2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate_wait.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/sched/debug.h>
39 #include <linux/nmi.h>
40 #include <linux/atomic.h>
41 #include <linux/bitops.h>
42 #include <linux/export.h>
43 #include <linux/completion.h>
44 #include <linux/moduleparam.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <uapi/linux/sched/types.h>
54 #include <linux/prefetch.h>
55 #include <linux/delay.h>
56 #include <linux/stop_machine.h>
57 #include <linux/random.h>
58 #include <linux/trace_events.h>
59 #include <linux/suspend.h>
60 #include <linux/ftrace.h>
65 #ifdef MODULE_PARAM_PREFIX
66 #undef MODULE_PARAM_PREFIX
68 #define MODULE_PARAM_PREFIX "rcutree."
70 /* Data structures. */
73 * In order to export the rcu_state name to the tracing tools, it
74 * needs to be added in the __tracepoint_string section.
75 * This requires defining a separate variable tp_<sname>_varname
76 * that points to the string being used, and this will allow
77 * the tracing userspace tools to be able to decipher the string
78 * address to the matching string.
81 # define DEFINE_RCU_TPS(sname) \
82 static char sname##_varname[] = #sname; \
83 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
84 # define RCU_STATE_NAME(sname) sname##_varname
86 # define DEFINE_RCU_TPS(sname)
87 # define RCU_STATE_NAME(sname) __stringify(sname)
90 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
91 DEFINE_RCU_TPS(sname) \
92 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
93 struct rcu_state sname##_state = { \
94 .level = { &sname##_state.node[0] }, \
95 .rda = &sname##_data, \
97 .gp_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
101 .name = RCU_STATE_NAME(sname), \
103 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
104 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
107 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
110 static struct rcu_state *const rcu_state_p;
111 LIST_HEAD(rcu_struct_flavors);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree;
115 module_param(dump_tree, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact;
118 module_param(rcu_fanout_exact, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121 module_param(rcu_fanout_leaf, int, 0444);
122 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123 /* Number of rcu_nodes at specified level. */
124 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126 /* panic() on RCU Stall sysctl. */
127 int sysctl_panic_on_rcu_stall __read_mostly;
130 * The rcu_scheduler_active variable is initialized to the value
131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
132 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
133 * RCU can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_rcu() to a simple barrier(). When this variable
135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
136 * to detect real grace periods. This variable is also used to suppress
137 * boot-time false positives from lockdep-RCU error checking. Finally, it
138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
139 * is fully initialized, including all of its kthreads having been spawned.
141 int rcu_scheduler_active __read_mostly;
142 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
156 static int rcu_scheduler_fully_active __read_mostly;
158 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
159 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163 static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
165 static void sync_sched_exp_online_cleanup(int cpu);
167 /* rcuc/rcub kthread realtime priority */
168 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
169 module_param(kthread_prio, int, 0644);
171 /* Delay in jiffies for grace-period initialization delays, debug only. */
173 static int gp_preinit_delay;
174 module_param(gp_preinit_delay, int, 0444);
175 static int gp_init_delay;
176 module_param(gp_init_delay, int, 0444);
177 static int gp_cleanup_delay;
178 module_param(gp_cleanup_delay, int, 0444);
181 * Number of grace periods between delays, normalized by the duration of
182 * the delay. The longer the delay, the more the grace periods between
183 * each delay. The reason for this normalization is that it means that,
184 * for non-zero delays, the overall slowdown of grace periods is constant
185 * regardless of the duration of the delay. This arrangement balances
186 * the need for long delays to increase some race probabilities with the
187 * need for fast grace periods to increase other race probabilities.
189 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
192 * Track the rcutorture test sequence number and the update version
193 * number within a given test. The rcutorture_testseq is incremented
194 * on every rcutorture module load and unload, so has an odd value
195 * when a test is running. The rcutorture_vernum is set to zero
196 * when rcutorture starts and is incremented on each rcutorture update.
197 * These variables enable correlating rcutorture output with the
198 * RCU tracing information.
200 unsigned long rcutorture_testseq;
201 unsigned long rcutorture_vernum;
204 * Compute the mask of online CPUs for the specified rcu_node structure.
205 * This will not be stable unless the rcu_node structure's ->lock is
206 * held, but the bit corresponding to the current CPU will be stable
209 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
211 return READ_ONCE(rnp->qsmaskinitnext);
215 * Return true if an RCU grace period is in progress. The READ_ONCE()s
216 * permit this function to be invoked without holding the root rcu_node
217 * structure's ->lock, but of course results can be subject to change.
219 static int rcu_gp_in_progress(struct rcu_state *rsp)
221 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
225 * Note a quiescent state. Because we do not need to know
226 * how many quiescent states passed, just if there was at least
227 * one since the start of the grace period, this just sets a flag.
228 * The caller must have disabled preemption.
230 void rcu_sched_qs(void)
232 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
233 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
235 trace_rcu_grace_period(TPS("rcu_sched"),
236 __this_cpu_read(rcu_sched_data.gpnum),
238 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
239 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
241 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
242 rcu_report_exp_rdp(&rcu_sched_state,
243 this_cpu_ptr(&rcu_sched_data), true);
248 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
249 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
250 trace_rcu_grace_period(TPS("rcu_bh"),
251 __this_cpu_read(rcu_bh_data.gpnum),
253 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
258 * Steal a bit from the bottom of ->dynticks for idle entry/exit
259 * control. Initially this is for TLB flushing.
261 #define RCU_DYNTICK_CTRL_MASK 0x1
262 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
263 #ifndef rcu_eqs_special_exit
264 #define rcu_eqs_special_exit() do { } while (0)
267 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
268 .dynticks_nesting = 1,
269 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
270 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
274 * Record entry into an extended quiescent state. This is only to be
275 * called when not already in an extended quiescent state.
277 static void rcu_dynticks_eqs_enter(void)
279 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
283 * CPUs seeing atomic_add_return() must see prior RCU read-side
284 * critical sections, and we also must force ordering with the
287 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
288 /* Better be in an extended quiescent state! */
289 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
290 (seq & RCU_DYNTICK_CTRL_CTR));
291 /* Better not have special action (TLB flush) pending! */
292 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
293 (seq & RCU_DYNTICK_CTRL_MASK));
297 * Record exit from an extended quiescent state. This is only to be
298 * called from an extended quiescent state.
300 static void rcu_dynticks_eqs_exit(void)
302 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
306 * CPUs seeing atomic_add_return() must see prior idle sojourns,
307 * and we also must force ordering with the next RCU read-side
310 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
311 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
312 !(seq & RCU_DYNTICK_CTRL_CTR));
313 if (seq & RCU_DYNTICK_CTRL_MASK) {
314 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
315 smp_mb__after_atomic(); /* _exit after clearing mask. */
316 /* Prefer duplicate flushes to losing a flush. */
317 rcu_eqs_special_exit();
322 * Reset the current CPU's ->dynticks counter to indicate that the
323 * newly onlined CPU is no longer in an extended quiescent state.
324 * This will either leave the counter unchanged, or increment it
325 * to the next non-quiescent value.
327 * The non-atomic test/increment sequence works because the upper bits
328 * of the ->dynticks counter are manipulated only by the corresponding CPU,
329 * or when the corresponding CPU is offline.
331 static void rcu_dynticks_eqs_online(void)
333 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
335 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
337 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
341 * Is the current CPU in an extended quiescent state?
343 * No ordering, as we are sampling CPU-local information.
345 bool rcu_dynticks_curr_cpu_in_eqs(void)
347 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
349 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
353 * Snapshot the ->dynticks counter with full ordering so as to allow
354 * stable comparison of this counter with past and future snapshots.
356 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
358 int snap = atomic_add_return(0, &rdtp->dynticks);
360 return snap & ~RCU_DYNTICK_CTRL_MASK;
364 * Return true if the snapshot returned from rcu_dynticks_snap()
365 * indicates that RCU is in an extended quiescent state.
367 static bool rcu_dynticks_in_eqs(int snap)
369 return !(snap & RCU_DYNTICK_CTRL_CTR);
373 * Return true if the CPU corresponding to the specified rcu_dynticks
374 * structure has spent some time in an extended quiescent state since
375 * rcu_dynticks_snap() returned the specified snapshot.
377 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
379 return snap != rcu_dynticks_snap(rdtp);
383 * Do a double-increment of the ->dynticks counter to emulate a
384 * momentary idle-CPU quiescent state.
386 static void rcu_dynticks_momentary_idle(void)
388 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
389 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
392 /* It is illegal to call this from idle state. */
393 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
397 * Set the special (bottom) bit of the specified CPU so that it
398 * will take special action (such as flushing its TLB) on the
399 * next exit from an extended quiescent state. Returns true if
400 * the bit was successfully set, or false if the CPU was not in
401 * an extended quiescent state.
403 bool rcu_eqs_special_set(int cpu)
407 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
410 old = atomic_read(&rdtp->dynticks);
411 if (old & RCU_DYNTICK_CTRL_CTR)
413 new = old | RCU_DYNTICK_CTRL_MASK;
414 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
419 * Let the RCU core know that this CPU has gone through the scheduler,
420 * which is a quiescent state. This is called when the need for a
421 * quiescent state is urgent, so we burn an atomic operation and full
422 * memory barriers to let the RCU core know about it, regardless of what
423 * this CPU might (or might not) do in the near future.
425 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
427 * The caller must have disabled interrupts.
429 static void rcu_momentary_dyntick_idle(void)
431 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
432 rcu_dynticks_momentary_idle();
436 * Note a context switch. This is a quiescent state for RCU-sched,
437 * and requires special handling for preemptible RCU.
438 * The caller must have disabled interrupts.
440 void rcu_note_context_switch(bool preempt)
442 barrier(); /* Avoid RCU read-side critical sections leaking down. */
443 trace_rcu_utilization(TPS("Start context switch"));
445 rcu_preempt_note_context_switch(preempt);
446 /* Load rcu_urgent_qs before other flags. */
447 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
449 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
450 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
451 rcu_momentary_dyntick_idle();
452 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
454 rcu_note_voluntary_context_switch_lite(current);
456 trace_rcu_utilization(TPS("End context switch"));
457 barrier(); /* Avoid RCU read-side critical sections leaking up. */
459 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
462 * Register a quiescent state for all RCU flavors. If there is an
463 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
464 * dyntick-idle quiescent state visible to other CPUs (but only for those
465 * RCU flavors in desperate need of a quiescent state, which will normally
466 * be none of them). Either way, do a lightweight quiescent state for
469 * The barrier() calls are redundant in the common case when this is
470 * called externally, but just in case this is called from within this
474 void rcu_all_qs(void)
478 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
481 /* Load rcu_urgent_qs before other flags. */
482 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
486 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
487 barrier(); /* Avoid RCU read-side critical sections leaking down. */
488 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
489 local_irq_save(flags);
490 rcu_momentary_dyntick_idle();
491 local_irq_restore(flags);
493 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
495 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
496 barrier(); /* Avoid RCU read-side critical sections leaking up. */
499 EXPORT_SYMBOL_GPL(rcu_all_qs);
501 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
502 static long blimit = DEFAULT_RCU_BLIMIT;
503 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
504 static long qhimark = DEFAULT_RCU_QHIMARK;
505 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
506 static long qlowmark = DEFAULT_RCU_QLOMARK;
508 module_param(blimit, long, 0444);
509 module_param(qhimark, long, 0444);
510 module_param(qlowmark, long, 0444);
512 static ulong jiffies_till_first_fqs = ULONG_MAX;
513 static ulong jiffies_till_next_fqs = ULONG_MAX;
514 static bool rcu_kick_kthreads;
516 module_param(jiffies_till_first_fqs, ulong, 0644);
517 module_param(jiffies_till_next_fqs, ulong, 0644);
518 module_param(rcu_kick_kthreads, bool, 0644);
521 * How long the grace period must be before we start recruiting
522 * quiescent-state help from rcu_note_context_switch().
524 static ulong jiffies_till_sched_qs = HZ / 10;
525 module_param(jiffies_till_sched_qs, ulong, 0444);
527 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
528 struct rcu_data *rdp);
529 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
530 static void force_quiescent_state(struct rcu_state *rsp);
531 static int rcu_pending(void);
534 * Return the number of RCU batches started thus far for debug & stats.
536 unsigned long rcu_batches_started(void)
538 return rcu_state_p->gpnum;
540 EXPORT_SYMBOL_GPL(rcu_batches_started);
543 * Return the number of RCU-sched batches started thus far for debug & stats.
545 unsigned long rcu_batches_started_sched(void)
547 return rcu_sched_state.gpnum;
549 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
552 * Return the number of RCU BH batches started thus far for debug & stats.
554 unsigned long rcu_batches_started_bh(void)
556 return rcu_bh_state.gpnum;
558 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
561 * Return the number of RCU batches completed thus far for debug & stats.
563 unsigned long rcu_batches_completed(void)
565 return rcu_state_p->completed;
567 EXPORT_SYMBOL_GPL(rcu_batches_completed);
570 * Return the number of RCU-sched batches completed thus far for debug & stats.
572 unsigned long rcu_batches_completed_sched(void)
574 return rcu_sched_state.completed;
576 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
579 * Return the number of RCU BH batches completed thus far for debug & stats.
581 unsigned long rcu_batches_completed_bh(void)
583 return rcu_bh_state.completed;
585 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
588 * Return the number of RCU expedited batches completed thus far for
589 * debug & stats. Odd numbers mean that a batch is in progress, even
590 * numbers mean idle. The value returned will thus be roughly double
591 * the cumulative batches since boot.
593 unsigned long rcu_exp_batches_completed(void)
595 return rcu_state_p->expedited_sequence;
597 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
600 * Return the number of RCU-sched expedited batches completed thus far
601 * for debug & stats. Similar to rcu_exp_batches_completed().
603 unsigned long rcu_exp_batches_completed_sched(void)
605 return rcu_sched_state.expedited_sequence;
607 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
610 * Force a quiescent state.
612 void rcu_force_quiescent_state(void)
614 force_quiescent_state(rcu_state_p);
616 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
619 * Force a quiescent state for RCU BH.
621 void rcu_bh_force_quiescent_state(void)
623 force_quiescent_state(&rcu_bh_state);
625 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
628 * Force a quiescent state for RCU-sched.
630 void rcu_sched_force_quiescent_state(void)
632 force_quiescent_state(&rcu_sched_state);
634 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
637 * Show the state of the grace-period kthreads.
639 void show_rcu_gp_kthreads(void)
641 struct rcu_state *rsp;
643 for_each_rcu_flavor(rsp) {
644 pr_info("%s: wait state: %d ->state: %#lx\n",
645 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
646 /* sched_show_task(rsp->gp_kthread); */
649 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
652 * Record the number of times rcutorture tests have been initiated and
653 * terminated. This information allows the debugfs tracing stats to be
654 * correlated to the rcutorture messages, even when the rcutorture module
655 * is being repeatedly loaded and unloaded. In other words, we cannot
656 * store this state in rcutorture itself.
658 void rcutorture_record_test_transition(void)
660 rcutorture_testseq++;
661 rcutorture_vernum = 0;
663 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
666 * Send along grace-period-related data for rcutorture diagnostics.
668 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
669 unsigned long *gpnum, unsigned long *completed)
671 struct rcu_state *rsp = NULL;
680 case RCU_SCHED_FLAVOR:
681 rsp = &rcu_sched_state;
688 *flags = READ_ONCE(rsp->gp_flags);
689 *gpnum = READ_ONCE(rsp->gpnum);
690 *completed = READ_ONCE(rsp->completed);
692 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
695 * Record the number of writer passes through the current rcutorture test.
696 * This is also used to correlate debugfs tracing stats with the rcutorture
699 void rcutorture_record_progress(unsigned long vernum)
703 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
706 * Return the root node of the specified rcu_state structure.
708 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
710 return &rsp->node[0];
714 * Is there any need for future grace periods?
715 * Interrupts must be disabled. If the caller does not hold the root
716 * rnp_node structure's ->lock, the results are advisory only.
718 static int rcu_future_needs_gp(struct rcu_state *rsp)
720 struct rcu_node *rnp = rcu_get_root(rsp);
721 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
722 int *fp = &rnp->need_future_gp[idx];
724 lockdep_assert_irqs_disabled();
725 return READ_ONCE(*fp);
729 * Does the current CPU require a not-yet-started grace period?
730 * The caller must have disabled interrupts to prevent races with
731 * normal callback registry.
734 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
736 lockdep_assert_irqs_disabled();
737 if (rcu_gp_in_progress(rsp))
738 return false; /* No, a grace period is already in progress. */
739 if (rcu_future_needs_gp(rsp))
740 return true; /* Yes, a no-CBs CPU needs one. */
741 if (!rcu_segcblist_is_enabled(&rdp->cblist))
742 return false; /* No, this is a no-CBs (or offline) CPU. */
743 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
744 return true; /* Yes, CPU has newly registered callbacks. */
745 if (rcu_segcblist_future_gp_needed(&rdp->cblist,
746 READ_ONCE(rsp->completed)))
747 return true; /* Yes, CBs for future grace period. */
748 return false; /* No grace period needed. */
752 * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
754 * Enter idle, doing appropriate accounting. The caller must have
755 * disabled interrupts.
757 static void rcu_eqs_enter_common(bool user)
759 struct rcu_state *rsp;
760 struct rcu_data *rdp;
761 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
763 lockdep_assert_irqs_disabled();
764 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
765 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
766 !user && !is_idle_task(current)) {
767 struct task_struct *idle __maybe_unused =
768 idle_task(smp_processor_id());
770 trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
771 rcu_ftrace_dump(DUMP_ORIG);
772 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
773 current->pid, current->comm,
774 idle->pid, idle->comm); /* must be idle task! */
776 for_each_rcu_flavor(rsp) {
777 rdp = this_cpu_ptr(rsp->rda);
778 do_nocb_deferred_wakeup(rdp);
780 rcu_prepare_for_idle();
781 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
782 rcu_dynticks_eqs_enter();
783 rcu_dynticks_task_enter();
787 * Enter an RCU extended quiescent state, which can be either the
788 * idle loop or adaptive-tickless usermode execution.
790 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
791 * the possibility of usermode upcalls having messed up our count
792 * of interrupt nesting level during the prior busy period.
794 static void rcu_eqs_enter(bool user)
796 struct rcu_dynticks *rdtp;
798 rdtp = this_cpu_ptr(&rcu_dynticks);
799 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
800 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
801 rdtp->dynticks_nesting == 0);
802 if (rdtp->dynticks_nesting == 1)
803 rcu_eqs_enter_common(user);
805 rdtp->dynticks_nesting--;
809 * rcu_idle_enter - inform RCU that current CPU is entering idle
811 * Enter idle mode, in other words, -leave- the mode in which RCU
812 * read-side critical sections can occur. (Though RCU read-side
813 * critical sections can occur in irq handlers in idle, a possibility
814 * handled by irq_enter() and irq_exit().)
816 * If you add or remove a call to rcu_idle_enter(), be sure to test with
817 * CONFIG_RCU_EQS_DEBUG=y.
819 void rcu_idle_enter(void)
821 lockdep_assert_irqs_disabled();
822 rcu_eqs_enter(false);
825 #ifdef CONFIG_NO_HZ_FULL
827 * rcu_user_enter - inform RCU that we are resuming userspace.
829 * Enter RCU idle mode right before resuming userspace. No use of RCU
830 * is permitted between this call and rcu_user_exit(). This way the
831 * CPU doesn't need to maintain the tick for RCU maintenance purposes
832 * when the CPU runs in userspace.
834 * If you add or remove a call to rcu_user_enter(), be sure to test with
835 * CONFIG_RCU_EQS_DEBUG=y.
837 void rcu_user_enter(void)
839 lockdep_assert_irqs_disabled();
842 #endif /* CONFIG_NO_HZ_FULL */
845 * rcu_nmi_exit - inform RCU of exit from NMI context
847 * If we are returning from the outermost NMI handler that interrupted an
848 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
849 * to let the RCU grace-period handling know that the CPU is back to
852 * If you add or remove a call to rcu_nmi_exit(), be sure to test
853 * with CONFIG_RCU_EQS_DEBUG=y.
855 void rcu_nmi_exit(void)
857 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
860 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
861 * (We are exiting an NMI handler, so RCU better be paying attention
864 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
865 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
868 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
869 * leave it in non-RCU-idle state.
871 if (rdtp->dynticks_nmi_nesting != 1) {
872 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
873 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
874 rdtp->dynticks_nmi_nesting - 2);
878 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
879 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
880 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
881 rcu_dynticks_eqs_enter();
885 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
887 * Exit from an interrupt handler, which might possibly result in entering
888 * idle mode, in other words, leaving the mode in which read-side critical
889 * sections can occur. The caller must have disabled interrupts.
891 * This code assumes that the idle loop never does anything that might
892 * result in unbalanced calls to irq_enter() and irq_exit(). If your
893 * architecture's idle loop violates this assumption, RCU will give you what
894 * you deserve, good and hard. But very infrequently and irreproducibly.
896 * Use things like work queues to work around this limitation.
898 * You have been warned.
900 * If you add or remove a call to rcu_irq_exit(), be sure to test with
901 * CONFIG_RCU_EQS_DEBUG=y.
903 void rcu_irq_exit(void)
905 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
907 lockdep_assert_irqs_disabled();
908 if (rdtp->dynticks_nmi_nesting == 1)
909 rcu_prepare_for_idle();
911 if (rdtp->dynticks_nmi_nesting == 0)
912 rcu_dynticks_task_enter();
916 * Wrapper for rcu_irq_exit() where interrupts are enabled.
918 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
919 * with CONFIG_RCU_EQS_DEBUG=y.
921 void rcu_irq_exit_irqson(void)
925 local_irq_save(flags);
927 local_irq_restore(flags);
931 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
933 * If the new value of the ->dynticks_nesting counter was previously zero,
934 * we really have exited idle, and must do the appropriate accounting.
935 * The caller must have disabled interrupts.
937 static void rcu_eqs_exit_common(long newval, int user)
939 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
941 rcu_dynticks_task_exit();
942 rcu_dynticks_eqs_exit();
943 rcu_cleanup_after_idle();
944 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, newval, rdtp->dynticks);
945 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
946 !user && !is_idle_task(current)) {
947 struct task_struct *idle __maybe_unused =
948 idle_task(smp_processor_id());
950 trace_rcu_dyntick(TPS("Error on exit: not idle task"), rdtp->dynticks_nesting, newval, rdtp->dynticks);
951 rcu_ftrace_dump(DUMP_ORIG);
952 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
953 current->pid, current->comm,
954 idle->pid, idle->comm); /* must be idle task! */
959 * Exit an RCU extended quiescent state, which can be either the
960 * idle loop or adaptive-tickless usermode execution.
962 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
963 * allow for the possibility of usermode upcalls messing up our count of
964 * interrupt nesting level during the busy period that is just now starting.
966 static void rcu_eqs_exit(bool user)
968 struct rcu_dynticks *rdtp;
971 lockdep_assert_irqs_disabled();
972 rdtp = this_cpu_ptr(&rcu_dynticks);
973 oldval = rdtp->dynticks_nesting;
974 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
976 rdtp->dynticks_nesting++;
978 rcu_eqs_exit_common(1, user);
979 WRITE_ONCE(rdtp->dynticks_nesting, 1);
980 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
985 * rcu_idle_exit - inform RCU that current CPU is leaving idle
987 * Exit idle mode, in other words, -enter- the mode in which RCU
988 * read-side critical sections can occur.
990 * If you add or remove a call to rcu_idle_exit(), be sure to test with
991 * CONFIG_RCU_EQS_DEBUG=y.
993 void rcu_idle_exit(void)
997 local_irq_save(flags);
999 local_irq_restore(flags);
1002 #ifdef CONFIG_NO_HZ_FULL
1004 * rcu_user_exit - inform RCU that we are exiting userspace.
1006 * Exit RCU idle mode while entering the kernel because it can
1007 * run a RCU read side critical section anytime.
1009 * If you add or remove a call to rcu_user_exit(), be sure to test with
1010 * CONFIG_RCU_EQS_DEBUG=y.
1012 void rcu_user_exit(void)
1016 #endif /* CONFIG_NO_HZ_FULL */
1019 * rcu_nmi_enter - inform RCU of entry to NMI context
1021 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1022 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1023 * that the CPU is active. This implementation permits nested NMIs, as
1024 * long as the nesting level does not overflow an int. (You will probably
1025 * run out of stack space first.)
1027 * If you add or remove a call to rcu_nmi_enter(), be sure to test
1028 * with CONFIG_RCU_EQS_DEBUG=y.
1030 void rcu_nmi_enter(void)
1032 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1035 /* Complain about underflow. */
1036 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
1039 * If idle from RCU viewpoint, atomically increment ->dynticks
1040 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1041 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1042 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1043 * to be in the outermost NMI handler that interrupted an RCU-idle
1044 * period (observation due to Andy Lutomirski).
1046 if (rcu_dynticks_curr_cpu_in_eqs()) {
1047 rcu_dynticks_eqs_exit();
1050 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1051 rdtp->dynticks_nmi_nesting,
1052 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
1053 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
1054 rdtp->dynticks_nmi_nesting + incby);
1059 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1061 * Enter an interrupt handler, which might possibly result in exiting
1062 * idle mode, in other words, entering the mode in which read-side critical
1063 * sections can occur. The caller must have disabled interrupts.
1065 * Note that the Linux kernel is fully capable of entering an interrupt
1066 * handler that it never exits, for example when doing upcalls to user mode!
1067 * This code assumes that the idle loop never does upcalls to user mode.
1068 * If your architecture's idle loop does do upcalls to user mode (or does
1069 * anything else that results in unbalanced calls to the irq_enter() and
1070 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1071 * But very infrequently and irreproducibly.
1073 * Use things like work queues to work around this limitation.
1075 * You have been warned.
1077 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1078 * CONFIG_RCU_EQS_DEBUG=y.
1080 void rcu_irq_enter(void)
1082 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1084 lockdep_assert_irqs_disabled();
1085 if (rdtp->dynticks_nmi_nesting == 0)
1086 rcu_dynticks_task_exit();
1088 if (rdtp->dynticks_nmi_nesting == 1)
1089 rcu_cleanup_after_idle();
1093 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1095 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1096 * with CONFIG_RCU_EQS_DEBUG=y.
1098 void rcu_irq_enter_irqson(void)
1100 unsigned long flags;
1102 local_irq_save(flags);
1104 local_irq_restore(flags);
1108 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1110 * Return true if RCU is watching the running CPU, which means that this
1111 * CPU can safely enter RCU read-side critical sections. In other words,
1112 * if the current CPU is in its idle loop and is neither in an interrupt
1113 * or NMI handler, return true.
1115 bool notrace rcu_is_watching(void)
1119 preempt_disable_notrace();
1120 ret = !rcu_dynticks_curr_cpu_in_eqs();
1121 preempt_enable_notrace();
1124 EXPORT_SYMBOL_GPL(rcu_is_watching);
1127 * If a holdout task is actually running, request an urgent quiescent
1128 * state from its CPU. This is unsynchronized, so migrations can cause
1129 * the request to go to the wrong CPU. Which is OK, all that will happen
1130 * is that the CPU's next context switch will be a bit slower and next
1131 * time around this task will generate another request.
1133 void rcu_request_urgent_qs_task(struct task_struct *t)
1140 return; /* This task is not running on that CPU. */
1141 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1144 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1147 * Is the current CPU online? Disable preemption to avoid false positives
1148 * that could otherwise happen due to the current CPU number being sampled,
1149 * this task being preempted, its old CPU being taken offline, resuming
1150 * on some other CPU, then determining that its old CPU is now offline.
1151 * It is OK to use RCU on an offline processor during initial boot, hence
1152 * the check for rcu_scheduler_fully_active. Note also that it is OK
1153 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1154 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1155 * offline to continue to use RCU for one jiffy after marking itself
1156 * offline in the cpu_online_mask. This leniency is necessary given the
1157 * non-atomic nature of the online and offline processing, for example,
1158 * the fact that a CPU enters the scheduler after completing the teardown
1161 * This is also why RCU internally marks CPUs online during in the
1162 * preparation phase and offline after the CPU has been taken down.
1164 * Disable checking if in an NMI handler because we cannot safely report
1165 * errors from NMI handlers anyway.
1167 bool rcu_lockdep_current_cpu_online(void)
1169 struct rcu_data *rdp;
1170 struct rcu_node *rnp;
1176 rdp = this_cpu_ptr(&rcu_sched_data);
1178 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1179 !rcu_scheduler_fully_active;
1183 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1185 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1188 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1190 * If the current CPU is idle or running at a first-level (not nested)
1191 * interrupt from idle, return true. The caller must have at least
1192 * disabled preemption.
1194 static int rcu_is_cpu_rrupt_from_idle(void)
1196 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
1197 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
1201 * We are reporting a quiescent state on behalf of some other CPU, so
1202 * it is our responsibility to check for and handle potential overflow
1203 * of the rcu_node ->gpnum counter with respect to the rcu_data counters.
1204 * After all, the CPU might be in deep idle state, and thus executing no
1207 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1209 lockdep_assert_held(&rnp->lock);
1210 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, rnp->gpnum))
1211 WRITE_ONCE(rdp->gpwrap, true);
1212 if (ULONG_CMP_LT(rdp->rcu_iw_gpnum + ULONG_MAX / 4, rnp->gpnum))
1213 rdp->rcu_iw_gpnum = rnp->gpnum + ULONG_MAX / 4;
1217 * Snapshot the specified CPU's dynticks counter so that we can later
1218 * credit them with an implicit quiescent state. Return 1 if this CPU
1219 * is in dynticks idle mode, which is an extended quiescent state.
1221 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1223 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1224 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1225 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1226 rcu_gpnum_ovf(rdp->mynode, rdp);
1233 * Handler for the irq_work request posted when a grace period has
1234 * gone on for too long, but not yet long enough for an RCU CPU
1235 * stall warning. Set state appropriately, but just complain if
1236 * there is unexpected state on entry.
1238 static void rcu_iw_handler(struct irq_work *iwp)
1240 struct rcu_data *rdp;
1241 struct rcu_node *rnp;
1243 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1245 raw_spin_lock_rcu_node(rnp);
1246 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1247 rdp->rcu_iw_gpnum = rnp->gpnum;
1248 rdp->rcu_iw_pending = false;
1250 raw_spin_unlock_rcu_node(rnp);
1254 * Return true if the specified CPU has passed through a quiescent
1255 * state by virtue of being in or having passed through an dynticks
1256 * idle state since the last call to dyntick_save_progress_counter()
1257 * for this same CPU, or by virtue of having been offline.
1259 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1264 struct rcu_node *rnp = rdp->mynode;
1267 * If the CPU passed through or entered a dynticks idle phase with
1268 * no active irq/NMI handlers, then we can safely pretend that the CPU
1269 * already acknowledged the request to pass through a quiescent
1270 * state. Either way, that CPU cannot possibly be in an RCU
1271 * read-side critical section that started before the beginning
1272 * of the current RCU grace period.
1274 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1275 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1276 rdp->dynticks_fqs++;
1277 rcu_gpnum_ovf(rnp, rdp);
1282 * Has this CPU encountered a cond_resched_rcu_qs() since the
1283 * beginning of the grace period? For this to be the case,
1284 * the CPU has to have noticed the current grace period. This
1285 * might not be the case for nohz_full CPUs looping in the kernel.
1287 jtsq = jiffies_till_sched_qs;
1288 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1289 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1290 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1291 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1292 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1293 rcu_gpnum_ovf(rnp, rdp);
1295 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1296 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1297 smp_store_release(ruqp, true);
1300 /* Check for the CPU being offline. */
1301 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1302 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1304 rcu_gpnum_ovf(rnp, rdp);
1309 * A CPU running for an extended time within the kernel can
1310 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1311 * even context-switching back and forth between a pair of
1312 * in-kernel CPU-bound tasks cannot advance grace periods.
1313 * So if the grace period is old enough, make the CPU pay attention.
1314 * Note that the unsynchronized assignments to the per-CPU
1315 * rcu_need_heavy_qs variable are safe. Yes, setting of
1316 * bits can be lost, but they will be set again on the next
1317 * force-quiescent-state pass. So lost bit sets do not result
1318 * in incorrect behavior, merely in a grace period lasting
1319 * a few jiffies longer than it might otherwise. Because
1320 * there are at most four threads involved, and because the
1321 * updates are only once every few jiffies, the probability of
1322 * lossage (and thus of slight grace-period extension) is
1325 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1326 if (!READ_ONCE(*rnhqp) &&
1327 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1328 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1329 WRITE_ONCE(*rnhqp, true);
1330 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1331 smp_store_release(ruqp, true);
1332 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1336 * If more than halfway to RCU CPU stall-warning time, do a
1337 * resched_cpu() to try to loosen things up a bit. Also check to
1338 * see if the CPU is getting hammered with interrupts, but only
1339 * once per grace period, just to keep the IPIs down to a dull roar.
1341 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1342 resched_cpu(rdp->cpu);
1343 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1344 !rdp->rcu_iw_pending && rdp->rcu_iw_gpnum != rnp->gpnum &&
1345 (rnp->ffmask & rdp->grpmask)) {
1346 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1347 rdp->rcu_iw_pending = true;
1348 rdp->rcu_iw_gpnum = rnp->gpnum;
1349 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1356 static void record_gp_stall_check_time(struct rcu_state *rsp)
1358 unsigned long j = jiffies;
1362 smp_wmb(); /* Record start time before stall time. */
1363 j1 = rcu_jiffies_till_stall_check();
1364 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1365 rsp->jiffies_resched = j + j1 / 2;
1366 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1370 * Convert a ->gp_state value to a character string.
1372 static const char *gp_state_getname(short gs)
1374 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1376 return gp_state_names[gs];
1380 * Complain about starvation of grace-period kthread.
1382 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1388 gpa = READ_ONCE(rsp->gp_activity);
1389 if (j - gpa > 2 * HZ) {
1390 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1392 rsp->gpnum, rsp->completed,
1394 gp_state_getname(rsp->gp_state), rsp->gp_state,
1395 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1396 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1397 if (rsp->gp_kthread) {
1398 sched_show_task(rsp->gp_kthread);
1399 wake_up_process(rsp->gp_kthread);
1405 * Dump stacks of all tasks running on stalled CPUs. First try using
1406 * NMIs, but fall back to manual remote stack tracing on architectures
1407 * that don't support NMI-based stack dumps. The NMI-triggered stack
1408 * traces are more accurate because they are printed by the target CPU.
1410 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1413 unsigned long flags;
1414 struct rcu_node *rnp;
1416 rcu_for_each_leaf_node(rsp, rnp) {
1417 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1418 for_each_leaf_node_possible_cpu(rnp, cpu)
1419 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1420 if (!trigger_single_cpu_backtrace(cpu))
1422 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1427 * If too much time has passed in the current grace period, and if
1428 * so configured, go kick the relevant kthreads.
1430 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1434 if (!rcu_kick_kthreads)
1436 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1437 if (time_after(jiffies, j) && rsp->gp_kthread &&
1438 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1439 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1440 rcu_ftrace_dump(DUMP_ALL);
1441 wake_up_process(rsp->gp_kthread);
1442 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1446 static inline void panic_on_rcu_stall(void)
1448 if (sysctl_panic_on_rcu_stall)
1449 panic("RCU Stall\n");
1452 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1456 unsigned long flags;
1460 struct rcu_node *rnp = rcu_get_root(rsp);
1463 /* Kick and suppress, if so configured. */
1464 rcu_stall_kick_kthreads(rsp);
1465 if (rcu_cpu_stall_suppress)
1468 /* Only let one CPU complain about others per time interval. */
1470 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1471 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1472 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1473 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1476 WRITE_ONCE(rsp->jiffies_stall,
1477 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1478 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1481 * OK, time to rat on our buddy...
1482 * See Documentation/RCU/stallwarn.txt for info on how to debug
1483 * RCU CPU stall warnings.
1485 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1487 print_cpu_stall_info_begin();
1488 rcu_for_each_leaf_node(rsp, rnp) {
1489 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1490 ndetected += rcu_print_task_stall(rnp);
1491 if (rnp->qsmask != 0) {
1492 for_each_leaf_node_possible_cpu(rnp, cpu)
1493 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1494 print_cpu_stall_info(rsp, cpu);
1498 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1501 print_cpu_stall_info_end();
1502 for_each_possible_cpu(cpu)
1503 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1505 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1506 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1507 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1509 rcu_dump_cpu_stacks(rsp);
1511 /* Complain about tasks blocking the grace period. */
1512 rcu_print_detail_task_stall(rsp);
1514 if (READ_ONCE(rsp->gpnum) != gpnum ||
1515 READ_ONCE(rsp->completed) == gpnum) {
1516 pr_err("INFO: Stall ended before state dump start\n");
1519 gpa = READ_ONCE(rsp->gp_activity);
1520 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1521 rsp->name, j - gpa, j, gpa,
1522 jiffies_till_next_fqs,
1523 rcu_get_root(rsp)->qsmask);
1524 /* In this case, the current CPU might be at fault. */
1525 sched_show_task(current);
1529 rcu_check_gp_kthread_starvation(rsp);
1531 panic_on_rcu_stall();
1533 force_quiescent_state(rsp); /* Kick them all. */
1536 static void print_cpu_stall(struct rcu_state *rsp)
1539 unsigned long flags;
1540 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1541 struct rcu_node *rnp = rcu_get_root(rsp);
1544 /* Kick and suppress, if so configured. */
1545 rcu_stall_kick_kthreads(rsp);
1546 if (rcu_cpu_stall_suppress)
1550 * OK, time to rat on ourselves...
1551 * See Documentation/RCU/stallwarn.txt for info on how to debug
1552 * RCU CPU stall warnings.
1554 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1555 print_cpu_stall_info_begin();
1556 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1557 print_cpu_stall_info(rsp, smp_processor_id());
1558 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1559 print_cpu_stall_info_end();
1560 for_each_possible_cpu(cpu)
1561 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1563 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1564 jiffies - rsp->gp_start,
1565 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1567 rcu_check_gp_kthread_starvation(rsp);
1569 rcu_dump_cpu_stacks(rsp);
1571 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1572 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1573 WRITE_ONCE(rsp->jiffies_stall,
1574 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1575 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1577 panic_on_rcu_stall();
1580 * Attempt to revive the RCU machinery by forcing a context switch.
1582 * A context switch would normally allow the RCU state machine to make
1583 * progress and it could be we're stuck in kernel space without context
1584 * switches for an entirely unreasonable amount of time.
1586 resched_cpu(smp_processor_id());
1589 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1591 unsigned long completed;
1592 unsigned long gpnum;
1596 struct rcu_node *rnp;
1598 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1599 !rcu_gp_in_progress(rsp))
1601 rcu_stall_kick_kthreads(rsp);
1605 * Lots of memory barriers to reject false positives.
1607 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1608 * then rsp->gp_start, and finally rsp->completed. These values
1609 * are updated in the opposite order with memory barriers (or
1610 * equivalent) during grace-period initialization and cleanup.
1611 * Now, a false positive can occur if we get an new value of
1612 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1613 * the memory barriers, the only way that this can happen is if one
1614 * grace period ends and another starts between these two fetches.
1615 * Detect this by comparing rsp->completed with the previous fetch
1618 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1619 * and rsp->gp_start suffice to forestall false positives.
1621 gpnum = READ_ONCE(rsp->gpnum);
1622 smp_rmb(); /* Pick up ->gpnum first... */
1623 js = READ_ONCE(rsp->jiffies_stall);
1624 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1625 gps = READ_ONCE(rsp->gp_start);
1626 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1627 completed = READ_ONCE(rsp->completed);
1628 if (ULONG_CMP_GE(completed, gpnum) ||
1629 ULONG_CMP_LT(j, js) ||
1630 ULONG_CMP_GE(gps, js))
1631 return; /* No stall or GP completed since entering function. */
1633 if (rcu_gp_in_progress(rsp) &&
1634 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1636 /* We haven't checked in, so go dump stack. */
1637 print_cpu_stall(rsp);
1639 } else if (rcu_gp_in_progress(rsp) &&
1640 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1642 /* They had a few time units to dump stack, so complain. */
1643 print_other_cpu_stall(rsp, gpnum);
1648 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1650 * Set the stall-warning timeout way off into the future, thus preventing
1651 * any RCU CPU stall-warning messages from appearing in the current set of
1652 * RCU grace periods.
1654 * The caller must disable hard irqs.
1656 void rcu_cpu_stall_reset(void)
1658 struct rcu_state *rsp;
1660 for_each_rcu_flavor(rsp)
1661 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1665 * Determine the value that ->completed will have at the end of the
1666 * next subsequent grace period. This is used to tag callbacks so that
1667 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1668 * been dyntick-idle for an extended period with callbacks under the
1669 * influence of RCU_FAST_NO_HZ.
1671 * The caller must hold rnp->lock with interrupts disabled.
1673 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1674 struct rcu_node *rnp)
1676 lockdep_assert_held(&rnp->lock);
1679 * If RCU is idle, we just wait for the next grace period.
1680 * But we can only be sure that RCU is idle if we are looking
1681 * at the root rcu_node structure -- otherwise, a new grace
1682 * period might have started, but just not yet gotten around
1683 * to initializing the current non-root rcu_node structure.
1685 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1686 return rnp->completed + 1;
1689 * Otherwise, wait for a possible partial grace period and
1690 * then the subsequent full grace period.
1692 return rnp->completed + 2;
1696 * Trace-event helper function for rcu_start_future_gp() and
1697 * rcu_nocb_wait_gp().
1699 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1700 unsigned long c, const char *s)
1702 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1703 rnp->completed, c, rnp->level,
1704 rnp->grplo, rnp->grphi, s);
1708 * Start some future grace period, as needed to handle newly arrived
1709 * callbacks. The required future grace periods are recorded in each
1710 * rcu_node structure's ->need_future_gp field. Returns true if there
1711 * is reason to awaken the grace-period kthread.
1713 * The caller must hold the specified rcu_node structure's ->lock.
1715 static bool __maybe_unused
1716 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1717 unsigned long *c_out)
1721 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1723 lockdep_assert_held(&rnp->lock);
1726 * Pick up grace-period number for new callbacks. If this
1727 * grace period is already marked as needed, return to the caller.
1729 c = rcu_cbs_completed(rdp->rsp, rnp);
1730 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1731 if (rnp->need_future_gp[c & 0x1]) {
1732 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1737 * If either this rcu_node structure or the root rcu_node structure
1738 * believe that a grace period is in progress, then we must wait
1739 * for the one following, which is in "c". Because our request
1740 * will be noticed at the end of the current grace period, we don't
1741 * need to explicitly start one. We only do the lockless check
1742 * of rnp_root's fields if the current rcu_node structure thinks
1743 * there is no grace period in flight, and because we hold rnp->lock,
1744 * the only possible change is when rnp_root's two fields are
1745 * equal, in which case rnp_root->gpnum might be concurrently
1746 * incremented. But that is OK, as it will just result in our
1747 * doing some extra useless work.
1749 if (rnp->gpnum != rnp->completed ||
1750 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1751 rnp->need_future_gp[c & 0x1]++;
1752 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1757 * There might be no grace period in progress. If we don't already
1758 * hold it, acquire the root rcu_node structure's lock in order to
1759 * start one (if needed).
1761 if (rnp != rnp_root)
1762 raw_spin_lock_rcu_node(rnp_root);
1765 * Get a new grace-period number. If there really is no grace
1766 * period in progress, it will be smaller than the one we obtained
1767 * earlier. Adjust callbacks as needed.
1769 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1770 if (!rcu_is_nocb_cpu(rdp->cpu))
1771 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1774 * If the needed for the required grace period is already
1775 * recorded, trace and leave.
1777 if (rnp_root->need_future_gp[c & 0x1]) {
1778 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1782 /* Record the need for the future grace period. */
1783 rnp_root->need_future_gp[c & 0x1]++;
1785 /* If a grace period is not already in progress, start one. */
1786 if (rnp_root->gpnum != rnp_root->completed) {
1787 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1789 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1790 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1793 if (rnp != rnp_root)
1794 raw_spin_unlock_rcu_node(rnp_root);
1802 * Clean up any old requests for the just-ended grace period. Also return
1803 * whether any additional grace periods have been requested.
1805 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1807 int c = rnp->completed;
1809 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1811 rnp->need_future_gp[c & 0x1] = 0;
1812 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1813 trace_rcu_future_gp(rnp, rdp, c,
1814 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1819 * Awaken the grace-period kthread for the specified flavor of RCU.
1820 * Don't do a self-awaken, and don't bother awakening when there is
1821 * nothing for the grace-period kthread to do (as in several CPUs
1822 * raced to awaken, and we lost), and finally don't try to awaken
1823 * a kthread that has not yet been created.
1825 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1827 if (current == rsp->gp_kthread ||
1828 !READ_ONCE(rsp->gp_flags) ||
1831 swake_up(&rsp->gp_wq);
1835 * If there is room, assign a ->completed number to any callbacks on
1836 * this CPU that have not already been assigned. Also accelerate any
1837 * callbacks that were previously assigned a ->completed number that has
1838 * since proven to be too conservative, which can happen if callbacks get
1839 * assigned a ->completed number while RCU is idle, but with reference to
1840 * a non-root rcu_node structure. This function is idempotent, so it does
1841 * not hurt to call it repeatedly. Returns an flag saying that we should
1842 * awaken the RCU grace-period kthread.
1844 * The caller must hold rnp->lock with interrupts disabled.
1846 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1847 struct rcu_data *rdp)
1851 lockdep_assert_held(&rnp->lock);
1853 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1854 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1858 * Callbacks are often registered with incomplete grace-period
1859 * information. Something about the fact that getting exact
1860 * information requires acquiring a global lock... RCU therefore
1861 * makes a conservative estimate of the grace period number at which
1862 * a given callback will become ready to invoke. The following
1863 * code checks this estimate and improves it when possible, thus
1864 * accelerating callback invocation to an earlier grace-period
1867 if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
1868 ret = rcu_start_future_gp(rnp, rdp, NULL);
1870 /* Trace depending on how much we were able to accelerate. */
1871 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1872 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1874 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1879 * Move any callbacks whose grace period has completed to the
1880 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1881 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1882 * sublist. This function is idempotent, so it does not hurt to
1883 * invoke it repeatedly. As long as it is not invoked -too- often...
1884 * Returns true if the RCU grace-period kthread needs to be awakened.
1886 * The caller must hold rnp->lock with interrupts disabled.
1888 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1889 struct rcu_data *rdp)
1891 lockdep_assert_held(&rnp->lock);
1893 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1894 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1898 * Find all callbacks whose ->completed numbers indicate that they
1899 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1901 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1903 /* Classify any remaining callbacks. */
1904 return rcu_accelerate_cbs(rsp, rnp, rdp);
1908 * Update CPU-local rcu_data state to record the beginnings and ends of
1909 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1910 * structure corresponding to the current CPU, and must have irqs disabled.
1911 * Returns true if the grace-period kthread needs to be awakened.
1913 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1914 struct rcu_data *rdp)
1919 lockdep_assert_held(&rnp->lock);
1921 /* Handle the ends of any preceding grace periods first. */
1922 if (rdp->completed == rnp->completed &&
1923 !unlikely(READ_ONCE(rdp->gpwrap))) {
1925 /* No grace period end, so just accelerate recent callbacks. */
1926 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1930 /* Advance callbacks. */
1931 ret = rcu_advance_cbs(rsp, rnp, rdp);
1933 /* Remember that we saw this grace-period completion. */
1934 rdp->completed = rnp->completed;
1935 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1938 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1940 * If the current grace period is waiting for this CPU,
1941 * set up to detect a quiescent state, otherwise don't
1942 * go looking for one.
1944 rdp->gpnum = rnp->gpnum;
1945 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1946 need_gp = !!(rnp->qsmask & rdp->grpmask);
1947 rdp->cpu_no_qs.b.norm = need_gp;
1948 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1949 rdp->core_needs_qs = need_gp;
1950 zero_cpu_stall_ticks(rdp);
1951 WRITE_ONCE(rdp->gpwrap, false);
1952 rcu_gpnum_ovf(rnp, rdp);
1957 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1959 unsigned long flags;
1961 struct rcu_node *rnp;
1963 local_irq_save(flags);
1965 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1966 rdp->completed == READ_ONCE(rnp->completed) &&
1967 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1968 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1969 local_irq_restore(flags);
1972 needwake = __note_gp_changes(rsp, rnp, rdp);
1973 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1975 rcu_gp_kthread_wake(rsp);
1978 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1981 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1982 schedule_timeout_uninterruptible(delay);
1986 * Initialize a new grace period. Return false if no grace period required.
1988 static bool rcu_gp_init(struct rcu_state *rsp)
1990 unsigned long oldmask;
1991 struct rcu_data *rdp;
1992 struct rcu_node *rnp = rcu_get_root(rsp);
1994 WRITE_ONCE(rsp->gp_activity, jiffies);
1995 raw_spin_lock_irq_rcu_node(rnp);
1996 if (!READ_ONCE(rsp->gp_flags)) {
1997 /* Spurious wakeup, tell caller to go back to sleep. */
1998 raw_spin_unlock_irq_rcu_node(rnp);
2001 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
2003 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
2005 * Grace period already in progress, don't start another.
2006 * Not supposed to be able to happen.
2008 raw_spin_unlock_irq_rcu_node(rnp);
2012 /* Advance to a new grace period and initialize state. */
2013 record_gp_stall_check_time(rsp);
2014 /* Record GP times before starting GP, hence smp_store_release(). */
2015 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
2016 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
2017 raw_spin_unlock_irq_rcu_node(rnp);
2020 * Apply per-leaf buffered online and offline operations to the
2021 * rcu_node tree. Note that this new grace period need not wait
2022 * for subsequent online CPUs, and that quiescent-state forcing
2023 * will handle subsequent offline CPUs.
2025 rcu_for_each_leaf_node(rsp, rnp) {
2026 rcu_gp_slow(rsp, gp_preinit_delay);
2027 raw_spin_lock_irq_rcu_node(rnp);
2028 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2029 !rnp->wait_blkd_tasks) {
2030 /* Nothing to do on this leaf rcu_node structure. */
2031 raw_spin_unlock_irq_rcu_node(rnp);
2035 /* Record old state, apply changes to ->qsmaskinit field. */
2036 oldmask = rnp->qsmaskinit;
2037 rnp->qsmaskinit = rnp->qsmaskinitnext;
2039 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2040 if (!oldmask != !rnp->qsmaskinit) {
2041 if (!oldmask) /* First online CPU for this rcu_node. */
2042 rcu_init_new_rnp(rnp);
2043 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2044 rnp->wait_blkd_tasks = true;
2045 else /* Last offline CPU and can propagate. */
2046 rcu_cleanup_dead_rnp(rnp);
2050 * If all waited-on tasks from prior grace period are
2051 * done, and if all this rcu_node structure's CPUs are
2052 * still offline, propagate up the rcu_node tree and
2053 * clear ->wait_blkd_tasks. Otherwise, if one of this
2054 * rcu_node structure's CPUs has since come back online,
2055 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2056 * checks for this, so just call it unconditionally).
2058 if (rnp->wait_blkd_tasks &&
2059 (!rcu_preempt_has_tasks(rnp) ||
2061 rnp->wait_blkd_tasks = false;
2062 rcu_cleanup_dead_rnp(rnp);
2065 raw_spin_unlock_irq_rcu_node(rnp);
2069 * Set the quiescent-state-needed bits in all the rcu_node
2070 * structures for all currently online CPUs in breadth-first order,
2071 * starting from the root rcu_node structure, relying on the layout
2072 * of the tree within the rsp->node[] array. Note that other CPUs
2073 * will access only the leaves of the hierarchy, thus seeing that no
2074 * grace period is in progress, at least until the corresponding
2075 * leaf node has been initialized.
2077 * The grace period cannot complete until the initialization
2078 * process finishes, because this kthread handles both.
2080 rcu_for_each_node_breadth_first(rsp, rnp) {
2081 rcu_gp_slow(rsp, gp_init_delay);
2082 raw_spin_lock_irq_rcu_node(rnp);
2083 rdp = this_cpu_ptr(rsp->rda);
2084 rcu_preempt_check_blocked_tasks(rnp);
2085 rnp->qsmask = rnp->qsmaskinit;
2086 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2087 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2088 WRITE_ONCE(rnp->completed, rsp->completed);
2089 if (rnp == rdp->mynode)
2090 (void)__note_gp_changes(rsp, rnp, rdp);
2091 rcu_preempt_boost_start_gp(rnp);
2092 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2093 rnp->level, rnp->grplo,
2094 rnp->grphi, rnp->qsmask);
2095 raw_spin_unlock_irq_rcu_node(rnp);
2096 cond_resched_rcu_qs();
2097 WRITE_ONCE(rsp->gp_activity, jiffies);
2104 * Helper function for swait_event_idle() wakeup at force-quiescent-state
2107 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2109 struct rcu_node *rnp = rcu_get_root(rsp);
2111 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2112 *gfp = READ_ONCE(rsp->gp_flags);
2113 if (*gfp & RCU_GP_FLAG_FQS)
2116 /* The current grace period has completed. */
2117 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2124 * Do one round of quiescent-state forcing.
2126 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2128 struct rcu_node *rnp = rcu_get_root(rsp);
2130 WRITE_ONCE(rsp->gp_activity, jiffies);
2133 /* Collect dyntick-idle snapshots. */
2134 force_qs_rnp(rsp, dyntick_save_progress_counter);
2136 /* Handle dyntick-idle and offline CPUs. */
2137 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2139 /* Clear flag to prevent immediate re-entry. */
2140 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2141 raw_spin_lock_irq_rcu_node(rnp);
2142 WRITE_ONCE(rsp->gp_flags,
2143 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2144 raw_spin_unlock_irq_rcu_node(rnp);
2149 * Clean up after the old grace period.
2151 static void rcu_gp_cleanup(struct rcu_state *rsp)
2153 unsigned long gp_duration;
2154 bool needgp = false;
2156 struct rcu_data *rdp;
2157 struct rcu_node *rnp = rcu_get_root(rsp);
2158 struct swait_queue_head *sq;
2160 WRITE_ONCE(rsp->gp_activity, jiffies);
2161 raw_spin_lock_irq_rcu_node(rnp);
2162 gp_duration = jiffies - rsp->gp_start;
2163 if (gp_duration > rsp->gp_max)
2164 rsp->gp_max = gp_duration;
2167 * We know the grace period is complete, but to everyone else
2168 * it appears to still be ongoing. But it is also the case
2169 * that to everyone else it looks like there is nothing that
2170 * they can do to advance the grace period. It is therefore
2171 * safe for us to drop the lock in order to mark the grace
2172 * period as completed in all of the rcu_node structures.
2174 raw_spin_unlock_irq_rcu_node(rnp);
2177 * Propagate new ->completed value to rcu_node structures so
2178 * that other CPUs don't have to wait until the start of the next
2179 * grace period to process their callbacks. This also avoids
2180 * some nasty RCU grace-period initialization races by forcing
2181 * the end of the current grace period to be completely recorded in
2182 * all of the rcu_node structures before the beginning of the next
2183 * grace period is recorded in any of the rcu_node structures.
2185 rcu_for_each_node_breadth_first(rsp, rnp) {
2186 raw_spin_lock_irq_rcu_node(rnp);
2187 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2188 WARN_ON_ONCE(rnp->qsmask);
2189 WRITE_ONCE(rnp->completed, rsp->gpnum);
2190 rdp = this_cpu_ptr(rsp->rda);
2191 if (rnp == rdp->mynode)
2192 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2193 /* smp_mb() provided by prior unlock-lock pair. */
2194 nocb += rcu_future_gp_cleanup(rsp, rnp);
2195 sq = rcu_nocb_gp_get(rnp);
2196 raw_spin_unlock_irq_rcu_node(rnp);
2197 rcu_nocb_gp_cleanup(sq);
2198 cond_resched_rcu_qs();
2199 WRITE_ONCE(rsp->gp_activity, jiffies);
2200 rcu_gp_slow(rsp, gp_cleanup_delay);
2202 rnp = rcu_get_root(rsp);
2203 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2204 rcu_nocb_gp_set(rnp, nocb);
2206 /* Declare grace period done. */
2207 WRITE_ONCE(rsp->completed, rsp->gpnum);
2208 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2209 rsp->gp_state = RCU_GP_IDLE;
2210 rdp = this_cpu_ptr(rsp->rda);
2211 /* Advance CBs to reduce false positives below. */
2212 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2213 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2214 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2215 trace_rcu_grace_period(rsp->name,
2216 READ_ONCE(rsp->gpnum),
2219 raw_spin_unlock_irq_rcu_node(rnp);
2223 * Body of kthread that handles grace periods.
2225 static int __noreturn rcu_gp_kthread(void *arg)
2231 struct rcu_state *rsp = arg;
2232 struct rcu_node *rnp = rcu_get_root(rsp);
2234 rcu_bind_gp_kthread();
2237 /* Handle grace-period start. */
2239 trace_rcu_grace_period(rsp->name,
2240 READ_ONCE(rsp->gpnum),
2242 rsp->gp_state = RCU_GP_WAIT_GPS;
2243 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2245 rsp->gp_state = RCU_GP_DONE_GPS;
2246 /* Locking provides needed memory barrier. */
2247 if (rcu_gp_init(rsp))
2249 cond_resched_rcu_qs();
2250 WRITE_ONCE(rsp->gp_activity, jiffies);
2251 WARN_ON(signal_pending(current));
2252 trace_rcu_grace_period(rsp->name,
2253 READ_ONCE(rsp->gpnum),
2257 /* Handle quiescent-state forcing. */
2258 first_gp_fqs = true;
2259 j = jiffies_till_first_fqs;
2262 jiffies_till_first_fqs = HZ;
2267 rsp->jiffies_force_qs = jiffies + j;
2268 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2271 trace_rcu_grace_period(rsp->name,
2272 READ_ONCE(rsp->gpnum),
2274 rsp->gp_state = RCU_GP_WAIT_FQS;
2275 ret = swait_event_idle_timeout(rsp->gp_wq,
2276 rcu_gp_fqs_check_wake(rsp, &gf), j);
2277 rsp->gp_state = RCU_GP_DOING_FQS;
2278 /* Locking provides needed memory barriers. */
2279 /* If grace period done, leave loop. */
2280 if (!READ_ONCE(rnp->qsmask) &&
2281 !rcu_preempt_blocked_readers_cgp(rnp))
2283 /* If time for quiescent-state forcing, do it. */
2284 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2285 (gf & RCU_GP_FLAG_FQS)) {
2286 trace_rcu_grace_period(rsp->name,
2287 READ_ONCE(rsp->gpnum),
2289 rcu_gp_fqs(rsp, first_gp_fqs);
2290 first_gp_fqs = false;
2291 trace_rcu_grace_period(rsp->name,
2292 READ_ONCE(rsp->gpnum),
2294 cond_resched_rcu_qs();
2295 WRITE_ONCE(rsp->gp_activity, jiffies);
2296 ret = 0; /* Force full wait till next FQS. */
2297 j = jiffies_till_next_fqs;
2300 jiffies_till_next_fqs = HZ;
2303 jiffies_till_next_fqs = 1;
2306 /* Deal with stray signal. */
2307 cond_resched_rcu_qs();
2308 WRITE_ONCE(rsp->gp_activity, jiffies);
2309 WARN_ON(signal_pending(current));
2310 trace_rcu_grace_period(rsp->name,
2311 READ_ONCE(rsp->gpnum),
2313 ret = 1; /* Keep old FQS timing. */
2315 if (time_after(jiffies, rsp->jiffies_force_qs))
2318 j = rsp->jiffies_force_qs - j;
2322 /* Handle grace-period end. */
2323 rsp->gp_state = RCU_GP_CLEANUP;
2324 rcu_gp_cleanup(rsp);
2325 rsp->gp_state = RCU_GP_CLEANED;
2330 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2331 * in preparation for detecting the next grace period. The caller must hold
2332 * the root node's ->lock and hard irqs must be disabled.
2334 * Note that it is legal for a dying CPU (which is marked as offline) to
2335 * invoke this function. This can happen when the dying CPU reports its
2338 * Returns true if the grace-period kthread must be awakened.
2341 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2342 struct rcu_data *rdp)
2344 lockdep_assert_held(&rnp->lock);
2345 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2347 * Either we have not yet spawned the grace-period
2348 * task, this CPU does not need another grace period,
2349 * or a grace period is already in progress.
2350 * Either way, don't start a new grace period.
2354 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2355 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2359 * We can't do wakeups while holding the rnp->lock, as that
2360 * could cause possible deadlocks with the rq->lock. Defer
2361 * the wakeup to our caller.
2367 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2368 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2369 * is invoked indirectly from rcu_advance_cbs(), which would result in
2370 * endless recursion -- or would do so if it wasn't for the self-deadlock
2371 * that is encountered beforehand.
2373 * Returns true if the grace-period kthread needs to be awakened.
2375 static bool rcu_start_gp(struct rcu_state *rsp)
2377 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2378 struct rcu_node *rnp = rcu_get_root(rsp);
2382 * If there is no grace period in progress right now, any
2383 * callbacks we have up to this point will be satisfied by the
2384 * next grace period. Also, advancing the callbacks reduces the
2385 * probability of false positives from cpu_needs_another_gp()
2386 * resulting in pointless grace periods. So, advance callbacks
2387 * then start the grace period!
2389 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2390 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2395 * Report a full set of quiescent states to the specified rcu_state data
2396 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2397 * kthread if another grace period is required. Whether we wake
2398 * the grace-period kthread or it awakens itself for the next round
2399 * of quiescent-state forcing, that kthread will clean up after the
2400 * just-completed grace period. Note that the caller must hold rnp->lock,
2401 * which is released before return.
2403 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2404 __releases(rcu_get_root(rsp)->lock)
2406 lockdep_assert_held(&rcu_get_root(rsp)->lock);
2407 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2408 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2409 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2410 rcu_gp_kthread_wake(rsp);
2414 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2415 * Allows quiescent states for a group of CPUs to be reported at one go
2416 * to the specified rcu_node structure, though all the CPUs in the group
2417 * must be represented by the same rcu_node structure (which need not be a
2418 * leaf rcu_node structure, though it often will be). The gps parameter
2419 * is the grace-period snapshot, which means that the quiescent states
2420 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2421 * must be held upon entry, and it is released before return.
2424 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2425 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2426 __releases(rnp->lock)
2428 unsigned long oldmask = 0;
2429 struct rcu_node *rnp_c;
2431 lockdep_assert_held(&rnp->lock);
2433 /* Walk up the rcu_node hierarchy. */
2435 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2438 * Our bit has already been cleared, or the
2439 * relevant grace period is already over, so done.
2441 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2444 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2445 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1 &&
2446 rcu_preempt_blocked_readers_cgp(rnp));
2447 rnp->qsmask &= ~mask;
2448 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2449 mask, rnp->qsmask, rnp->level,
2450 rnp->grplo, rnp->grphi,
2452 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2454 /* Other bits still set at this level, so done. */
2455 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2458 mask = rnp->grpmask;
2459 if (rnp->parent == NULL) {
2461 /* No more levels. Exit loop holding root lock. */
2465 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2468 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2469 oldmask = rnp_c->qsmask;
2473 * Get here if we are the last CPU to pass through a quiescent
2474 * state for this grace period. Invoke rcu_report_qs_rsp()
2475 * to clean up and start the next grace period if one is needed.
2477 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2481 * Record a quiescent state for all tasks that were previously queued
2482 * on the specified rcu_node structure and that were blocking the current
2483 * RCU grace period. The caller must hold the specified rnp->lock with
2484 * irqs disabled, and this lock is released upon return, but irqs remain
2487 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2488 struct rcu_node *rnp, unsigned long flags)
2489 __releases(rnp->lock)
2493 struct rcu_node *rnp_p;
2495 lockdep_assert_held(&rnp->lock);
2496 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2497 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2498 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2499 return; /* Still need more quiescent states! */
2502 rnp_p = rnp->parent;
2503 if (rnp_p == NULL) {
2505 * Only one rcu_node structure in the tree, so don't
2506 * try to report up to its nonexistent parent!
2508 rcu_report_qs_rsp(rsp, flags);
2512 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2514 mask = rnp->grpmask;
2515 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2516 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2517 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2521 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2522 * structure. This must be called from the specified CPU.
2525 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2527 unsigned long flags;
2530 struct rcu_node *rnp;
2533 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2534 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2535 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2538 * The grace period in which this quiescent state was
2539 * recorded has ended, so don't report it upwards.
2540 * We will instead need a new quiescent state that lies
2541 * within the current grace period.
2543 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2544 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2545 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2548 mask = rdp->grpmask;
2549 if ((rnp->qsmask & mask) == 0) {
2550 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2552 rdp->core_needs_qs = false;
2555 * This GP can't end until cpu checks in, so all of our
2556 * callbacks can be processed during the next GP.
2558 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2560 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2561 /* ^^^ Released rnp->lock */
2563 rcu_gp_kthread_wake(rsp);
2568 * Check to see if there is a new grace period of which this CPU
2569 * is not yet aware, and if so, set up local rcu_data state for it.
2570 * Otherwise, see if this CPU has just passed through its first
2571 * quiescent state for this grace period, and record that fact if so.
2574 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2576 /* Check for grace-period ends and beginnings. */
2577 note_gp_changes(rsp, rdp);
2580 * Does this CPU still need to do its part for current grace period?
2581 * If no, return and let the other CPUs do their part as well.
2583 if (!rdp->core_needs_qs)
2587 * Was there a quiescent state since the beginning of the grace
2588 * period? If no, then exit and wait for the next call.
2590 if (rdp->cpu_no_qs.b.norm)
2594 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2597 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2601 * Trace the fact that this CPU is going offline.
2603 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2605 RCU_TRACE(unsigned long mask;)
2606 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2607 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2609 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2612 RCU_TRACE(mask = rdp->grpmask;)
2613 trace_rcu_grace_period(rsp->name,
2614 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2619 * All CPUs for the specified rcu_node structure have gone offline,
2620 * and all tasks that were preempted within an RCU read-side critical
2621 * section while running on one of those CPUs have since exited their RCU
2622 * read-side critical section. Some other CPU is reporting this fact with
2623 * the specified rcu_node structure's ->lock held and interrupts disabled.
2624 * This function therefore goes up the tree of rcu_node structures,
2625 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2626 * the leaf rcu_node structure's ->qsmaskinit field has already been
2629 * This function does check that the specified rcu_node structure has
2630 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2631 * prematurely. That said, invoking it after the fact will cost you
2632 * a needless lock acquisition. So once it has done its work, don't
2635 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2638 struct rcu_node *rnp = rnp_leaf;
2640 lockdep_assert_held(&rnp->lock);
2641 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2642 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2645 mask = rnp->grpmask;
2649 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2650 rnp->qsmaskinit &= ~mask;
2651 rnp->qsmask &= ~mask;
2652 if (rnp->qsmaskinit) {
2653 raw_spin_unlock_rcu_node(rnp);
2654 /* irqs remain disabled. */
2657 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2662 * The CPU has been completely removed, and some other CPU is reporting
2663 * this fact from process context. Do the remainder of the cleanup.
2664 * There can only be one CPU hotplug operation at a time, so no need for
2667 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2669 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2670 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2672 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2675 /* Adjust any no-longer-needed kthreads. */
2676 rcu_boost_kthread_setaffinity(rnp, -1);
2680 * Invoke any RCU callbacks that have made it to the end of their grace
2681 * period. Thottle as specified by rdp->blimit.
2683 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2685 unsigned long flags;
2686 struct rcu_head *rhp;
2687 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2690 /* If no callbacks are ready, just return. */
2691 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2692 trace_rcu_batch_start(rsp->name,
2693 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2694 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2695 trace_rcu_batch_end(rsp->name, 0,
2696 !rcu_segcblist_empty(&rdp->cblist),
2697 need_resched(), is_idle_task(current),
2698 rcu_is_callbacks_kthread());
2703 * Extract the list of ready callbacks, disabling to prevent
2704 * races with call_rcu() from interrupt handlers. Leave the
2705 * callback counts, as rcu_barrier() needs to be conservative.
2707 local_irq_save(flags);
2708 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2710 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2711 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2712 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2713 local_irq_restore(flags);
2715 /* Invoke callbacks. */
2716 rhp = rcu_cblist_dequeue(&rcl);
2717 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2718 debug_rcu_head_unqueue(rhp);
2719 if (__rcu_reclaim(rsp->name, rhp))
2720 rcu_cblist_dequeued_lazy(&rcl);
2722 * Stop only if limit reached and CPU has something to do.
2723 * Note: The rcl structure counts down from zero.
2725 if (-rcl.len >= bl &&
2727 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2731 local_irq_save(flags);
2733 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2734 is_idle_task(current), rcu_is_callbacks_kthread());
2736 /* Update counts and requeue any remaining callbacks. */
2737 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2738 smp_mb(); /* List handling before counting for rcu_barrier(). */
2739 rdp->n_cbs_invoked += count;
2740 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2742 /* Reinstate batch limit if we have worked down the excess. */
2743 count = rcu_segcblist_n_cbs(&rdp->cblist);
2744 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2745 rdp->blimit = blimit;
2747 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2748 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2749 rdp->qlen_last_fqs_check = 0;
2750 rdp->n_force_qs_snap = rsp->n_force_qs;
2751 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2752 rdp->qlen_last_fqs_check = count;
2753 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2755 local_irq_restore(flags);
2757 /* Re-invoke RCU core processing if there are callbacks remaining. */
2758 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2763 * Check to see if this CPU is in a non-context-switch quiescent state
2764 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2765 * Also schedule RCU core processing.
2767 * This function must be called from hardirq context. It is normally
2768 * invoked from the scheduling-clock interrupt.
2770 void rcu_check_callbacks(int user)
2772 trace_rcu_utilization(TPS("Start scheduler-tick"));
2773 increment_cpu_stall_ticks();
2774 if (user || rcu_is_cpu_rrupt_from_idle()) {
2777 * Get here if this CPU took its interrupt from user
2778 * mode or from the idle loop, and if this is not a
2779 * nested interrupt. In this case, the CPU is in
2780 * a quiescent state, so note it.
2782 * No memory barrier is required here because both
2783 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2784 * variables that other CPUs neither access nor modify,
2785 * at least not while the corresponding CPU is online.
2791 } else if (!in_softirq()) {
2794 * Get here if this CPU did not take its interrupt from
2795 * softirq, in other words, if it is not interrupting
2796 * a rcu_bh read-side critical section. This is an _bh
2797 * critical section, so note it.
2802 rcu_preempt_check_callbacks();
2806 rcu_note_voluntary_context_switch(current);
2807 trace_rcu_utilization(TPS("End scheduler-tick"));
2811 * Scan the leaf rcu_node structures, processing dyntick state for any that
2812 * have not yet encountered a quiescent state, using the function specified.
2813 * Also initiate boosting for any threads blocked on the root rcu_node.
2815 * The caller must have suppressed start of new grace periods.
2817 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2820 unsigned long flags;
2822 struct rcu_node *rnp;
2824 rcu_for_each_leaf_node(rsp, rnp) {
2825 cond_resched_rcu_qs();
2827 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2828 if (rnp->qsmask == 0) {
2829 if (rcu_state_p == &rcu_sched_state ||
2830 rsp != rcu_state_p ||
2831 rcu_preempt_blocked_readers_cgp(rnp)) {
2833 * No point in scanning bits because they
2834 * are all zero. But we might need to
2835 * priority-boost blocked readers.
2837 rcu_initiate_boost(rnp, flags);
2838 /* rcu_initiate_boost() releases rnp->lock */
2842 (rnp->parent->qsmask & rnp->grpmask)) {
2844 * Race between grace-period
2845 * initialization and task exiting RCU
2846 * read-side critical section: Report.
2848 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2849 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2853 for_each_leaf_node_possible_cpu(rnp, cpu) {
2854 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2855 if ((rnp->qsmask & bit) != 0) {
2856 if (f(per_cpu_ptr(rsp->rda, cpu)))
2861 /* Idle/offline CPUs, report (releases rnp->lock. */
2862 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2864 /* Nothing to do here, so just drop the lock. */
2865 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2871 * Force quiescent states on reluctant CPUs, and also detect which
2872 * CPUs are in dyntick-idle mode.
2874 static void force_quiescent_state(struct rcu_state *rsp)
2876 unsigned long flags;
2878 struct rcu_node *rnp;
2879 struct rcu_node *rnp_old = NULL;
2881 /* Funnel through hierarchy to reduce memory contention. */
2882 rnp = __this_cpu_read(rsp->rda->mynode);
2883 for (; rnp != NULL; rnp = rnp->parent) {
2884 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2885 !raw_spin_trylock(&rnp->fqslock);
2886 if (rnp_old != NULL)
2887 raw_spin_unlock(&rnp_old->fqslock);
2889 rsp->n_force_qs_lh++;
2894 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2896 /* Reached the root of the rcu_node tree, acquire lock. */
2897 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2898 raw_spin_unlock(&rnp_old->fqslock);
2899 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2900 rsp->n_force_qs_lh++;
2901 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2902 return; /* Someone beat us to it. */
2904 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2905 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2906 rcu_gp_kthread_wake(rsp);
2910 * This does the RCU core processing work for the specified rcu_state
2911 * and rcu_data structures. This may be called only from the CPU to
2912 * whom the rdp belongs.
2915 __rcu_process_callbacks(struct rcu_state *rsp)
2917 unsigned long flags;
2919 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2921 WARN_ON_ONCE(!rdp->beenonline);
2923 /* Update RCU state based on any recent quiescent states. */
2924 rcu_check_quiescent_state(rsp, rdp);
2926 /* Does this CPU require a not-yet-started grace period? */
2927 local_irq_save(flags);
2928 if (cpu_needs_another_gp(rsp, rdp)) {
2929 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2930 needwake = rcu_start_gp(rsp);
2931 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2933 rcu_gp_kthread_wake(rsp);
2935 local_irq_restore(flags);
2938 /* If there are callbacks ready, invoke them. */
2939 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2940 invoke_rcu_callbacks(rsp, rdp);
2942 /* Do any needed deferred wakeups of rcuo kthreads. */
2943 do_nocb_deferred_wakeup(rdp);
2947 * Do RCU core processing for the current CPU.
2949 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2951 struct rcu_state *rsp;
2953 if (cpu_is_offline(smp_processor_id()))
2955 trace_rcu_utilization(TPS("Start RCU core"));
2956 for_each_rcu_flavor(rsp)
2957 __rcu_process_callbacks(rsp);
2958 trace_rcu_utilization(TPS("End RCU core"));
2962 * Schedule RCU callback invocation. If the specified type of RCU
2963 * does not support RCU priority boosting, just do a direct call,
2964 * otherwise wake up the per-CPU kernel kthread. Note that because we
2965 * are running on the current CPU with softirqs disabled, the
2966 * rcu_cpu_kthread_task cannot disappear out from under us.
2968 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2970 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2972 if (likely(!rsp->boost)) {
2973 rcu_do_batch(rsp, rdp);
2976 invoke_rcu_callbacks_kthread();
2979 static void invoke_rcu_core(void)
2981 if (cpu_online(smp_processor_id()))
2982 raise_softirq(RCU_SOFTIRQ);
2986 * Handle any core-RCU processing required by a call_rcu() invocation.
2988 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2989 struct rcu_head *head, unsigned long flags)
2994 * If called from an extended quiescent state, invoke the RCU
2995 * core in order to force a re-evaluation of RCU's idleness.
2997 if (!rcu_is_watching())
3000 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3001 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3005 * Force the grace period if too many callbacks or too long waiting.
3006 * Enforce hysteresis, and don't invoke force_quiescent_state()
3007 * if some other CPU has recently done so. Also, don't bother
3008 * invoking force_quiescent_state() if the newly enqueued callback
3009 * is the only one waiting for a grace period to complete.
3011 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
3012 rdp->qlen_last_fqs_check + qhimark)) {
3014 /* Are we ignoring a completed grace period? */
3015 note_gp_changes(rsp, rdp);
3017 /* Start a new grace period if one not already started. */
3018 if (!rcu_gp_in_progress(rsp)) {
3019 struct rcu_node *rnp_root = rcu_get_root(rsp);
3021 raw_spin_lock_rcu_node(rnp_root);
3022 needwake = rcu_start_gp(rsp);
3023 raw_spin_unlock_rcu_node(rnp_root);
3025 rcu_gp_kthread_wake(rsp);
3027 /* Give the grace period a kick. */
3028 rdp->blimit = LONG_MAX;
3029 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3030 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
3031 force_quiescent_state(rsp);
3032 rdp->n_force_qs_snap = rsp->n_force_qs;
3033 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
3039 * RCU callback function to leak a callback.
3041 static void rcu_leak_callback(struct rcu_head *rhp)
3046 * Helper function for call_rcu() and friends. The cpu argument will
3047 * normally be -1, indicating "currently running CPU". It may specify
3048 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3049 * is expected to specify a CPU.
3052 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3053 struct rcu_state *rsp, int cpu, bool lazy)
3055 unsigned long flags;
3056 struct rcu_data *rdp;
3058 /* Misaligned rcu_head! */
3059 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3061 if (debug_rcu_head_queue(head)) {
3063 * Probable double call_rcu(), so leak the callback.
3064 * Use rcu:rcu_callback trace event to find the previous
3065 * time callback was passed to __call_rcu().
3067 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
3069 WRITE_ONCE(head->func, rcu_leak_callback);
3074 local_irq_save(flags);
3075 rdp = this_cpu_ptr(rsp->rda);
3077 /* Add the callback to our list. */
3078 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
3082 rdp = per_cpu_ptr(rsp->rda, cpu);
3083 if (likely(rdp->mynode)) {
3084 /* Post-boot, so this should be for a no-CBs CPU. */
3085 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3086 WARN_ON_ONCE(offline);
3087 /* Offline CPU, _call_rcu() illegal, leak callback. */
3088 local_irq_restore(flags);
3092 * Very early boot, before rcu_init(). Initialize if needed
3093 * and then drop through to queue the callback.
3096 WARN_ON_ONCE(!rcu_is_watching());
3097 if (rcu_segcblist_empty(&rdp->cblist))
3098 rcu_segcblist_init(&rdp->cblist);
3100 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3102 rcu_idle_count_callbacks_posted();
3104 if (__is_kfree_rcu_offset((unsigned long)func))
3105 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3106 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3107 rcu_segcblist_n_cbs(&rdp->cblist));
3109 trace_rcu_callback(rsp->name, head,
3110 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3111 rcu_segcblist_n_cbs(&rdp->cblist));
3113 /* Go handle any RCU core processing required. */
3114 __call_rcu_core(rsp, rdp, head, flags);
3115 local_irq_restore(flags);
3119 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3120 * @head: structure to be used for queueing the RCU updates.
3121 * @func: actual callback function to be invoked after the grace period
3123 * The callback function will be invoked some time after a full grace
3124 * period elapses, in other words after all currently executing RCU
3125 * read-side critical sections have completed. call_rcu_sched() assumes
3126 * that the read-side critical sections end on enabling of preemption
3127 * or on voluntary preemption.
3128 * RCU read-side critical sections are delimited by:
3130 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3131 * - anything that disables preemption.
3133 * These may be nested.
3135 * See the description of call_rcu() for more detailed information on
3136 * memory ordering guarantees.
3138 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3140 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3142 EXPORT_SYMBOL_GPL(call_rcu_sched);
3145 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3146 * @head: structure to be used for queueing the RCU updates.
3147 * @func: actual callback function to be invoked after the grace period
3149 * The callback function will be invoked some time after a full grace
3150 * period elapses, in other words after all currently executing RCU
3151 * read-side critical sections have completed. call_rcu_bh() assumes
3152 * that the read-side critical sections end on completion of a softirq
3153 * handler. This means that read-side critical sections in process
3154 * context must not be interrupted by softirqs. This interface is to be
3155 * used when most of the read-side critical sections are in softirq context.
3156 * RCU read-side critical sections are delimited by:
3158 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3159 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3161 * These may be nested.
3163 * See the description of call_rcu() for more detailed information on
3164 * memory ordering guarantees.
3166 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3168 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3170 EXPORT_SYMBOL_GPL(call_rcu_bh);
3173 * Queue an RCU callback for lazy invocation after a grace period.
3174 * This will likely be later named something like "call_rcu_lazy()",
3175 * but this change will require some way of tagging the lazy RCU
3176 * callbacks in the list of pending callbacks. Until then, this
3177 * function may only be called from __kfree_rcu().
3179 void kfree_call_rcu(struct rcu_head *head,
3180 rcu_callback_t func)
3182 __call_rcu(head, func, rcu_state_p, -1, 1);
3184 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3187 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3188 * any blocking grace-period wait automatically implies a grace period
3189 * if there is only one CPU online at any point time during execution
3190 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3191 * occasionally incorrectly indicate that there are multiple CPUs online
3192 * when there was in fact only one the whole time, as this just adds
3193 * some overhead: RCU still operates correctly.
3195 static inline int rcu_blocking_is_gp(void)
3199 might_sleep(); /* Check for RCU read-side critical section. */
3201 ret = num_online_cpus() <= 1;
3207 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3209 * Control will return to the caller some time after a full rcu-sched
3210 * grace period has elapsed, in other words after all currently executing
3211 * rcu-sched read-side critical sections have completed. These read-side
3212 * critical sections are delimited by rcu_read_lock_sched() and
3213 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3214 * local_irq_disable(), and so on may be used in place of
3215 * rcu_read_lock_sched().
3217 * This means that all preempt_disable code sequences, including NMI and
3218 * non-threaded hardware-interrupt handlers, in progress on entry will
3219 * have completed before this primitive returns. However, this does not
3220 * guarantee that softirq handlers will have completed, since in some
3221 * kernels, these handlers can run in process context, and can block.
3223 * Note that this guarantee implies further memory-ordering guarantees.
3224 * On systems with more than one CPU, when synchronize_sched() returns,
3225 * each CPU is guaranteed to have executed a full memory barrier since the
3226 * end of its last RCU-sched read-side critical section whose beginning
3227 * preceded the call to synchronize_sched(). In addition, each CPU having
3228 * an RCU read-side critical section that extends beyond the return from
3229 * synchronize_sched() is guaranteed to have executed a full memory barrier
3230 * after the beginning of synchronize_sched() and before the beginning of
3231 * that RCU read-side critical section. Note that these guarantees include
3232 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3233 * that are executing in the kernel.
3235 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3236 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3237 * to have executed a full memory barrier during the execution of
3238 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3239 * again only if the system has more than one CPU).
3241 void synchronize_sched(void)
3243 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3244 lock_is_held(&rcu_lock_map) ||
3245 lock_is_held(&rcu_sched_lock_map),
3246 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3247 if (rcu_blocking_is_gp())
3249 if (rcu_gp_is_expedited())
3250 synchronize_sched_expedited();
3252 wait_rcu_gp(call_rcu_sched);
3254 EXPORT_SYMBOL_GPL(synchronize_sched);
3257 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3259 * Control will return to the caller some time after a full rcu_bh grace
3260 * period has elapsed, in other words after all currently executing rcu_bh
3261 * read-side critical sections have completed. RCU read-side critical
3262 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3263 * and may be nested.
3265 * See the description of synchronize_sched() for more detailed information
3266 * on memory ordering guarantees.
3268 void synchronize_rcu_bh(void)
3270 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3271 lock_is_held(&rcu_lock_map) ||
3272 lock_is_held(&rcu_sched_lock_map),
3273 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3274 if (rcu_blocking_is_gp())
3276 if (rcu_gp_is_expedited())
3277 synchronize_rcu_bh_expedited();
3279 wait_rcu_gp(call_rcu_bh);
3281 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3284 * get_state_synchronize_rcu - Snapshot current RCU state
3286 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3287 * to determine whether or not a full grace period has elapsed in the
3290 unsigned long get_state_synchronize_rcu(void)
3293 * Any prior manipulation of RCU-protected data must happen
3294 * before the load from ->gpnum.
3299 * Make sure this load happens before the purportedly
3300 * time-consuming work between get_state_synchronize_rcu()
3301 * and cond_synchronize_rcu().
3303 return smp_load_acquire(&rcu_state_p->gpnum);
3305 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3308 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3310 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3312 * If a full RCU grace period has elapsed since the earlier call to
3313 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3314 * synchronize_rcu() to wait for a full grace period.
3316 * Yes, this function does not take counter wrap into account. But
3317 * counter wrap is harmless. If the counter wraps, we have waited for
3318 * more than 2 billion grace periods (and way more on a 64-bit system!),
3319 * so waiting for one additional grace period should be just fine.
3321 void cond_synchronize_rcu(unsigned long oldstate)
3323 unsigned long newstate;
3326 * Ensure that this load happens before any RCU-destructive
3327 * actions the caller might carry out after we return.
3329 newstate = smp_load_acquire(&rcu_state_p->completed);
3330 if (ULONG_CMP_GE(oldstate, newstate))
3333 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3336 * get_state_synchronize_sched - Snapshot current RCU-sched state
3338 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3339 * to determine whether or not a full grace period has elapsed in the
3342 unsigned long get_state_synchronize_sched(void)
3345 * Any prior manipulation of RCU-protected data must happen
3346 * before the load from ->gpnum.
3351 * Make sure this load happens before the purportedly
3352 * time-consuming work between get_state_synchronize_sched()
3353 * and cond_synchronize_sched().
3355 return smp_load_acquire(&rcu_sched_state.gpnum);
3357 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3360 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3362 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3364 * If a full RCU-sched grace period has elapsed since the earlier call to
3365 * get_state_synchronize_sched(), just return. Otherwise, invoke
3366 * synchronize_sched() to wait for a full grace period.
3368 * Yes, this function does not take counter wrap into account. But
3369 * counter wrap is harmless. If the counter wraps, we have waited for
3370 * more than 2 billion grace periods (and way more on a 64-bit system!),
3371 * so waiting for one additional grace period should be just fine.
3373 void cond_synchronize_sched(unsigned long oldstate)
3375 unsigned long newstate;
3378 * Ensure that this load happens before any RCU-destructive
3379 * actions the caller might carry out after we return.
3381 newstate = smp_load_acquire(&rcu_sched_state.completed);
3382 if (ULONG_CMP_GE(oldstate, newstate))
3383 synchronize_sched();
3385 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3388 * Check to see if there is any immediate RCU-related work to be done
3389 * by the current CPU, for the specified type of RCU, returning 1 if so.
3390 * The checks are in order of increasing expense: checks that can be
3391 * carried out against CPU-local state are performed first. However,
3392 * we must check for CPU stalls first, else we might not get a chance.
3394 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3396 struct rcu_node *rnp = rdp->mynode;
3398 rdp->n_rcu_pending++;
3400 /* Check for CPU stalls, if enabled. */
3401 check_cpu_stall(rsp, rdp);
3403 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3404 if (rcu_nohz_full_cpu(rsp))
3407 /* Is the RCU core waiting for a quiescent state from this CPU? */
3408 if (rcu_scheduler_fully_active &&
3409 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3410 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
3411 rdp->n_rp_core_needs_qs++;
3412 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3413 rdp->n_rp_report_qs++;
3417 /* Does this CPU have callbacks ready to invoke? */
3418 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
3419 rdp->n_rp_cb_ready++;
3423 /* Has RCU gone idle with this CPU needing another grace period? */
3424 if (cpu_needs_another_gp(rsp, rdp)) {
3425 rdp->n_rp_cpu_needs_gp++;
3429 /* Has another RCU grace period completed? */
3430 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3431 rdp->n_rp_gp_completed++;
3435 /* Has a new RCU grace period started? */
3436 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3437 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3438 rdp->n_rp_gp_started++;
3442 /* Does this CPU need a deferred NOCB wakeup? */
3443 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3444 rdp->n_rp_nocb_defer_wakeup++;
3449 rdp->n_rp_need_nothing++;
3454 * Check to see if there is any immediate RCU-related work to be done
3455 * by the current CPU, returning 1 if so. This function is part of the
3456 * RCU implementation; it is -not- an exported member of the RCU API.
3458 static int rcu_pending(void)
3460 struct rcu_state *rsp;
3462 for_each_rcu_flavor(rsp)
3463 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3469 * Return true if the specified CPU has any callback. If all_lazy is
3470 * non-NULL, store an indication of whether all callbacks are lazy.
3471 * (If there are no callbacks, all of them are deemed to be lazy.)
3473 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3477 struct rcu_data *rdp;
3478 struct rcu_state *rsp;
3480 for_each_rcu_flavor(rsp) {
3481 rdp = this_cpu_ptr(rsp->rda);
3482 if (rcu_segcblist_empty(&rdp->cblist))
3485 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3496 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3497 * the compiler is expected to optimize this away.
3499 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3500 int cpu, unsigned long done)
3502 trace_rcu_barrier(rsp->name, s, cpu,
3503 atomic_read(&rsp->barrier_cpu_count), done);
3507 * RCU callback function for _rcu_barrier(). If we are last, wake
3508 * up the task executing _rcu_barrier().
3510 static void rcu_barrier_callback(struct rcu_head *rhp)
3512 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3513 struct rcu_state *rsp = rdp->rsp;
3515 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3516 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3517 rsp->barrier_sequence);
3518 complete(&rsp->barrier_completion);
3520 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3525 * Called with preemption disabled, and from cross-cpu IRQ context.
3527 static void rcu_barrier_func(void *type)
3529 struct rcu_state *rsp = type;
3530 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3532 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3533 rdp->barrier_head.func = rcu_barrier_callback;
3534 debug_rcu_head_queue(&rdp->barrier_head);
3535 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3536 atomic_inc(&rsp->barrier_cpu_count);
3538 debug_rcu_head_unqueue(&rdp->barrier_head);
3539 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3540 rsp->barrier_sequence);
3545 * Orchestrate the specified type of RCU barrier, waiting for all
3546 * RCU callbacks of the specified type to complete.
3548 static void _rcu_barrier(struct rcu_state *rsp)
3551 struct rcu_data *rdp;
3552 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3554 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3556 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3557 mutex_lock(&rsp->barrier_mutex);
3559 /* Did someone else do our work for us? */
3560 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3561 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3562 rsp->barrier_sequence);
3563 smp_mb(); /* caller's subsequent code after above check. */
3564 mutex_unlock(&rsp->barrier_mutex);
3568 /* Mark the start of the barrier operation. */
3569 rcu_seq_start(&rsp->barrier_sequence);
3570 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3573 * Initialize the count to one rather than to zero in order to
3574 * avoid a too-soon return to zero in case of a short grace period
3575 * (or preemption of this task). Exclude CPU-hotplug operations
3576 * to ensure that no offline CPU has callbacks queued.
3578 init_completion(&rsp->barrier_completion);
3579 atomic_set(&rsp->barrier_cpu_count, 1);
3583 * Force each CPU with callbacks to register a new callback.
3584 * When that callback is invoked, we will know that all of the
3585 * corresponding CPU's preceding callbacks have been invoked.
3587 for_each_possible_cpu(cpu) {
3588 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3590 rdp = per_cpu_ptr(rsp->rda, cpu);
3591 if (rcu_is_nocb_cpu(cpu)) {
3592 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3593 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3594 rsp->barrier_sequence);
3596 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3597 rsp->barrier_sequence);
3598 smp_mb__before_atomic();
3599 atomic_inc(&rsp->barrier_cpu_count);
3600 __call_rcu(&rdp->barrier_head,
3601 rcu_barrier_callback, rsp, cpu, 0);
3603 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3604 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3605 rsp->barrier_sequence);
3606 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3608 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3609 rsp->barrier_sequence);
3615 * Now that we have an rcu_barrier_callback() callback on each
3616 * CPU, and thus each counted, remove the initial count.
3618 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3619 complete(&rsp->barrier_completion);
3621 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3622 wait_for_completion(&rsp->barrier_completion);
3624 /* Mark the end of the barrier operation. */
3625 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3626 rcu_seq_end(&rsp->barrier_sequence);
3628 /* Other rcu_barrier() invocations can now safely proceed. */
3629 mutex_unlock(&rsp->barrier_mutex);
3633 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3635 void rcu_barrier_bh(void)
3637 _rcu_barrier(&rcu_bh_state);
3639 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3642 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3644 void rcu_barrier_sched(void)
3646 _rcu_barrier(&rcu_sched_state);
3648 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3651 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3652 * first CPU in a given leaf rcu_node structure coming online. The caller
3653 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3656 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3659 struct rcu_node *rnp = rnp_leaf;
3661 lockdep_assert_held(&rnp->lock);
3663 mask = rnp->grpmask;
3667 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3668 rnp->qsmaskinit |= mask;
3669 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3674 * Do boot-time initialization of a CPU's per-CPU RCU data.
3677 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3679 unsigned long flags;
3680 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3681 struct rcu_node *rnp = rcu_get_root(rsp);
3683 /* Set up local state, ensuring consistent view of global state. */
3684 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3685 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3686 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3687 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3688 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3691 rcu_boot_init_nocb_percpu_data(rdp);
3692 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3696 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3697 * offline event can be happening at a given time. Note also that we
3698 * can accept some slop in the rsp->completed access due to the fact
3699 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3702 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3704 unsigned long flags;
3705 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3706 struct rcu_node *rnp = rcu_get_root(rsp);
3708 /* Set up local state, ensuring consistent view of global state. */
3709 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3710 rdp->qlen_last_fqs_check = 0;
3711 rdp->n_force_qs_snap = rsp->n_force_qs;
3712 rdp->blimit = blimit;
3713 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3714 !init_nocb_callback_list(rdp))
3715 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3716 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3717 rcu_dynticks_eqs_online();
3718 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3721 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3722 * propagation up the rcu_node tree will happen at the beginning
3723 * of the next grace period.
3726 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3727 rdp->beenonline = true; /* We have now been online. */
3728 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3729 rdp->completed = rnp->completed;
3730 rdp->cpu_no_qs.b.norm = true;
3731 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3732 rdp->core_needs_qs = false;
3733 rdp->rcu_iw_pending = false;
3734 rdp->rcu_iw_gpnum = rnp->gpnum - 1;
3735 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3736 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3740 * Invoked early in the CPU-online process, when pretty much all
3741 * services are available. The incoming CPU is not present.
3743 int rcutree_prepare_cpu(unsigned int cpu)
3745 struct rcu_state *rsp;
3747 for_each_rcu_flavor(rsp)
3748 rcu_init_percpu_data(cpu, rsp);
3750 rcu_prepare_kthreads(cpu);
3751 rcu_spawn_all_nocb_kthreads(cpu);
3757 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3759 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3761 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3763 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3767 * Near the end of the CPU-online process. Pretty much all services
3768 * enabled, and the CPU is now very much alive.
3770 int rcutree_online_cpu(unsigned int cpu)
3772 unsigned long flags;
3773 struct rcu_data *rdp;
3774 struct rcu_node *rnp;
3775 struct rcu_state *rsp;
3777 for_each_rcu_flavor(rsp) {
3778 rdp = per_cpu_ptr(rsp->rda, cpu);
3780 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3781 rnp->ffmask |= rdp->grpmask;
3782 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3784 if (IS_ENABLED(CONFIG_TREE_SRCU))
3785 srcu_online_cpu(cpu);
3786 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3787 return 0; /* Too early in boot for scheduler work. */
3788 sync_sched_exp_online_cleanup(cpu);
3789 rcutree_affinity_setting(cpu, -1);
3794 * Near the beginning of the process. The CPU is still very much alive
3795 * with pretty much all services enabled.
3797 int rcutree_offline_cpu(unsigned int cpu)
3799 unsigned long flags;
3800 struct rcu_data *rdp;
3801 struct rcu_node *rnp;
3802 struct rcu_state *rsp;
3804 for_each_rcu_flavor(rsp) {
3805 rdp = per_cpu_ptr(rsp->rda, cpu);
3807 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3808 rnp->ffmask &= ~rdp->grpmask;
3809 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3812 rcutree_affinity_setting(cpu, cpu);
3813 if (IS_ENABLED(CONFIG_TREE_SRCU))
3814 srcu_offline_cpu(cpu);
3819 * Near the end of the offline process. We do only tracing here.
3821 int rcutree_dying_cpu(unsigned int cpu)
3823 struct rcu_state *rsp;
3825 for_each_rcu_flavor(rsp)
3826 rcu_cleanup_dying_cpu(rsp);
3831 * The outgoing CPU is gone and we are running elsewhere.
3833 int rcutree_dead_cpu(unsigned int cpu)
3835 struct rcu_state *rsp;
3837 for_each_rcu_flavor(rsp) {
3838 rcu_cleanup_dead_cpu(cpu, rsp);
3839 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3845 * Mark the specified CPU as being online so that subsequent grace periods
3846 * (both expedited and normal) will wait on it. Note that this means that
3847 * incoming CPUs are not allowed to use RCU read-side critical sections
3848 * until this function is called. Failing to observe this restriction
3849 * will result in lockdep splats.
3851 * Note that this function is special in that it is invoked directly
3852 * from the incoming CPU rather than from the cpuhp_step mechanism.
3853 * This is because this function must be invoked at a precise location.
3855 void rcu_cpu_starting(unsigned int cpu)
3857 unsigned long flags;
3860 unsigned long oldmask;
3861 struct rcu_data *rdp;
3862 struct rcu_node *rnp;
3863 struct rcu_state *rsp;
3865 for_each_rcu_flavor(rsp) {
3866 rdp = per_cpu_ptr(rsp->rda, cpu);
3868 mask = rdp->grpmask;
3869 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3870 rnp->qsmaskinitnext |= mask;
3871 oldmask = rnp->expmaskinitnext;
3872 rnp->expmaskinitnext |= mask;
3873 oldmask ^= rnp->expmaskinitnext;
3874 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3875 /* Allow lockless access for expedited grace periods. */
3876 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3877 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3879 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3882 #ifdef CONFIG_HOTPLUG_CPU
3884 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3885 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3888 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3890 unsigned long flags;
3892 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3893 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3895 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3896 mask = rdp->grpmask;
3897 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3898 rnp->qsmaskinitnext &= ~mask;
3899 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3903 * The outgoing function has no further need of RCU, so remove it from
3904 * the list of CPUs that RCU must track.
3906 * Note that this function is special in that it is invoked directly
3907 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3908 * This is because this function must be invoked at a precise location.
3910 void rcu_report_dead(unsigned int cpu)
3912 struct rcu_state *rsp;
3914 /* QS for any half-done expedited RCU-sched GP. */
3916 rcu_report_exp_rdp(&rcu_sched_state,
3917 this_cpu_ptr(rcu_sched_state.rda), true);
3919 for_each_rcu_flavor(rsp)
3920 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3923 /* Migrate the dead CPU's callbacks to the current CPU. */
3924 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3926 unsigned long flags;
3927 struct rcu_data *my_rdp;
3928 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3929 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3931 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3932 return; /* No callbacks to migrate. */
3934 local_irq_save(flags);
3935 my_rdp = this_cpu_ptr(rsp->rda);
3936 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3937 local_irq_restore(flags);
3940 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3941 rcu_advance_cbs(rsp, rnp_root, rdp); /* Leverage recent GPs. */
3942 rcu_advance_cbs(rsp, rnp_root, my_rdp); /* Assign GP to pending CBs. */
3943 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3944 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3945 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3946 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3947 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3948 !rcu_segcblist_empty(&rdp->cblist),
3949 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3950 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3951 rcu_segcblist_first_cb(&rdp->cblist));
3955 * The outgoing CPU has just passed through the dying-idle state,
3956 * and we are being invoked from the CPU that was IPIed to continue the
3957 * offline operation. We need to migrate the outgoing CPU's callbacks.
3959 void rcutree_migrate_callbacks(int cpu)
3961 struct rcu_state *rsp;
3963 for_each_rcu_flavor(rsp)
3964 rcu_migrate_callbacks(cpu, rsp);
3969 * On non-huge systems, use expedited RCU grace periods to make suspend
3970 * and hibernation run faster.
3972 static int rcu_pm_notify(struct notifier_block *self,
3973 unsigned long action, void *hcpu)
3976 case PM_HIBERNATION_PREPARE:
3977 case PM_SUSPEND_PREPARE:
3978 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3981 case PM_POST_HIBERNATION:
3982 case PM_POST_SUSPEND:
3983 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3984 rcu_unexpedite_gp();
3993 * Spawn the kthreads that handle each RCU flavor's grace periods.
3995 static int __init rcu_spawn_gp_kthread(void)
3997 unsigned long flags;
3998 int kthread_prio_in = kthread_prio;
3999 struct rcu_node *rnp;
4000 struct rcu_state *rsp;
4001 struct sched_param sp;
4002 struct task_struct *t;
4004 /* Force priority into range. */
4005 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4007 else if (kthread_prio < 0)
4009 else if (kthread_prio > 99)
4011 if (kthread_prio != kthread_prio_in)
4012 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4013 kthread_prio, kthread_prio_in);
4015 rcu_scheduler_fully_active = 1;
4016 for_each_rcu_flavor(rsp) {
4017 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4019 rnp = rcu_get_root(rsp);
4020 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4021 rsp->gp_kthread = t;
4023 sp.sched_priority = kthread_prio;
4024 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4026 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4029 rcu_spawn_nocb_kthreads();
4030 rcu_spawn_boost_kthreads();
4033 early_initcall(rcu_spawn_gp_kthread);
4036 * This function is invoked towards the end of the scheduler's
4037 * initialization process. Before this is called, the idle task might
4038 * contain synchronous grace-period primitives (during which time, this idle
4039 * task is booting the system, and such primitives are no-ops). After this
4040 * function is called, any synchronous grace-period primitives are run as
4041 * expedited, with the requesting task driving the grace period forward.
4042 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4043 * runtime RCU functionality.
4045 void rcu_scheduler_starting(void)
4047 WARN_ON(num_online_cpus() != 1);
4048 WARN_ON(nr_context_switches() > 0);
4049 rcu_test_sync_prims();
4050 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4051 rcu_test_sync_prims();
4055 * Helper function for rcu_init() that initializes one rcu_state structure.
4057 static void __init rcu_init_one(struct rcu_state *rsp)
4059 static const char * const buf[] = RCU_NODE_NAME_INIT;
4060 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4061 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4062 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4064 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4068 struct rcu_node *rnp;
4070 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4072 /* Silence gcc 4.8 false positive about array index out of range. */
4073 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4074 panic("rcu_init_one: rcu_num_lvls out of range");
4076 /* Initialize the level-tracking arrays. */
4078 for (i = 1; i < rcu_num_lvls; i++)
4079 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
4080 rcu_init_levelspread(levelspread, num_rcu_lvl);
4082 /* Initialize the elements themselves, starting from the leaves. */
4084 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4085 cpustride *= levelspread[i];
4086 rnp = rsp->level[i];
4087 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4088 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4089 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4090 &rcu_node_class[i], buf[i]);
4091 raw_spin_lock_init(&rnp->fqslock);
4092 lockdep_set_class_and_name(&rnp->fqslock,
4093 &rcu_fqs_class[i], fqs[i]);
4094 rnp->gpnum = rsp->gpnum;
4095 rnp->completed = rsp->completed;
4097 rnp->qsmaskinit = 0;
4098 rnp->grplo = j * cpustride;
4099 rnp->grphi = (j + 1) * cpustride - 1;
4100 if (rnp->grphi >= nr_cpu_ids)
4101 rnp->grphi = nr_cpu_ids - 1;
4107 rnp->grpnum = j % levelspread[i - 1];
4108 rnp->grpmask = 1UL << rnp->grpnum;
4109 rnp->parent = rsp->level[i - 1] +
4110 j / levelspread[i - 1];
4113 INIT_LIST_HEAD(&rnp->blkd_tasks);
4114 rcu_init_one_nocb(rnp);
4115 init_waitqueue_head(&rnp->exp_wq[0]);
4116 init_waitqueue_head(&rnp->exp_wq[1]);
4117 init_waitqueue_head(&rnp->exp_wq[2]);
4118 init_waitqueue_head(&rnp->exp_wq[3]);
4119 spin_lock_init(&rnp->exp_lock);
4123 init_swait_queue_head(&rsp->gp_wq);
4124 init_swait_queue_head(&rsp->expedited_wq);
4125 rnp = rsp->level[rcu_num_lvls - 1];
4126 for_each_possible_cpu(i) {
4127 while (i > rnp->grphi)
4129 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4130 rcu_boot_init_percpu_data(i, rsp);
4132 list_add(&rsp->flavors, &rcu_struct_flavors);
4136 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4137 * replace the definitions in tree.h because those are needed to size
4138 * the ->node array in the rcu_state structure.
4140 static void __init rcu_init_geometry(void)
4144 int rcu_capacity[RCU_NUM_LVLS];
4147 * Initialize any unspecified boot parameters.
4148 * The default values of jiffies_till_first_fqs and
4149 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4150 * value, which is a function of HZ, then adding one for each
4151 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4153 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4154 if (jiffies_till_first_fqs == ULONG_MAX)
4155 jiffies_till_first_fqs = d;
4156 if (jiffies_till_next_fqs == ULONG_MAX)
4157 jiffies_till_next_fqs = d;
4159 /* If the compile-time values are accurate, just leave. */
4160 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4161 nr_cpu_ids == NR_CPUS)
4163 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4164 rcu_fanout_leaf, nr_cpu_ids);
4167 * The boot-time rcu_fanout_leaf parameter must be at least two
4168 * and cannot exceed the number of bits in the rcu_node masks.
4169 * Complain and fall back to the compile-time values if this
4170 * limit is exceeded.
4172 if (rcu_fanout_leaf < 2 ||
4173 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4174 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4180 * Compute number of nodes that can be handled an rcu_node tree
4181 * with the given number of levels.
4183 rcu_capacity[0] = rcu_fanout_leaf;
4184 for (i = 1; i < RCU_NUM_LVLS; i++)
4185 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4188 * The tree must be able to accommodate the configured number of CPUs.
4189 * If this limit is exceeded, fall back to the compile-time values.
4191 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4192 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4197 /* Calculate the number of levels in the tree. */
4198 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4200 rcu_num_lvls = i + 1;
4202 /* Calculate the number of rcu_nodes at each level of the tree. */
4203 for (i = 0; i < rcu_num_lvls; i++) {
4204 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4205 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4208 /* Calculate the total number of rcu_node structures. */
4210 for (i = 0; i < rcu_num_lvls; i++)
4211 rcu_num_nodes += num_rcu_lvl[i];
4215 * Dump out the structure of the rcu_node combining tree associated
4216 * with the rcu_state structure referenced by rsp.
4218 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4221 struct rcu_node *rnp;
4223 pr_info("rcu_node tree layout dump\n");
4225 rcu_for_each_node_breadth_first(rsp, rnp) {
4226 if (rnp->level != level) {
4231 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4236 void __init rcu_init(void)
4240 rcu_early_boot_tests();
4242 rcu_bootup_announce();
4243 rcu_init_geometry();
4244 rcu_init_one(&rcu_bh_state);
4245 rcu_init_one(&rcu_sched_state);
4247 rcu_dump_rcu_node_tree(&rcu_sched_state);
4248 __rcu_init_preempt();
4249 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4252 * We don't need protection against CPU-hotplug here because
4253 * this is called early in boot, before either interrupts
4254 * or the scheduler are operational.
4256 pm_notifier(rcu_pm_notify, 0);
4257 for_each_online_cpu(cpu) {
4258 rcutree_prepare_cpu(cpu);
4259 rcu_cpu_starting(cpu);
4260 rcutree_online_cpu(cpu);
4264 #include "tree_exp.h"
4265 #include "tree_plugin.h"