2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate_wait.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/sched/debug.h>
39 #include <linux/nmi.h>
40 #include <linux/atomic.h>
41 #include <linux/bitops.h>
42 #include <linux/export.h>
43 #include <linux/completion.h>
44 #include <linux/moduleparam.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <uapi/linux/sched/types.h>
54 #include <linux/prefetch.h>
55 #include <linux/delay.h>
56 #include <linux/stop_machine.h>
57 #include <linux/random.h>
58 #include <linux/trace_events.h>
59 #include <linux/suspend.h>
60 #include <linux/ftrace.h>
65 #ifdef MODULE_PARAM_PREFIX
66 #undef MODULE_PARAM_PREFIX
68 #define MODULE_PARAM_PREFIX "rcutree."
70 /* Data structures. */
73 * In order to export the rcu_state name to the tracing tools, it
74 * needs to be added in the __tracepoint_string section.
75 * This requires defining a separate variable tp_<sname>_varname
76 * that points to the string being used, and this will allow
77 * the tracing userspace tools to be able to decipher the string
78 * address to the matching string.
81 # define DEFINE_RCU_TPS(sname) \
82 static char sname##_varname[] = #sname; \
83 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
84 # define RCU_STATE_NAME(sname) sname##_varname
86 # define DEFINE_RCU_TPS(sname)
87 # define RCU_STATE_NAME(sname) __stringify(sname)
90 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
91 DEFINE_RCU_TPS(sname) \
92 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
93 struct rcu_state sname##_state = { \
94 .level = { &sname##_state.node[0] }, \
95 .rda = &sname##_data, \
97 .gp_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
101 .orphan_pend = RCU_CBLIST_INITIALIZER(sname##_state.orphan_pend), \
102 .orphan_done = RCU_CBLIST_INITIALIZER(sname##_state.orphan_done), \
103 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
104 .name = RCU_STATE_NAME(sname), \
106 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
107 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
110 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
111 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
113 static struct rcu_state *const rcu_state_p;
114 LIST_HEAD(rcu_struct_flavors);
116 /* Dump rcu_node combining tree at boot to verify correct setup. */
117 static bool dump_tree;
118 module_param(dump_tree, bool, 0444);
119 /* Control rcu_node-tree auto-balancing at boot time. */
120 static bool rcu_fanout_exact;
121 module_param(rcu_fanout_exact, bool, 0444);
122 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
123 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
124 module_param(rcu_fanout_leaf, int, 0444);
125 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
126 /* Number of rcu_nodes at specified level. */
127 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
128 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
129 /* panic() on RCU Stall sysctl. */
130 int sysctl_panic_on_rcu_stall __read_mostly;
133 * The rcu_scheduler_active variable is initialized to the value
134 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
135 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
136 * RCU can assume that there is but one task, allowing RCU to (for example)
137 * optimize synchronize_rcu() to a simple barrier(). When this variable
138 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
139 * to detect real grace periods. This variable is also used to suppress
140 * boot-time false positives from lockdep-RCU error checking. Finally, it
141 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
142 * is fully initialized, including all of its kthreads having been spawned.
144 int rcu_scheduler_active __read_mostly;
145 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
148 * The rcu_scheduler_fully_active variable transitions from zero to one
149 * during the early_initcall() processing, which is after the scheduler
150 * is capable of creating new tasks. So RCU processing (for example,
151 * creating tasks for RCU priority boosting) must be delayed until after
152 * rcu_scheduler_fully_active transitions from zero to one. We also
153 * currently delay invocation of any RCU callbacks until after this point.
155 * It might later prove better for people registering RCU callbacks during
156 * early boot to take responsibility for these callbacks, but one step at
159 static int rcu_scheduler_fully_active __read_mostly;
161 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
162 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
163 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
164 static void invoke_rcu_core(void);
165 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
166 static void rcu_report_exp_rdp(struct rcu_state *rsp,
167 struct rcu_data *rdp, bool wake);
168 static void sync_sched_exp_online_cleanup(int cpu);
170 /* rcuc/rcub kthread realtime priority */
171 #ifdef CONFIG_RCU_KTHREAD_PRIO
172 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
173 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
174 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
175 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
176 module_param(kthread_prio, int, 0644);
178 /* Delay in jiffies for grace-period initialization delays, debug only. */
180 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
181 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
182 module_param(gp_preinit_delay, int, 0644);
183 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
184 static const int gp_preinit_delay;
185 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
187 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
188 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
189 module_param(gp_init_delay, int, 0644);
190 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
191 static const int gp_init_delay;
192 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
194 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
195 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
196 module_param(gp_cleanup_delay, int, 0644);
197 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
198 static const int gp_cleanup_delay;
199 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
202 * Number of grace periods between delays, normalized by the duration of
203 * the delay. The longer the delay, the more the grace periods between
204 * each delay. The reason for this normalization is that it means that,
205 * for non-zero delays, the overall slowdown of grace periods is constant
206 * regardless of the duration of the delay. This arrangement balances
207 * the need for long delays to increase some race probabilities with the
208 * need for fast grace periods to increase other race probabilities.
210 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
213 * Track the rcutorture test sequence number and the update version
214 * number within a given test. The rcutorture_testseq is incremented
215 * on every rcutorture module load and unload, so has an odd value
216 * when a test is running. The rcutorture_vernum is set to zero
217 * when rcutorture starts and is incremented on each rcutorture update.
218 * These variables enable correlating rcutorture output with the
219 * RCU tracing information.
221 unsigned long rcutorture_testseq;
222 unsigned long rcutorture_vernum;
225 * Compute the mask of online CPUs for the specified rcu_node structure.
226 * This will not be stable unless the rcu_node structure's ->lock is
227 * held, but the bit corresponding to the current CPU will be stable
230 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
232 return READ_ONCE(rnp->qsmaskinitnext);
236 * Return true if an RCU grace period is in progress. The READ_ONCE()s
237 * permit this function to be invoked without holding the root rcu_node
238 * structure's ->lock, but of course results can be subject to change.
240 static int rcu_gp_in_progress(struct rcu_state *rsp)
242 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
246 * Note a quiescent state. Because we do not need to know
247 * how many quiescent states passed, just if there was at least
248 * one since the start of the grace period, this just sets a flag.
249 * The caller must have disabled preemption.
251 void rcu_sched_qs(void)
253 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
255 trace_rcu_grace_period(TPS("rcu_sched"),
256 __this_cpu_read(rcu_sched_data.gpnum),
258 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
259 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
261 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
262 rcu_report_exp_rdp(&rcu_sched_state,
263 this_cpu_ptr(&rcu_sched_data), true);
268 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
269 trace_rcu_grace_period(TPS("rcu_bh"),
270 __this_cpu_read(rcu_bh_data.gpnum),
272 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
277 * Steal a bit from the bottom of ->dynticks for idle entry/exit
278 * control. Initially this is for TLB flushing.
280 #define RCU_DYNTICK_CTRL_MASK 0x1
281 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
282 #ifndef rcu_eqs_special_exit
283 #define rcu_eqs_special_exit() do { } while (0)
286 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
287 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
288 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
289 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
290 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
291 .dynticks_idle = ATOMIC_INIT(1),
292 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
296 * There's a few places, currently just in the tracing infrastructure,
297 * that uses rcu_irq_enter() to make sure RCU is watching. But there's
298 * a small location where that will not even work. In those cases
299 * rcu_irq_enter_disabled() needs to be checked to make sure rcu_irq_enter()
302 static DEFINE_PER_CPU(bool, disable_rcu_irq_enter);
304 bool rcu_irq_enter_disabled(void)
306 return this_cpu_read(disable_rcu_irq_enter);
310 * Record entry into an extended quiescent state. This is only to be
311 * called when not already in an extended quiescent state.
313 static void rcu_dynticks_eqs_enter(void)
315 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
319 * CPUs seeing atomic_add_return() must see prior RCU read-side
320 * critical sections, and we also must force ordering with the
323 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
324 /* Better be in an extended quiescent state! */
325 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
326 (seq & RCU_DYNTICK_CTRL_CTR));
327 /* Better not have special action (TLB flush) pending! */
328 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
329 (seq & RCU_DYNTICK_CTRL_MASK));
333 * Record exit from an extended quiescent state. This is only to be
334 * called from an extended quiescent state.
336 static void rcu_dynticks_eqs_exit(void)
338 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
342 * CPUs seeing atomic_add_return() must see prior idle sojourns,
343 * and we also must force ordering with the next RCU read-side
346 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
347 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
348 !(seq & RCU_DYNTICK_CTRL_CTR));
349 if (seq & RCU_DYNTICK_CTRL_MASK) {
350 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
351 smp_mb__after_atomic(); /* _exit after clearing mask. */
352 /* Prefer duplicate flushes to losing a flush. */
353 rcu_eqs_special_exit();
358 * Reset the current CPU's ->dynticks counter to indicate that the
359 * newly onlined CPU is no longer in an extended quiescent state.
360 * This will either leave the counter unchanged, or increment it
361 * to the next non-quiescent value.
363 * The non-atomic test/increment sequence works because the upper bits
364 * of the ->dynticks counter are manipulated only by the corresponding CPU,
365 * or when the corresponding CPU is offline.
367 static void rcu_dynticks_eqs_online(void)
369 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
371 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
373 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
377 * Is the current CPU in an extended quiescent state?
379 * No ordering, as we are sampling CPU-local information.
381 bool rcu_dynticks_curr_cpu_in_eqs(void)
383 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
385 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
389 * Snapshot the ->dynticks counter with full ordering so as to allow
390 * stable comparison of this counter with past and future snapshots.
392 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
394 int snap = atomic_add_return(0, &rdtp->dynticks);
396 return snap & ~RCU_DYNTICK_CTRL_MASK;
400 * Return true if the snapshot returned from rcu_dynticks_snap()
401 * indicates that RCU is in an extended quiescent state.
403 static bool rcu_dynticks_in_eqs(int snap)
405 return !(snap & RCU_DYNTICK_CTRL_CTR);
409 * Return true if the CPU corresponding to the specified rcu_dynticks
410 * structure has spent some time in an extended quiescent state since
411 * rcu_dynticks_snap() returned the specified snapshot.
413 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
415 return snap != rcu_dynticks_snap(rdtp);
419 * Do a double-increment of the ->dynticks counter to emulate a
420 * momentary idle-CPU quiescent state.
422 static void rcu_dynticks_momentary_idle(void)
424 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
425 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
428 /* It is illegal to call this from idle state. */
429 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
433 * Set the special (bottom) bit of the specified CPU so that it
434 * will take special action (such as flushing its TLB) on the
435 * next exit from an extended quiescent state. Returns true if
436 * the bit was successfully set, or false if the CPU was not in
437 * an extended quiescent state.
439 bool rcu_eqs_special_set(int cpu)
443 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
446 old = atomic_read(&rdtp->dynticks);
447 if (old & RCU_DYNTICK_CTRL_CTR)
449 new = old | RCU_DYNTICK_CTRL_MASK;
450 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
455 * Let the RCU core know that this CPU has gone through the scheduler,
456 * which is a quiescent state. This is called when the need for a
457 * quiescent state is urgent, so we burn an atomic operation and full
458 * memory barriers to let the RCU core know about it, regardless of what
459 * this CPU might (or might not) do in the near future.
461 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
463 * The caller must have disabled interrupts.
465 static void rcu_momentary_dyntick_idle(void)
467 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
468 rcu_dynticks_momentary_idle();
472 * Note a context switch. This is a quiescent state for RCU-sched,
473 * and requires special handling for preemptible RCU.
474 * The caller must have disabled interrupts.
476 void rcu_note_context_switch(bool preempt)
478 barrier(); /* Avoid RCU read-side critical sections leaking down. */
479 trace_rcu_utilization(TPS("Start context switch"));
481 rcu_preempt_note_context_switch();
482 /* Load rcu_urgent_qs before other flags. */
483 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
485 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
486 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
487 rcu_momentary_dyntick_idle();
488 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
490 rcu_note_voluntary_context_switch_lite(current);
492 trace_rcu_utilization(TPS("End context switch"));
493 barrier(); /* Avoid RCU read-side critical sections leaking up. */
495 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
498 * Register a quiescent state for all RCU flavors. If there is an
499 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
500 * dyntick-idle quiescent state visible to other CPUs (but only for those
501 * RCU flavors in desperate need of a quiescent state, which will normally
502 * be none of them). Either way, do a lightweight quiescent state for
505 * The barrier() calls are redundant in the common case when this is
506 * called externally, but just in case this is called from within this
510 void rcu_all_qs(void)
514 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
517 /* Load rcu_urgent_qs before other flags. */
518 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
522 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
523 barrier(); /* Avoid RCU read-side critical sections leaking down. */
524 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
525 local_irq_save(flags);
526 rcu_momentary_dyntick_idle();
527 local_irq_restore(flags);
529 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
531 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
532 barrier(); /* Avoid RCU read-side critical sections leaking up. */
535 EXPORT_SYMBOL_GPL(rcu_all_qs);
537 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
538 static long qhimark = 10000; /* If this many pending, ignore blimit. */
539 static long qlowmark = 100; /* Once only this many pending, use blimit. */
541 module_param(blimit, long, 0444);
542 module_param(qhimark, long, 0444);
543 module_param(qlowmark, long, 0444);
545 static ulong jiffies_till_first_fqs = ULONG_MAX;
546 static ulong jiffies_till_next_fqs = ULONG_MAX;
547 static bool rcu_kick_kthreads;
549 module_param(jiffies_till_first_fqs, ulong, 0644);
550 module_param(jiffies_till_next_fqs, ulong, 0644);
551 module_param(rcu_kick_kthreads, bool, 0644);
554 * How long the grace period must be before we start recruiting
555 * quiescent-state help from rcu_note_context_switch().
557 static ulong jiffies_till_sched_qs = HZ / 20;
558 module_param(jiffies_till_sched_qs, ulong, 0644);
560 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
561 struct rcu_data *rdp);
562 static void force_qs_rnp(struct rcu_state *rsp,
563 int (*f)(struct rcu_data *rsp, bool *isidle,
564 unsigned long *maxj),
565 bool *isidle, unsigned long *maxj);
566 static void force_quiescent_state(struct rcu_state *rsp);
567 static int rcu_pending(void);
570 * Return the number of RCU batches started thus far for debug & stats.
572 unsigned long rcu_batches_started(void)
574 return rcu_state_p->gpnum;
576 EXPORT_SYMBOL_GPL(rcu_batches_started);
579 * Return the number of RCU-sched batches started thus far for debug & stats.
581 unsigned long rcu_batches_started_sched(void)
583 return rcu_sched_state.gpnum;
585 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
588 * Return the number of RCU BH batches started thus far for debug & stats.
590 unsigned long rcu_batches_started_bh(void)
592 return rcu_bh_state.gpnum;
594 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
597 * Return the number of RCU batches completed thus far for debug & stats.
599 unsigned long rcu_batches_completed(void)
601 return rcu_state_p->completed;
603 EXPORT_SYMBOL_GPL(rcu_batches_completed);
606 * Return the number of RCU-sched batches completed thus far for debug & stats.
608 unsigned long rcu_batches_completed_sched(void)
610 return rcu_sched_state.completed;
612 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
615 * Return the number of RCU BH batches completed thus far for debug & stats.
617 unsigned long rcu_batches_completed_bh(void)
619 return rcu_bh_state.completed;
621 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
624 * Return the number of RCU expedited batches completed thus far for
625 * debug & stats. Odd numbers mean that a batch is in progress, even
626 * numbers mean idle. The value returned will thus be roughly double
627 * the cumulative batches since boot.
629 unsigned long rcu_exp_batches_completed(void)
631 return rcu_state_p->expedited_sequence;
633 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
636 * Return the number of RCU-sched expedited batches completed thus far
637 * for debug & stats. Similar to rcu_exp_batches_completed().
639 unsigned long rcu_exp_batches_completed_sched(void)
641 return rcu_sched_state.expedited_sequence;
643 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
646 * Force a quiescent state.
648 void rcu_force_quiescent_state(void)
650 force_quiescent_state(rcu_state_p);
652 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
655 * Force a quiescent state for RCU BH.
657 void rcu_bh_force_quiescent_state(void)
659 force_quiescent_state(&rcu_bh_state);
661 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
664 * Force a quiescent state for RCU-sched.
666 void rcu_sched_force_quiescent_state(void)
668 force_quiescent_state(&rcu_sched_state);
670 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
673 * Show the state of the grace-period kthreads.
675 void show_rcu_gp_kthreads(void)
677 struct rcu_state *rsp;
679 for_each_rcu_flavor(rsp) {
680 pr_info("%s: wait state: %d ->state: %#lx\n",
681 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
682 /* sched_show_task(rsp->gp_kthread); */
685 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
688 * Record the number of times rcutorture tests have been initiated and
689 * terminated. This information allows the debugfs tracing stats to be
690 * correlated to the rcutorture messages, even when the rcutorture module
691 * is being repeatedly loaded and unloaded. In other words, we cannot
692 * store this state in rcutorture itself.
694 void rcutorture_record_test_transition(void)
696 rcutorture_testseq++;
697 rcutorture_vernum = 0;
699 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
702 * Send along grace-period-related data for rcutorture diagnostics.
704 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
705 unsigned long *gpnum, unsigned long *completed)
707 struct rcu_state *rsp = NULL;
716 case RCU_SCHED_FLAVOR:
717 rsp = &rcu_sched_state;
724 *flags = READ_ONCE(rsp->gp_flags);
725 *gpnum = READ_ONCE(rsp->gpnum);
726 *completed = READ_ONCE(rsp->completed);
728 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
731 * Record the number of writer passes through the current rcutorture test.
732 * This is also used to correlate debugfs tracing stats with the rcutorture
735 void rcutorture_record_progress(unsigned long vernum)
739 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
742 * Return the root node of the specified rcu_state structure.
744 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
746 return &rsp->node[0];
750 * Is there any need for future grace periods?
751 * Interrupts must be disabled. If the caller does not hold the root
752 * rnp_node structure's ->lock, the results are advisory only.
754 static int rcu_future_needs_gp(struct rcu_state *rsp)
756 struct rcu_node *rnp = rcu_get_root(rsp);
757 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
758 int *fp = &rnp->need_future_gp[idx];
760 return READ_ONCE(*fp);
764 * Does the current CPU require a not-yet-started grace period?
765 * The caller must have disabled interrupts to prevent races with
766 * normal callback registry.
769 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
771 if (rcu_gp_in_progress(rsp))
772 return false; /* No, a grace period is already in progress. */
773 if (rcu_future_needs_gp(rsp))
774 return true; /* Yes, a no-CBs CPU needs one. */
775 if (!rcu_segcblist_is_enabled(&rdp->cblist))
776 return false; /* No, this is a no-CBs (or offline) CPU. */
777 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
778 return true; /* Yes, CPU has newly registered callbacks. */
779 if (rcu_segcblist_future_gp_needed(&rdp->cblist,
780 READ_ONCE(rsp->completed)))
781 return true; /* Yes, CBs for future grace period. */
782 return false; /* No grace period needed. */
786 * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
788 * Enter idle, doing appropriate accounting. The caller must have
789 * disabled interrupts.
791 static void rcu_eqs_enter_common(bool user)
793 struct rcu_state *rsp;
794 struct rcu_data *rdp;
795 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
797 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0);
798 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
799 !user && !is_idle_task(current)) {
800 struct task_struct *idle __maybe_unused =
801 idle_task(smp_processor_id());
803 trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0);
804 rcu_ftrace_dump(DUMP_ORIG);
805 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
806 current->pid, current->comm,
807 idle->pid, idle->comm); /* must be idle task! */
809 for_each_rcu_flavor(rsp) {
810 rdp = this_cpu_ptr(rsp->rda);
811 do_nocb_deferred_wakeup(rdp);
813 rcu_prepare_for_idle();
814 __this_cpu_inc(disable_rcu_irq_enter);
815 rdtp->dynticks_nesting = 0; /* Breaks tracing momentarily. */
816 rcu_dynticks_eqs_enter(); /* After this, tracing works again. */
817 __this_cpu_dec(disable_rcu_irq_enter);
818 rcu_dynticks_task_enter();
821 * It is illegal to enter an extended quiescent state while
822 * in an RCU read-side critical section.
824 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
825 "Illegal idle entry in RCU read-side critical section.");
826 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
827 "Illegal idle entry in RCU-bh read-side critical section.");
828 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
829 "Illegal idle entry in RCU-sched read-side critical section.");
833 * Enter an RCU extended quiescent state, which can be either the
834 * idle loop or adaptive-tickless usermode execution.
836 static void rcu_eqs_enter(bool user)
838 struct rcu_dynticks *rdtp;
840 rdtp = this_cpu_ptr(&rcu_dynticks);
841 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
842 (rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
843 if ((rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
844 rcu_eqs_enter_common(user);
846 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
850 * rcu_idle_enter - inform RCU that current CPU is entering idle
852 * Enter idle mode, in other words, -leave- the mode in which RCU
853 * read-side critical sections can occur. (Though RCU read-side
854 * critical sections can occur in irq handlers in idle, a possibility
855 * handled by irq_enter() and irq_exit().)
857 * We crowbar the ->dynticks_nesting field to zero to allow for
858 * the possibility of usermode upcalls having messed up our count
859 * of interrupt nesting level during the prior busy period.
861 void rcu_idle_enter(void)
865 local_irq_save(flags);
866 rcu_eqs_enter(false);
867 rcu_sysidle_enter(0);
868 local_irq_restore(flags);
870 EXPORT_SYMBOL_GPL(rcu_idle_enter);
872 #ifdef CONFIG_NO_HZ_FULL
874 * rcu_user_enter - inform RCU that we are resuming userspace.
876 * Enter RCU idle mode right before resuming userspace. No use of RCU
877 * is permitted between this call and rcu_user_exit(). This way the
878 * CPU doesn't need to maintain the tick for RCU maintenance purposes
879 * when the CPU runs in userspace.
881 void rcu_user_enter(void)
885 #endif /* CONFIG_NO_HZ_FULL */
888 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
890 * Exit from an interrupt handler, which might possibly result in entering
891 * idle mode, in other words, leaving the mode in which read-side critical
892 * sections can occur. The caller must have disabled interrupts.
894 * This code assumes that the idle loop never does anything that might
895 * result in unbalanced calls to irq_enter() and irq_exit(). If your
896 * architecture violates this assumption, RCU will give you what you
897 * deserve, good and hard. But very infrequently and irreproducibly.
899 * Use things like work queues to work around this limitation.
901 * You have been warned.
903 void rcu_irq_exit(void)
905 struct rcu_dynticks *rdtp;
907 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
908 rdtp = this_cpu_ptr(&rcu_dynticks);
909 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
910 rdtp->dynticks_nesting < 1);
911 if (rdtp->dynticks_nesting <= 1) {
912 rcu_eqs_enter_common(true);
914 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nesting, rdtp->dynticks_nesting - 1);
915 rdtp->dynticks_nesting--;
917 rcu_sysidle_enter(1);
921 * Wrapper for rcu_irq_exit() where interrupts are enabled.
923 void rcu_irq_exit_irqson(void)
927 local_irq_save(flags);
929 local_irq_restore(flags);
933 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
935 * If the new value of the ->dynticks_nesting counter was previously zero,
936 * we really have exited idle, and must do the appropriate accounting.
937 * The caller must have disabled interrupts.
939 static void rcu_eqs_exit_common(long long oldval, int user)
941 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
943 rcu_dynticks_task_exit();
944 rcu_dynticks_eqs_exit();
945 rcu_cleanup_after_idle();
946 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
947 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
948 !user && !is_idle_task(current)) {
949 struct task_struct *idle __maybe_unused =
950 idle_task(smp_processor_id());
952 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
953 oldval, rdtp->dynticks_nesting);
954 rcu_ftrace_dump(DUMP_ORIG);
955 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
956 current->pid, current->comm,
957 idle->pid, idle->comm); /* must be idle task! */
962 * Exit an RCU extended quiescent state, which can be either the
963 * idle loop or adaptive-tickless usermode execution.
965 static void rcu_eqs_exit(bool user)
967 struct rcu_dynticks *rdtp;
970 rdtp = this_cpu_ptr(&rcu_dynticks);
971 oldval = rdtp->dynticks_nesting;
972 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
973 if (oldval & DYNTICK_TASK_NEST_MASK) {
974 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
976 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
977 rcu_eqs_exit_common(oldval, user);
982 * rcu_idle_exit - inform RCU that current CPU is leaving idle
984 * Exit idle mode, in other words, -enter- the mode in which RCU
985 * read-side critical sections can occur.
987 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
988 * allow for the possibility of usermode upcalls messing up our count
989 * of interrupt nesting level during the busy period that is just
992 void rcu_idle_exit(void)
996 local_irq_save(flags);
999 local_irq_restore(flags);
1001 EXPORT_SYMBOL_GPL(rcu_idle_exit);
1003 #ifdef CONFIG_NO_HZ_FULL
1005 * rcu_user_exit - inform RCU that we are exiting userspace.
1007 * Exit RCU idle mode while entering the kernel because it can
1008 * run a RCU read side critical section anytime.
1010 void rcu_user_exit(void)
1014 #endif /* CONFIG_NO_HZ_FULL */
1017 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1019 * Enter an interrupt handler, which might possibly result in exiting
1020 * idle mode, in other words, entering the mode in which read-side critical
1021 * sections can occur. The caller must have disabled interrupts.
1023 * Note that the Linux kernel is fully capable of entering an interrupt
1024 * handler that it never exits, for example when doing upcalls to
1025 * user mode! This code assumes that the idle loop never does upcalls to
1026 * user mode. If your architecture does do upcalls from the idle loop (or
1027 * does anything else that results in unbalanced calls to the irq_enter()
1028 * and irq_exit() functions), RCU will give you what you deserve, good
1029 * and hard. But very infrequently and irreproducibly.
1031 * Use things like work queues to work around this limitation.
1033 * You have been warned.
1035 void rcu_irq_enter(void)
1037 struct rcu_dynticks *rdtp;
1040 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
1041 rdtp = this_cpu_ptr(&rcu_dynticks);
1042 oldval = rdtp->dynticks_nesting;
1043 rdtp->dynticks_nesting++;
1044 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
1045 rdtp->dynticks_nesting == 0);
1047 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
1049 rcu_eqs_exit_common(oldval, true);
1050 rcu_sysidle_exit(1);
1054 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1056 void rcu_irq_enter_irqson(void)
1058 unsigned long flags;
1060 local_irq_save(flags);
1062 local_irq_restore(flags);
1066 * rcu_nmi_enter - inform RCU of entry to NMI context
1068 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1069 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1070 * that the CPU is active. This implementation permits nested NMIs, as
1071 * long as the nesting level does not overflow an int. (You will probably
1072 * run out of stack space first.)
1074 void rcu_nmi_enter(void)
1076 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1079 /* Complain about underflow. */
1080 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
1083 * If idle from RCU viewpoint, atomically increment ->dynticks
1084 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1085 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1086 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1087 * to be in the outermost NMI handler that interrupted an RCU-idle
1088 * period (observation due to Andy Lutomirski).
1090 if (rcu_dynticks_curr_cpu_in_eqs()) {
1091 rcu_dynticks_eqs_exit();
1094 rdtp->dynticks_nmi_nesting += incby;
1099 * rcu_nmi_exit - inform RCU of exit from NMI context
1101 * If we are returning from the outermost NMI handler that interrupted an
1102 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
1103 * to let the RCU grace-period handling know that the CPU is back to
1106 void rcu_nmi_exit(void)
1108 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1111 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1112 * (We are exiting an NMI handler, so RCU better be paying attention
1115 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
1116 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
1119 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1120 * leave it in non-RCU-idle state.
1122 if (rdtp->dynticks_nmi_nesting != 1) {
1123 rdtp->dynticks_nmi_nesting -= 2;
1127 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1128 rdtp->dynticks_nmi_nesting = 0;
1129 rcu_dynticks_eqs_enter();
1133 * __rcu_is_watching - are RCU read-side critical sections safe?
1135 * Return true if RCU is watching the running CPU, which means that
1136 * this CPU can safely enter RCU read-side critical sections. Unlike
1137 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1138 * least disabled preemption.
1140 bool notrace __rcu_is_watching(void)
1142 return !rcu_dynticks_curr_cpu_in_eqs();
1146 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1148 * If the current CPU is in its idle loop and is neither in an interrupt
1149 * or NMI handler, return true.
1151 bool notrace rcu_is_watching(void)
1155 preempt_disable_notrace();
1156 ret = __rcu_is_watching();
1157 preempt_enable_notrace();
1160 EXPORT_SYMBOL_GPL(rcu_is_watching);
1163 * If a holdout task is actually running, request an urgent quiescent
1164 * state from its CPU. This is unsynchronized, so migrations can cause
1165 * the request to go to the wrong CPU. Which is OK, all that will happen
1166 * is that the CPU's next context switch will be a bit slower and next
1167 * time around this task will generate another request.
1169 void rcu_request_urgent_qs_task(struct task_struct *t)
1176 return; /* This task is not running on that CPU. */
1177 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1180 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1183 * Is the current CPU online? Disable preemption to avoid false positives
1184 * that could otherwise happen due to the current CPU number being sampled,
1185 * this task being preempted, its old CPU being taken offline, resuming
1186 * on some other CPU, then determining that its old CPU is now offline.
1187 * It is OK to use RCU on an offline processor during initial boot, hence
1188 * the check for rcu_scheduler_fully_active. Note also that it is OK
1189 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1190 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1191 * offline to continue to use RCU for one jiffy after marking itself
1192 * offline in the cpu_online_mask. This leniency is necessary given the
1193 * non-atomic nature of the online and offline processing, for example,
1194 * the fact that a CPU enters the scheduler after completing the teardown
1197 * This is also why RCU internally marks CPUs online during in the
1198 * preparation phase and offline after the CPU has been taken down.
1200 * Disable checking if in an NMI handler because we cannot safely report
1201 * errors from NMI handlers anyway.
1203 bool rcu_lockdep_current_cpu_online(void)
1205 struct rcu_data *rdp;
1206 struct rcu_node *rnp;
1212 rdp = this_cpu_ptr(&rcu_sched_data);
1214 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1215 !rcu_scheduler_fully_active;
1219 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1221 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1224 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1226 * If the current CPU is idle or running at a first-level (not nested)
1227 * interrupt from idle, return true. The caller must have at least
1228 * disabled preemption.
1230 static int rcu_is_cpu_rrupt_from_idle(void)
1232 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1236 * Snapshot the specified CPU's dynticks counter so that we can later
1237 * credit them with an implicit quiescent state. Return 1 if this CPU
1238 * is in dynticks idle mode, which is an extended quiescent state.
1240 static int dyntick_save_progress_counter(struct rcu_data *rdp,
1241 bool *isidle, unsigned long *maxj)
1243 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1244 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1245 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1246 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1247 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1248 rdp->mynode->gpnum))
1249 WRITE_ONCE(rdp->gpwrap, true);
1256 * Return true if the specified CPU has passed through a quiescent
1257 * state by virtue of being in or having passed through an dynticks
1258 * idle state since the last call to dyntick_save_progress_counter()
1259 * for this same CPU, or by virtue of having been offline.
1261 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1262 bool *isidle, unsigned long *maxj)
1267 unsigned long rjtsc;
1268 struct rcu_node *rnp;
1271 * If the CPU passed through or entered a dynticks idle phase with
1272 * no active irq/NMI handlers, then we can safely pretend that the CPU
1273 * already acknowledged the request to pass through a quiescent
1274 * state. Either way, that CPU cannot possibly be in an RCU
1275 * read-side critical section that started before the beginning
1276 * of the current RCU grace period.
1278 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1279 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1280 rdp->dynticks_fqs++;
1284 /* Compute and saturate jiffies_till_sched_qs. */
1285 jtsq = jiffies_till_sched_qs;
1286 rjtsc = rcu_jiffies_till_stall_check();
1287 if (jtsq > rjtsc / 2) {
1288 WRITE_ONCE(jiffies_till_sched_qs, rjtsc);
1290 } else if (jtsq < 1) {
1291 WRITE_ONCE(jiffies_till_sched_qs, 1);
1296 * Has this CPU encountered a cond_resched_rcu_qs() since the
1297 * beginning of the grace period? For this to be the case,
1298 * the CPU has to have noticed the current grace period. This
1299 * might not be the case for nohz_full CPUs looping in the kernel.
1302 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1303 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1304 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1305 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1306 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1309 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1310 smp_store_release(ruqp, true);
1313 /* Check for the CPU being offline. */
1314 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1315 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1321 * A CPU running for an extended time within the kernel can
1322 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1323 * even context-switching back and forth between a pair of
1324 * in-kernel CPU-bound tasks cannot advance grace periods.
1325 * So if the grace period is old enough, make the CPU pay attention.
1326 * Note that the unsynchronized assignments to the per-CPU
1327 * rcu_need_heavy_qs variable are safe. Yes, setting of
1328 * bits can be lost, but they will be set again on the next
1329 * force-quiescent-state pass. So lost bit sets do not result
1330 * in incorrect behavior, merely in a grace period lasting
1331 * a few jiffies longer than it might otherwise. Because
1332 * there are at most four threads involved, and because the
1333 * updates are only once every few jiffies, the probability of
1334 * lossage (and thus of slight grace-period extension) is
1337 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1338 * is set too high, we override with half of the RCU CPU stall
1341 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1342 if (!READ_ONCE(*rnhqp) &&
1343 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1344 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1345 WRITE_ONCE(*rnhqp, true);
1346 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1347 smp_store_release(ruqp, true);
1348 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1352 * If more than halfway to RCU CPU stall-warning time, do
1353 * a resched_cpu() to try to loosen things up a bit.
1355 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2)
1356 resched_cpu(rdp->cpu);
1361 static void record_gp_stall_check_time(struct rcu_state *rsp)
1363 unsigned long j = jiffies;
1367 smp_wmb(); /* Record start time before stall time. */
1368 j1 = rcu_jiffies_till_stall_check();
1369 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1370 rsp->jiffies_resched = j + j1 / 2;
1371 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1375 * Convert a ->gp_state value to a character string.
1377 static const char *gp_state_getname(short gs)
1379 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1381 return gp_state_names[gs];
1385 * Complain about starvation of grace-period kthread.
1387 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1393 gpa = READ_ONCE(rsp->gp_activity);
1394 if (j - gpa > 2 * HZ) {
1395 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1397 rsp->gpnum, rsp->completed,
1399 gp_state_getname(rsp->gp_state), rsp->gp_state,
1400 rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
1401 if (rsp->gp_kthread) {
1402 sched_show_task(rsp->gp_kthread);
1403 wake_up_process(rsp->gp_kthread);
1409 * Dump stacks of all tasks running on stalled CPUs. First try using
1410 * NMIs, but fall back to manual remote stack tracing on architectures
1411 * that don't support NMI-based stack dumps. The NMI-triggered stack
1412 * traces are more accurate because they are printed by the target CPU.
1414 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1417 unsigned long flags;
1418 struct rcu_node *rnp;
1420 rcu_for_each_leaf_node(rsp, rnp) {
1421 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1422 for_each_leaf_node_possible_cpu(rnp, cpu)
1423 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1424 if (!trigger_single_cpu_backtrace(cpu))
1426 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1431 * If too much time has passed in the current grace period, and if
1432 * so configured, go kick the relevant kthreads.
1434 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1438 if (!rcu_kick_kthreads)
1440 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1441 if (time_after(jiffies, j) && rsp->gp_kthread &&
1442 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1443 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1444 rcu_ftrace_dump(DUMP_ALL);
1445 wake_up_process(rsp->gp_kthread);
1446 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1450 static inline void panic_on_rcu_stall(void)
1452 if (sysctl_panic_on_rcu_stall)
1453 panic("RCU Stall\n");
1456 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1460 unsigned long flags;
1464 struct rcu_node *rnp = rcu_get_root(rsp);
1467 /* Kick and suppress, if so configured. */
1468 rcu_stall_kick_kthreads(rsp);
1469 if (rcu_cpu_stall_suppress)
1472 /* Only let one CPU complain about others per time interval. */
1474 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1475 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1476 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1477 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1480 WRITE_ONCE(rsp->jiffies_stall,
1481 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1482 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1485 * OK, time to rat on our buddy...
1486 * See Documentation/RCU/stallwarn.txt for info on how to debug
1487 * RCU CPU stall warnings.
1489 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1491 print_cpu_stall_info_begin();
1492 rcu_for_each_leaf_node(rsp, rnp) {
1493 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1494 ndetected += rcu_print_task_stall(rnp);
1495 if (rnp->qsmask != 0) {
1496 for_each_leaf_node_possible_cpu(rnp, cpu)
1497 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1498 print_cpu_stall_info(rsp, cpu);
1502 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1505 print_cpu_stall_info_end();
1506 for_each_possible_cpu(cpu)
1507 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1509 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1510 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1511 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1513 rcu_dump_cpu_stacks(rsp);
1515 /* Complain about tasks blocking the grace period. */
1516 rcu_print_detail_task_stall(rsp);
1518 if (READ_ONCE(rsp->gpnum) != gpnum ||
1519 READ_ONCE(rsp->completed) == gpnum) {
1520 pr_err("INFO: Stall ended before state dump start\n");
1523 gpa = READ_ONCE(rsp->gp_activity);
1524 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1525 rsp->name, j - gpa, j, gpa,
1526 jiffies_till_next_fqs,
1527 rcu_get_root(rsp)->qsmask);
1528 /* In this case, the current CPU might be at fault. */
1529 sched_show_task(current);
1533 rcu_check_gp_kthread_starvation(rsp);
1535 panic_on_rcu_stall();
1537 force_quiescent_state(rsp); /* Kick them all. */
1540 static void print_cpu_stall(struct rcu_state *rsp)
1543 unsigned long flags;
1544 struct rcu_node *rnp = rcu_get_root(rsp);
1547 /* Kick and suppress, if so configured. */
1548 rcu_stall_kick_kthreads(rsp);
1549 if (rcu_cpu_stall_suppress)
1553 * OK, time to rat on ourselves...
1554 * See Documentation/RCU/stallwarn.txt for info on how to debug
1555 * RCU CPU stall warnings.
1557 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1558 print_cpu_stall_info_begin();
1559 print_cpu_stall_info(rsp, smp_processor_id());
1560 print_cpu_stall_info_end();
1561 for_each_possible_cpu(cpu)
1562 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1564 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1565 jiffies - rsp->gp_start,
1566 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1568 rcu_check_gp_kthread_starvation(rsp);
1570 rcu_dump_cpu_stacks(rsp);
1572 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1573 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1574 WRITE_ONCE(rsp->jiffies_stall,
1575 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1576 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1578 panic_on_rcu_stall();
1581 * Attempt to revive the RCU machinery by forcing a context switch.
1583 * A context switch would normally allow the RCU state machine to make
1584 * progress and it could be we're stuck in kernel space without context
1585 * switches for an entirely unreasonable amount of time.
1587 resched_cpu(smp_processor_id());
1590 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1592 unsigned long completed;
1593 unsigned long gpnum;
1597 struct rcu_node *rnp;
1599 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1600 !rcu_gp_in_progress(rsp))
1602 rcu_stall_kick_kthreads(rsp);
1606 * Lots of memory barriers to reject false positives.
1608 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1609 * then rsp->gp_start, and finally rsp->completed. These values
1610 * are updated in the opposite order with memory barriers (or
1611 * equivalent) during grace-period initialization and cleanup.
1612 * Now, a false positive can occur if we get an new value of
1613 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1614 * the memory barriers, the only way that this can happen is if one
1615 * grace period ends and another starts between these two fetches.
1616 * Detect this by comparing rsp->completed with the previous fetch
1619 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1620 * and rsp->gp_start suffice to forestall false positives.
1622 gpnum = READ_ONCE(rsp->gpnum);
1623 smp_rmb(); /* Pick up ->gpnum first... */
1624 js = READ_ONCE(rsp->jiffies_stall);
1625 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1626 gps = READ_ONCE(rsp->gp_start);
1627 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1628 completed = READ_ONCE(rsp->completed);
1629 if (ULONG_CMP_GE(completed, gpnum) ||
1630 ULONG_CMP_LT(j, js) ||
1631 ULONG_CMP_GE(gps, js))
1632 return; /* No stall or GP completed since entering function. */
1634 if (rcu_gp_in_progress(rsp) &&
1635 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1637 /* We haven't checked in, so go dump stack. */
1638 print_cpu_stall(rsp);
1640 } else if (rcu_gp_in_progress(rsp) &&
1641 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1643 /* They had a few time units to dump stack, so complain. */
1644 print_other_cpu_stall(rsp, gpnum);
1649 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1651 * Set the stall-warning timeout way off into the future, thus preventing
1652 * any RCU CPU stall-warning messages from appearing in the current set of
1653 * RCU grace periods.
1655 * The caller must disable hard irqs.
1657 void rcu_cpu_stall_reset(void)
1659 struct rcu_state *rsp;
1661 for_each_rcu_flavor(rsp)
1662 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1666 * Determine the value that ->completed will have at the end of the
1667 * next subsequent grace period. This is used to tag callbacks so that
1668 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1669 * been dyntick-idle for an extended period with callbacks under the
1670 * influence of RCU_FAST_NO_HZ.
1672 * The caller must hold rnp->lock with interrupts disabled.
1674 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1675 struct rcu_node *rnp)
1678 * If RCU is idle, we just wait for the next grace period.
1679 * But we can only be sure that RCU is idle if we are looking
1680 * at the root rcu_node structure -- otherwise, a new grace
1681 * period might have started, but just not yet gotten around
1682 * to initializing the current non-root rcu_node structure.
1684 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1685 return rnp->completed + 1;
1688 * Otherwise, wait for a possible partial grace period and
1689 * then the subsequent full grace period.
1691 return rnp->completed + 2;
1695 * Trace-event helper function for rcu_start_future_gp() and
1696 * rcu_nocb_wait_gp().
1698 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1699 unsigned long c, const char *s)
1701 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1702 rnp->completed, c, rnp->level,
1703 rnp->grplo, rnp->grphi, s);
1707 * Start some future grace period, as needed to handle newly arrived
1708 * callbacks. The required future grace periods are recorded in each
1709 * rcu_node structure's ->need_future_gp field. Returns true if there
1710 * is reason to awaken the grace-period kthread.
1712 * The caller must hold the specified rcu_node structure's ->lock.
1714 static bool __maybe_unused
1715 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1716 unsigned long *c_out)
1720 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1723 * Pick up grace-period number for new callbacks. If this
1724 * grace period is already marked as needed, return to the caller.
1726 c = rcu_cbs_completed(rdp->rsp, rnp);
1727 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1728 if (rnp->need_future_gp[c & 0x1]) {
1729 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1734 * If either this rcu_node structure or the root rcu_node structure
1735 * believe that a grace period is in progress, then we must wait
1736 * for the one following, which is in "c". Because our request
1737 * will be noticed at the end of the current grace period, we don't
1738 * need to explicitly start one. We only do the lockless check
1739 * of rnp_root's fields if the current rcu_node structure thinks
1740 * there is no grace period in flight, and because we hold rnp->lock,
1741 * the only possible change is when rnp_root's two fields are
1742 * equal, in which case rnp_root->gpnum might be concurrently
1743 * incremented. But that is OK, as it will just result in our
1744 * doing some extra useless work.
1746 if (rnp->gpnum != rnp->completed ||
1747 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1748 rnp->need_future_gp[c & 0x1]++;
1749 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1754 * There might be no grace period in progress. If we don't already
1755 * hold it, acquire the root rcu_node structure's lock in order to
1756 * start one (if needed).
1758 if (rnp != rnp_root)
1759 raw_spin_lock_rcu_node(rnp_root);
1762 * Get a new grace-period number. If there really is no grace
1763 * period in progress, it will be smaller than the one we obtained
1764 * earlier. Adjust callbacks as needed.
1766 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1767 if (!rcu_is_nocb_cpu(rdp->cpu))
1768 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1771 * If the needed for the required grace period is already
1772 * recorded, trace and leave.
1774 if (rnp_root->need_future_gp[c & 0x1]) {
1775 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1779 /* Record the need for the future grace period. */
1780 rnp_root->need_future_gp[c & 0x1]++;
1782 /* If a grace period is not already in progress, start one. */
1783 if (rnp_root->gpnum != rnp_root->completed) {
1784 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1786 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1787 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1790 if (rnp != rnp_root)
1791 raw_spin_unlock_rcu_node(rnp_root);
1799 * Clean up any old requests for the just-ended grace period. Also return
1800 * whether any additional grace periods have been requested.
1802 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1804 int c = rnp->completed;
1806 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1808 rnp->need_future_gp[c & 0x1] = 0;
1809 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1810 trace_rcu_future_gp(rnp, rdp, c,
1811 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1816 * Awaken the grace-period kthread for the specified flavor of RCU.
1817 * Don't do a self-awaken, and don't bother awakening when there is
1818 * nothing for the grace-period kthread to do (as in several CPUs
1819 * raced to awaken, and we lost), and finally don't try to awaken
1820 * a kthread that has not yet been created.
1822 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1824 if (current == rsp->gp_kthread ||
1825 !READ_ONCE(rsp->gp_flags) ||
1828 swake_up(&rsp->gp_wq);
1832 * If there is room, assign a ->completed number to any callbacks on
1833 * this CPU that have not already been assigned. Also accelerate any
1834 * callbacks that were previously assigned a ->completed number that has
1835 * since proven to be too conservative, which can happen if callbacks get
1836 * assigned a ->completed number while RCU is idle, but with reference to
1837 * a non-root rcu_node structure. This function is idempotent, so it does
1838 * not hurt to call it repeatedly. Returns an flag saying that we should
1839 * awaken the RCU grace-period kthread.
1841 * The caller must hold rnp->lock with interrupts disabled.
1843 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1844 struct rcu_data *rdp)
1848 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1849 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1853 * Callbacks are often registered with incomplete grace-period
1854 * information. Something about the fact that getting exact
1855 * information requires acquiring a global lock... RCU therefore
1856 * makes a conservative estimate of the grace period number at which
1857 * a given callback will become ready to invoke. The following
1858 * code checks this estimate and improves it when possible, thus
1859 * accelerating callback invocation to an earlier grace-period
1862 if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
1863 ret = rcu_start_future_gp(rnp, rdp, NULL);
1865 /* Trace depending on how much we were able to accelerate. */
1866 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1867 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1869 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1874 * Move any callbacks whose grace period has completed to the
1875 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1876 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1877 * sublist. This function is idempotent, so it does not hurt to
1878 * invoke it repeatedly. As long as it is not invoked -too- often...
1879 * Returns true if the RCU grace-period kthread needs to be awakened.
1881 * The caller must hold rnp->lock with interrupts disabled.
1883 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1884 struct rcu_data *rdp)
1886 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1887 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1891 * Find all callbacks whose ->completed numbers indicate that they
1892 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1894 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1896 /* Classify any remaining callbacks. */
1897 return rcu_accelerate_cbs(rsp, rnp, rdp);
1901 * Update CPU-local rcu_data state to record the beginnings and ends of
1902 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1903 * structure corresponding to the current CPU, and must have irqs disabled.
1904 * Returns true if the grace-period kthread needs to be awakened.
1906 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1907 struct rcu_data *rdp)
1912 /* Handle the ends of any preceding grace periods first. */
1913 if (rdp->completed == rnp->completed &&
1914 !unlikely(READ_ONCE(rdp->gpwrap))) {
1916 /* No grace period end, so just accelerate recent callbacks. */
1917 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1921 /* Advance callbacks. */
1922 ret = rcu_advance_cbs(rsp, rnp, rdp);
1924 /* Remember that we saw this grace-period completion. */
1925 rdp->completed = rnp->completed;
1926 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1929 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1931 * If the current grace period is waiting for this CPU,
1932 * set up to detect a quiescent state, otherwise don't
1933 * go looking for one.
1935 rdp->gpnum = rnp->gpnum;
1936 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1937 need_gp = !!(rnp->qsmask & rdp->grpmask);
1938 rdp->cpu_no_qs.b.norm = need_gp;
1939 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1940 rdp->core_needs_qs = need_gp;
1941 zero_cpu_stall_ticks(rdp);
1942 WRITE_ONCE(rdp->gpwrap, false);
1947 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1949 unsigned long flags;
1951 struct rcu_node *rnp;
1953 local_irq_save(flags);
1955 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1956 rdp->completed == READ_ONCE(rnp->completed) &&
1957 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1958 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1959 local_irq_restore(flags);
1962 needwake = __note_gp_changes(rsp, rnp, rdp);
1963 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1965 rcu_gp_kthread_wake(rsp);
1968 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1971 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1972 schedule_timeout_uninterruptible(delay);
1976 * Initialize a new grace period. Return false if no grace period required.
1978 static bool rcu_gp_init(struct rcu_state *rsp)
1980 unsigned long oldmask;
1981 struct rcu_data *rdp;
1982 struct rcu_node *rnp = rcu_get_root(rsp);
1984 WRITE_ONCE(rsp->gp_activity, jiffies);
1985 raw_spin_lock_irq_rcu_node(rnp);
1986 if (!READ_ONCE(rsp->gp_flags)) {
1987 /* Spurious wakeup, tell caller to go back to sleep. */
1988 raw_spin_unlock_irq_rcu_node(rnp);
1991 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1993 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1995 * Grace period already in progress, don't start another.
1996 * Not supposed to be able to happen.
1998 raw_spin_unlock_irq_rcu_node(rnp);
2002 /* Advance to a new grace period and initialize state. */
2003 record_gp_stall_check_time(rsp);
2004 /* Record GP times before starting GP, hence smp_store_release(). */
2005 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
2006 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
2007 raw_spin_unlock_irq_rcu_node(rnp);
2010 * Apply per-leaf buffered online and offline operations to the
2011 * rcu_node tree. Note that this new grace period need not wait
2012 * for subsequent online CPUs, and that quiescent-state forcing
2013 * will handle subsequent offline CPUs.
2015 rcu_for_each_leaf_node(rsp, rnp) {
2016 rcu_gp_slow(rsp, gp_preinit_delay);
2017 raw_spin_lock_irq_rcu_node(rnp);
2018 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2019 !rnp->wait_blkd_tasks) {
2020 /* Nothing to do on this leaf rcu_node structure. */
2021 raw_spin_unlock_irq_rcu_node(rnp);
2025 /* Record old state, apply changes to ->qsmaskinit field. */
2026 oldmask = rnp->qsmaskinit;
2027 rnp->qsmaskinit = rnp->qsmaskinitnext;
2029 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2030 if (!oldmask != !rnp->qsmaskinit) {
2031 if (!oldmask) /* First online CPU for this rcu_node. */
2032 rcu_init_new_rnp(rnp);
2033 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2034 rnp->wait_blkd_tasks = true;
2035 else /* Last offline CPU and can propagate. */
2036 rcu_cleanup_dead_rnp(rnp);
2040 * If all waited-on tasks from prior grace period are
2041 * done, and if all this rcu_node structure's CPUs are
2042 * still offline, propagate up the rcu_node tree and
2043 * clear ->wait_blkd_tasks. Otherwise, if one of this
2044 * rcu_node structure's CPUs has since come back online,
2045 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2046 * checks for this, so just call it unconditionally).
2048 if (rnp->wait_blkd_tasks &&
2049 (!rcu_preempt_has_tasks(rnp) ||
2051 rnp->wait_blkd_tasks = false;
2052 rcu_cleanup_dead_rnp(rnp);
2055 raw_spin_unlock_irq_rcu_node(rnp);
2059 * Set the quiescent-state-needed bits in all the rcu_node
2060 * structures for all currently online CPUs in breadth-first order,
2061 * starting from the root rcu_node structure, relying on the layout
2062 * of the tree within the rsp->node[] array. Note that other CPUs
2063 * will access only the leaves of the hierarchy, thus seeing that no
2064 * grace period is in progress, at least until the corresponding
2065 * leaf node has been initialized.
2067 * The grace period cannot complete until the initialization
2068 * process finishes, because this kthread handles both.
2070 rcu_for_each_node_breadth_first(rsp, rnp) {
2071 rcu_gp_slow(rsp, gp_init_delay);
2072 raw_spin_lock_irq_rcu_node(rnp);
2073 rdp = this_cpu_ptr(rsp->rda);
2074 rcu_preempt_check_blocked_tasks(rnp);
2075 rnp->qsmask = rnp->qsmaskinit;
2076 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2077 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2078 WRITE_ONCE(rnp->completed, rsp->completed);
2079 if (rnp == rdp->mynode)
2080 (void)__note_gp_changes(rsp, rnp, rdp);
2081 rcu_preempt_boost_start_gp(rnp);
2082 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2083 rnp->level, rnp->grplo,
2084 rnp->grphi, rnp->qsmask);
2085 raw_spin_unlock_irq_rcu_node(rnp);
2086 cond_resched_rcu_qs();
2087 WRITE_ONCE(rsp->gp_activity, jiffies);
2094 * Helper function for wait_event_interruptible_timeout() wakeup
2095 * at force-quiescent-state time.
2097 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2099 struct rcu_node *rnp = rcu_get_root(rsp);
2101 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2102 *gfp = READ_ONCE(rsp->gp_flags);
2103 if (*gfp & RCU_GP_FLAG_FQS)
2106 /* The current grace period has completed. */
2107 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2114 * Do one round of quiescent-state forcing.
2116 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2118 bool isidle = false;
2120 struct rcu_node *rnp = rcu_get_root(rsp);
2122 WRITE_ONCE(rsp->gp_activity, jiffies);
2125 /* Collect dyntick-idle snapshots. */
2126 if (is_sysidle_rcu_state(rsp)) {
2128 maxj = jiffies - ULONG_MAX / 4;
2130 force_qs_rnp(rsp, dyntick_save_progress_counter,
2132 rcu_sysidle_report_gp(rsp, isidle, maxj);
2134 /* Handle dyntick-idle and offline CPUs. */
2136 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
2138 /* Clear flag to prevent immediate re-entry. */
2139 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2140 raw_spin_lock_irq_rcu_node(rnp);
2141 WRITE_ONCE(rsp->gp_flags,
2142 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2143 raw_spin_unlock_irq_rcu_node(rnp);
2148 * Clean up after the old grace period.
2150 static void rcu_gp_cleanup(struct rcu_state *rsp)
2152 unsigned long gp_duration;
2153 bool needgp = false;
2155 struct rcu_data *rdp;
2156 struct rcu_node *rnp = rcu_get_root(rsp);
2157 struct swait_queue_head *sq;
2159 WRITE_ONCE(rsp->gp_activity, jiffies);
2160 raw_spin_lock_irq_rcu_node(rnp);
2161 gp_duration = jiffies - rsp->gp_start;
2162 if (gp_duration > rsp->gp_max)
2163 rsp->gp_max = gp_duration;
2166 * We know the grace period is complete, but to everyone else
2167 * it appears to still be ongoing. But it is also the case
2168 * that to everyone else it looks like there is nothing that
2169 * they can do to advance the grace period. It is therefore
2170 * safe for us to drop the lock in order to mark the grace
2171 * period as completed in all of the rcu_node structures.
2173 raw_spin_unlock_irq_rcu_node(rnp);
2176 * Propagate new ->completed value to rcu_node structures so
2177 * that other CPUs don't have to wait until the start of the next
2178 * grace period to process their callbacks. This also avoids
2179 * some nasty RCU grace-period initialization races by forcing
2180 * the end of the current grace period to be completely recorded in
2181 * all of the rcu_node structures before the beginning of the next
2182 * grace period is recorded in any of the rcu_node structures.
2184 rcu_for_each_node_breadth_first(rsp, rnp) {
2185 raw_spin_lock_irq_rcu_node(rnp);
2186 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2187 WARN_ON_ONCE(rnp->qsmask);
2188 WRITE_ONCE(rnp->completed, rsp->gpnum);
2189 rdp = this_cpu_ptr(rsp->rda);
2190 if (rnp == rdp->mynode)
2191 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2192 /* smp_mb() provided by prior unlock-lock pair. */
2193 nocb += rcu_future_gp_cleanup(rsp, rnp);
2194 sq = rcu_nocb_gp_get(rnp);
2195 raw_spin_unlock_irq_rcu_node(rnp);
2196 rcu_nocb_gp_cleanup(sq);
2197 cond_resched_rcu_qs();
2198 WRITE_ONCE(rsp->gp_activity, jiffies);
2199 rcu_gp_slow(rsp, gp_cleanup_delay);
2201 rnp = rcu_get_root(rsp);
2202 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2203 rcu_nocb_gp_set(rnp, nocb);
2205 /* Declare grace period done. */
2206 WRITE_ONCE(rsp->completed, rsp->gpnum);
2207 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2208 rsp->gp_state = RCU_GP_IDLE;
2209 rdp = this_cpu_ptr(rsp->rda);
2210 /* Advance CBs to reduce false positives below. */
2211 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2212 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2213 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2214 trace_rcu_grace_period(rsp->name,
2215 READ_ONCE(rsp->gpnum),
2218 raw_spin_unlock_irq_rcu_node(rnp);
2222 * Body of kthread that handles grace periods.
2224 static int __noreturn rcu_gp_kthread(void *arg)
2230 struct rcu_state *rsp = arg;
2231 struct rcu_node *rnp = rcu_get_root(rsp);
2233 rcu_bind_gp_kthread();
2236 /* Handle grace-period start. */
2238 trace_rcu_grace_period(rsp->name,
2239 READ_ONCE(rsp->gpnum),
2241 rsp->gp_state = RCU_GP_WAIT_GPS;
2242 swait_event_interruptible(rsp->gp_wq,
2243 READ_ONCE(rsp->gp_flags) &
2245 rsp->gp_state = RCU_GP_DONE_GPS;
2246 /* Locking provides needed memory barrier. */
2247 if (rcu_gp_init(rsp))
2249 cond_resched_rcu_qs();
2250 WRITE_ONCE(rsp->gp_activity, jiffies);
2251 WARN_ON(signal_pending(current));
2252 trace_rcu_grace_period(rsp->name,
2253 READ_ONCE(rsp->gpnum),
2257 /* Handle quiescent-state forcing. */
2258 first_gp_fqs = true;
2259 j = jiffies_till_first_fqs;
2262 jiffies_till_first_fqs = HZ;
2267 rsp->jiffies_force_qs = jiffies + j;
2268 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2271 trace_rcu_grace_period(rsp->name,
2272 READ_ONCE(rsp->gpnum),
2274 rsp->gp_state = RCU_GP_WAIT_FQS;
2275 ret = swait_event_interruptible_timeout(rsp->gp_wq,
2276 rcu_gp_fqs_check_wake(rsp, &gf), j);
2277 rsp->gp_state = RCU_GP_DOING_FQS;
2278 /* Locking provides needed memory barriers. */
2279 /* If grace period done, leave loop. */
2280 if (!READ_ONCE(rnp->qsmask) &&
2281 !rcu_preempt_blocked_readers_cgp(rnp))
2283 /* If time for quiescent-state forcing, do it. */
2284 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2285 (gf & RCU_GP_FLAG_FQS)) {
2286 trace_rcu_grace_period(rsp->name,
2287 READ_ONCE(rsp->gpnum),
2289 rcu_gp_fqs(rsp, first_gp_fqs);
2290 first_gp_fqs = false;
2291 trace_rcu_grace_period(rsp->name,
2292 READ_ONCE(rsp->gpnum),
2294 cond_resched_rcu_qs();
2295 WRITE_ONCE(rsp->gp_activity, jiffies);
2296 ret = 0; /* Force full wait till next FQS. */
2297 j = jiffies_till_next_fqs;
2300 jiffies_till_next_fqs = HZ;
2303 jiffies_till_next_fqs = 1;
2306 /* Deal with stray signal. */
2307 cond_resched_rcu_qs();
2308 WRITE_ONCE(rsp->gp_activity, jiffies);
2309 WARN_ON(signal_pending(current));
2310 trace_rcu_grace_period(rsp->name,
2311 READ_ONCE(rsp->gpnum),
2313 ret = 1; /* Keep old FQS timing. */
2315 if (time_after(jiffies, rsp->jiffies_force_qs))
2318 j = rsp->jiffies_force_qs - j;
2322 /* Handle grace-period end. */
2323 rsp->gp_state = RCU_GP_CLEANUP;
2324 rcu_gp_cleanup(rsp);
2325 rsp->gp_state = RCU_GP_CLEANED;
2330 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2331 * in preparation for detecting the next grace period. The caller must hold
2332 * the root node's ->lock and hard irqs must be disabled.
2334 * Note that it is legal for a dying CPU (which is marked as offline) to
2335 * invoke this function. This can happen when the dying CPU reports its
2338 * Returns true if the grace-period kthread must be awakened.
2341 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2342 struct rcu_data *rdp)
2344 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2346 * Either we have not yet spawned the grace-period
2347 * task, this CPU does not need another grace period,
2348 * or a grace period is already in progress.
2349 * Either way, don't start a new grace period.
2353 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2354 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2358 * We can't do wakeups while holding the rnp->lock, as that
2359 * could cause possible deadlocks with the rq->lock. Defer
2360 * the wakeup to our caller.
2366 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2367 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2368 * is invoked indirectly from rcu_advance_cbs(), which would result in
2369 * endless recursion -- or would do so if it wasn't for the self-deadlock
2370 * that is encountered beforehand.
2372 * Returns true if the grace-period kthread needs to be awakened.
2374 static bool rcu_start_gp(struct rcu_state *rsp)
2376 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2377 struct rcu_node *rnp = rcu_get_root(rsp);
2381 * If there is no grace period in progress right now, any
2382 * callbacks we have up to this point will be satisfied by the
2383 * next grace period. Also, advancing the callbacks reduces the
2384 * probability of false positives from cpu_needs_another_gp()
2385 * resulting in pointless grace periods. So, advance callbacks
2386 * then start the grace period!
2388 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2389 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2394 * Report a full set of quiescent states to the specified rcu_state data
2395 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2396 * kthread if another grace period is required. Whether we wake
2397 * the grace-period kthread or it awakens itself for the next round
2398 * of quiescent-state forcing, that kthread will clean up after the
2399 * just-completed grace period. Note that the caller must hold rnp->lock,
2400 * which is released before return.
2402 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2403 __releases(rcu_get_root(rsp)->lock)
2405 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2406 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2407 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2408 rcu_gp_kthread_wake(rsp);
2412 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2413 * Allows quiescent states for a group of CPUs to be reported at one go
2414 * to the specified rcu_node structure, though all the CPUs in the group
2415 * must be represented by the same rcu_node structure (which need not be a
2416 * leaf rcu_node structure, though it often will be). The gps parameter
2417 * is the grace-period snapshot, which means that the quiescent states
2418 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2419 * must be held upon entry, and it is released before return.
2422 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2423 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2424 __releases(rnp->lock)
2426 unsigned long oldmask = 0;
2427 struct rcu_node *rnp_c;
2429 /* Walk up the rcu_node hierarchy. */
2431 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2434 * Our bit has already been cleared, or the
2435 * relevant grace period is already over, so done.
2437 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2440 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2441 rnp->qsmask &= ~mask;
2442 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2443 mask, rnp->qsmask, rnp->level,
2444 rnp->grplo, rnp->grphi,
2446 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2448 /* Other bits still set at this level, so done. */
2449 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2452 mask = rnp->grpmask;
2453 if (rnp->parent == NULL) {
2455 /* No more levels. Exit loop holding root lock. */
2459 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2462 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2463 oldmask = rnp_c->qsmask;
2467 * Get here if we are the last CPU to pass through a quiescent
2468 * state for this grace period. Invoke rcu_report_qs_rsp()
2469 * to clean up and start the next grace period if one is needed.
2471 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2475 * Record a quiescent state for all tasks that were previously queued
2476 * on the specified rcu_node structure and that were blocking the current
2477 * RCU grace period. The caller must hold the specified rnp->lock with
2478 * irqs disabled, and this lock is released upon return, but irqs remain
2481 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2482 struct rcu_node *rnp, unsigned long flags)
2483 __releases(rnp->lock)
2487 struct rcu_node *rnp_p;
2489 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2490 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2491 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2492 return; /* Still need more quiescent states! */
2495 rnp_p = rnp->parent;
2496 if (rnp_p == NULL) {
2498 * Only one rcu_node structure in the tree, so don't
2499 * try to report up to its nonexistent parent!
2501 rcu_report_qs_rsp(rsp, flags);
2505 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2507 mask = rnp->grpmask;
2508 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2509 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2510 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2514 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2515 * structure. This must be called from the specified CPU.
2518 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2520 unsigned long flags;
2523 struct rcu_node *rnp;
2526 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2527 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2528 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2531 * The grace period in which this quiescent state was
2532 * recorded has ended, so don't report it upwards.
2533 * We will instead need a new quiescent state that lies
2534 * within the current grace period.
2536 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2537 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2538 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2541 mask = rdp->grpmask;
2542 if ((rnp->qsmask & mask) == 0) {
2543 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2545 rdp->core_needs_qs = false;
2548 * This GP can't end until cpu checks in, so all of our
2549 * callbacks can be processed during the next GP.
2551 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2553 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2554 /* ^^^ Released rnp->lock */
2556 rcu_gp_kthread_wake(rsp);
2561 * Check to see if there is a new grace period of which this CPU
2562 * is not yet aware, and if so, set up local rcu_data state for it.
2563 * Otherwise, see if this CPU has just passed through its first
2564 * quiescent state for this grace period, and record that fact if so.
2567 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2569 /* Check for grace-period ends and beginnings. */
2570 note_gp_changes(rsp, rdp);
2573 * Does this CPU still need to do its part for current grace period?
2574 * If no, return and let the other CPUs do their part as well.
2576 if (!rdp->core_needs_qs)
2580 * Was there a quiescent state since the beginning of the grace
2581 * period? If no, then exit and wait for the next call.
2583 if (rdp->cpu_no_qs.b.norm)
2587 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2590 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2594 * Send the specified CPU's RCU callbacks to the orphanage. The
2595 * specified CPU must be offline, and the caller must hold the
2599 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2600 struct rcu_node *rnp, struct rcu_data *rdp)
2602 /* No-CBs CPUs do not have orphanable callbacks. */
2603 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2607 * Orphan the callbacks. First adjust the counts. This is safe
2608 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2609 * cannot be running now. Thus no memory barrier is required.
2611 rdp->n_cbs_orphaned += rcu_segcblist_n_cbs(&rdp->cblist);
2612 rcu_segcblist_extract_count(&rdp->cblist, &rsp->orphan_done);
2615 * Next, move those callbacks still needing a grace period to
2616 * the orphanage, where some other CPU will pick them up.
2617 * Some of the callbacks might have gone partway through a grace
2618 * period, but that is too bad. They get to start over because we
2619 * cannot assume that grace periods are synchronized across CPUs.
2621 rcu_segcblist_extract_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
2624 * Then move the ready-to-invoke callbacks to the orphanage,
2625 * where some other CPU will pick them up. These will not be
2626 * required to pass though another grace period: They are done.
2628 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rsp->orphan_done);
2630 /* Finally, disallow further callbacks on this CPU. */
2631 rcu_segcblist_disable(&rdp->cblist);
2635 * Adopt the RCU callbacks from the specified rcu_state structure's
2636 * orphanage. The caller must hold the ->orphan_lock.
2638 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2640 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2642 /* No-CBs CPUs are handled specially. */
2643 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2644 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2647 /* Do the accounting first. */
2648 rdp->n_cbs_adopted += rsp->orphan_done.len;
2649 if (rsp->orphan_done.len_lazy != rsp->orphan_done.len)
2650 rcu_idle_count_callbacks_posted();
2651 rcu_segcblist_insert_count(&rdp->cblist, &rsp->orphan_done);
2654 * We do not need a memory barrier here because the only way we
2655 * can get here if there is an rcu_barrier() in flight is if
2656 * we are the task doing the rcu_barrier().
2659 /* First adopt the ready-to-invoke callbacks, then the done ones. */
2660 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rsp->orphan_done);
2661 WARN_ON_ONCE(rsp->orphan_done.head);
2662 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
2663 WARN_ON_ONCE(rsp->orphan_pend.head);
2664 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) !=
2665 !rcu_segcblist_n_cbs(&rdp->cblist));
2669 * Trace the fact that this CPU is going offline.
2671 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2673 RCU_TRACE(unsigned long mask;)
2674 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2675 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2677 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2680 RCU_TRACE(mask = rdp->grpmask;)
2681 trace_rcu_grace_period(rsp->name,
2682 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2687 * All CPUs for the specified rcu_node structure have gone offline,
2688 * and all tasks that were preempted within an RCU read-side critical
2689 * section while running on one of those CPUs have since exited their RCU
2690 * read-side critical section. Some other CPU is reporting this fact with
2691 * the specified rcu_node structure's ->lock held and interrupts disabled.
2692 * This function therefore goes up the tree of rcu_node structures,
2693 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2694 * the leaf rcu_node structure's ->qsmaskinit field has already been
2697 * This function does check that the specified rcu_node structure has
2698 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2699 * prematurely. That said, invoking it after the fact will cost you
2700 * a needless lock acquisition. So once it has done its work, don't
2703 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2706 struct rcu_node *rnp = rnp_leaf;
2708 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2709 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2712 mask = rnp->grpmask;
2716 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2717 rnp->qsmaskinit &= ~mask;
2718 rnp->qsmask &= ~mask;
2719 if (rnp->qsmaskinit) {
2720 raw_spin_unlock_rcu_node(rnp);
2721 /* irqs remain disabled. */
2724 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2729 * The CPU has been completely removed, and some other CPU is reporting
2730 * this fact from process context. Do the remainder of the cleanup,
2731 * including orphaning the outgoing CPU's RCU callbacks, and also
2732 * adopting them. There can only be one CPU hotplug operation at a time,
2733 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2735 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2737 unsigned long flags;
2738 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2739 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2741 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2744 /* Adjust any no-longer-needed kthreads. */
2745 rcu_boost_kthread_setaffinity(rnp, -1);
2747 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2748 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2749 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2750 rcu_adopt_orphan_cbs(rsp, flags);
2751 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2753 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
2754 !rcu_segcblist_empty(&rdp->cblist),
2755 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
2756 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
2757 rcu_segcblist_first_cb(&rdp->cblist));
2761 * Invoke any RCU callbacks that have made it to the end of their grace
2762 * period. Thottle as specified by rdp->blimit.
2764 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2766 unsigned long flags;
2767 struct rcu_head *rhp;
2768 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2771 /* If no callbacks are ready, just return. */
2772 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2773 trace_rcu_batch_start(rsp->name,
2774 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2775 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2776 trace_rcu_batch_end(rsp->name, 0,
2777 !rcu_segcblist_empty(&rdp->cblist),
2778 need_resched(), is_idle_task(current),
2779 rcu_is_callbacks_kthread());
2784 * Extract the list of ready callbacks, disabling to prevent
2785 * races with call_rcu() from interrupt handlers. Leave the
2786 * callback counts, as rcu_barrier() needs to be conservative.
2788 local_irq_save(flags);
2789 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2791 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2792 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2793 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2794 local_irq_restore(flags);
2796 /* Invoke callbacks. */
2797 rhp = rcu_cblist_dequeue(&rcl);
2798 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2799 debug_rcu_head_unqueue(rhp);
2800 if (__rcu_reclaim(rsp->name, rhp))
2801 rcu_cblist_dequeued_lazy(&rcl);
2803 * Stop only if limit reached and CPU has something to do.
2804 * Note: The rcl structure counts down from zero.
2806 if (-rcl.len >= bl &&
2808 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2812 local_irq_save(flags);
2814 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2815 is_idle_task(current), rcu_is_callbacks_kthread());
2817 /* Update counts and requeue any remaining callbacks. */
2818 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2819 smp_mb(); /* List handling before counting for rcu_barrier(). */
2820 rdp->n_cbs_invoked += count;
2821 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2823 /* Reinstate batch limit if we have worked down the excess. */
2824 count = rcu_segcblist_n_cbs(&rdp->cblist);
2825 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2826 rdp->blimit = blimit;
2828 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2829 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2830 rdp->qlen_last_fqs_check = 0;
2831 rdp->n_force_qs_snap = rsp->n_force_qs;
2832 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2833 rdp->qlen_last_fqs_check = count;
2834 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2836 local_irq_restore(flags);
2838 /* Re-invoke RCU core processing if there are callbacks remaining. */
2839 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2844 * Check to see if this CPU is in a non-context-switch quiescent state
2845 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2846 * Also schedule RCU core processing.
2848 * This function must be called from hardirq context. It is normally
2849 * invoked from the scheduling-clock interrupt.
2851 void rcu_check_callbacks(int user)
2853 trace_rcu_utilization(TPS("Start scheduler-tick"));
2854 increment_cpu_stall_ticks();
2855 if (user || rcu_is_cpu_rrupt_from_idle()) {
2858 * Get here if this CPU took its interrupt from user
2859 * mode or from the idle loop, and if this is not a
2860 * nested interrupt. In this case, the CPU is in
2861 * a quiescent state, so note it.
2863 * No memory barrier is required here because both
2864 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2865 * variables that other CPUs neither access nor modify,
2866 * at least not while the corresponding CPU is online.
2872 } else if (!in_softirq()) {
2875 * Get here if this CPU did not take its interrupt from
2876 * softirq, in other words, if it is not interrupting
2877 * a rcu_bh read-side critical section. This is an _bh
2878 * critical section, so note it.
2883 rcu_preempt_check_callbacks();
2887 rcu_note_voluntary_context_switch(current);
2888 trace_rcu_utilization(TPS("End scheduler-tick"));
2892 * Scan the leaf rcu_node structures, processing dyntick state for any that
2893 * have not yet encountered a quiescent state, using the function specified.
2894 * Also initiate boosting for any threads blocked on the root rcu_node.
2896 * The caller must have suppressed start of new grace periods.
2898 static void force_qs_rnp(struct rcu_state *rsp,
2899 int (*f)(struct rcu_data *rsp, bool *isidle,
2900 unsigned long *maxj),
2901 bool *isidle, unsigned long *maxj)
2904 unsigned long flags;
2906 struct rcu_node *rnp;
2908 rcu_for_each_leaf_node(rsp, rnp) {
2909 cond_resched_rcu_qs();
2911 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2912 if (rnp->qsmask == 0) {
2913 if (rcu_state_p == &rcu_sched_state ||
2914 rsp != rcu_state_p ||
2915 rcu_preempt_blocked_readers_cgp(rnp)) {
2917 * No point in scanning bits because they
2918 * are all zero. But we might need to
2919 * priority-boost blocked readers.
2921 rcu_initiate_boost(rnp, flags);
2922 /* rcu_initiate_boost() releases rnp->lock */
2926 (rnp->parent->qsmask & rnp->grpmask)) {
2928 * Race between grace-period
2929 * initialization and task exiting RCU
2930 * read-side critical section: Report.
2932 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2933 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2937 for_each_leaf_node_possible_cpu(rnp, cpu) {
2938 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2939 if ((rnp->qsmask & bit) != 0) {
2940 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2945 /* Idle/offline CPUs, report (releases rnp->lock. */
2946 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2948 /* Nothing to do here, so just drop the lock. */
2949 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2955 * Force quiescent states on reluctant CPUs, and also detect which
2956 * CPUs are in dyntick-idle mode.
2958 static void force_quiescent_state(struct rcu_state *rsp)
2960 unsigned long flags;
2962 struct rcu_node *rnp;
2963 struct rcu_node *rnp_old = NULL;
2965 /* Funnel through hierarchy to reduce memory contention. */
2966 rnp = __this_cpu_read(rsp->rda->mynode);
2967 for (; rnp != NULL; rnp = rnp->parent) {
2968 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2969 !raw_spin_trylock(&rnp->fqslock);
2970 if (rnp_old != NULL)
2971 raw_spin_unlock(&rnp_old->fqslock);
2973 rsp->n_force_qs_lh++;
2978 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2980 /* Reached the root of the rcu_node tree, acquire lock. */
2981 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2982 raw_spin_unlock(&rnp_old->fqslock);
2983 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2984 rsp->n_force_qs_lh++;
2985 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2986 return; /* Someone beat us to it. */
2988 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2989 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2990 rcu_gp_kthread_wake(rsp);
2994 * This does the RCU core processing work for the specified rcu_state
2995 * and rcu_data structures. This may be called only from the CPU to
2996 * whom the rdp belongs.
2999 __rcu_process_callbacks(struct rcu_state *rsp)
3001 unsigned long flags;
3003 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3005 WARN_ON_ONCE(!rdp->beenonline);
3007 /* Update RCU state based on any recent quiescent states. */
3008 rcu_check_quiescent_state(rsp, rdp);
3010 /* Does this CPU require a not-yet-started grace period? */
3011 local_irq_save(flags);
3012 if (cpu_needs_another_gp(rsp, rdp)) {
3013 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
3014 needwake = rcu_start_gp(rsp);
3015 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
3017 rcu_gp_kthread_wake(rsp);
3019 local_irq_restore(flags);
3022 /* If there are callbacks ready, invoke them. */
3023 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3024 invoke_rcu_callbacks(rsp, rdp);
3026 /* Do any needed deferred wakeups of rcuo kthreads. */
3027 do_nocb_deferred_wakeup(rdp);
3031 * Do RCU core processing for the current CPU.
3033 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
3035 struct rcu_state *rsp;
3037 if (cpu_is_offline(smp_processor_id()))
3039 trace_rcu_utilization(TPS("Start RCU core"));
3040 for_each_rcu_flavor(rsp)
3041 __rcu_process_callbacks(rsp);
3042 trace_rcu_utilization(TPS("End RCU core"));
3046 * Schedule RCU callback invocation. If the specified type of RCU
3047 * does not support RCU priority boosting, just do a direct call,
3048 * otherwise wake up the per-CPU kernel kthread. Note that because we
3049 * are running on the current CPU with softirqs disabled, the
3050 * rcu_cpu_kthread_task cannot disappear out from under us.
3052 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
3054 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
3056 if (likely(!rsp->boost)) {
3057 rcu_do_batch(rsp, rdp);
3060 invoke_rcu_callbacks_kthread();
3063 static void invoke_rcu_core(void)
3065 if (cpu_online(smp_processor_id()))
3066 raise_softirq(RCU_SOFTIRQ);
3070 * Handle any core-RCU processing required by a call_rcu() invocation.
3072 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
3073 struct rcu_head *head, unsigned long flags)
3078 * If called from an extended quiescent state, invoke the RCU
3079 * core in order to force a re-evaluation of RCU's idleness.
3081 if (!rcu_is_watching())
3084 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3085 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3089 * Force the grace period if too many callbacks or too long waiting.
3090 * Enforce hysteresis, and don't invoke force_quiescent_state()
3091 * if some other CPU has recently done so. Also, don't bother
3092 * invoking force_quiescent_state() if the newly enqueued callback
3093 * is the only one waiting for a grace period to complete.
3095 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
3096 rdp->qlen_last_fqs_check + qhimark)) {
3098 /* Are we ignoring a completed grace period? */
3099 note_gp_changes(rsp, rdp);
3101 /* Start a new grace period if one not already started. */
3102 if (!rcu_gp_in_progress(rsp)) {
3103 struct rcu_node *rnp_root = rcu_get_root(rsp);
3105 raw_spin_lock_rcu_node(rnp_root);
3106 needwake = rcu_start_gp(rsp);
3107 raw_spin_unlock_rcu_node(rnp_root);
3109 rcu_gp_kthread_wake(rsp);
3111 /* Give the grace period a kick. */
3112 rdp->blimit = LONG_MAX;
3113 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3114 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
3115 force_quiescent_state(rsp);
3116 rdp->n_force_qs_snap = rsp->n_force_qs;
3117 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
3123 * RCU callback function to leak a callback.
3125 static void rcu_leak_callback(struct rcu_head *rhp)
3130 * Helper function for call_rcu() and friends. The cpu argument will
3131 * normally be -1, indicating "currently running CPU". It may specify
3132 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3133 * is expected to specify a CPU.
3136 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3137 struct rcu_state *rsp, int cpu, bool lazy)
3139 unsigned long flags;
3140 struct rcu_data *rdp;
3142 /* Misaligned rcu_head! */
3143 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3145 if (debug_rcu_head_queue(head)) {
3146 /* Probable double call_rcu(), so leak the callback. */
3147 WRITE_ONCE(head->func, rcu_leak_callback);
3148 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3153 local_irq_save(flags);
3154 rdp = this_cpu_ptr(rsp->rda);
3156 /* Add the callback to our list. */
3157 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
3161 rdp = per_cpu_ptr(rsp->rda, cpu);
3162 if (likely(rdp->mynode)) {
3163 /* Post-boot, so this should be for a no-CBs CPU. */
3164 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3165 WARN_ON_ONCE(offline);
3166 /* Offline CPU, _call_rcu() illegal, leak callback. */
3167 local_irq_restore(flags);
3171 * Very early boot, before rcu_init(). Initialize if needed
3172 * and then drop through to queue the callback.
3175 WARN_ON_ONCE(!rcu_is_watching());
3176 if (rcu_segcblist_empty(&rdp->cblist))
3177 rcu_segcblist_init(&rdp->cblist);
3179 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3181 rcu_idle_count_callbacks_posted();
3183 if (__is_kfree_rcu_offset((unsigned long)func))
3184 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3185 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3186 rcu_segcblist_n_cbs(&rdp->cblist));
3188 trace_rcu_callback(rsp->name, head,
3189 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3190 rcu_segcblist_n_cbs(&rdp->cblist));
3192 /* Go handle any RCU core processing required. */
3193 __call_rcu_core(rsp, rdp, head, flags);
3194 local_irq_restore(flags);
3198 * Queue an RCU-sched callback for invocation after a grace period.
3200 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3202 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3204 EXPORT_SYMBOL_GPL(call_rcu_sched);
3207 * Queue an RCU callback for invocation after a quicker grace period.
3209 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3211 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3213 EXPORT_SYMBOL_GPL(call_rcu_bh);
3216 * Queue an RCU callback for lazy invocation after a grace period.
3217 * This will likely be later named something like "call_rcu_lazy()",
3218 * but this change will require some way of tagging the lazy RCU
3219 * callbacks in the list of pending callbacks. Until then, this
3220 * function may only be called from __kfree_rcu().
3222 void kfree_call_rcu(struct rcu_head *head,
3223 rcu_callback_t func)
3225 __call_rcu(head, func, rcu_state_p, -1, 1);
3227 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3230 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3231 * any blocking grace-period wait automatically implies a grace period
3232 * if there is only one CPU online at any point time during execution
3233 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3234 * occasionally incorrectly indicate that there are multiple CPUs online
3235 * when there was in fact only one the whole time, as this just adds
3236 * some overhead: RCU still operates correctly.
3238 static inline int rcu_blocking_is_gp(void)
3242 might_sleep(); /* Check for RCU read-side critical section. */
3244 ret = num_online_cpus() <= 1;
3250 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3252 * Control will return to the caller some time after a full rcu-sched
3253 * grace period has elapsed, in other words after all currently executing
3254 * rcu-sched read-side critical sections have completed. These read-side
3255 * critical sections are delimited by rcu_read_lock_sched() and
3256 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3257 * local_irq_disable(), and so on may be used in place of
3258 * rcu_read_lock_sched().
3260 * This means that all preempt_disable code sequences, including NMI and
3261 * non-threaded hardware-interrupt handlers, in progress on entry will
3262 * have completed before this primitive returns. However, this does not
3263 * guarantee that softirq handlers will have completed, since in some
3264 * kernels, these handlers can run in process context, and can block.
3266 * Note that this guarantee implies further memory-ordering guarantees.
3267 * On systems with more than one CPU, when synchronize_sched() returns,
3268 * each CPU is guaranteed to have executed a full memory barrier since the
3269 * end of its last RCU-sched read-side critical section whose beginning
3270 * preceded the call to synchronize_sched(). In addition, each CPU having
3271 * an RCU read-side critical section that extends beyond the return from
3272 * synchronize_sched() is guaranteed to have executed a full memory barrier
3273 * after the beginning of synchronize_sched() and before the beginning of
3274 * that RCU read-side critical section. Note that these guarantees include
3275 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3276 * that are executing in the kernel.
3278 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3279 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3280 * to have executed a full memory barrier during the execution of
3281 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3282 * again only if the system has more than one CPU).
3284 * This primitive provides the guarantees made by the (now removed)
3285 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3286 * guarantees that rcu_read_lock() sections will have completed.
3287 * In "classic RCU", these two guarantees happen to be one and
3288 * the same, but can differ in realtime RCU implementations.
3290 void synchronize_sched(void)
3292 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3293 lock_is_held(&rcu_lock_map) ||
3294 lock_is_held(&rcu_sched_lock_map),
3295 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3296 if (rcu_blocking_is_gp())
3298 if (rcu_gp_is_expedited())
3299 synchronize_sched_expedited();
3301 wait_rcu_gp(call_rcu_sched);
3303 EXPORT_SYMBOL_GPL(synchronize_sched);
3306 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3308 * Control will return to the caller some time after a full rcu_bh grace
3309 * period has elapsed, in other words after all currently executing rcu_bh
3310 * read-side critical sections have completed. RCU read-side critical
3311 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3312 * and may be nested.
3314 * See the description of synchronize_sched() for more detailed information
3315 * on memory ordering guarantees.
3317 void synchronize_rcu_bh(void)
3319 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3320 lock_is_held(&rcu_lock_map) ||
3321 lock_is_held(&rcu_sched_lock_map),
3322 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3323 if (rcu_blocking_is_gp())
3325 if (rcu_gp_is_expedited())
3326 synchronize_rcu_bh_expedited();
3328 wait_rcu_gp(call_rcu_bh);
3330 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3333 * get_state_synchronize_rcu - Snapshot current RCU state
3335 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3336 * to determine whether or not a full grace period has elapsed in the
3339 unsigned long get_state_synchronize_rcu(void)
3342 * Any prior manipulation of RCU-protected data must happen
3343 * before the load from ->gpnum.
3348 * Make sure this load happens before the purportedly
3349 * time-consuming work between get_state_synchronize_rcu()
3350 * and cond_synchronize_rcu().
3352 return smp_load_acquire(&rcu_state_p->gpnum);
3354 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3357 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3359 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3361 * If a full RCU grace period has elapsed since the earlier call to
3362 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3363 * synchronize_rcu() to wait for a full grace period.
3365 * Yes, this function does not take counter wrap into account. But
3366 * counter wrap is harmless. If the counter wraps, we have waited for
3367 * more than 2 billion grace periods (and way more on a 64-bit system!),
3368 * so waiting for one additional grace period should be just fine.
3370 void cond_synchronize_rcu(unsigned long oldstate)
3372 unsigned long newstate;
3375 * Ensure that this load happens before any RCU-destructive
3376 * actions the caller might carry out after we return.
3378 newstate = smp_load_acquire(&rcu_state_p->completed);
3379 if (ULONG_CMP_GE(oldstate, newstate))
3382 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3385 * get_state_synchronize_sched - Snapshot current RCU-sched state
3387 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3388 * to determine whether or not a full grace period has elapsed in the
3391 unsigned long get_state_synchronize_sched(void)
3394 * Any prior manipulation of RCU-protected data must happen
3395 * before the load from ->gpnum.
3400 * Make sure this load happens before the purportedly
3401 * time-consuming work between get_state_synchronize_sched()
3402 * and cond_synchronize_sched().
3404 return smp_load_acquire(&rcu_sched_state.gpnum);
3406 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3409 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3411 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3413 * If a full RCU-sched grace period has elapsed since the earlier call to
3414 * get_state_synchronize_sched(), just return. Otherwise, invoke
3415 * synchronize_sched() to wait for a full grace period.
3417 * Yes, this function does not take counter wrap into account. But
3418 * counter wrap is harmless. If the counter wraps, we have waited for
3419 * more than 2 billion grace periods (and way more on a 64-bit system!),
3420 * so waiting for one additional grace period should be just fine.
3422 void cond_synchronize_sched(unsigned long oldstate)
3424 unsigned long newstate;
3427 * Ensure that this load happens before any RCU-destructive
3428 * actions the caller might carry out after we return.
3430 newstate = smp_load_acquire(&rcu_sched_state.completed);
3431 if (ULONG_CMP_GE(oldstate, newstate))
3432 synchronize_sched();
3434 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3437 * Check to see if there is any immediate RCU-related work to be done
3438 * by the current CPU, for the specified type of RCU, returning 1 if so.
3439 * The checks are in order of increasing expense: checks that can be
3440 * carried out against CPU-local state are performed first. However,
3441 * we must check for CPU stalls first, else we might not get a chance.
3443 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3445 struct rcu_node *rnp = rdp->mynode;
3447 rdp->n_rcu_pending++;
3449 /* Check for CPU stalls, if enabled. */
3450 check_cpu_stall(rsp, rdp);
3452 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3453 if (rcu_nohz_full_cpu(rsp))
3456 /* Is the RCU core waiting for a quiescent state from this CPU? */
3457 if (rcu_scheduler_fully_active &&
3458 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3459 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
3460 rdp->n_rp_core_needs_qs++;
3461 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3462 rdp->n_rp_report_qs++;
3466 /* Does this CPU have callbacks ready to invoke? */
3467 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
3468 rdp->n_rp_cb_ready++;
3472 /* Has RCU gone idle with this CPU needing another grace period? */
3473 if (cpu_needs_another_gp(rsp, rdp)) {
3474 rdp->n_rp_cpu_needs_gp++;
3478 /* Has another RCU grace period completed? */
3479 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3480 rdp->n_rp_gp_completed++;
3484 /* Has a new RCU grace period started? */
3485 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3486 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3487 rdp->n_rp_gp_started++;
3491 /* Does this CPU need a deferred NOCB wakeup? */
3492 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3493 rdp->n_rp_nocb_defer_wakeup++;
3498 rdp->n_rp_need_nothing++;
3503 * Check to see if there is any immediate RCU-related work to be done
3504 * by the current CPU, returning 1 if so. This function is part of the
3505 * RCU implementation; it is -not- an exported member of the RCU API.
3507 static int rcu_pending(void)
3509 struct rcu_state *rsp;
3511 for_each_rcu_flavor(rsp)
3512 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3518 * Return true if the specified CPU has any callback. If all_lazy is
3519 * non-NULL, store an indication of whether all callbacks are lazy.
3520 * (If there are no callbacks, all of them are deemed to be lazy.)
3522 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3526 struct rcu_data *rdp;
3527 struct rcu_state *rsp;
3529 for_each_rcu_flavor(rsp) {
3530 rdp = this_cpu_ptr(rsp->rda);
3531 if (rcu_segcblist_empty(&rdp->cblist))
3534 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3545 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3546 * the compiler is expected to optimize this away.
3548 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3549 int cpu, unsigned long done)
3551 trace_rcu_barrier(rsp->name, s, cpu,
3552 atomic_read(&rsp->barrier_cpu_count), done);
3556 * RCU callback function for _rcu_barrier(). If we are last, wake
3557 * up the task executing _rcu_barrier().
3559 static void rcu_barrier_callback(struct rcu_head *rhp)
3561 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3562 struct rcu_state *rsp = rdp->rsp;
3564 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3565 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
3566 complete(&rsp->barrier_completion);
3568 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
3573 * Called with preemption disabled, and from cross-cpu IRQ context.
3575 static void rcu_barrier_func(void *type)
3577 struct rcu_state *rsp = type;
3578 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3580 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
3581 atomic_inc(&rsp->barrier_cpu_count);
3582 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3586 * Orchestrate the specified type of RCU barrier, waiting for all
3587 * RCU callbacks of the specified type to complete.
3589 static void _rcu_barrier(struct rcu_state *rsp)
3592 struct rcu_data *rdp;
3593 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3595 _rcu_barrier_trace(rsp, "Begin", -1, s);
3597 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3598 mutex_lock(&rsp->barrier_mutex);
3600 /* Did someone else do our work for us? */
3601 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3602 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
3603 smp_mb(); /* caller's subsequent code after above check. */
3604 mutex_unlock(&rsp->barrier_mutex);
3608 /* Mark the start of the barrier operation. */
3609 rcu_seq_start(&rsp->barrier_sequence);
3610 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
3613 * Initialize the count to one rather than to zero in order to
3614 * avoid a too-soon return to zero in case of a short grace period
3615 * (or preemption of this task). Exclude CPU-hotplug operations
3616 * to ensure that no offline CPU has callbacks queued.
3618 init_completion(&rsp->barrier_completion);
3619 atomic_set(&rsp->barrier_cpu_count, 1);
3623 * Force each CPU with callbacks to register a new callback.
3624 * When that callback is invoked, we will know that all of the
3625 * corresponding CPU's preceding callbacks have been invoked.
3627 for_each_possible_cpu(cpu) {
3628 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3630 rdp = per_cpu_ptr(rsp->rda, cpu);
3631 if (rcu_is_nocb_cpu(cpu)) {
3632 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3633 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3634 rsp->barrier_sequence);
3636 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3637 rsp->barrier_sequence);
3638 smp_mb__before_atomic();
3639 atomic_inc(&rsp->barrier_cpu_count);
3640 __call_rcu(&rdp->barrier_head,
3641 rcu_barrier_callback, rsp, cpu, 0);
3643 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3644 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3645 rsp->barrier_sequence);
3646 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3648 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3649 rsp->barrier_sequence);
3655 * Now that we have an rcu_barrier_callback() callback on each
3656 * CPU, and thus each counted, remove the initial count.
3658 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3659 complete(&rsp->barrier_completion);
3661 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3662 wait_for_completion(&rsp->barrier_completion);
3664 /* Mark the end of the barrier operation. */
3665 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
3666 rcu_seq_end(&rsp->barrier_sequence);
3668 /* Other rcu_barrier() invocations can now safely proceed. */
3669 mutex_unlock(&rsp->barrier_mutex);
3673 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3675 void rcu_barrier_bh(void)
3677 _rcu_barrier(&rcu_bh_state);
3679 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3682 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3684 void rcu_barrier_sched(void)
3686 _rcu_barrier(&rcu_sched_state);
3688 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3691 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3692 * first CPU in a given leaf rcu_node structure coming online. The caller
3693 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3696 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3699 struct rcu_node *rnp = rnp_leaf;
3702 mask = rnp->grpmask;
3706 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3707 rnp->qsmaskinit |= mask;
3708 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3713 * Do boot-time initialization of a CPU's per-CPU RCU data.
3716 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3718 unsigned long flags;
3719 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3720 struct rcu_node *rnp = rcu_get_root(rsp);
3722 /* Set up local state, ensuring consistent view of global state. */
3723 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3724 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3725 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3726 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3727 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3730 rcu_boot_init_nocb_percpu_data(rdp);
3731 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3735 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3736 * offline event can be happening at a given time. Note also that we
3737 * can accept some slop in the rsp->completed access due to the fact
3738 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3741 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3743 unsigned long flags;
3744 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3745 struct rcu_node *rnp = rcu_get_root(rsp);
3747 /* Set up local state, ensuring consistent view of global state. */
3748 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3749 rdp->qlen_last_fqs_check = 0;
3750 rdp->n_force_qs_snap = rsp->n_force_qs;
3751 rdp->blimit = blimit;
3752 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3753 !init_nocb_callback_list(rdp))
3754 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3755 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3756 rcu_sysidle_init_percpu_data(rdp->dynticks);
3757 rcu_dynticks_eqs_online();
3758 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3761 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3762 * propagation up the rcu_node tree will happen at the beginning
3763 * of the next grace period.
3766 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3767 if (!rdp->beenonline)
3768 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
3769 rdp->beenonline = true; /* We have now been online. */
3770 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3771 rdp->completed = rnp->completed;
3772 rdp->cpu_no_qs.b.norm = true;
3773 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3774 rdp->core_needs_qs = false;
3775 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3776 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3780 * Invoked early in the CPU-online process, when pretty much all
3781 * services are available. The incoming CPU is not present.
3783 int rcutree_prepare_cpu(unsigned int cpu)
3785 struct rcu_state *rsp;
3787 for_each_rcu_flavor(rsp)
3788 rcu_init_percpu_data(cpu, rsp);
3790 rcu_prepare_kthreads(cpu);
3791 rcu_spawn_all_nocb_kthreads(cpu);
3797 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3799 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3801 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3803 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3807 * Near the end of the CPU-online process. Pretty much all services
3808 * enabled, and the CPU is now very much alive.
3810 int rcutree_online_cpu(unsigned int cpu)
3812 sync_sched_exp_online_cleanup(cpu);
3813 rcutree_affinity_setting(cpu, -1);
3814 if (IS_ENABLED(CONFIG_TREE_SRCU))
3815 srcu_online_cpu(cpu);
3820 * Near the beginning of the process. The CPU is still very much alive
3821 * with pretty much all services enabled.
3823 int rcutree_offline_cpu(unsigned int cpu)
3825 rcutree_affinity_setting(cpu, cpu);
3826 if (IS_ENABLED(CONFIG_TREE_SRCU))
3827 srcu_offline_cpu(cpu);
3832 * Near the end of the offline process. We do only tracing here.
3834 int rcutree_dying_cpu(unsigned int cpu)
3836 struct rcu_state *rsp;
3838 for_each_rcu_flavor(rsp)
3839 rcu_cleanup_dying_cpu(rsp);
3844 * The outgoing CPU is gone and we are running elsewhere.
3846 int rcutree_dead_cpu(unsigned int cpu)
3848 struct rcu_state *rsp;
3850 for_each_rcu_flavor(rsp) {
3851 rcu_cleanup_dead_cpu(cpu, rsp);
3852 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3858 * Mark the specified CPU as being online so that subsequent grace periods
3859 * (both expedited and normal) will wait on it. Note that this means that
3860 * incoming CPUs are not allowed to use RCU read-side critical sections
3861 * until this function is called. Failing to observe this restriction
3862 * will result in lockdep splats.
3864 * Note that this function is special in that it is invoked directly
3865 * from the incoming CPU rather than from the cpuhp_step mechanism.
3866 * This is because this function must be invoked at a precise location.
3868 void rcu_cpu_starting(unsigned int cpu)
3870 unsigned long flags;
3872 struct rcu_data *rdp;
3873 struct rcu_node *rnp;
3874 struct rcu_state *rsp;
3876 for_each_rcu_flavor(rsp) {
3877 rdp = per_cpu_ptr(rsp->rda, cpu);
3879 mask = rdp->grpmask;
3880 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3881 rnp->qsmaskinitnext |= mask;
3882 rnp->expmaskinitnext |= mask;
3883 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3887 #ifdef CONFIG_HOTPLUG_CPU
3889 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3890 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3893 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3895 unsigned long flags;
3897 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3898 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3900 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3901 mask = rdp->grpmask;
3902 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3903 rnp->qsmaskinitnext &= ~mask;
3904 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3908 * The outgoing function has no further need of RCU, so remove it from
3909 * the list of CPUs that RCU must track.
3911 * Note that this function is special in that it is invoked directly
3912 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3913 * This is because this function must be invoked at a precise location.
3915 void rcu_report_dead(unsigned int cpu)
3917 struct rcu_state *rsp;
3919 /* QS for any half-done expedited RCU-sched GP. */
3921 rcu_report_exp_rdp(&rcu_sched_state,
3922 this_cpu_ptr(rcu_sched_state.rda), true);
3924 for_each_rcu_flavor(rsp)
3925 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3930 * On non-huge systems, use expedited RCU grace periods to make suspend
3931 * and hibernation run faster.
3933 static int rcu_pm_notify(struct notifier_block *self,
3934 unsigned long action, void *hcpu)
3937 case PM_HIBERNATION_PREPARE:
3938 case PM_SUSPEND_PREPARE:
3939 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3942 case PM_POST_HIBERNATION:
3943 case PM_POST_SUSPEND:
3944 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3945 rcu_unexpedite_gp();
3954 * Spawn the kthreads that handle each RCU flavor's grace periods.
3956 static int __init rcu_spawn_gp_kthread(void)
3958 unsigned long flags;
3959 int kthread_prio_in = kthread_prio;
3960 struct rcu_node *rnp;
3961 struct rcu_state *rsp;
3962 struct sched_param sp;
3963 struct task_struct *t;
3965 /* Force priority into range. */
3966 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3968 else if (kthread_prio < 0)
3970 else if (kthread_prio > 99)
3972 if (kthread_prio != kthread_prio_in)
3973 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3974 kthread_prio, kthread_prio_in);
3976 rcu_scheduler_fully_active = 1;
3977 for_each_rcu_flavor(rsp) {
3978 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3980 rnp = rcu_get_root(rsp);
3981 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3982 rsp->gp_kthread = t;
3984 sp.sched_priority = kthread_prio;
3985 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3987 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3990 rcu_spawn_nocb_kthreads();
3991 rcu_spawn_boost_kthreads();
3994 early_initcall(rcu_spawn_gp_kthread);
3997 * This function is invoked towards the end of the scheduler's
3998 * initialization process. Before this is called, the idle task might
3999 * contain synchronous grace-period primitives (during which time, this idle
4000 * task is booting the system, and such primitives are no-ops). After this
4001 * function is called, any synchronous grace-period primitives are run as
4002 * expedited, with the requesting task driving the grace period forward.
4003 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4004 * runtime RCU functionality.
4006 void rcu_scheduler_starting(void)
4008 WARN_ON(num_online_cpus() != 1);
4009 WARN_ON(nr_context_switches() > 0);
4010 rcu_test_sync_prims();
4011 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4012 rcu_test_sync_prims();
4016 * Helper function for rcu_init() that initializes one rcu_state structure.
4018 static void __init rcu_init_one(struct rcu_state *rsp)
4020 static const char * const buf[] = RCU_NODE_NAME_INIT;
4021 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4022 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4023 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4025 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4029 struct rcu_node *rnp;
4031 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4033 /* Silence gcc 4.8 false positive about array index out of range. */
4034 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4035 panic("rcu_init_one: rcu_num_lvls out of range");
4037 /* Initialize the level-tracking arrays. */
4039 for (i = 1; i < rcu_num_lvls; i++)
4040 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
4041 rcu_init_levelspread(levelspread, num_rcu_lvl);
4043 /* Initialize the elements themselves, starting from the leaves. */
4045 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4046 cpustride *= levelspread[i];
4047 rnp = rsp->level[i];
4048 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4049 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4050 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4051 &rcu_node_class[i], buf[i]);
4052 raw_spin_lock_init(&rnp->fqslock);
4053 lockdep_set_class_and_name(&rnp->fqslock,
4054 &rcu_fqs_class[i], fqs[i]);
4055 rnp->gpnum = rsp->gpnum;
4056 rnp->completed = rsp->completed;
4058 rnp->qsmaskinit = 0;
4059 rnp->grplo = j * cpustride;
4060 rnp->grphi = (j + 1) * cpustride - 1;
4061 if (rnp->grphi >= nr_cpu_ids)
4062 rnp->grphi = nr_cpu_ids - 1;
4068 rnp->grpnum = j % levelspread[i - 1];
4069 rnp->grpmask = 1UL << rnp->grpnum;
4070 rnp->parent = rsp->level[i - 1] +
4071 j / levelspread[i - 1];
4074 INIT_LIST_HEAD(&rnp->blkd_tasks);
4075 rcu_init_one_nocb(rnp);
4076 init_waitqueue_head(&rnp->exp_wq[0]);
4077 init_waitqueue_head(&rnp->exp_wq[1]);
4078 init_waitqueue_head(&rnp->exp_wq[2]);
4079 init_waitqueue_head(&rnp->exp_wq[3]);
4080 spin_lock_init(&rnp->exp_lock);
4084 init_swait_queue_head(&rsp->gp_wq);
4085 init_swait_queue_head(&rsp->expedited_wq);
4086 rnp = rsp->level[rcu_num_lvls - 1];
4087 for_each_possible_cpu(i) {
4088 while (i > rnp->grphi)
4090 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4091 rcu_boot_init_percpu_data(i, rsp);
4093 list_add(&rsp->flavors, &rcu_struct_flavors);
4097 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4098 * replace the definitions in tree.h because those are needed to size
4099 * the ->node array in the rcu_state structure.
4101 static void __init rcu_init_geometry(void)
4105 int rcu_capacity[RCU_NUM_LVLS];
4108 * Initialize any unspecified boot parameters.
4109 * The default values of jiffies_till_first_fqs and
4110 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4111 * value, which is a function of HZ, then adding one for each
4112 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4114 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4115 if (jiffies_till_first_fqs == ULONG_MAX)
4116 jiffies_till_first_fqs = d;
4117 if (jiffies_till_next_fqs == ULONG_MAX)
4118 jiffies_till_next_fqs = d;
4120 /* If the compile-time values are accurate, just leave. */
4121 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4122 nr_cpu_ids == NR_CPUS)
4124 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4125 rcu_fanout_leaf, nr_cpu_ids);
4128 * The boot-time rcu_fanout_leaf parameter must be at least two
4129 * and cannot exceed the number of bits in the rcu_node masks.
4130 * Complain and fall back to the compile-time values if this
4131 * limit is exceeded.
4133 if (rcu_fanout_leaf < 2 ||
4134 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4135 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4141 * Compute number of nodes that can be handled an rcu_node tree
4142 * with the given number of levels.
4144 rcu_capacity[0] = rcu_fanout_leaf;
4145 for (i = 1; i < RCU_NUM_LVLS; i++)
4146 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4149 * The tree must be able to accommodate the configured number of CPUs.
4150 * If this limit is exceeded, fall back to the compile-time values.
4152 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4153 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4158 /* Calculate the number of levels in the tree. */
4159 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4161 rcu_num_lvls = i + 1;
4163 /* Calculate the number of rcu_nodes at each level of the tree. */
4164 for (i = 0; i < rcu_num_lvls; i++) {
4165 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4166 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4169 /* Calculate the total number of rcu_node structures. */
4171 for (i = 0; i < rcu_num_lvls; i++)
4172 rcu_num_nodes += num_rcu_lvl[i];
4176 * Dump out the structure of the rcu_node combining tree associated
4177 * with the rcu_state structure referenced by rsp.
4179 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4182 struct rcu_node *rnp;
4184 pr_info("rcu_node tree layout dump\n");
4186 rcu_for_each_node_breadth_first(rsp, rnp) {
4187 if (rnp->level != level) {
4192 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4197 void __init rcu_init(void)
4201 rcu_early_boot_tests();
4203 rcu_bootup_announce();
4204 rcu_init_geometry();
4205 rcu_init_one(&rcu_bh_state);
4206 rcu_init_one(&rcu_sched_state);
4208 rcu_dump_rcu_node_tree(&rcu_sched_state);
4209 __rcu_init_preempt();
4210 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4213 * We don't need protection against CPU-hotplug here because
4214 * this is called early in boot, before either interrupts
4215 * or the scheduler are operational.
4217 pm_notifier(rcu_pm_notify, 0);
4218 for_each_online_cpu(cpu) {
4219 rcutree_prepare_cpu(cpu);
4220 rcu_cpu_starting(cpu);
4221 if (IS_ENABLED(CONFIG_TREE_SRCU))
4222 srcu_online_cpu(cpu);
4226 #include "tree_exp.h"
4227 #include "tree_plugin.h"