snap(&rcu_state_p->gp_seq);
3209 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3212 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3214 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3216 * If a full RCU grace period has elapsed since the earlier call to
3217 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3218 * synchronize_rcu() to wait for a full grace period.
3220 * Yes, this function does not take counter wrap into account. But
3221 * counter wrap is harmless. If the counter wraps, we have waited for
3222 * more than 2 billion grace periods (and way more on a 64-bit system!),
3223 * so waiting for one additional grace period should be just fine.
3225 void cond_synchronize_rcu(unsigned long oldstate)
3227 if (!rcu_seq_done(&rcu_state_p->gp_seq, oldstate))
3230 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3232 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3235 * get_state_synchronize_sched - Snapshot current RCU-sched state
3237 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3238 * to determine whether or not a full grace period has elapsed in the
3241 unsigned long get_state_synchronize_sched(void)
3244 * Any prior manipulation of RCU-protected data must happen
3245 * before the load from ->gp_seq.
3248 return rcu_seq_snap(&rcu_sched_state.gp_seq);
3250 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3253 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3255 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3257 * If a full RCU-sched grace period has elapsed since the earlier call to
3258 * get_state_synchronize_sched(), just return. Otherwise, invoke
3259 * synchronize_sched() to wait for a full grace period.
3261 * Yes, this function does not take counter wrap into account. But
3262 * counter wrap is harmless. If the counter wraps, we have waited for
3263 * more than 2 billion grace periods (and way more on a 64-bit system!),
3264 * so waiting for one additional grace period should be just fine.
3266 void cond_synchronize_sched(unsigned long oldstate)
3268 if (!rcu_seq_done(&rcu_sched_state.gp_seq, oldstate))
3269 synchronize_sched();
3271 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3273 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3276 * Check to see if there is any immediate RCU-related work to be done
3277 * by the current CPU, for the specified type of RCU, returning 1 if so.
3278 * The checks are in order of increasing expense: checks that can be
3279 * carried out against CPU-local state are performed first. However,
3280 * we must check for CPU stalls first, else we might not get a chance.
3282 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3284 struct rcu_node *rnp = rdp->mynode;
3286 /* Check for CPU stalls, if enabled. */
3287 check_cpu_stall(rsp, rdp);
3289 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3290 if (rcu_nohz_full_cpu(rsp))
3293 /* Is the RCU core waiting for a quiescent state from this CPU? */
3294 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
3297 /* Does this CPU have callbacks ready to invoke? */
3298 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3301 /* Has RCU gone idle with this CPU needing another grace period? */
3302 if (!rcu_gp_in_progress(rsp) &&
3303 rcu_segcblist_is_enabled(&rdp->cblist) &&
3304 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3307 /* Have RCU grace period completed or started? */
3308 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3309 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3312 /* Does this CPU need a deferred NOCB wakeup? */
3313 if (rcu_nocb_need_deferred_wakeup(rdp))
3321 * Check to see if there is any immediate RCU-related work to be done
3322 * by the current CPU, returning 1 if so. This function is part of the
3323 * RCU implementation; it is -not- an exported member of the RCU API.
3325 static int rcu_pending(void)
3327 struct rcu_state *rsp;
3329 for_each_rcu_flavor(rsp)
3330 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3336 * Return true if the specified CPU has any callback. If all_lazy is
3337 * non-NULL, store an indication of whether all callbacks are lazy.
3338 * (If there are no callbacks, all of them are deemed to be lazy.)
3340 static bool rcu_cpu_has_callbacks(bool *all_lazy)
3344 struct rcu_data *rdp;
3345 struct rcu_state *rsp;
3347 for_each_rcu_flavor(rsp) {
3348 rdp = this_cpu_ptr(rsp->rda);
3349 if (rcu_segcblist_empty(&rdp->cblist))
3352 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3363 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3364 * the compiler is expected to optimize this away.
3366 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3367 int cpu, unsigned long done)
3369 trace_rcu_barrier(rsp->name, s, cpu,
3370 atomic_read(&rsp->barrier_cpu_count), done);
3374 * RCU callback function for _rcu_barrier(). If we are last, wake
3375 * up the task executing _rcu_barrier().
3377 static void rcu_barrier_callback(struct rcu_head *rhp)
3379 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3380 struct rcu_state *rsp = rdp->rsp;
3382 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3383 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3384 rsp->barrier_sequence);
3385 complete(&rsp->barrier_completion);
3387 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3392 * Called with preemption disabled, and from cross-cpu IRQ context.
3394 static void rcu_barrier_func(void *type)
3396 struct rcu_state *rsp = type;
3397 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3399 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3400 rdp->barrier_head.func = rcu_barrier_callback;
3401 debug_rcu_head_queue(&rdp->barrier_head);
3402 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3403 atomic_inc(&rsp->barrier_cpu_count);
3405 debug_rcu_head_unqueue(&rdp->barrier_head);
3406 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3407 rsp->barrier_sequence);
3412 * Orchestrate the specified type of RCU barrier, waiting for all
3413 * RCU callbacks of the specified type to complete.
3415 static void _rcu_barrier(struct rcu_state *rsp)
3418 struct rcu_data *rdp;
3419 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3421 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3423 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3424 mutex_lock(&rsp->barrier_mutex);
3426 /* Did someone else do our work for us? */
3427 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3428 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3429 rsp->barrier_sequence);
3430 smp_mb(); /* caller's subsequent code after above check. */
3431 mutex_unlock(&rsp->barrier_mutex);
3435 /* Mark the start of the barrier operation. */
3436 rcu_seq_start(&rsp->barrier_sequence);
3437 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3440 * Initialize the count to one rather than to zero in order to
3441 * avoid a too-soon return to zero in case of a short grace period
3442 * (or preemption of this task). Exclude CPU-hotplug operations
3443 * to ensure that no offline CPU has callbacks queued.
3445 init_completion(&rsp->barrier_completion);
3446 atomic_set(&rsp->barrier_cpu_count, 1);
3450 * Force each CPU with callbacks to register a new callback.
3451 * When that callback is invoked, we will know that all of the
3452 * corresponding CPU's preceding callbacks have been invoked.
3454 for_each_possible_cpu(cpu) {
3455 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3457 rdp = per_cpu_ptr(rsp->rda, cpu);
3458 if (rcu_is_nocb_cpu(cpu)) {
3459 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3460 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3461 rsp->barrier_sequence);
3463 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3464 rsp->barrier_sequence);
3465 smp_mb__before_atomic();
3466 atomic_inc(&rsp->barrier_cpu_count);
3467 __call_rcu(&rdp->barrier_head,
3468 rcu_barrier_callback, rsp, cpu, 0);
3470 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3471 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3472 rsp->barrier_sequence);
3473 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3475 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3476 rsp->barrier_sequence);
3482 * Now that we have an rcu_barrier_callback() callback on each
3483 * CPU, and thus each counted, remove the initial count.
3485 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3486 complete(&rsp->barrier_completion);
3488 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3489 wait_for_completion(&rsp->barrier_completion);
3491 /* Mark the end of the barrier operation. */
3492 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3493 rcu_seq_end(&rsp->barrier_sequence);
3495 /* Other rcu_barrier() invocations can now safely proceed. */
3496 mutex_unlock(&rsp->barrier_mutex);
3500 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3502 void rcu_barrier_bh(void)
3504 _rcu_barrier(&rcu_bh_state);
3506 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3509 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3511 void rcu_barrier_sched(void)
3513 _rcu_barrier(&rcu_sched_state);
3515 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3518 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3519 * first CPU in a given leaf rcu_node structure coming online. The caller
3520 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3523 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3527 struct rcu_node *rnp = rnp_leaf;
3529 raw_lockdep_assert_held_rcu_node(rnp_leaf);
3530 WARN_ON_ONCE(rnp->wait_blkd_tasks);
3532 mask = rnp->grpmask;
3536 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3537 oldmask = rnp->qsmaskinit;
3538 rnp->qsmaskinit |= mask;
3539 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3546 * Do boot-time initialization of a CPU's per-CPU RCU data.
3549 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3551 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3553 /* Set up local state, ensuring consistent view of global state. */
3554 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3555 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3556 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3557 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3558 rdp->rcu_ofl_gp_seq = rsp->gp_seq;
3559 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3560 rdp->rcu_onl_gp_seq = rsp->gp_seq;
3561 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3564 rcu_boot_init_nocb_percpu_data(rdp);
3568 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3569 * offline event can be happening at a given time. Note also that we can
3570 * accept some slop in the rsp->gp_seq access due to the fact that this
3571 * CPU cannot possibly have any RCU callbacks in flight yet.
3574 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3576 unsigned long flags;
3577 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3578 struct rcu_node *rnp = rcu_get_root(rsp);
3580 /* Set up local state, ensuring consistent view of global state. */
3581 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3582 rdp->qlen_last_fqs_check = 0;
3583 rdp->n_force_qs_snap = rsp->n_force_qs;
3584 rdp->blimit = blimit;
3585 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3586 !init_nocb_callback_list(rdp))
3587 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3588 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3589 rcu_dynticks_eqs_online();
3590 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3593 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3594 * propagation up the rcu_node tree will happen at the beginning
3595 * of the next grace period.
3598 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3599 rdp->beenonline = true; /* We have now been online. */
3600 rdp->gp_seq = rnp->gp_seq;
3601 rdp->gp_seq_needed = rnp->gp_seq;
3602 rdp->cpu_no_qs.b.norm = true;
3603 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3604 rdp->core_needs_qs = false;
3605 rdp->rcu_iw_pending = false;
3606 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3607 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpuonl"));
3608 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3612 * Invoked early in the CPU-online process, when pretty much all
3613 * services are available. The incoming CPU is not present.
3615 int rcutree_prepare_cpu(unsigned int cpu)
3617 struct rcu_state *rsp;
3619 for_each_rcu_flavor(rsp)
3620 rcu_init_percpu_data(cpu, rsp);
3622 rcu_prepare_kthreads(cpu);
3623 rcu_spawn_all_nocb_kthreads(cpu);
3629 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3631 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3633 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3635 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3639 * Near the end of the CPU-online process. Pretty much all services
3640 * enabled, and the CPU is now very much alive.
3642 int rcutree_online_cpu(unsigned int cpu)
3644 unsigned long flags;
3645 struct rcu_data *rdp;
3646 struct rcu_node *rnp;
3647 struct rcu_state *rsp;
3649 for_each_rcu_flavor(rsp) {
3650 rdp = per_cpu_ptr(rsp->rda, cpu);
3652 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3653 rnp->ffmask |= rdp->grpmask;
3654 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3656 if (IS_ENABLED(CONFIG_TREE_SRCU))
3657 srcu_online_cpu(cpu);
3658 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3659 return 0; /* Too early in boot for scheduler work. */
3660 sync_sched_exp_online_cleanup(cpu);
3661 rcutree_affinity_setting(cpu, -1);
3666 * Near the beginning of the process. The CPU is still very much alive
3667 * with pretty much all services enabled.
3669 int rcutree_offline_cpu(unsigned int cpu)
3671 unsigned long flags;
3672 struct rcu_data *rdp;
3673 struct rcu_node *rnp;
3674 struct rcu_state *rsp;
3676 for_each_rcu_flavor(rsp) {
3677 rdp = per_cpu_ptr(rsp->rda, cpu);
3679 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3680 rnp->ffmask &= ~rdp->grpmask;
3681 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3684 rcutree_affinity_setting(cpu, cpu);
3685 if (IS_ENABLED(CONFIG_TREE_SRCU))
3686 srcu_offline_cpu(cpu);
3691 * Near the end of the offline process. We do only tracing here.
3693 int rcutree_dying_cpu(unsigned int cpu)
3695 struct rcu_state *rsp;
3697 for_each_rcu_flavor(rsp)
3698 rcu_cleanup_dying_cpu(rsp);
3703 * The outgoing CPU is gone and we are running elsewhere.
3705 int rcutree_dead_cpu(unsigned int cpu)
3707 struct rcu_state *rsp;
3709 for_each_rcu_flavor(rsp) {
3710 rcu_cleanup_dead_cpu(cpu, rsp);
3711 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3716 static DEFINE_PER_CPU(int, rcu_cpu_started);
3719 * Mark the specified CPU as being online so that subsequent grace periods
3720 * (both expedited and normal) will wait on it. Note that this means that
3721 * incoming CPUs are not allowed to use RCU read-side critical sections
3722 * until this function is called. Failing to observe this restriction
3723 * will result in lockdep splats.
3725 * Note that this function is special in that it is invoked directly
3726 * from the incoming CPU rather than from the cpuhp_step mechanism.
3727 * This is because this function must be invoked at a precise location.
3729 void rcu_cpu_starting(unsigned int cpu)
3731 unsigned long flags;
3734 unsigned long oldmask;
3735 struct rcu_data *rdp;
3736 struct rcu_node *rnp;
3737 struct rcu_state *rsp;
3739 if (per_cpu(rcu_cpu_started, cpu))
3742 per_cpu(rcu_cpu_started, cpu) = 1;
3744 for_each_rcu_flavor(rsp) {
3745 rdp = per_cpu_ptr(rsp->rda, cpu);
3747 mask = rdp->grpmask;
3748 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3749 rnp->qsmaskinitnext |= mask;
3750 oldmask = rnp->expmaskinitnext;
3751 rnp->expmaskinitnext |= mask;
3752 oldmask ^= rnp->expmaskinitnext;
3753 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3754 /* Allow lockless access for expedited grace periods. */
3755 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3756 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3757 rdp->rcu_onl_gp_seq = READ_ONCE(rsp->gp_seq);
3758 rdp->rcu_onl_gp_flags = READ_ONCE(rsp->gp_flags);
3759 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3760 /* Report QS -after- changing ->qsmaskinitnext! */
3761 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
3763 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3766 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3769 #ifdef CONFIG_HOTPLUG_CPU
3771 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3772 * function. We now remove it from the rcu_node tree's ->qsmaskinitnext
3775 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3777 unsigned long flags;
3779 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3780 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3782 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3783 mask = rdp->grpmask;
3784 spin_lock(&rsp->ofl_lock);
3785 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3786 rdp->rcu_ofl_gp_seq = READ_ONCE(rsp->gp_seq);
3787 rdp->rcu_ofl_gp_flags = READ_ONCE(rsp->gp_flags);
3788 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3789 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3790 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
3791 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3793 rnp->qsmaskinitnext &= ~mask;
3794 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3795 spin_unlock(&rsp->ofl_lock);
3799 * The outgoing function has no further need of RCU, so remove it from
3800 * the list of CPUs that RCU must track.
3802 * Note that this function is special in that it is invoked directly
3803 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3804 * This is because this function must be invoked at a precise location.
3806 void rcu_report_dead(unsigned int cpu)
3808 struct rcu_state *rsp;
3810 /* QS for any half-done expedited RCU-sched GP. */
3812 rcu_report_exp_rdp(&rcu_sched_state,
3813 this_cpu_ptr(rcu_sched_state.rda), true);
3815 for_each_rcu_flavor(rsp)
3816 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3818 per_cpu(rcu_cpu_started, cpu) = 0;
3821 /* Migrate the dead CPU's callbacks to the current CPU. */
3822 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3824 unsigned long flags;
3825 struct rcu_data *my_rdp;
3826 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3827 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3830 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3831 return; /* No callbacks to migrate. */
3833 local_irq_save(flags);
3834 my_rdp = this_cpu_ptr(rsp->rda);
3835 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3836 local_irq_restore(flags);
3839 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3840 /* Leverage recent GPs and set GP for new callbacks. */
3841 needwake = rcu_advance_cbs(rsp, rnp_root, rdp) ||
3842 rcu_advance_cbs(rsp, rnp_root, my_rdp);
3843 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3844 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3845 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3846 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3848 rcu_gp_kthread_wake(rsp);
3849 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3850 !rcu_segcblist_empty(&rdp->cblist),
3851 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3852 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3853 rcu_segcblist_first_cb(&rdp->cblist));
3857 * The outgoing CPU has just passed through the dying-idle state,
3858 * and we are being invoked from the CPU that was IPIed to continue the
3859 * offline operation. We need to migrate the outgoing CPU's callbacks.
3861 void rcutree_migrate_callbacks(int cpu)
3863 struct rcu_state *rsp;
3865 for_each_rcu_flavor(rsp)
3866 rcu_migrate_callbacks(cpu, rsp);
3871 * On non-huge systems, use expedited RCU grace periods to make suspend
3872 * and hibernation run faster.
3874 static int rcu_pm_notify(struct notifier_block *self,
3875 unsigned long action, void *hcpu)
3878 case PM_HIBERNATION_PREPARE:
3879 case PM_SUSPEND_PREPARE:
3880 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3883 case PM_POST_HIBERNATION:
3884 case PM_POST_SUSPEND:
3885 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3886 rcu_unexpedite_gp();
3895 * Spawn the kthreads that handle each RCU flavor's grace periods.
3897 static int __init rcu_spawn_gp_kthread(void)
3899 unsigned long flags;
3900 int kthread_prio_in = kthread_prio;
3901 struct rcu_node *rnp;
3902 struct rcu_state *rsp;
3903 struct sched_param sp;
3904 struct task_struct *t;
3906 /* Force priority into range. */
3907 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3909 else if (kthread_prio < 0)
3911 else if (kthread_prio > 99)
3913 if (kthread_prio != kthread_prio_in)
3914 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3915 kthread_prio, kthread_prio_in);
3917 rcu_scheduler_fully_active = 1;
3918 for_each_rcu_flavor(rsp) {
3919 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3921 rnp = rcu_get_root(rsp);
3922 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3923 rsp->gp_kthread = t;
3925 sp.sched_priority = kthread_prio;
3926 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3928 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3931 rcu_spawn_nocb_kthreads();
3932 rcu_spawn_boost_kthreads();
3935 early_initcall(rcu_spawn_gp_kthread);
3938 * This function is invoked towards the end of the scheduler's
3939 * initialization process. Before this is called, the idle task might
3940 * contain synchronous grace-period primitives (during which time, this idle
3941 * task is booting the system, and such primitives are no-ops). After this
3942 * function is called, any synchronous grace-period primitives are run as
3943 * expedited, with the requesting task driving the grace period forward.
3944 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3945 * runtime RCU functionality.
3947 void rcu_scheduler_starting(void)
3949 WARN_ON(num_online_cpus() != 1);
3950 WARN_ON(nr_context_switches() > 0);
3951 rcu_test_sync_prims();
3952 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3953 rcu_test_sync_prims();
3957 * Helper function for rcu_init() that initializes one rcu_state structure.
3959 static void __init rcu_init_one(struct rcu_state *rsp)
3961 static const char * const buf[] = RCU_NODE_NAME_INIT;
3962 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3963 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3964 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3966 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3970 struct rcu_node *rnp;
3972 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3974 /* Silence gcc 4.8 false positive about array index out of range. */
3975 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3976 panic("rcu_init_one: rcu_num_lvls out of range");
3978 /* Initialize the level-tracking arrays. */
3980 for (i = 1; i < rcu_num_lvls; i++)
3981 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
3982 rcu_init_levelspread(levelspread, num_rcu_lvl);
3984 /* Initialize the elements themselves, starting from the leaves. */
3986 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3987 cpustride *= levelspread[i];
3988 rnp = rsp->level[i];
3989 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3990 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3991 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3992 &rcu_node_class[i], buf[i]);
3993 raw_spin_lock_init(&rnp->fqslock);
3994 lockdep_set_class_and_name(&rnp->fqslock,
3995 &rcu_fqs_class[i], fqs[i]);
3996 rnp->gp_seq = rsp->gp_seq;
3997 rnp->gp_seq_needed = rsp->gp_seq;
3998 rnp->completedqs = rsp->gp_seq;
4000 rnp->qsmaskinit = 0;
4001 rnp->grplo = j * cpustride;
4002 rnp->grphi = (j + 1) * cpustride - 1;
4003 if (rnp->grphi >= nr_cpu_ids)
4004 rnp->grphi = nr_cpu_ids - 1;
4010 rnp->grpnum = j % levelspread[i - 1];
4011 rnp->grpmask = 1UL << rnp->grpnum;
4012 rnp->parent = rsp->level[i - 1] +
4013 j / levelspread[i - 1];
4016 INIT_LIST_HEAD(&rnp->blkd_tasks);
4017 rcu_init_one_nocb(rnp);
4018 init_waitqueue_head(&rnp->exp_wq[0]);
4019 init_waitqueue_head(&rnp->exp_wq[1]);
4020 init_waitqueue_head(&rnp->exp_wq[2]);
4021 init_waitqueue_head(&rnp->exp_wq[3]);
4022 spin_lock_init(&rnp->exp_lock);
4026 init_swait_queue_head(&rsp->gp_wq);
4027 init_swait_queue_head(&rsp->expedited_wq);
4028 rnp = rcu_first_leaf_node(rsp);
4029 for_each_possible_cpu(i) {
4030 while (i > rnp->grphi)
4032 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4033 rcu_boot_init_percpu_data(i, rsp);
4035 list_add(&rsp->flavors, &rcu_struct_flavors);
4039 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4040 * replace the definitions in tree.h because those are needed to size
4041 * the ->node array in the rcu_state structure.
4043 static void __init rcu_init_geometry(void)
4047 int rcu_capacity[RCU_NUM_LVLS];
4050 * Initialize any unspecified boot parameters.
4051 * The default values of jiffies_till_first_fqs and
4052 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4053 * value, which is a function of HZ, then adding one for each
4054 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4056 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4057 if (jiffies_till_first_fqs == ULONG_MAX)
4058 jiffies_till_first_fqs = d;
4059 if (jiffies_till_next_fqs == ULONG_MAX)
4060 jiffies_till_next_fqs = d;
4062 /* If the compile-time values are accurate, just leave. */
4063 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4064 nr_cpu_ids == NR_CPUS)
4066 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4067 rcu_fanout_leaf, nr_cpu_ids);
4070 * The boot-time rcu_fanout_leaf parameter must be at least two
4071 * and cannot exceed the number of bits in the rcu_node masks.
4072 * Complain and fall back to the compile-time values if this
4073 * limit is exceeded.
4075 if (rcu_fanout_leaf < 2 ||
4076 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4077 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4083 * Compute number of nodes that can be handled an rcu_node tree
4084 * with the given number of levels.
4086 rcu_capacity[0] = rcu_fanout_leaf;
4087 for (i = 1; i < RCU_NUM_LVLS; i++)
4088 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4091 * The tree must be able to accommodate the configured number of CPUs.
4092 * If this limit is exceeded, fall back to the compile-time values.
4094 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4095 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4100 /* Calculate the number of levels in the tree. */
4101 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4103 rcu_num_lvls = i + 1;
4105 /* Calculate the number of rcu_nodes at each level of the tree. */
4106 for (i = 0; i < rcu_num_lvls; i++) {
4107 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4108 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4111 /* Calculate the total number of rcu_node structures. */
4113 for (i = 0; i < rcu_num_lvls; i++)
4114 rcu_num_nodes += num_rcu_lvl[i];
4118 * Dump out the structure of the rcu_node combining tree associated
4119 * with the rcu_state structure referenced by rsp.
4121 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4124 struct rcu_node *rnp;
4126 pr_info("rcu_node tree layout dump\n");
4128 rcu_for_each_node_breadth_first(rsp, rnp) {
4129 if (rnp->level != level) {
4134 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4139 struct workqueue_struct *rcu_gp_wq;
4140 struct workqueue_struct *rcu_par_gp_wq;
4142 void __init rcu_init(void)
4146 rcu_early_boot_tests();
4148 rcu_bootup_announce();
4149 rcu_init_geometry();
4150 rcu_init_one(&rcu_bh_state);
4151 rcu_init_one(&rcu_sched_state);
4153 rcu_dump_rcu_node_tree(&rcu_sched_state);
4154 __rcu_init_preempt();
4155 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4158 * We don't need protection against CPU-hotplug here because
4159 * this is called early in boot, before either interrupts
4160 * or the scheduler are operational.
4162 pm_notifier(rcu_pm_notify, 0);
4163 for_each_online_cpu(cpu) {
4164 rcutree_prepare_cpu(cpu);
4165 rcu_cpu_starting(cpu);
4166 rcutree_online_cpu(cpu);
4169 /* Create workqueue for expedited GPs and for Tree SRCU. */
4170 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4171 WARN_ON(!rcu_gp_wq);
4172 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4173 WARN_ON(!rcu_par_gp_wq);
4176 #include "tree_exp.h"
4177 #include "tree_plugin.h"