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>
64 #ifdef MODULE_PARAM_PREFIX
65 #undef MODULE_PARAM_PREFIX
67 #define MODULE_PARAM_PREFIX "rcutree."
69 /* Data structures. */
72 * In order to export the rcu_state name to the tracing tools, it
73 * needs to be added in the __tracepoint_string section.
74 * This requires defining a separate variable tp_<sname>_varname
75 * that points to the string being used, and this will allow
76 * the tracing userspace tools to be able to decipher the string
77 * address to the matching string.
80 # define DEFINE_RCU_TPS(sname) \
81 static char sname##_varname[] = #sname; \
82 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
83 # define RCU_STATE_NAME(sname) sname##_varname
85 # define DEFINE_RCU_TPS(sname)
86 # define RCU_STATE_NAME(sname) __stringify(sname)
89 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
90 DEFINE_RCU_TPS(sname) \
91 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
92 struct rcu_state sname##_state = { \
93 .level = { &sname##_state.node[0] }, \
94 .rda = &sname##_data, \
96 .gp_state = RCU_GP_IDLE, \
97 .gpnum = 0UL - 300UL, \
98 .completed = 0UL - 300UL, \
99 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
100 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
101 .orphan_donetail = &sname##_state.orphan_donelist, \
102 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
103 .name = RCU_STATE_NAME(sname), \
105 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
106 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
109 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
110 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
112 static struct rcu_state *const rcu_state_p;
113 LIST_HEAD(rcu_struct_flavors);
115 /* Dump rcu_node combining tree at boot to verify correct setup. */
116 static bool dump_tree;
117 module_param(dump_tree, bool, 0444);
118 /* Control rcu_node-tree auto-balancing at boot time. */
119 static bool rcu_fanout_exact;
120 module_param(rcu_fanout_exact, bool, 0444);
121 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
122 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
123 module_param(rcu_fanout_leaf, int, 0444);
124 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
125 /* Number of rcu_nodes at specified level. */
126 static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
127 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
128 /* panic() on RCU Stall sysctl. */
129 int sysctl_panic_on_rcu_stall __read_mostly;
132 * The rcu_scheduler_active variable is initialized to the value
133 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
134 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
135 * RCU can assume that there is but one task, allowing RCU to (for example)
136 * optimize synchronize_rcu() to a simple barrier(). When this variable
137 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
138 * to detect real grace periods. This variable is also used to suppress
139 * boot-time false positives from lockdep-RCU error checking. Finally, it
140 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
141 * is fully initialized, including all of its kthreads having been spawned.
143 int rcu_scheduler_active __read_mostly;
144 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
147 * The rcu_scheduler_fully_active variable transitions from zero to one
148 * during the early_initcall() processing, which is after the scheduler
149 * is capable of creating new tasks. So RCU processing (for example,
150 * creating tasks for RCU priority boosting) must be delayed until after
151 * rcu_scheduler_fully_active transitions from zero to one. We also
152 * currently delay invocation of any RCU callbacks until after this point.
154 * It might later prove better for people registering RCU callbacks during
155 * early boot to take responsibility for these callbacks, but one step at
158 static int rcu_scheduler_fully_active __read_mostly;
160 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
161 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
162 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
163 static void invoke_rcu_core(void);
164 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
165 static void rcu_report_exp_rdp(struct rcu_state *rsp,
166 struct rcu_data *rdp, bool wake);
167 static void sync_sched_exp_online_cleanup(int cpu);
169 /* rcuc/rcub kthread realtime priority */
170 #ifdef CONFIG_RCU_KTHREAD_PRIO
171 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
172 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
173 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
174 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
175 module_param(kthread_prio, int, 0644);
177 /* Delay in jiffies for grace-period initialization delays, debug only. */
179 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
180 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
181 module_param(gp_preinit_delay, int, 0644);
182 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
183 static const int gp_preinit_delay;
184 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
186 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
187 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
188 module_param(gp_init_delay, int, 0644);
189 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
190 static const int gp_init_delay;
191 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
193 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
194 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
195 module_param(gp_cleanup_delay, int, 0644);
196 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
197 static const int gp_cleanup_delay;
198 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
201 * Number of grace periods between delays, normalized by the duration of
202 * the delay. The longer the the delay, the more the grace periods between
203 * each delay. The reason for this normalization is that it means that,
204 * for non-zero delays, the overall slowdown of grace periods is constant
205 * regardless of the duration of the delay. This arrangement balances
206 * the need for long delays to increase some race probabilities with the
207 * need for fast grace periods to increase other race probabilities.
209 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
212 * Track the rcutorture test sequence number and the update version
213 * number within a given test. The rcutorture_testseq is incremented
214 * on every rcutorture module load and unload, so has an odd value
215 * when a test is running. The rcutorture_vernum is set to zero
216 * when rcutorture starts and is incremented on each rcutorture update.
217 * These variables enable correlating rcutorture output with the
218 * RCU tracing information.
220 unsigned long rcutorture_testseq;
221 unsigned long rcutorture_vernum;
224 * Compute the mask of online CPUs for the specified rcu_node structure.
225 * This will not be stable unless the rcu_node structure's ->lock is
226 * held, but the bit corresponding to the current CPU will be stable
229 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
231 return READ_ONCE(rnp->qsmaskinitnext);
235 * Return true if an RCU grace period is in progress. The READ_ONCE()s
236 * permit this function to be invoked without holding the root rcu_node
237 * structure's ->lock, but of course results can be subject to change.
239 static int rcu_gp_in_progress(struct rcu_state *rsp)
241 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
245 * Note a quiescent state. Because we do not need to know
246 * how many quiescent states passed, just if there was at least
247 * one since the start of the grace period, this just sets a flag.
248 * The caller must have disabled preemption.
250 void rcu_sched_qs(void)
252 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
254 trace_rcu_grace_period(TPS("rcu_sched"),
255 __this_cpu_read(rcu_sched_data.gpnum),
257 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
258 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
260 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
261 rcu_report_exp_rdp(&rcu_sched_state,
262 this_cpu_ptr(&rcu_sched_data), true);
267 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
268 trace_rcu_grace_period(TPS("rcu_bh"),
269 __this_cpu_read(rcu_bh_data.gpnum),
271 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
276 * Steal a bit from the bottom of ->dynticks for idle entry/exit
277 * control. Initially this is for TLB flushing.
279 #define RCU_DYNTICK_CTRL_MASK 0x1
280 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
281 #ifndef rcu_eqs_special_exit
282 #define rcu_eqs_special_exit() do { } while (0)
285 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
286 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
287 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
288 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
289 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
290 .dynticks_idle = ATOMIC_INIT(1),
291 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
295 * Record entry into an extended quiescent state. This is only to be
296 * called when not already in an extended quiescent state.
298 static void rcu_dynticks_eqs_enter(void)
300 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
304 * CPUs seeing atomic_add_return() must see prior RCU read-side
305 * critical sections, and we also must force ordering with the
308 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
309 /* Better be in an extended quiescent state! */
310 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
311 (seq & RCU_DYNTICK_CTRL_CTR));
312 /* Better not have special action (TLB flush) pending! */
313 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
314 (seq & RCU_DYNTICK_CTRL_MASK));
318 * Record exit from an extended quiescent state. This is only to be
319 * called from an extended quiescent state.
321 static void rcu_dynticks_eqs_exit(void)
323 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
327 * CPUs seeing atomic_add_return() must see prior idle sojourns,
328 * and we also must force ordering with the next RCU read-side
331 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
332 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
333 !(seq & RCU_DYNTICK_CTRL_CTR));
334 if (seq & RCU_DYNTICK_CTRL_MASK) {
335 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
336 smp_mb__after_atomic(); /* _exit after clearing mask. */
337 /* Prefer duplicate flushes to losing a flush. */
338 rcu_eqs_special_exit();
343 * Reset the current CPU's ->dynticks counter to indicate that the
344 * newly onlined CPU is no longer in an extended quiescent state.
345 * This will either leave the counter unchanged, or increment it
346 * to the next non-quiescent value.
348 * The non-atomic test/increment sequence works because the upper bits
349 * of the ->dynticks counter are manipulated only by the corresponding CPU,
350 * or when the corresponding CPU is offline.
352 static void rcu_dynticks_eqs_online(void)
354 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
356 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
358 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
362 * Is the current CPU in an extended quiescent state?
364 * No ordering, as we are sampling CPU-local information.
366 bool rcu_dynticks_curr_cpu_in_eqs(void)
368 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
370 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
374 * Snapshot the ->dynticks counter with full ordering so as to allow
375 * stable comparison of this counter with past and future snapshots.
377 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
379 int snap = atomic_add_return(0, &rdtp->dynticks);
381 return snap & ~RCU_DYNTICK_CTRL_MASK;
385 * Return true if the snapshot returned from rcu_dynticks_snap()
386 * indicates that RCU is in an extended quiescent state.
388 static bool rcu_dynticks_in_eqs(int snap)
390 return !(snap & RCU_DYNTICK_CTRL_CTR);
394 * Return true if the CPU corresponding to the specified rcu_dynticks
395 * structure has spent some time in an extended quiescent state since
396 * rcu_dynticks_snap() returned the specified snapshot.
398 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
400 return snap != rcu_dynticks_snap(rdtp);
404 * Do a double-increment of the ->dynticks counter to emulate a
405 * momentary idle-CPU quiescent state.
407 static void rcu_dynticks_momentary_idle(void)
409 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
410 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
413 /* It is illegal to call this from idle state. */
414 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
418 * Set the special (bottom) bit of the specified CPU so that it
419 * will take special action (such as flushing its TLB) on the
420 * next exit from an extended quiescent state. Returns true if
421 * the bit was successfully set, or false if the CPU was not in
422 * an extended quiescent state.
424 bool rcu_eqs_special_set(int cpu)
428 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
431 old = atomic_read(&rdtp->dynticks);
432 if (old & RCU_DYNTICK_CTRL_CTR)
434 new = old | RCU_DYNTICK_CTRL_MASK;
435 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
440 * Let the RCU core know that this CPU has gone through the scheduler,
441 * which is a quiescent state. This is called when the need for a
442 * quiescent state is urgent, so we burn an atomic operation and full
443 * memory barriers to let the RCU core know about it, regardless of what
444 * this CPU might (or might not) do in the near future.
446 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
448 * The caller must have disabled interrupts.
450 static void rcu_momentary_dyntick_idle(void)
452 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
453 rcu_dynticks_momentary_idle();
457 * Note a context switch. This is a quiescent state for RCU-sched,
458 * and requires special handling for preemptible RCU.
459 * The caller must have disabled interrupts.
461 void rcu_note_context_switch(void)
463 barrier(); /* Avoid RCU read-side critical sections leaking down. */
464 trace_rcu_utilization(TPS("Start context switch"));
466 rcu_preempt_note_context_switch();
467 /* Load rcu_urgent_qs before other flags. */
468 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
470 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
471 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
472 rcu_momentary_dyntick_idle();
473 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
475 trace_rcu_utilization(TPS("End context switch"));
476 barrier(); /* Avoid RCU read-side critical sections leaking up. */
478 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
481 * Register a quiescent state for all RCU flavors. If there is an
482 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
483 * dyntick-idle quiescent state visible to other CPUs (but only for those
484 * RCU flavors in desperate need of a quiescent state, which will normally
485 * be none of them). Either way, do a lightweight quiescent state for
488 * The barrier() calls are redundant in the common case when this is
489 * called externally, but just in case this is called from within this
493 void rcu_all_qs(void)
497 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
500 /* Load rcu_urgent_qs before other flags. */
501 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
505 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
506 barrier(); /* Avoid RCU read-side critical sections leaking down. */
507 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
508 local_irq_save(flags);
509 rcu_momentary_dyntick_idle();
510 local_irq_restore(flags);
512 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
514 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
515 barrier(); /* Avoid RCU read-side critical sections leaking up. */
518 EXPORT_SYMBOL_GPL(rcu_all_qs);
520 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
521 static long qhimark = 10000; /* If this many pending, ignore blimit. */
522 static long qlowmark = 100; /* Once only this many pending, use blimit. */
524 module_param(blimit, long, 0444);
525 module_param(qhimark, long, 0444);
526 module_param(qlowmark, long, 0444);
528 static ulong jiffies_till_first_fqs = ULONG_MAX;
529 static ulong jiffies_till_next_fqs = ULONG_MAX;
530 static bool rcu_kick_kthreads;
532 module_param(jiffies_till_first_fqs, ulong, 0644);
533 module_param(jiffies_till_next_fqs, ulong, 0644);
534 module_param(rcu_kick_kthreads, bool, 0644);
537 * How long the grace period must be before we start recruiting
538 * quiescent-state help from rcu_note_context_switch().
540 static ulong jiffies_till_sched_qs = HZ / 20;
541 module_param(jiffies_till_sched_qs, ulong, 0644);
543 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
544 struct rcu_data *rdp);
545 static void force_qs_rnp(struct rcu_state *rsp,
546 int (*f)(struct rcu_data *rsp, bool *isidle,
547 unsigned long *maxj),
548 bool *isidle, unsigned long *maxj);
549 static void force_quiescent_state(struct rcu_state *rsp);
550 static int rcu_pending(void);
553 * Return the number of RCU batches started thus far for debug & stats.
555 unsigned long rcu_batches_started(void)
557 return rcu_state_p->gpnum;
559 EXPORT_SYMBOL_GPL(rcu_batches_started);
562 * Return the number of RCU-sched batches started thus far for debug & stats.
564 unsigned long rcu_batches_started_sched(void)
566 return rcu_sched_state.gpnum;
568 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
571 * Return the number of RCU BH batches started thus far for debug & stats.
573 unsigned long rcu_batches_started_bh(void)
575 return rcu_bh_state.gpnum;
577 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
580 * Return the number of RCU batches completed thus far for debug & stats.
582 unsigned long rcu_batches_completed(void)
584 return rcu_state_p->completed;
586 EXPORT_SYMBOL_GPL(rcu_batches_completed);
589 * Return the number of RCU-sched batches completed thus far for debug & stats.
591 unsigned long rcu_batches_completed_sched(void)
593 return rcu_sched_state.completed;
595 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
598 * Return the number of RCU BH batches completed thus far for debug & stats.
600 unsigned long rcu_batches_completed_bh(void)
602 return rcu_bh_state.completed;
604 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
607 * Return the number of RCU expedited batches completed thus far for
608 * debug & stats. Odd numbers mean that a batch is in progress, even
609 * numbers mean idle. The value returned will thus be roughly double
610 * the cumulative batches since boot.
612 unsigned long rcu_exp_batches_completed(void)
614 return rcu_state_p->expedited_sequence;
616 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
619 * Return the number of RCU-sched expedited batches completed thus far
620 * for debug & stats. Similar to rcu_exp_batches_completed().
622 unsigned long rcu_exp_batches_completed_sched(void)
624 return rcu_sched_state.expedited_sequence;
626 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
629 * Force a quiescent state.
631 void rcu_force_quiescent_state(void)
633 force_quiescent_state(rcu_state_p);
635 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
638 * Force a quiescent state for RCU BH.
640 void rcu_bh_force_quiescent_state(void)
642 force_quiescent_state(&rcu_bh_state);
644 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
647 * Force a quiescent state for RCU-sched.
649 void rcu_sched_force_quiescent_state(void)
651 force_quiescent_state(&rcu_sched_state);
653 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
656 * Show the state of the grace-period kthreads.
658 void show_rcu_gp_kthreads(void)
660 struct rcu_state *rsp;
662 for_each_rcu_flavor(rsp) {
663 pr_info("%s: wait state: %d ->state: %#lx\n",
664 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
665 /* sched_show_task(rsp->gp_kthread); */
668 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
671 * Record the number of times rcutorture tests have been initiated and
672 * terminated. This information allows the debugfs tracing stats to be
673 * correlated to the rcutorture messages, even when the rcutorture module
674 * is being repeatedly loaded and unloaded. In other words, we cannot
675 * store this state in rcutorture itself.
677 void rcutorture_record_test_transition(void)
679 rcutorture_testseq++;
680 rcutorture_vernum = 0;
682 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
685 * Send along grace-period-related data for rcutorture diagnostics.
687 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
688 unsigned long *gpnum, unsigned long *completed)
690 struct rcu_state *rsp = NULL;
699 case RCU_SCHED_FLAVOR:
700 rsp = &rcu_sched_state;
706 *flags = READ_ONCE(rsp->gp_flags);
707 *gpnum = READ_ONCE(rsp->gpnum);
708 *completed = READ_ONCE(rsp->completed);
715 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
718 * Record the number of writer passes through the current rcutorture test.
719 * This is also used to correlate debugfs tracing stats with the rcutorture
722 void rcutorture_record_progress(unsigned long vernum)
726 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
729 * Does the CPU have callbacks ready to be invoked?
732 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
734 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
735 rdp->nxttail[RCU_NEXT_TAIL] != NULL;
739 * Return the root node of the specified rcu_state structure.
741 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
743 return &rsp->node[0];
747 * Is there any need for future grace periods?
748 * Interrupts must be disabled. If the caller does not hold the root
749 * rnp_node structure's ->lock, the results are advisory only.
751 static int rcu_future_needs_gp(struct rcu_state *rsp)
753 struct rcu_node *rnp = rcu_get_root(rsp);
754 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
755 int *fp = &rnp->need_future_gp[idx];
757 return READ_ONCE(*fp);
761 * Does the current CPU require a not-yet-started grace period?
762 * The caller must have disabled interrupts to prevent races with
763 * normal callback registry.
766 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
770 if (rcu_gp_in_progress(rsp))
771 return false; /* No, a grace period is already in progress. */
772 if (rcu_future_needs_gp(rsp))
773 return true; /* Yes, a no-CBs CPU needs one. */
774 if (!rdp->nxttail[RCU_NEXT_TAIL])
775 return false; /* No, this is a no-CBs (or offline) CPU. */
776 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
777 return true; /* Yes, CPU has newly registered callbacks. */
778 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
779 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
780 ULONG_CMP_LT(READ_ONCE(rsp->completed),
781 rdp->nxtcompleted[i]))
782 return true; /* Yes, CBs for future grace period. */
783 return false; /* No grace period needed. */
787 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
789 * If the new value of the ->dynticks_nesting counter now is zero,
790 * we really have entered idle, and must do the appropriate accounting.
791 * The caller must have disabled interrupts.
793 static void rcu_eqs_enter_common(long long oldval, bool user)
795 struct rcu_state *rsp;
796 struct rcu_data *rdp;
797 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
799 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
800 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
801 !user && !is_idle_task(current)) {
802 struct task_struct *idle __maybe_unused =
803 idle_task(smp_processor_id());
805 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
806 rcu_ftrace_dump(DUMP_ORIG);
807 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
808 current->pid, current->comm,
809 idle->pid, idle->comm); /* must be idle task! */
811 for_each_rcu_flavor(rsp) {
812 rdp = this_cpu_ptr(rsp->rda);
813 do_nocb_deferred_wakeup(rdp);
815 rcu_prepare_for_idle();
816 rcu_dynticks_eqs_enter();
817 rcu_dynticks_task_enter();
820 * It is illegal to enter an extended quiescent state while
821 * in an RCU read-side critical section.
823 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
824 "Illegal idle entry in RCU read-side critical section.");
825 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
826 "Illegal idle entry in RCU-bh read-side critical section.");
827 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
828 "Illegal idle entry in RCU-sched read-side critical section.");
832 * Enter an RCU extended quiescent state, which can be either the
833 * idle loop or adaptive-tickless usermode execution.
835 static void rcu_eqs_enter(bool user)
838 struct rcu_dynticks *rdtp;
840 rdtp = this_cpu_ptr(&rcu_dynticks);
841 oldval = rdtp->dynticks_nesting;
842 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
843 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
844 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
845 rdtp->dynticks_nesting = 0;
846 rcu_eqs_enter_common(oldval, user);
848 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
853 * rcu_idle_enter - inform RCU that current CPU is entering idle
855 * Enter idle mode, in other words, -leave- the mode in which RCU
856 * read-side critical sections can occur. (Though RCU read-side
857 * critical sections can occur in irq handlers in idle, a possibility
858 * handled by irq_enter() and irq_exit().)
860 * We crowbar the ->dynticks_nesting field to zero to allow for
861 * the possibility of usermode upcalls having messed up our count
862 * of interrupt nesting level during the prior busy period.
864 void rcu_idle_enter(void)
868 local_irq_save(flags);
869 rcu_eqs_enter(false);
870 rcu_sysidle_enter(0);
871 local_irq_restore(flags);
873 EXPORT_SYMBOL_GPL(rcu_idle_enter);
875 #ifdef CONFIG_NO_HZ_FULL
877 * rcu_user_enter - inform RCU that we are resuming userspace.
879 * Enter RCU idle mode right before resuming userspace. No use of RCU
880 * is permitted between this call and rcu_user_exit(). This way the
881 * CPU doesn't need to maintain the tick for RCU maintenance purposes
882 * when the CPU runs in userspace.
884 void rcu_user_enter(void)
888 #endif /* CONFIG_NO_HZ_FULL */
891 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
893 * Exit from an interrupt handler, which might possibly result in entering
894 * idle mode, in other words, leaving the mode in which read-side critical
895 * sections can occur. The caller must have disabled interrupts.
897 * This code assumes that the idle loop never does anything that might
898 * result in unbalanced calls to irq_enter() and irq_exit(). If your
899 * architecture violates this assumption, RCU will give you what you
900 * deserve, good and hard. But very infrequently and irreproducibly.
902 * Use things like work queues to work around this limitation.
904 * You have been warned.
906 void rcu_irq_exit(void)
909 struct rcu_dynticks *rdtp;
911 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
912 rdtp = this_cpu_ptr(&rcu_dynticks);
913 oldval = rdtp->dynticks_nesting;
914 rdtp->dynticks_nesting--;
915 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
916 rdtp->dynticks_nesting < 0);
917 if (rdtp->dynticks_nesting)
918 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
920 rcu_eqs_enter_common(oldval, true);
921 rcu_sysidle_enter(1);
925 * Wrapper for rcu_irq_exit() where interrupts are enabled.
927 void rcu_irq_exit_irqson(void)
931 local_irq_save(flags);
933 local_irq_restore(flags);
937 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
939 * If the new value of the ->dynticks_nesting counter was previously zero,
940 * we really have exited idle, and must do the appropriate accounting.
941 * The caller must have disabled interrupts.
943 static void rcu_eqs_exit_common(long long oldval, int user)
945 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
947 rcu_dynticks_task_exit();
948 rcu_dynticks_eqs_exit();
949 rcu_cleanup_after_idle();
950 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
951 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
952 !user && !is_idle_task(current)) {
953 struct task_struct *idle __maybe_unused =
954 idle_task(smp_processor_id());
956 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
957 oldval, rdtp->dynticks_nesting);
958 rcu_ftrace_dump(DUMP_ORIG);
959 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
960 current->pid, current->comm,
961 idle->pid, idle->comm); /* must be idle task! */
966 * Exit an RCU extended quiescent state, which can be either the
967 * idle loop or adaptive-tickless usermode execution.
969 static void rcu_eqs_exit(bool user)
971 struct rcu_dynticks *rdtp;
974 rdtp = this_cpu_ptr(&rcu_dynticks);
975 oldval = rdtp->dynticks_nesting;
976 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
977 if (oldval & DYNTICK_TASK_NEST_MASK) {
978 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
980 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
981 rcu_eqs_exit_common(oldval, user);
986 * rcu_idle_exit - inform RCU that current CPU is leaving idle
988 * Exit idle mode, in other words, -enter- the mode in which RCU
989 * read-side critical sections can occur.
991 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
992 * allow for the possibility of usermode upcalls messing up our count
993 * of interrupt nesting level during the busy period that is just
996 void rcu_idle_exit(void)
1000 local_irq_save(flags);
1001 rcu_eqs_exit(false);
1002 rcu_sysidle_exit(0);
1003 local_irq_restore(flags);
1005 EXPORT_SYMBOL_GPL(rcu_idle_exit);
1007 #ifdef CONFIG_NO_HZ_FULL
1009 * rcu_user_exit - inform RCU that we are exiting userspace.
1011 * Exit RCU idle mode while entering the kernel because it can
1012 * run a RCU read side critical section anytime.
1014 void rcu_user_exit(void)
1018 #endif /* CONFIG_NO_HZ_FULL */
1021 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1023 * Enter an interrupt handler, which might possibly result in exiting
1024 * idle mode, in other words, entering the mode in which read-side critical
1025 * sections can occur. The caller must have disabled interrupts.
1027 * Note that the Linux kernel is fully capable of entering an interrupt
1028 * handler that it never exits, for example when doing upcalls to
1029 * user mode! This code assumes that the idle loop never does upcalls to
1030 * user mode. If your architecture does do upcalls from the idle loop (or
1031 * does anything else that results in unbalanced calls to the irq_enter()
1032 * and irq_exit() functions), RCU will give you what you deserve, good
1033 * and hard. But very infrequently and irreproducibly.
1035 * Use things like work queues to work around this limitation.
1037 * You have been warned.
1039 void rcu_irq_enter(void)
1041 struct rcu_dynticks *rdtp;
1044 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
1045 rdtp = this_cpu_ptr(&rcu_dynticks);
1046 oldval = rdtp->dynticks_nesting;
1047 rdtp->dynticks_nesting++;
1048 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
1049 rdtp->dynticks_nesting == 0);
1051 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
1053 rcu_eqs_exit_common(oldval, true);
1054 rcu_sysidle_exit(1);
1058 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1060 void rcu_irq_enter_irqson(void)
1062 unsigned long flags;
1064 local_irq_save(flags);
1066 local_irq_restore(flags);
1070 * rcu_nmi_enter - inform RCU of entry to NMI context
1072 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1073 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1074 * that the CPU is active. This implementation permits nested NMIs, as
1075 * long as the nesting level does not overflow an int. (You will probably
1076 * run out of stack space first.)
1078 void rcu_nmi_enter(void)
1080 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1083 /* Complain about underflow. */
1084 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
1087 * If idle from RCU viewpoint, atomically increment ->dynticks
1088 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1089 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1090 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1091 * to be in the outermost NMI handler that interrupted an RCU-idle
1092 * period (observation due to Andy Lutomirski).
1094 if (rcu_dynticks_curr_cpu_in_eqs()) {
1095 rcu_dynticks_eqs_exit();
1098 rdtp->dynticks_nmi_nesting += incby;
1103 * rcu_nmi_exit - inform RCU of exit from NMI context
1105 * If we are returning from the outermost NMI handler that interrupted an
1106 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
1107 * to let the RCU grace-period handling know that the CPU is back to
1110 void rcu_nmi_exit(void)
1112 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1115 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1116 * (We are exiting an NMI handler, so RCU better be paying attention
1119 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
1120 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
1123 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1124 * leave it in non-RCU-idle state.
1126 if (rdtp->dynticks_nmi_nesting != 1) {
1127 rdtp->dynticks_nmi_nesting -= 2;
1131 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1132 rdtp->dynticks_nmi_nesting = 0;
1133 rcu_dynticks_eqs_enter();
1137 * __rcu_is_watching - are RCU read-side critical sections safe?
1139 * Return true if RCU is watching the running CPU, which means that
1140 * this CPU can safely enter RCU read-side critical sections. Unlike
1141 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1142 * least disabled preemption.
1144 bool notrace __rcu_is_watching(void)
1146 return !rcu_dynticks_curr_cpu_in_eqs();
1150 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1152 * If the current CPU is in its idle loop and is neither in an interrupt
1153 * or NMI handler, return true.
1155 bool notrace rcu_is_watching(void)
1159 preempt_disable_notrace();
1160 ret = __rcu_is_watching();
1161 preempt_enable_notrace();
1164 EXPORT_SYMBOL_GPL(rcu_is_watching);
1166 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1169 * Is the current CPU online? Disable preemption to avoid false positives
1170 * that could otherwise happen due to the current CPU number being sampled,
1171 * this task being preempted, its old CPU being taken offline, resuming
1172 * on some other CPU, then determining that its old CPU is now offline.
1173 * It is OK to use RCU on an offline processor during initial boot, hence
1174 * the check for rcu_scheduler_fully_active. Note also that it is OK
1175 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1176 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1177 * offline to continue to use RCU for one jiffy after marking itself
1178 * offline in the cpu_online_mask. This leniency is necessary given the
1179 * non-atomic nature of the online and offline processing, for example,
1180 * the fact that a CPU enters the scheduler after completing the teardown
1183 * This is also why RCU internally marks CPUs online during in the
1184 * preparation phase and offline after the CPU has been taken down.
1186 * Disable checking if in an NMI handler because we cannot safely report
1187 * errors from NMI handlers anyway.
1189 bool rcu_lockdep_current_cpu_online(void)
1191 struct rcu_data *rdp;
1192 struct rcu_node *rnp;
1198 rdp = this_cpu_ptr(&rcu_sched_data);
1200 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1201 !rcu_scheduler_fully_active;
1205 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1207 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1210 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1212 * If the current CPU is idle or running at a first-level (not nested)
1213 * interrupt from idle, return true. The caller must have at least
1214 * disabled preemption.
1216 static int rcu_is_cpu_rrupt_from_idle(void)
1218 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1222 * Snapshot the specified CPU's dynticks counter so that we can later
1223 * credit them with an implicit quiescent state. Return 1 if this CPU
1224 * is in dynticks idle mode, which is an extended quiescent state.
1226 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1227 bool *isidle, unsigned long *maxj)
1229 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1230 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1231 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1232 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1233 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1234 rdp->mynode->gpnum))
1235 WRITE_ONCE(rdp->gpwrap, true);
1242 * Return true if the specified CPU has passed through a quiescent
1243 * state by virtue of being in or having passed through an dynticks
1244 * idle state since the last call to dyntick_save_progress_counter()
1245 * for this same CPU, or by virtue of having been offline.
1247 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1248 bool *isidle, unsigned long *maxj)
1253 unsigned long rjtsc;
1254 struct rcu_node *rnp;
1257 * If the CPU passed through or entered a dynticks idle phase with
1258 * no active irq/NMI handlers, then we can safely pretend that the CPU
1259 * already acknowledged the request to pass through a quiescent
1260 * state. Either way, that CPU cannot possibly be in an RCU
1261 * read-side critical section that started before the beginning
1262 * of the current RCU grace period.
1264 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1265 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1266 rdp->dynticks_fqs++;
1270 /* Compute and saturate jiffies_till_sched_qs. */
1271 jtsq = jiffies_till_sched_qs;
1272 rjtsc = rcu_jiffies_till_stall_check();
1273 if (jtsq > rjtsc / 2) {
1274 WRITE_ONCE(jiffies_till_sched_qs, rjtsc);
1276 } else if (jtsq < 1) {
1277 WRITE_ONCE(jiffies_till_sched_qs, 1);
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.
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"));
1295 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1296 smp_store_release(ruqp, true);
1299 /* Check for the CPU being offline. */
1300 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1301 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1307 * A CPU running for an extended time within the kernel can
1308 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1309 * even context-switching back and forth between a pair of
1310 * in-kernel CPU-bound tasks cannot advance grace periods.
1311 * So if the grace period is old enough, make the CPU pay attention.
1312 * Note that the unsynchronized assignments to the per-CPU
1313 * rcu_need_heavy_qs variable are safe. Yes, setting of
1314 * bits can be lost, but they will be set again on the next
1315 * force-quiescent-state pass. So lost bit sets do not result
1316 * in incorrect behavior, merely in a grace period lasting
1317 * a few jiffies longer than it might otherwise. Because
1318 * there are at most four threads involved, and because the
1319 * updates are only once every few jiffies, the probability of
1320 * lossage (and thus of slight grace-period extension) is
1323 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1324 * is set too high, we override with half of the RCU CPU stall
1327 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1328 if (!READ_ONCE(*rnhqp) &&
1329 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1330 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1331 WRITE_ONCE(*rnhqp, true);
1332 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1333 smp_store_release(ruqp, true);
1334 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1338 * If more than halfway to RCU CPU stall-warning time, do
1339 * a resched_cpu() to try to loosen things up a bit.
1341 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2)
1342 resched_cpu(rdp->cpu);
1347 static void record_gp_stall_check_time(struct rcu_state *rsp)
1349 unsigned long j = jiffies;
1353 smp_wmb(); /* Record start time before stall time. */
1354 j1 = rcu_jiffies_till_stall_check();
1355 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1356 rsp->jiffies_resched = j + j1 / 2;
1357 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1361 * Convert a ->gp_state value to a character string.
1363 static const char *gp_state_getname(short gs)
1365 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1367 return gp_state_names[gs];
1371 * Complain about starvation of grace-period kthread.
1373 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1379 gpa = READ_ONCE(rsp->gp_activity);
1380 if (j - gpa > 2 * HZ) {
1381 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1383 rsp->gpnum, rsp->completed,
1385 gp_state_getname(rsp->gp_state), rsp->gp_state,
1386 rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
1387 if (rsp->gp_kthread) {
1388 sched_show_task(rsp->gp_kthread);
1389 wake_up_process(rsp->gp_kthread);
1395 * Dump stacks of all tasks running on stalled CPUs. First try using
1396 * NMIs, but fall back to manual remote stack tracing on architectures
1397 * that don't support NMI-based stack dumps. The NMI-triggered stack
1398 * traces are more accurate because they are printed by the target CPU.
1400 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1403 unsigned long flags;
1404 struct rcu_node *rnp;
1406 rcu_for_each_leaf_node(rsp, rnp) {
1407 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1408 for_each_leaf_node_possible_cpu(rnp, cpu)
1409 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1410 if (!trigger_single_cpu_backtrace(cpu))
1412 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1417 * If too much time has passed in the current grace period, and if
1418 * so configured, go kick the relevant kthreads.
1420 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1424 if (!rcu_kick_kthreads)
1426 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1427 if (time_after(jiffies, j) && rsp->gp_kthread &&
1428 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1429 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1430 rcu_ftrace_dump(DUMP_ALL);
1431 wake_up_process(rsp->gp_kthread);
1432 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1436 static inline void panic_on_rcu_stall(void)
1438 if (sysctl_panic_on_rcu_stall)
1439 panic("RCU Stall\n");
1442 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1446 unsigned long flags;
1450 struct rcu_node *rnp = rcu_get_root(rsp);
1453 /* Kick and suppress, if so configured. */
1454 rcu_stall_kick_kthreads(rsp);
1455 if (rcu_cpu_stall_suppress)
1458 /* Only let one CPU complain about others per time interval. */
1460 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1461 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1462 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1463 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1466 WRITE_ONCE(rsp->jiffies_stall,
1467 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1468 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1471 * OK, time to rat on our buddy...
1472 * See Documentation/RCU/stallwarn.txt for info on how to debug
1473 * RCU CPU stall warnings.
1475 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1477 print_cpu_stall_info_begin();
1478 rcu_for_each_leaf_node(rsp, rnp) {
1479 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1480 ndetected += rcu_print_task_stall(rnp);
1481 if (rnp->qsmask != 0) {
1482 for_each_leaf_node_possible_cpu(rnp, cpu)
1483 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1484 print_cpu_stall_info(rsp, cpu);
1488 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1491 print_cpu_stall_info_end();
1492 for_each_possible_cpu(cpu)
1493 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1494 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1495 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1496 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1498 rcu_dump_cpu_stacks(rsp);
1500 /* Complain about tasks blocking the grace period. */
1501 rcu_print_detail_task_stall(rsp);
1503 if (READ_ONCE(rsp->gpnum) != gpnum ||
1504 READ_ONCE(rsp->completed) == gpnum) {
1505 pr_err("INFO: Stall ended before state dump start\n");
1508 gpa = READ_ONCE(rsp->gp_activity);
1509 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1510 rsp->name, j - gpa, j, gpa,
1511 jiffies_till_next_fqs,
1512 rcu_get_root(rsp)->qsmask);
1513 /* In this case, the current CPU might be at fault. */
1514 sched_show_task(current);
1518 rcu_check_gp_kthread_starvation(rsp);
1520 panic_on_rcu_stall();
1522 force_quiescent_state(rsp); /* Kick them all. */
1525 static void print_cpu_stall(struct rcu_state *rsp)
1528 unsigned long flags;
1529 struct rcu_node *rnp = rcu_get_root(rsp);
1532 /* Kick and suppress, if so configured. */
1533 rcu_stall_kick_kthreads(rsp);
1534 if (rcu_cpu_stall_suppress)
1538 * OK, time to rat on ourselves...
1539 * See Documentation/RCU/stallwarn.txt for info on how to debug
1540 * RCU CPU stall warnings.
1542 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1543 print_cpu_stall_info_begin();
1544 print_cpu_stall_info(rsp, smp_processor_id());
1545 print_cpu_stall_info_end();
1546 for_each_possible_cpu(cpu)
1547 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1548 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1549 jiffies - rsp->gp_start,
1550 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1552 rcu_check_gp_kthread_starvation(rsp);
1554 rcu_dump_cpu_stacks(rsp);
1556 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1557 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1558 WRITE_ONCE(rsp->jiffies_stall,
1559 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1560 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1562 panic_on_rcu_stall();
1565 * Attempt to revive the RCU machinery by forcing a context switch.
1567 * A context switch would normally allow the RCU state machine to make
1568 * progress and it could be we're stuck in kernel space without context
1569 * switches for an entirely unreasonable amount of time.
1571 resched_cpu(smp_processor_id());
1574 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1576 unsigned long completed;
1577 unsigned long gpnum;
1581 struct rcu_node *rnp;
1583 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1584 !rcu_gp_in_progress(rsp))
1586 rcu_stall_kick_kthreads(rsp);
1590 * Lots of memory barriers to reject false positives.
1592 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1593 * then rsp->gp_start, and finally rsp->completed. These values
1594 * are updated in the opposite order with memory barriers (or
1595 * equivalent) during grace-period initialization and cleanup.
1596 * Now, a false positive can occur if we get an new value of
1597 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1598 * the memory barriers, the only way that this can happen is if one
1599 * grace period ends and another starts between these two fetches.
1600 * Detect this by comparing rsp->completed with the previous fetch
1603 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1604 * and rsp->gp_start suffice to forestall false positives.
1606 gpnum = READ_ONCE(rsp->gpnum);
1607 smp_rmb(); /* Pick up ->gpnum first... */
1608 js = READ_ONCE(rsp->jiffies_stall);
1609 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1610 gps = READ_ONCE(rsp->gp_start);
1611 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1612 completed = READ_ONCE(rsp->completed);
1613 if (ULONG_CMP_GE(completed, gpnum) ||
1614 ULONG_CMP_LT(j, js) ||
1615 ULONG_CMP_GE(gps, js))
1616 return; /* No stall or GP completed since entering function. */
1618 if (rcu_gp_in_progress(rsp) &&
1619 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1621 /* We haven't checked in, so go dump stack. */
1622 print_cpu_stall(rsp);
1624 } else if (rcu_gp_in_progress(rsp) &&
1625 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1627 /* They had a few time units to dump stack, so complain. */
1628 print_other_cpu_stall(rsp, gpnum);
1633 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1635 * Set the stall-warning timeout way off into the future, thus preventing
1636 * any RCU CPU stall-warning messages from appearing in the current set of
1637 * RCU grace periods.
1639 * The caller must disable hard irqs.
1641 void rcu_cpu_stall_reset(void)
1643 struct rcu_state *rsp;
1645 for_each_rcu_flavor(rsp)
1646 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1650 * Initialize the specified rcu_data structure's default callback list
1651 * to empty. The default callback list is the one that is not used by
1652 * no-callbacks CPUs.
1654 static void init_default_callback_list(struct rcu_data *rdp)
1658 rdp->nxtlist = NULL;
1659 for (i = 0; i < RCU_NEXT_SIZE; i++)
1660 rdp->nxttail[i] = &rdp->nxtlist;
1664 * Initialize the specified rcu_data structure's callback list to empty.
1666 static void init_callback_list(struct rcu_data *rdp)
1668 if (init_nocb_callback_list(rdp))
1670 init_default_callback_list(rdp);
1674 * Determine the value that ->completed will have at the end of the
1675 * next subsequent grace period. This is used to tag callbacks so that
1676 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1677 * been dyntick-idle for an extended period with callbacks under the
1678 * influence of RCU_FAST_NO_HZ.
1680 * The caller must hold rnp->lock with interrupts disabled.
1682 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1683 struct rcu_node *rnp)
1686 * If RCU is idle, we just wait for the next grace period.
1687 * But we can only be sure that RCU is idle if we are looking
1688 * at the root rcu_node structure -- otherwise, a new grace
1689 * period might have started, but just not yet gotten around
1690 * to initializing the current non-root rcu_node structure.
1692 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1693 return rnp->completed + 1;
1696 * Otherwise, wait for a possible partial grace period and
1697 * then the subsequent full grace period.
1699 return rnp->completed + 2;
1703 * Trace-event helper function for rcu_start_future_gp() and
1704 * rcu_nocb_wait_gp().
1706 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1707 unsigned long c, const char *s)
1709 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1710 rnp->completed, c, rnp->level,
1711 rnp->grplo, rnp->grphi, s);
1715 * Start some future grace period, as needed to handle newly arrived
1716 * callbacks. The required future grace periods are recorded in each
1717 * rcu_node structure's ->need_future_gp field. Returns true if there
1718 * is reason to awaken the grace-period kthread.
1720 * The caller must hold the specified rcu_node structure's ->lock.
1722 static bool __maybe_unused
1723 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1724 unsigned long *c_out)
1729 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1732 * Pick up grace-period number for new callbacks. If this
1733 * grace period is already marked as needed, return to the caller.
1735 c = rcu_cbs_completed(rdp->rsp, rnp);
1736 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1737 if (rnp->need_future_gp[c & 0x1]) {
1738 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1743 * If either this rcu_node structure or the root rcu_node structure
1744 * believe that a grace period is in progress, then we must wait
1745 * for the one following, which is in "c". Because our request
1746 * will be noticed at the end of the current grace period, we don't
1747 * need to explicitly start one. We only do the lockless check
1748 * of rnp_root's fields if the current rcu_node structure thinks
1749 * there is no grace period in flight, and because we hold rnp->lock,
1750 * the only possible change is when rnp_root's two fields are
1751 * equal, in which case rnp_root->gpnum might be concurrently
1752 * incremented. But that is OK, as it will just result in our
1753 * doing some extra useless work.
1755 if (rnp->gpnum != rnp->completed ||
1756 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1757 rnp->need_future_gp[c & 0x1]++;
1758 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1763 * There might be no grace period in progress. If we don't already
1764 * hold it, acquire the root rcu_node structure's lock in order to
1765 * start one (if needed).
1767 if (rnp != rnp_root)
1768 raw_spin_lock_rcu_node(rnp_root);
1771 * Get a new grace-period number. If there really is no grace
1772 * period in progress, it will be smaller than the one we obtained
1773 * earlier. Adjust callbacks as needed. Note that even no-CBs
1774 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1776 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1777 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1778 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1779 rdp->nxtcompleted[i] = c;
1782 * If the needed for the required grace period is already
1783 * recorded, trace and leave.
1785 if (rnp_root->need_future_gp[c & 0x1]) {
1786 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1790 /* Record the need for the future grace period. */
1791 rnp_root->need_future_gp[c & 0x1]++;
1793 /* If a grace period is not already in progress, start one. */
1794 if (rnp_root->gpnum != rnp_root->completed) {
1795 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1797 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1798 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1801 if (rnp != rnp_root)
1802 raw_spin_unlock_rcu_node(rnp_root);
1810 * Clean up any old requests for the just-ended grace period. Also return
1811 * whether any additional grace periods have been requested. Also invoke
1812 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1813 * waiting for this grace period to complete.
1815 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1817 int c = rnp->completed;
1819 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1821 rnp->need_future_gp[c & 0x1] = 0;
1822 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1823 trace_rcu_future_gp(rnp, rdp, c,
1824 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1829 * Awaken the grace-period kthread for the specified flavor of RCU.
1830 * Don't do a self-awaken, and don't bother awakening when there is
1831 * nothing for the grace-period kthread to do (as in several CPUs
1832 * raced to awaken, and we lost), and finally don't try to awaken
1833 * a kthread that has not yet been created.
1835 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1837 if (current == rsp->gp_kthread ||
1838 !READ_ONCE(rsp->gp_flags) ||
1841 swake_up(&rsp->gp_wq);
1845 * If there is room, assign a ->completed number to any callbacks on
1846 * this CPU that have not already been assigned. Also accelerate any
1847 * callbacks that were previously assigned a ->completed number that has
1848 * since proven to be too conservative, which can happen if callbacks get
1849 * assigned a ->completed number while RCU is idle, but with reference to
1850 * a non-root rcu_node structure. This function is idempotent, so it does
1851 * not hurt to call it repeatedly. Returns an flag saying that we should
1852 * awaken the RCU grace-period kthread.
1854 * The caller must hold rnp->lock with interrupts disabled.
1856 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1857 struct rcu_data *rdp)
1863 /* If the CPU has no callbacks, nothing to do. */
1864 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1868 * Starting from the sublist containing the callbacks most
1869 * recently assigned a ->completed number and working down, find the
1870 * first sublist that is not assignable to an upcoming grace period.
1871 * Such a sublist has something in it (first two tests) and has
1872 * a ->completed number assigned that will complete sooner than
1873 * the ->completed number for newly arrived callbacks (last test).
1875 * The key point is that any later sublist can be assigned the
1876 * same ->completed number as the newly arrived callbacks, which
1877 * means that the callbacks in any of these later sublist can be
1878 * grouped into a single sublist, whether or not they have already
1879 * been assigned a ->completed number.
1881 c = rcu_cbs_completed(rsp, rnp);
1882 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1883 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1884 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1888 * If there are no sublist for unassigned callbacks, leave.
1889 * At the same time, advance "i" one sublist, so that "i" will
1890 * index into the sublist where all the remaining callbacks should
1893 if (++i >= RCU_NEXT_TAIL)
1897 * Assign all subsequent callbacks' ->completed number to the next
1898 * full grace period and group them all in the sublist initially
1901 for (; i <= RCU_NEXT_TAIL; i++) {
1902 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1903 rdp->nxtcompleted[i] = c;
1905 /* Record any needed additional grace periods. */
1906 ret = rcu_start_future_gp(rnp, rdp, NULL);
1908 /* Trace depending on how much we were able to accelerate. */
1909 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1910 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1912 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1917 * Move any callbacks whose grace period has completed to the
1918 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1919 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1920 * sublist. This function is idempotent, so it does not hurt to
1921 * invoke it repeatedly. As long as it is not invoked -too- often...
1922 * Returns true if the RCU grace-period kthread needs to be awakened.
1924 * The caller must hold rnp->lock with interrupts disabled.
1926 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1927 struct rcu_data *rdp)
1931 /* If the CPU has no callbacks, nothing to do. */
1932 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1936 * Find all callbacks whose ->completed numbers indicate that they
1937 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1939 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1940 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1942 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1944 /* Clean up any sublist tail pointers that were misordered above. */
1945 for (j = RCU_WAIT_TAIL; j < i; j++)
1946 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1948 /* Copy down callbacks to fill in empty sublists. */
1949 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1950 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1952 rdp->nxttail[j] = rdp->nxttail[i];
1953 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1956 /* Classify any remaining callbacks. */
1957 return rcu_accelerate_cbs(rsp, rnp, rdp);
1961 * Update CPU-local rcu_data state to record the beginnings and ends of
1962 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1963 * structure corresponding to the current CPU, and must have irqs disabled.
1964 * Returns true if the grace-period kthread needs to be awakened.
1966 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1967 struct rcu_data *rdp)
1972 /* Handle the ends of any preceding grace periods first. */
1973 if (rdp->completed == rnp->completed &&
1974 !unlikely(READ_ONCE(rdp->gpwrap))) {
1976 /* No grace period end, so just accelerate recent callbacks. */
1977 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1981 /* Advance callbacks. */
1982 ret = rcu_advance_cbs(rsp, rnp, rdp);
1984 /* Remember that we saw this grace-period completion. */
1985 rdp->completed = rnp->completed;
1986 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1989 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1991 * If the current grace period is waiting for this CPU,
1992 * set up to detect a quiescent state, otherwise don't
1993 * go looking for one.
1995 rdp->gpnum = rnp->gpnum;
1996 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1997 need_gp = !!(rnp->qsmask & rdp->grpmask);
1998 rdp->cpu_no_qs.b.norm = need_gp;
1999 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2000 rdp->core_needs_qs = need_gp;
2001 zero_cpu_stall_ticks(rdp);
2002 WRITE_ONCE(rdp->gpwrap, false);
2007 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
2009 unsigned long flags;
2011 struct rcu_node *rnp;
2013 local_irq_save(flags);
2015 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
2016 rdp->completed == READ_ONCE(rnp->completed) &&
2017 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
2018 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
2019 local_irq_restore(flags);
2022 needwake = __note_gp_changes(rsp, rnp, rdp);
2023 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2025 rcu_gp_kthread_wake(rsp);
2028 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
2031 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
2032 schedule_timeout_uninterruptible(delay);
2036 * Initialize a new grace period. Return false if no grace period required.
2038 static bool rcu_gp_init(struct rcu_state *rsp)
2040 unsigned long oldmask;
2041 struct rcu_data *rdp;
2042 struct rcu_node *rnp = rcu_get_root(rsp);
2044 WRITE_ONCE(rsp->gp_activity, jiffies);
2045 raw_spin_lock_irq_rcu_node(rnp);
2046 if (!READ_ONCE(rsp->gp_flags)) {
2047 /* Spurious wakeup, tell caller to go back to sleep. */
2048 raw_spin_unlock_irq_rcu_node(rnp);
2051 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
2053 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
2055 * Grace period already in progress, don't start another.
2056 * Not supposed to be able to happen.
2058 raw_spin_unlock_irq_rcu_node(rnp);
2062 /* Advance to a new grace period and initialize state. */
2063 record_gp_stall_check_time(rsp);
2064 /* Record GP times before starting GP, hence smp_store_release(). */
2065 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
2066 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
2067 raw_spin_unlock_irq_rcu_node(rnp);
2070 * Apply per-leaf buffered online and offline operations to the
2071 * rcu_node tree. Note that this new grace period need not wait
2072 * for subsequent online CPUs, and that quiescent-state forcing
2073 * will handle subsequent offline CPUs.
2075 rcu_for_each_leaf_node(rsp, rnp) {
2076 rcu_gp_slow(rsp, gp_preinit_delay);
2077 raw_spin_lock_irq_rcu_node(rnp);
2078 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2079 !rnp->wait_blkd_tasks) {
2080 /* Nothing to do on this leaf rcu_node structure. */
2081 raw_spin_unlock_irq_rcu_node(rnp);
2085 /* Record old state, apply changes to ->qsmaskinit field. */
2086 oldmask = rnp->qsmaskinit;
2087 rnp->qsmaskinit = rnp->qsmaskinitnext;
2089 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2090 if (!oldmask != !rnp->qsmaskinit) {
2091 if (!oldmask) /* First online CPU for this rcu_node. */
2092 rcu_init_new_rnp(rnp);
2093 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2094 rnp->wait_blkd_tasks = true;
2095 else /* Last offline CPU and can propagate. */
2096 rcu_cleanup_dead_rnp(rnp);
2100 * If all waited-on tasks from prior grace period are
2101 * done, and if all this rcu_node structure's CPUs are
2102 * still offline, propagate up the rcu_node tree and
2103 * clear ->wait_blkd_tasks. Otherwise, if one of this
2104 * rcu_node structure's CPUs has since come back online,
2105 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2106 * checks for this, so just call it unconditionally).
2108 if (rnp->wait_blkd_tasks &&
2109 (!rcu_preempt_has_tasks(rnp) ||
2111 rnp->wait_blkd_tasks = false;
2112 rcu_cleanup_dead_rnp(rnp);
2115 raw_spin_unlock_irq_rcu_node(rnp);
2119 * Set the quiescent-state-needed bits in all the rcu_node
2120 * structures for all currently online CPUs in breadth-first order,
2121 * starting from the root rcu_node structure, relying on the layout
2122 * of the tree within the rsp->node[] array. Note that other CPUs
2123 * will access only the leaves of the hierarchy, thus seeing that no
2124 * grace period is in progress, at least until the corresponding
2125 * leaf node has been initialized.
2127 * The grace period cannot complete until the initialization
2128 * process finishes, because this kthread handles both.
2130 rcu_for_each_node_breadth_first(rsp, rnp) {
2131 rcu_gp_slow(rsp, gp_init_delay);
2132 raw_spin_lock_irq_rcu_node(rnp);
2133 rdp = this_cpu_ptr(rsp->rda);
2134 rcu_preempt_check_blocked_tasks(rnp);
2135 rnp->qsmask = rnp->qsmaskinit;
2136 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2137 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2138 WRITE_ONCE(rnp->completed, rsp->completed);
2139 if (rnp == rdp->mynode)
2140 (void)__note_gp_changes(rsp, rnp, rdp);
2141 rcu_preempt_boost_start_gp(rnp);
2142 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2143 rnp->level, rnp->grplo,
2144 rnp->grphi, rnp->qsmask);
2145 raw_spin_unlock_irq_rcu_node(rnp);
2146 cond_resched_rcu_qs();
2147 WRITE_ONCE(rsp->gp_activity, jiffies);
2154 * Helper function for wait_event_interruptible_timeout() wakeup
2155 * at force-quiescent-state time.
2157 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2159 struct rcu_node *rnp = rcu_get_root(rsp);
2161 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2162 *gfp = READ_ONCE(rsp->gp_flags);
2163 if (*gfp & RCU_GP_FLAG_FQS)
2166 /* The current grace period has completed. */
2167 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2174 * Do one round of quiescent-state forcing.
2176 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2178 bool isidle = false;
2180 struct rcu_node *rnp = rcu_get_root(rsp);
2182 WRITE_ONCE(rsp->gp_activity, jiffies);
2185 /* Collect dyntick-idle snapshots. */
2186 if (is_sysidle_rcu_state(rsp)) {
2188 maxj = jiffies - ULONG_MAX / 4;
2190 force_qs_rnp(rsp, dyntick_save_progress_counter,
2192 rcu_sysidle_report_gp(rsp, isidle, maxj);
2194 /* Handle dyntick-idle and offline CPUs. */
2196 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
2198 /* Clear flag to prevent immediate re-entry. */
2199 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2200 raw_spin_lock_irq_rcu_node(rnp);
2201 WRITE_ONCE(rsp->gp_flags,
2202 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2203 raw_spin_unlock_irq_rcu_node(rnp);
2208 * Clean up after the old grace period.
2210 static void rcu_gp_cleanup(struct rcu_state *rsp)
2212 unsigned long gp_duration;
2213 bool needgp = false;
2215 struct rcu_data *rdp;
2216 struct rcu_node *rnp = rcu_get_root(rsp);
2217 struct swait_queue_head *sq;
2219 WRITE_ONCE(rsp->gp_activity, jiffies);
2220 raw_spin_lock_irq_rcu_node(rnp);
2221 gp_duration = jiffies - rsp->gp_start;
2222 if (gp_duration > rsp->gp_max)
2223 rsp->gp_max = gp_duration;
2226 * We know the grace period is complete, but to everyone else
2227 * it appears to still be ongoing. But it is also the case
2228 * that to everyone else it looks like there is nothing that
2229 * they can do to advance the grace period. It is therefore
2230 * safe for us to drop the lock in order to mark the grace
2231 * period as completed in all of the rcu_node structures.
2233 raw_spin_unlock_irq_rcu_node(rnp);
2236 * Propagate new ->completed value to rcu_node structures so
2237 * that other CPUs don't have to wait until the start of the next
2238 * grace period to process their callbacks. This also avoids
2239 * some nasty RCU grace-period initialization races by forcing
2240 * the end of the current grace period to be completely recorded in
2241 * all of the rcu_node structures before the beginning of the next
2242 * grace period is recorded in any of the rcu_node structures.
2244 rcu_for_each_node_breadth_first(rsp, rnp) {
2245 raw_spin_lock_irq_rcu_node(rnp);
2246 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2247 WARN_ON_ONCE(rnp->qsmask);
2248 WRITE_ONCE(rnp->completed, rsp->gpnum);
2249 rdp = this_cpu_ptr(rsp->rda);
2250 if (rnp == rdp->mynode)
2251 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2252 /* smp_mb() provided by prior unlock-lock pair. */
2253 nocb += rcu_future_gp_cleanup(rsp, rnp);
2254 sq = rcu_nocb_gp_get(rnp);
2255 raw_spin_unlock_irq_rcu_node(rnp);
2256 rcu_nocb_gp_cleanup(sq);
2257 cond_resched_rcu_qs();
2258 WRITE_ONCE(rsp->gp_activity, jiffies);
2259 rcu_gp_slow(rsp, gp_cleanup_delay);
2261 rnp = rcu_get_root(rsp);
2262 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2263 rcu_nocb_gp_set(rnp, nocb);
2265 /* Declare grace period done. */
2266 WRITE_ONCE(rsp->completed, rsp->gpnum);
2267 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2268 rsp->gp_state = RCU_GP_IDLE;
2269 rdp = this_cpu_ptr(rsp->rda);
2270 /* Advance CBs to reduce false positives below. */
2271 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2272 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2273 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2274 trace_rcu_grace_period(rsp->name,
2275 READ_ONCE(rsp->gpnum),
2278 raw_spin_unlock_irq_rcu_node(rnp);
2282 * Body of kthread that handles grace periods.
2284 static int __noreturn rcu_gp_kthread(void *arg)
2290 struct rcu_state *rsp = arg;
2291 struct rcu_node *rnp = rcu_get_root(rsp);
2293 rcu_bind_gp_kthread();
2296 /* Handle grace-period start. */
2298 trace_rcu_grace_period(rsp->name,
2299 READ_ONCE(rsp->gpnum),
2301 rsp->gp_state = RCU_GP_WAIT_GPS;
2302 swait_event_interruptible(rsp->gp_wq,
2303 READ_ONCE(rsp->gp_flags) &
2305 rsp->gp_state = RCU_GP_DONE_GPS;
2306 /* Locking provides needed memory barrier. */
2307 if (rcu_gp_init(rsp))
2309 cond_resched_rcu_qs();
2310 WRITE_ONCE(rsp->gp_activity, jiffies);
2311 WARN_ON(signal_pending(current));
2312 trace_rcu_grace_period(rsp->name,
2313 READ_ONCE(rsp->gpnum),
2317 /* Handle quiescent-state forcing. */
2318 first_gp_fqs = true;
2319 j = jiffies_till_first_fqs;
2322 jiffies_till_first_fqs = HZ;
2327 rsp->jiffies_force_qs = jiffies + j;
2328 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2331 trace_rcu_grace_period(rsp->name,
2332 READ_ONCE(rsp->gpnum),
2334 rsp->gp_state = RCU_GP_WAIT_FQS;
2335 ret = swait_event_interruptible_timeout(rsp->gp_wq,
2336 rcu_gp_fqs_check_wake(rsp, &gf), j);
2337 rsp->gp_state = RCU_GP_DOING_FQS;
2338 /* Locking provides needed memory barriers. */
2339 /* If grace period done, leave loop. */
2340 if (!READ_ONCE(rnp->qsmask) &&
2341 !rcu_preempt_blocked_readers_cgp(rnp))
2343 /* If time for quiescent-state forcing, do it. */
2344 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2345 (gf & RCU_GP_FLAG_FQS)) {
2346 trace_rcu_grace_period(rsp->name,
2347 READ_ONCE(rsp->gpnum),
2349 rcu_gp_fqs(rsp, first_gp_fqs);
2350 first_gp_fqs = false;
2351 trace_rcu_grace_period(rsp->name,
2352 READ_ONCE(rsp->gpnum),
2354 cond_resched_rcu_qs();
2355 WRITE_ONCE(rsp->gp_activity, jiffies);
2356 ret = 0; /* Force full wait till next FQS. */
2357 j = jiffies_till_next_fqs;
2360 jiffies_till_next_fqs = HZ;
2363 jiffies_till_next_fqs = 1;
2366 /* Deal with stray signal. */
2367 cond_resched_rcu_qs();
2368 WRITE_ONCE(rsp->gp_activity, jiffies);
2369 WARN_ON(signal_pending(current));
2370 trace_rcu_grace_period(rsp->name,
2371 READ_ONCE(rsp->gpnum),
2373 ret = 1; /* Keep old FQS timing. */
2375 if (time_after(jiffies, rsp->jiffies_force_qs))
2378 j = rsp->jiffies_force_qs - j;
2382 /* Handle grace-period end. */
2383 rsp->gp_state = RCU_GP_CLEANUP;
2384 rcu_gp_cleanup(rsp);
2385 rsp->gp_state = RCU_GP_CLEANED;
2390 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2391 * in preparation for detecting the next grace period. The caller must hold
2392 * the root node's ->lock and hard irqs must be disabled.
2394 * Note that it is legal for a dying CPU (which is marked as offline) to
2395 * invoke this function. This can happen when the dying CPU reports its
2398 * Returns true if the grace-period kthread must be awakened.
2401 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2402 struct rcu_data *rdp)
2404 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2406 * Either we have not yet spawned the grace-period
2407 * task, this CPU does not need another grace period,
2408 * or a grace period is already in progress.
2409 * Either way, don't start a new grace period.
2413 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2414 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2418 * We can't do wakeups while holding the rnp->lock, as that
2419 * could cause possible deadlocks with the rq->lock. Defer
2420 * the wakeup to our caller.
2426 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2427 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2428 * is invoked indirectly from rcu_advance_cbs(), which would result in
2429 * endless recursion -- or would do so if it wasn't for the self-deadlock
2430 * that is encountered beforehand.
2432 * Returns true if the grace-period kthread needs to be awakened.
2434 static bool rcu_start_gp(struct rcu_state *rsp)
2436 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2437 struct rcu_node *rnp = rcu_get_root(rsp);
2441 * If there is no grace period in progress right now, any
2442 * callbacks we have up to this point will be satisfied by the
2443 * next grace period. Also, advancing the callbacks reduces the
2444 * probability of false positives from cpu_needs_another_gp()
2445 * resulting in pointless grace periods. So, advance callbacks
2446 * then start the grace period!
2448 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2449 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2454 * Report a full set of quiescent states to the specified rcu_state data
2455 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2456 * kthread if another grace period is required. Whether we wake
2457 * the grace-period kthread or it awakens itself for the next round
2458 * of quiescent-state forcing, that kthread will clean up after the
2459 * just-completed grace period. Note that the caller must hold rnp->lock,
2460 * which is released before return.
2462 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2463 __releases(rcu_get_root(rsp)->lock)
2465 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2466 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2467 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2468 rcu_gp_kthread_wake(rsp);
2472 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2473 * Allows quiescent states for a group of CPUs to be reported at one go
2474 * to the specified rcu_node structure, though all the CPUs in the group
2475 * must be represented by the same rcu_node structure (which need not be a
2476 * leaf rcu_node structure, though it often will be). The gps parameter
2477 * is the grace-period snapshot, which means that the quiescent states
2478 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2479 * must be held upon entry, and it is released before return.
2482 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2483 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2484 __releases(rnp->lock)
2486 unsigned long oldmask = 0;
2487 struct rcu_node *rnp_c;
2489 /* Walk up the rcu_node hierarchy. */
2491 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2494 * Our bit has already been cleared, or the
2495 * relevant grace period is already over, so done.
2497 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2500 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2501 rnp->qsmask &= ~mask;
2502 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2503 mask, rnp->qsmask, rnp->level,
2504 rnp->grplo, rnp->grphi,
2506 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2508 /* Other bits still set at this level, so done. */
2509 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2512 mask = rnp->grpmask;
2513 if (rnp->parent == NULL) {
2515 /* No more levels. Exit loop holding root lock. */
2519 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2522 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2523 oldmask = rnp_c->qsmask;
2527 * Get here if we are the last CPU to pass through a quiescent
2528 * state for this grace period. Invoke rcu_report_qs_rsp()
2529 * to clean up and start the next grace period if one is needed.
2531 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2535 * Record a quiescent state for all tasks that were previously queued
2536 * on the specified rcu_node structure and that were blocking the current
2537 * RCU grace period. The caller must hold the specified rnp->lock with
2538 * irqs disabled, and this lock is released upon return, but irqs remain
2541 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2542 struct rcu_node *rnp, unsigned long flags)
2543 __releases(rnp->lock)
2547 struct rcu_node *rnp_p;
2549 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2550 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2551 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2552 return; /* Still need more quiescent states! */
2555 rnp_p = rnp->parent;
2556 if (rnp_p == NULL) {
2558 * Only one rcu_node structure in the tree, so don't
2559 * try to report up to its nonexistent parent!
2561 rcu_report_qs_rsp(rsp, flags);
2565 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2567 mask = rnp->grpmask;
2568 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2569 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2570 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2574 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2575 * structure. This must be called from the specified CPU.
2578 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2580 unsigned long flags;
2583 struct rcu_node *rnp;
2586 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2587 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2588 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2591 * The grace period in which this quiescent state was
2592 * recorded has ended, so don't report it upwards.
2593 * We will instead need a new quiescent state that lies
2594 * within the current grace period.
2596 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2597 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2598 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2601 mask = rdp->grpmask;
2602 if ((rnp->qsmask & mask) == 0) {
2603 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2605 rdp->core_needs_qs = false;
2608 * This GP can't end until cpu checks in, so all of our
2609 * callbacks can be processed during the next GP.
2611 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2613 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2614 /* ^^^ Released rnp->lock */
2616 rcu_gp_kthread_wake(rsp);
2621 * Check to see if there is a new grace period of which this CPU
2622 * is not yet aware, and if so, set up local rcu_data state for it.
2623 * Otherwise, see if this CPU has just passed through its first
2624 * quiescent state for this grace period, and record that fact if so.
2627 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2629 /* Check for grace-period ends and beginnings. */
2630 note_gp_changes(rsp, rdp);
2633 * Does this CPU still need to do its part for current grace period?
2634 * If no, return and let the other CPUs do their part as well.
2636 if (!rdp->core_needs_qs)
2640 * Was there a quiescent state since the beginning of the grace
2641 * period? If no, then exit and wait for the next call.
2643 if (rdp->cpu_no_qs.b.norm)
2647 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2650 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2654 * Send the specified CPU's RCU callbacks to the orphanage. The
2655 * specified CPU must be offline, and the caller must hold the
2659 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2660 struct rcu_node *rnp, struct rcu_data *rdp)
2662 /* No-CBs CPUs do not have orphanable callbacks. */
2663 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2667 * Orphan the callbacks. First adjust the counts. This is safe
2668 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2669 * cannot be running now. Thus no memory barrier is required.
2671 if (rdp->nxtlist != NULL) {
2672 rsp->qlen_lazy += rdp->qlen_lazy;
2673 rsp->qlen += rdp->qlen;
2674 rdp->n_cbs_orphaned += rdp->qlen;
2676 WRITE_ONCE(rdp->qlen, 0);
2680 * Next, move those callbacks still needing a grace period to
2681 * the orphanage, where some other CPU will pick them up.
2682 * Some of the callbacks might have gone partway through a grace
2683 * period, but that is too bad. They get to start over because we
2684 * cannot assume that grace periods are synchronized across CPUs.
2685 * We don't bother updating the ->nxttail[] array yet, instead
2686 * we just reset the whole thing later on.
2688 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2689 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2690 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2691 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2695 * Then move the ready-to-invoke callbacks to the orphanage,
2696 * where some other CPU will pick them up. These will not be
2697 * required to pass though another grace period: They are done.
2699 if (rdp->nxtlist != NULL) {
2700 *rsp->orphan_donetail = rdp->nxtlist;
2701 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2705 * Finally, initialize the rcu_data structure's list to empty and
2706 * disallow further callbacks on this CPU.
2708 init_callback_list(rdp);
2709 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2713 * Adopt the RCU callbacks from the specified rcu_state structure's
2714 * orphanage. The caller must hold the ->orphan_lock.
2716 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2719 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2721 /* No-CBs CPUs are handled specially. */
2722 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2723 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2726 /* Do the accounting first. */
2727 rdp->qlen_lazy += rsp->qlen_lazy;
2728 rdp->qlen += rsp->qlen;
2729 rdp->n_cbs_adopted += rsp->qlen;
2730 if (rsp->qlen_lazy != rsp->qlen)
2731 rcu_idle_count_callbacks_posted();
2736 * We do not need a memory barrier here because the only way we
2737 * can get here if there is an rcu_barrier() in flight is if
2738 * we are the task doing the rcu_barrier().
2741 /* First adopt the ready-to-invoke callbacks. */
2742 if (rsp->orphan_donelist != NULL) {
2743 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2744 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2745 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2746 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2747 rdp->nxttail[i] = rsp->orphan_donetail;
2748 rsp->orphan_donelist = NULL;
2749 rsp->orphan_donetail = &rsp->orphan_donelist;
2752 /* And then adopt the callbacks that still need a grace period. */
2753 if (rsp->orphan_nxtlist != NULL) {
2754 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2755 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2756 rsp->orphan_nxtlist = NULL;
2757 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2762 * Trace the fact that this CPU is going offline.
2764 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2766 RCU_TRACE(unsigned long mask;)
2767 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2768 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2770 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2773 RCU_TRACE(mask = rdp->grpmask;)
2774 trace_rcu_grace_period(rsp->name,
2775 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2780 * All CPUs for the specified rcu_node structure have gone offline,
2781 * and all tasks that were preempted within an RCU read-side critical
2782 * section while running on one of those CPUs have since exited their RCU
2783 * read-side critical section. Some other CPU is reporting this fact with
2784 * the specified rcu_node structure's ->lock held and interrupts disabled.
2785 * This function therefore goes up the tree of rcu_node structures,
2786 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2787 * the leaf rcu_node structure's ->qsmaskinit field has already been
2790 * This function does check that the specified rcu_node structure has
2791 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2792 * prematurely. That said, invoking it after the fact will cost you
2793 * a needless lock acquisition. So once it has done its work, don't
2796 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2799 struct rcu_node *rnp = rnp_leaf;
2801 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2802 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2805 mask = rnp->grpmask;
2809 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2810 rnp->qsmaskinit &= ~mask;
2811 rnp->qsmask &= ~mask;
2812 if (rnp->qsmaskinit) {
2813 raw_spin_unlock_rcu_node(rnp);
2814 /* irqs remain disabled. */
2817 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2822 * The CPU has been completely removed, and some other CPU is reporting
2823 * this fact from process context. Do the remainder of the cleanup,
2824 * including orphaning the outgoing CPU's RCU callbacks, and also
2825 * adopting them. There can only be one CPU hotplug operation at a time,
2826 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2828 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2830 unsigned long flags;
2831 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2832 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2834 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2837 /* Adjust any no-longer-needed kthreads. */
2838 rcu_boost_kthread_setaffinity(rnp, -1);
2840 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2841 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2842 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2843 rcu_adopt_orphan_cbs(rsp, flags);
2844 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2846 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2847 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2848 cpu, rdp->qlen, rdp->nxtlist);
2852 * Invoke any RCU callbacks that have made it to the end of their grace
2853 * period. Thottle as specified by rdp->blimit.
2855 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2857 unsigned long flags;
2858 struct rcu_head *next, *list, **tail;
2859 long bl, count, count_lazy;
2862 /* If no callbacks are ready, just return. */
2863 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2864 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2865 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2866 need_resched(), is_idle_task(current),
2867 rcu_is_callbacks_kthread());
2872 * Extract the list of ready callbacks, disabling to prevent
2873 * races with call_rcu() from interrupt handlers.
2875 local_irq_save(flags);
2876 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2878 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2879 list = rdp->nxtlist;
2880 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2881 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2882 tail = rdp->nxttail[RCU_DONE_TAIL];
2883 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2884 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2885 rdp->nxttail[i] = &rdp->nxtlist;
2886 local_irq_restore(flags);
2888 /* Invoke callbacks. */
2889 count = count_lazy = 0;
2893 debug_rcu_head_unqueue(list);
2894 if (__rcu_reclaim(rsp->name, list))
2897 /* Stop only if limit reached and CPU has something to do. */
2898 if (++count >= bl &&
2900 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2904 local_irq_save(flags);
2905 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2906 is_idle_task(current),
2907 rcu_is_callbacks_kthread());
2909 /* Update count, and requeue any remaining callbacks. */
2911 *tail = rdp->nxtlist;
2912 rdp->nxtlist = list;
2913 for (i = 0; i < RCU_NEXT_SIZE; i++)
2914 if (&rdp->nxtlist == rdp->nxttail[i])
2915 rdp->nxttail[i] = tail;
2919 smp_mb(); /* List handling before counting for rcu_barrier(). */
2920 rdp->qlen_lazy -= count_lazy;
2921 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2922 rdp->n_cbs_invoked += count;
2924 /* Reinstate batch limit if we have worked down the excess. */
2925 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2926 rdp->blimit = blimit;
2928 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2929 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2930 rdp->qlen_last_fqs_check = 0;
2931 rdp->n_force_qs_snap = rsp->n_force_qs;
2932 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2933 rdp->qlen_last_fqs_check = rdp->qlen;
2934 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2936 local_irq_restore(flags);
2938 /* Re-invoke RCU core processing if there are callbacks remaining. */
2939 if (cpu_has_callbacks_ready_to_invoke(rdp))
2944 * Check to see if this CPU is in a non-context-switch quiescent state
2945 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2946 * Also schedule RCU core processing.
2948 * This function must be called from hardirq context. It is normally
2949 * invoked from the scheduling-clock interrupt.
2951 void rcu_check_callbacks(int user)
2953 trace_rcu_utilization(TPS("Start scheduler-tick"));
2954 increment_cpu_stall_ticks();
2955 if (user || rcu_is_cpu_rrupt_from_idle()) {
2958 * Get here if this CPU took its interrupt from user
2959 * mode or from the idle loop, and if this is not a
2960 * nested interrupt. In this case, the CPU is in
2961 * a quiescent state, so note it.
2963 * No memory barrier is required here because both
2964 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2965 * variables that other CPUs neither access nor modify,
2966 * at least not while the corresponding CPU is online.
2972 } else if (!in_softirq()) {
2975 * Get here if this CPU did not take its interrupt from
2976 * softirq, in other words, if it is not interrupting
2977 * a rcu_bh read-side critical section. This is an _bh
2978 * critical section, so note it.
2983 rcu_preempt_check_callbacks();
2987 rcu_note_voluntary_context_switch(current);
2988 trace_rcu_utilization(TPS("End scheduler-tick"));
2992 * Scan the leaf rcu_node structures, processing dyntick state for any that
2993 * have not yet encountered a quiescent state, using the function specified.
2994 * Also initiate boosting for any threads blocked on the root rcu_node.
2996 * The caller must have suppressed start of new grace periods.
2998 static void force_qs_rnp(struct rcu_state *rsp,
2999 int (*f)(struct rcu_data *rsp, bool *isidle,
3000 unsigned long *maxj),
3001 bool *isidle, unsigned long *maxj)
3004 unsigned long flags;
3006 struct rcu_node *rnp;
3008 rcu_for_each_leaf_node(rsp, rnp) {
3009 cond_resched_rcu_qs();
3011 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3012 if (rnp->qsmask == 0) {
3013 if (rcu_state_p == &rcu_sched_state ||
3014 rsp != rcu_state_p ||
3015 rcu_preempt_blocked_readers_cgp(rnp)) {
3017 * No point in scanning bits because they
3018 * are all zero. But we might need to
3019 * priority-boost blocked readers.
3021 rcu_initiate_boost(rnp, flags);
3022 /* rcu_initiate_boost() releases rnp->lock */
3026 (rnp->parent->qsmask & rnp->grpmask)) {
3028 * Race between grace-period
3029 * initialization and task exiting RCU
3030 * read-side critical section: Report.
3032 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
3033 /* rcu_report_unblock_qs_rnp() rlses ->lock */
3037 for_each_leaf_node_possible_cpu(rnp, cpu) {
3038 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
3039 if ((rnp->qsmask & bit) != 0) {
3040 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
3045 /* Idle/offline CPUs, report (releases rnp->lock. */
3046 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
3048 /* Nothing to do here, so just drop the lock. */
3049 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3055 * Force quiescent states on reluctant CPUs, and also detect which
3056 * CPUs are in dyntick-idle mode.
3058 static void force_quiescent_state(struct rcu_state *rsp)
3060 unsigned long flags;
3062 struct rcu_node *rnp;
3063 struct rcu_node *rnp_old = NULL;
3065 /* Funnel through hierarchy to reduce memory contention. */
3066 rnp = __this_cpu_read(rsp->rda->mynode);
3067 for (; rnp != NULL; rnp = rnp->parent) {
3068 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
3069 !raw_spin_trylock(&rnp->fqslock);
3070 if (rnp_old != NULL)
3071 raw_spin_unlock(&rnp_old->fqslock);
3073 rsp->n_force_qs_lh++;
3078 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
3080 /* Reached the root of the rcu_node tree, acquire lock. */
3081 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
3082 raw_spin_unlock(&rnp_old->fqslock);
3083 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
3084 rsp->n_force_qs_lh++;
3085 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
3086 return; /* Someone beat us to it. */
3088 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
3089 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
3090 rcu_gp_kthread_wake(rsp);
3094 * This does the RCU core processing work for the specified rcu_state
3095 * and rcu_data structures. This may be called only from the CPU to
3096 * whom the rdp belongs.
3099 __rcu_process_callbacks(struct rcu_state *rsp)
3101 unsigned long flags;
3103 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3105 WARN_ON_ONCE(rdp->beenonline == 0);
3107 /* Update RCU state based on any recent quiescent states. */
3108 rcu_check_quiescent_state(rsp, rdp);
3110 /* Does this CPU require a not-yet-started grace period? */
3111 local_irq_save(flags);
3112 if (cpu_needs_another_gp(rsp, rdp)) {
3113 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
3114 needwake = rcu_start_gp(rsp);
3115 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
3117 rcu_gp_kthread_wake(rsp);
3119 local_irq_restore(flags);
3122 /* If there are callbacks ready, invoke them. */
3123 if (cpu_has_callbacks_ready_to_invoke(rdp))
3124 invoke_rcu_callbacks(rsp, rdp);
3126 /* Do any needed deferred wakeups of rcuo kthreads. */
3127 do_nocb_deferred_wakeup(rdp);
3131 * Do RCU core processing for the current CPU.
3133 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
3135 struct rcu_state *rsp;
3137 if (cpu_is_offline(smp_processor_id()))
3139 trace_rcu_utilization(TPS("Start RCU core"));
3140 for_each_rcu_flavor(rsp)
3141 __rcu_process_callbacks(rsp);
3142 trace_rcu_utilization(TPS("End RCU core"));
3146 * Schedule RCU callback invocation. If the specified type of RCU
3147 * does not support RCU priority boosting, just do a direct call,
3148 * otherwise wake up the per-CPU kernel kthread. Note that because we
3149 * are running on the current CPU with softirqs disabled, the
3150 * rcu_cpu_kthread_task cannot disappear out from under us.
3152 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
3154 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
3156 if (likely(!rsp->boost)) {
3157 rcu_do_batch(rsp, rdp);
3160 invoke_rcu_callbacks_kthread();
3163 static void invoke_rcu_core(void)
3165 if (cpu_online(smp_processor_id()))
3166 raise_softirq(RCU_SOFTIRQ);
3170 * Handle any core-RCU processing required by a call_rcu() invocation.
3172 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
3173 struct rcu_head *head, unsigned long flags)
3178 * If called from an extended quiescent state, invoke the RCU
3179 * core in order to force a re-evaluation of RCU's idleness.
3181 if (!rcu_is_watching())
3184 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3185 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3189 * Force the grace period if too many callbacks or too long waiting.
3190 * Enforce hysteresis, and don't invoke force_quiescent_state()
3191 * if some other CPU has recently done so. Also, don't bother
3192 * invoking force_quiescent_state() if the newly enqueued callback
3193 * is the only one waiting for a grace period to complete.
3195 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
3197 /* Are we ignoring a completed grace period? */
3198 note_gp_changes(rsp, rdp);
3200 /* Start a new grace period if one not already started. */
3201 if (!rcu_gp_in_progress(rsp)) {
3202 struct rcu_node *rnp_root = rcu_get_root(rsp);
3204 raw_spin_lock_rcu_node(rnp_root);
3205 needwake = rcu_start_gp(rsp);
3206 raw_spin_unlock_rcu_node(rnp_root);
3208 rcu_gp_kthread_wake(rsp);
3210 /* Give the grace period a kick. */
3211 rdp->blimit = LONG_MAX;
3212 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3213 *rdp->nxttail[RCU_DONE_TAIL] != head)
3214 force_quiescent_state(rsp);
3215 rdp->n_force_qs_snap = rsp->n_force_qs;
3216 rdp->qlen_last_fqs_check = rdp->qlen;
3222 * RCU callback function to leak a callback.
3224 static void rcu_leak_callback(struct rcu_head *rhp)
3229 * Helper function for call_rcu() and friends. The cpu argument will
3230 * normally be -1, indicating "currently running CPU". It may specify
3231 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3232 * is expected to specify a CPU.
3235 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3236 struct rcu_state *rsp, int cpu, bool lazy)
3238 unsigned long flags;
3239 struct rcu_data *rdp;
3241 /* Misaligned rcu_head! */
3242 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3244 if (debug_rcu_head_queue(head)) {
3245 /* Probable double call_rcu(), so leak the callback. */
3246 WRITE_ONCE(head->func, rcu_leak_callback);
3247 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3252 local_irq_save(flags);
3253 rdp = this_cpu_ptr(rsp->rda);
3255 /* Add the callback to our list. */
3256 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3260 rdp = per_cpu_ptr(rsp->rda, cpu);
3261 if (likely(rdp->mynode)) {
3262 /* Post-boot, so this should be for a no-CBs CPU. */
3263 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3264 WARN_ON_ONCE(offline);
3265 /* Offline CPU, _call_rcu() illegal, leak callback. */
3266 local_irq_restore(flags);
3270 * Very early boot, before rcu_init(). Initialize if needed
3271 * and then drop through to queue the callback.
3274 WARN_ON_ONCE(!rcu_is_watching());
3275 if (!likely(rdp->nxtlist))
3276 init_default_callback_list(rdp);
3278 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3282 rcu_idle_count_callbacks_posted();
3283 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3284 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3285 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3287 if (__is_kfree_rcu_offset((unsigned long)func))
3288 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3289 rdp->qlen_lazy, rdp->qlen);
3291 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3293 /* Go handle any RCU core processing required. */
3294 __call_rcu_core(rsp, rdp, head, flags);
3295 local_irq_restore(flags);
3299 * Queue an RCU-sched callback for invocation after a grace period.
3301 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3303 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3305 EXPORT_SYMBOL_GPL(call_rcu_sched);
3308 * Queue an RCU callback for invocation after a quicker grace period.
3310 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3312 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3314 EXPORT_SYMBOL_GPL(call_rcu_bh);
3317 * Queue an RCU callback for lazy invocation after a grace period.
3318 * This will likely be later named something like "call_rcu_lazy()",
3319 * but this change will require some way of tagging the lazy RCU
3320 * callbacks in the list of pending callbacks. Until then, this
3321 * function may only be called from __kfree_rcu().
3323 void kfree_call_rcu(struct rcu_head *head,
3324 rcu_callback_t func)
3326 __call_rcu(head, func, rcu_state_p, -1, 1);
3328 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3331 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3332 * any blocking grace-period wait automatically implies a grace period
3333 * if there is only one CPU online at any point time during execution
3334 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3335 * occasionally incorrectly indicate that there are multiple CPUs online
3336 * when there was in fact only one the whole time, as this just adds
3337 * some overhead: RCU still operates correctly.
3339 static inline int rcu_blocking_is_gp(void)
3343 might_sleep(); /* Check for RCU read-side critical section. */
3345 ret = num_online_cpus() <= 1;
3351 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3353 * Control will return to the caller some time after a full rcu-sched
3354 * grace period has elapsed, in other words after all currently executing
3355 * rcu-sched read-side critical sections have completed. These read-side
3356 * critical sections are delimited by rcu_read_lock_sched() and
3357 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3358 * local_irq_disable(), and so on may be used in place of
3359 * rcu_read_lock_sched().
3361 * This means that all preempt_disable code sequences, including NMI and
3362 * non-threaded hardware-interrupt handlers, in progress on entry will
3363 * have completed before this primitive returns. However, this does not
3364 * guarantee that softirq handlers will have completed, since in some
3365 * kernels, these handlers can run in process context, and can block.
3367 * Note that this guarantee implies further memory-ordering guarantees.
3368 * On systems with more than one CPU, when synchronize_sched() returns,
3369 * each CPU is guaranteed to have executed a full memory barrier since the
3370 * end of its last RCU-sched read-side critical section whose beginning
3371 * preceded the call to synchronize_sched(). In addition, each CPU having
3372 * an RCU read-side critical section that extends beyond the return from
3373 * synchronize_sched() is guaranteed to have executed a full memory barrier
3374 * after the beginning of synchronize_sched() and before the beginning of
3375 * that RCU read-side critical section. Note that these guarantees include
3376 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3377 * that are executing in the kernel.
3379 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3380 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3381 * to have executed a full memory barrier during the execution of
3382 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3383 * again only if the system has more than one CPU).
3385 * This primitive provides the guarantees made by the (now removed)
3386 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3387 * guarantees that rcu_read_lock() sections will have completed.
3388 * In "classic RCU", these two guarantees happen to be one and
3389 * the same, but can differ in realtime RCU implementations.
3391 void synchronize_sched(void)
3393 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3394 lock_is_held(&rcu_lock_map) ||
3395 lock_is_held(&rcu_sched_lock_map),
3396 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3397 if (rcu_blocking_is_gp())
3399 if (rcu_gp_is_expedited())
3400 synchronize_sched_expedited();
3402 wait_rcu_gp(call_rcu_sched);
3404 EXPORT_SYMBOL_GPL(synchronize_sched);
3407 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3409 * Control will return to the caller some time after a full rcu_bh grace
3410 * period has elapsed, in other words after all currently executing rcu_bh
3411 * read-side critical sections have completed. RCU read-side critical
3412 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3413 * and may be nested.
3415 * See the description of synchronize_sched() for more detailed information
3416 * on memory ordering guarantees.
3418 void synchronize_rcu_bh(void)
3420 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3421 lock_is_held(&rcu_lock_map) ||
3422 lock_is_held(&rcu_sched_lock_map),
3423 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3424 if (rcu_blocking_is_gp())
3426 if (rcu_gp_is_expedited())
3427 synchronize_rcu_bh_expedited();
3429 wait_rcu_gp(call_rcu_bh);
3431 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3434 * get_state_synchronize_rcu - Snapshot current RCU state
3436 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3437 * to determine whether or not a full grace period has elapsed in the
3440 unsigned long get_state_synchronize_rcu(void)
3443 * Any prior manipulation of RCU-protected data must happen
3444 * before the load from ->gpnum.
3449 * Make sure this load happens before the purportedly
3450 * time-consuming work between get_state_synchronize_rcu()
3451 * and cond_synchronize_rcu().
3453 return smp_load_acquire(&rcu_state_p->gpnum);
3455 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3458 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3460 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3462 * If a full RCU grace period has elapsed since the earlier call to
3463 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3464 * synchronize_rcu() to wait for a full grace period.
3466 * Yes, this function does not take counter wrap into account. But
3467 * counter wrap is harmless. If the counter wraps, we have waited for
3468 * more than 2 billion grace periods (and way more on a 64-bit system!),
3469 * so waiting for one additional grace period should be just fine.
3471 void cond_synchronize_rcu(unsigned long oldstate)
3473 unsigned long newstate;
3476 * Ensure that this load happens before any RCU-destructive
3477 * actions the caller might carry out after we return.
3479 newstate = smp_load_acquire(&rcu_state_p->completed);
3480 if (ULONG_CMP_GE(oldstate, newstate))
3483 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3486 * get_state_synchronize_sched - Snapshot current RCU-sched state
3488 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3489 * to determine whether or not a full grace period has elapsed in the
3492 unsigned long get_state_synchronize_sched(void)
3495 * Any prior manipulation of RCU-protected data must happen
3496 * before the load from ->gpnum.
3501 * Make sure this load happens before the purportedly
3502 * time-consuming work between get_state_synchronize_sched()
3503 * and cond_synchronize_sched().
3505 return smp_load_acquire(&rcu_sched_state.gpnum);
3507 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3510 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3512 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3514 * If a full RCU-sched grace period has elapsed since the earlier call to
3515 * get_state_synchronize_sched(), just return. Otherwise, invoke
3516 * synchronize_sched() to wait for a full grace period.
3518 * Yes, this function does not take counter wrap into account. But
3519 * counter wrap is harmless. If the counter wraps, we have waited for
3520 * more than 2 billion grace periods (and way more on a 64-bit system!),
3521 * so waiting for one additional grace period should be just fine.
3523 void cond_synchronize_sched(unsigned long oldstate)
3525 unsigned long newstate;
3528 * Ensure that this load happens before any RCU-destructive
3529 * actions the caller might carry out after we return.
3531 newstate = smp_load_acquire(&rcu_sched_state.completed);
3532 if (ULONG_CMP_GE(oldstate, newstate))
3533 synchronize_sched();
3535 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3537 /* Adjust sequence number for start of update-side operation. */
3538 static void rcu_seq_start(unsigned long *sp)
3540 WRITE_ONCE(*sp, *sp + 1);
3541 smp_mb(); /* Ensure update-side operation after counter increment. */
3542 WARN_ON_ONCE(!(*sp & 0x1));
3545 /* Adjust sequence number for end of update-side operation. */
3546 static void rcu_seq_end(unsigned long *sp)
3548 smp_mb(); /* Ensure update-side operation before counter increment. */
3549 WRITE_ONCE(*sp, *sp + 1);
3550 WARN_ON_ONCE(*sp & 0x1);
3553 /* Take a snapshot of the update side's sequence number. */
3554 static unsigned long rcu_seq_snap(unsigned long *sp)
3558 s = (READ_ONCE(*sp) + 3) & ~0x1;
3559 smp_mb(); /* Above access must not bleed into critical section. */
3564 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3565 * full update-side operation has occurred.
3567 static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3569 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3573 * Check to see if there is any immediate RCU-related work to be done
3574 * by the current CPU, for the specified type of RCU, returning 1 if so.
3575 * The checks are in order of increasing expense: checks that can be
3576 * carried out against CPU-local state are performed first. However,
3577 * we must check for CPU stalls first, else we might not get a chance.
3579 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3581 struct rcu_node *rnp = rdp->mynode;
3583 rdp->n_rcu_pending++;
3585 /* Check for CPU stalls, if enabled. */
3586 check_cpu_stall(rsp, rdp);
3588 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3589 if (rcu_nohz_full_cpu(rsp))
3592 /* Is the RCU core waiting for a quiescent state from this CPU? */
3593 if (rcu_scheduler_fully_active &&
3594 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3595 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
3596 rdp->n_rp_core_needs_qs++;
3597 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3598 rdp->n_rp_report_qs++;
3602 /* Does this CPU have callbacks ready to invoke? */
3603 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3604 rdp->n_rp_cb_ready++;
3608 /* Has RCU gone idle with this CPU needing another grace period? */
3609 if (cpu_needs_another_gp(rsp, rdp)) {
3610 rdp->n_rp_cpu_needs_gp++;
3614 /* Has another RCU grace period completed? */
3615 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3616 rdp->n_rp_gp_completed++;
3620 /* Has a new RCU grace period started? */
3621 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3622 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3623 rdp->n_rp_gp_started++;
3627 /* Does this CPU need a deferred NOCB wakeup? */
3628 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3629 rdp->n_rp_nocb_defer_wakeup++;
3634 rdp->n_rp_need_nothing++;
3639 * Check to see if there is any immediate RCU-related work to be done
3640 * by the current CPU, returning 1 if so. This function is part of the
3641 * RCU implementation; it is -not- an exported member of the RCU API.
3643 static int rcu_pending(void)
3645 struct rcu_state *rsp;
3647 for_each_rcu_flavor(rsp)
3648 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3654 * Return true if the specified CPU has any callback. If all_lazy is
3655 * non-NULL, store an indication of whether all callbacks are lazy.
3656 * (If there are no callbacks, all of them are deemed to be lazy.)
3658 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3662 struct rcu_data *rdp;
3663 struct rcu_state *rsp;
3665 for_each_rcu_flavor(rsp) {
3666 rdp = this_cpu_ptr(rsp->rda);
3670 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3681 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3682 * the compiler is expected to optimize this away.
3684 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3685 int cpu, unsigned long done)
3687 trace_rcu_barrier(rsp->name, s, cpu,
3688 atomic_read(&rsp->barrier_cpu_count), done);
3692 * RCU callback function for _rcu_barrier(). If we are last, wake
3693 * up the task executing _rcu_barrier().
3695 static void rcu_barrier_callback(struct rcu_head *rhp)
3697 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3698 struct rcu_state *rsp = rdp->rsp;
3700 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3701 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
3702 complete(&rsp->barrier_completion);
3704 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
3709 * Called with preemption disabled, and from cross-cpu IRQ context.
3711 static void rcu_barrier_func(void *type)
3713 struct rcu_state *rsp = type;
3714 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3716 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
3717 atomic_inc(&rsp->barrier_cpu_count);
3718 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3722 * Orchestrate the specified type of RCU barrier, waiting for all
3723 * RCU callbacks of the specified type to complete.
3725 static void _rcu_barrier(struct rcu_state *rsp)
3728 struct rcu_data *rdp;
3729 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3731 _rcu_barrier_trace(rsp, "Begin", -1, s);
3733 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3734 mutex_lock(&rsp->barrier_mutex);
3736 /* Did someone else do our work for us? */
3737 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3738 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
3739 smp_mb(); /* caller's subsequent code after above check. */
3740 mutex_unlock(&rsp->barrier_mutex);
3744 /* Mark the start of the barrier operation. */
3745 rcu_seq_start(&rsp->barrier_sequence);
3746 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
3749 * Initialize the count to one rather than to zero in order to
3750 * avoid a too-soon return to zero in case of a short grace period
3751 * (or preemption of this task). Exclude CPU-hotplug operations
3752 * to ensure that no offline CPU has callbacks queued.
3754 init_completion(&rsp->barrier_completion);
3755 atomic_set(&rsp->barrier_cpu_count, 1);
3759 * Force each CPU with callbacks to register a new callback.
3760 * When that callback is invoked, we will know that all of the
3761 * corresponding CPU's preceding callbacks have been invoked.
3763 for_each_possible_cpu(cpu) {
3764 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3766 rdp = per_cpu_ptr(rsp->rda, cpu);
3767 if (rcu_is_nocb_cpu(cpu)) {
3768 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3769 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3770 rsp->barrier_sequence);
3772 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3773 rsp->barrier_sequence);
3774 smp_mb__before_atomic();
3775 atomic_inc(&rsp->barrier_cpu_count);
3776 __call_rcu(&rdp->barrier_head,
3777 rcu_barrier_callback, rsp, cpu, 0);
3779 } else if (READ_ONCE(rdp->qlen)) {
3780 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3781 rsp->barrier_sequence);
3782 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3784 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3785 rsp->barrier_sequence);
3791 * Now that we have an rcu_barrier_callback() callback on each
3792 * CPU, and thus each counted, remove the initial count.
3794 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3795 complete(&rsp->barrier_completion);
3797 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3798 wait_for_completion(&rsp->barrier_completion);
3800 /* Mark the end of the barrier operation. */
3801 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
3802 rcu_seq_end(&rsp->barrier_sequence);
3804 /* Other rcu_barrier() invocations can now safely proceed. */
3805 mutex_unlock(&rsp->barrier_mutex);
3809 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3811 void rcu_barrier_bh(void)
3813 _rcu_barrier(&rcu_bh_state);
3815 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3818 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3820 void rcu_barrier_sched(void)
3822 _rcu_barrier(&rcu_sched_state);
3824 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3827 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3828 * first CPU in a given leaf rcu_node structure coming online. The caller
3829 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3832 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3835 struct rcu_node *rnp = rnp_leaf;
3838 mask = rnp->grpmask;
3842 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3843 rnp->qsmaskinit |= mask;
3844 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3849 * Do boot-time initialization of a CPU's per-CPU RCU data.
3852 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3854 unsigned long flags;
3855 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3856 struct rcu_node *rnp = rcu_get_root(rsp);
3858 /* Set up local state, ensuring consistent view of global state. */
3859 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3860 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3861 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3862 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3863 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3866 rcu_boot_init_nocb_percpu_data(rdp);
3867 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3871 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3872 * offline event can be happening at a given time. Note also that we
3873 * can accept some slop in the rsp->completed access due to the fact
3874 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3877 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3879 unsigned long flags;
3880 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3881 struct rcu_node *rnp = rcu_get_root(rsp);
3883 /* Set up local state, ensuring consistent view of global state. */
3884 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3885 rdp->qlen_last_fqs_check = 0;
3886 rdp->n_force_qs_snap = rsp->n_force_qs;
3887 rdp->blimit = blimit;
3889 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3890 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3891 rcu_sysidle_init_percpu_data(rdp->dynticks);
3892 rcu_dynticks_eqs_online();
3893 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3896 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3897 * propagation up the rcu_node tree will happen at the beginning
3898 * of the next grace period.
3901 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3902 if (!rdp->beenonline)
3903 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
3904 rdp->beenonline = true; /* We have now been online. */
3905 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3906 rdp->completed = rnp->completed;
3907 rdp->cpu_no_qs.b.norm = true;
3908 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3909 rdp->core_needs_qs = false;
3910 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3911 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3914 int rcutree_prepare_cpu(unsigned int cpu)
3916 struct rcu_state *rsp;
3918 for_each_rcu_flavor(rsp)
3919 rcu_init_percpu_data(cpu, rsp);
3921 rcu_prepare_kthreads(cpu);
3922 rcu_spawn_all_nocb_kthreads(cpu);
3927 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3929 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3931 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3934 int rcutree_online_cpu(unsigned int cpu)
3936 sync_sched_exp_online_cleanup(cpu);
3937 rcutree_affinity_setting(cpu, -1);
3941 int rcutree_offline_cpu(unsigned int cpu)
3943 rcutree_affinity_setting(cpu, cpu);
3948 int rcutree_dying_cpu(unsigned int cpu)
3950 struct rcu_state *rsp;
3952 for_each_rcu_flavor(rsp)
3953 rcu_cleanup_dying_cpu(rsp);
3957 int rcutree_dead_cpu(unsigned int cpu)
3959 struct rcu_state *rsp;
3961 for_each_rcu_flavor(rsp) {
3962 rcu_cleanup_dead_cpu(cpu, rsp);
3963 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3969 * Mark the specified CPU as being online so that subsequent grace periods
3970 * (both expedited and normal) will wait on it. Note that this means that
3971 * incoming CPUs are not allowed to use RCU read-side critical sections
3972 * until this function is called. Failing to observe this restriction
3973 * will result in lockdep splats.
3975 void rcu_cpu_starting(unsigned int cpu)
3977 unsigned long flags;
3979 struct rcu_data *rdp;
3980 struct rcu_node *rnp;
3981 struct rcu_state *rsp;
3983 for_each_rcu_flavor(rsp) {
3984 rdp = per_cpu_ptr(rsp->rda, cpu);
3986 mask = rdp->grpmask;
3987 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3988 rnp->qsmaskinitnext |= mask;
3989 rnp->expmaskinitnext |= mask;
3990 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3994 #ifdef CONFIG_HOTPLUG_CPU
3996 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3997 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3999 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4000 * function. We now remove it from the rcu_node tree's ->qsmaskinit
4003 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
4005 unsigned long flags;
4007 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4008 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4010 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4011 mask = rdp->grpmask;
4012 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4013 rnp->qsmaskinitnext &= ~mask;
4014 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4017 void rcu_report_dead(unsigned int cpu)
4019 struct rcu_state *rsp;
4021 /* QS for any half-done expedited RCU-sched GP. */
4023 rcu_report_exp_rdp(&rcu_sched_state,
4024 this_cpu_ptr(rcu_sched_state.rda), true);
4026 for_each_rcu_flavor(rsp)
4027 rcu_cleanup_dying_idle_cpu(cpu, rsp);
4031 static int rcu_pm_notify(struct notifier_block *self,
4032 unsigned long action, void *hcpu)
4035 case PM_HIBERNATION_PREPARE:
4036 case PM_SUSPEND_PREPARE:
4037 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4040 case PM_POST_HIBERNATION:
4041 case PM_POST_SUSPEND:
4042 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4043 rcu_unexpedite_gp();
4052 * Spawn the kthreads that handle each RCU flavor's grace periods.
4054 static int __init rcu_spawn_gp_kthread(void)
4056 unsigned long flags;
4057 int kthread_prio_in = kthread_prio;
4058 struct rcu_node *rnp;
4059 struct rcu_state *rsp;
4060 struct sched_param sp;
4061 struct task_struct *t;
4063 /* Force priority into range. */
4064 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4066 else if (kthread_prio < 0)
4068 else if (kthread_prio > 99)
4070 if (kthread_prio != kthread_prio_in)
4071 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4072 kthread_prio, kthread_prio_in);
4074 rcu_scheduler_fully_active = 1;
4075 for_each_rcu_flavor(rsp) {
4076 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4078 rnp = rcu_get_root(rsp);
4079 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4080 rsp->gp_kthread = t;
4082 sp.sched_priority = kthread_prio;
4083 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4085 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4088 rcu_spawn_nocb_kthreads();
4089 rcu_spawn_boost_kthreads();
4092 early_initcall(rcu_spawn_gp_kthread);
4095 * This function is invoked towards the end of the scheduler's
4096 * initialization process. Before this is called, the idle task might
4097 * contain synchronous grace-period primitives (during which time, this idle
4098 * task is booting the system, and such primitives are no-ops). After this
4099 * function is called, any synchronous grace-period primitives are run as
4100 * expedited, with the requesting task driving the grace period forward.
4101 * A later core_initcall() rcu_exp_runtime_mode() will switch to full
4102 * runtime RCU functionality.
4104 void rcu_scheduler_starting(void)
4106 WARN_ON(num_online_cpus() != 1);
4107 WARN_ON(nr_context_switches() > 0);
4108 rcu_test_sync_prims();
4109 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4110 rcu_test_sync_prims();
4114 * Compute the per-level fanout, either using the exact fanout specified
4115 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4117 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4121 if (rcu_fanout_exact) {
4122 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4123 for (i = rcu_num_lvls - 2; i >= 0; i--)
4124 levelspread[i] = RCU_FANOUT;
4130 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4132 levelspread[i] = (cprv + ccur - 1) / ccur;
4139 * Helper function for rcu_init() that initializes one rcu_state structure.
4141 static void __init rcu_init_one(struct rcu_state *rsp)
4143 static const char * const buf[] = RCU_NODE_NAME_INIT;
4144 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4145 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4146 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4148 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4149 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4153 struct rcu_node *rnp;
4155 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4157 /* Silence gcc 4.8 false positive about array index out of range. */
4158 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4159 panic("rcu_init_one: rcu_num_lvls out of range");
4161 /* Initialize the level-tracking arrays. */
4163 for (i = 0; i < rcu_num_lvls; i++)
4164 levelcnt[i] = num_rcu_lvl[i];
4165 for (i = 1; i < rcu_num_lvls; i++)
4166 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4167 rcu_init_levelspread(levelspread, levelcnt);
4169 /* Initialize the elements themselves, starting from the leaves. */
4171 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4172 cpustride *= levelspread[i];
4173 rnp = rsp->level[i];
4174 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4175 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4176 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4177 &rcu_node_class[i], buf[i]);
4178 raw_spin_lock_init(&rnp->fqslock);
4179 lockdep_set_class_and_name(&rnp->fqslock,
4180 &rcu_fqs_class[i], fqs[i]);
4181 rnp->gpnum = rsp->gpnum;
4182 rnp->completed = rsp->completed;
4184 rnp->qsmaskinit = 0;
4185 rnp->grplo = j * cpustride;
4186 rnp->grphi = (j + 1) * cpustride - 1;
4187 if (rnp->grphi >= nr_cpu_ids)
4188 rnp->grphi = nr_cpu_ids - 1;
4194 rnp->grpnum = j % levelspread[i - 1];
4195 rnp->grpmask = 1UL << rnp->grpnum;
4196 rnp->parent = rsp->level[i - 1] +
4197 j / levelspread[i - 1];
4200 INIT_LIST_HEAD(&rnp->blkd_tasks);
4201 rcu_init_one_nocb(rnp);
4202 init_waitqueue_head(&rnp->exp_wq[0]);
4203 init_waitqueue_head(&rnp->exp_wq[1]);
4204 init_waitqueue_head(&rnp->exp_wq[2]);
4205 init_waitqueue_head(&rnp->exp_wq[3]);
4206 spin_lock_init(&rnp->exp_lock);
4210 init_swait_queue_head(&rsp->gp_wq);
4211 init_swait_queue_head(&rsp->expedited_wq);
4212 rnp = rsp->level[rcu_num_lvls - 1];
4213 for_each_possible_cpu(i) {
4214 while (i > rnp->grphi)
4216 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4217 rcu_boot_init_percpu_data(i, rsp);
4219 list_add(&rsp->flavors, &rcu_struct_flavors);
4223 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4224 * replace the definitions in tree.h because those are needed to size
4225 * the ->node array in the rcu_state structure.
4227 static void __init rcu_init_geometry(void)
4231 int rcu_capacity[RCU_NUM_LVLS];
4234 * Initialize any unspecified boot parameters.
4235 * The default values of jiffies_till_first_fqs and
4236 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4237 * value, which is a function of HZ, then adding one for each
4238 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4240 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4241 if (jiffies_till_first_fqs == ULONG_MAX)
4242 jiffies_till_first_fqs = d;
4243 if (jiffies_till_next_fqs == ULONG_MAX)
4244 jiffies_till_next_fqs = d;
4246 /* If the compile-time values are accurate, just leave. */
4247 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4248 nr_cpu_ids == NR_CPUS)
4250 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4251 rcu_fanout_leaf, nr_cpu_ids);
4254 * The boot-time rcu_fanout_leaf parameter must be at least two
4255 * and cannot exceed the number of bits in the rcu_node masks.
4256 * Complain and fall back to the compile-time values if this
4257 * limit is exceeded.
4259 if (rcu_fanout_leaf < 2 ||
4260 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4261 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4267 * Compute number of nodes that can be handled an rcu_node tree
4268 * with the given number of levels.
4270 rcu_capacity[0] = rcu_fanout_leaf;
4271 for (i = 1; i < RCU_NUM_LVLS; i++)
4272 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4275 * The tree must be able to accommodate the configured number of CPUs.
4276 * If this limit is exceeded, fall back to the compile-time values.
4278 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4279 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4284 /* Calculate the number of levels in the tree. */
4285 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4287 rcu_num_lvls = i + 1;
4289 /* Calculate the number of rcu_nodes at each level of the tree. */
4290 for (i = 0; i < rcu_num_lvls; i++) {
4291 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4292 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4295 /* Calculate the total number of rcu_node structures. */
4297 for (i = 0; i < rcu_num_lvls; i++)
4298 rcu_num_nodes += num_rcu_lvl[i];
4302 * Dump out the structure of the rcu_node combining tree associated
4303 * with the rcu_state structure referenced by rsp.
4305 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4308 struct rcu_node *rnp;
4310 pr_info("rcu_node tree layout dump\n");
4312 rcu_for_each_node_breadth_first(rsp, rnp) {
4313 if (rnp->level != level) {
4318 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4323 void __init rcu_init(void)
4327 rcu_early_boot_tests();
4329 rcu_bootup_announce();
4330 rcu_init_geometry();
4331 rcu_init_one(&rcu_bh_state);
4332 rcu_init_one(&rcu_sched_state);
4334 rcu_dump_rcu_node_tree(&rcu_sched_state);
4335 __rcu_init_preempt();
4336 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4339 * We don't need protection against CPU-hotplug here because
4340 * this is called early in boot, before either interrupts
4341 * or the scheduler are operational.
4343 pm_notifier(rcu_pm_notify, 0);
4344 for_each_online_cpu(cpu) {
4345 rcutree_prepare_cpu(cpu);
4346 rcu_cpu_starting(cpu);
4350 #include "tree_exp.h"
4351 #include "tree_plugin.h"