2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate_wait.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/sched/debug.h>
39 #include <linux/nmi.h>
40 #include <linux/atomic.h>
41 #include <linux/bitops.h>
42 #include <linux/export.h>
43 #include <linux/completion.h>
44 #include <linux/moduleparam.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/mutex.h>
49 #include <linux/time.h>
50 #include <linux/kernel_stat.h>
51 #include <linux/wait.h>
52 #include <linux/kthread.h>
53 #include <uapi/linux/sched/types.h>
54 #include <linux/prefetch.h>
55 #include <linux/delay.h>
56 #include <linux/stop_machine.h>
57 #include <linux/random.h>
58 #include <linux/trace_events.h>
59 #include <linux/suspend.h>
60 #include <linux/ftrace.h>
65 #ifdef MODULE_PARAM_PREFIX
66 #undef MODULE_PARAM_PREFIX
68 #define MODULE_PARAM_PREFIX "rcutree."
70 /* Data structures. */
73 * In order to export the rcu_state name to the tracing tools, it
74 * needs to be added in the __tracepoint_string section.
75 * This requires defining a separate variable tp_<sname>_varname
76 * that points to the string being used, and this will allow
77 * the tracing userspace tools to be able to decipher the string
78 * address to the matching string.
81 # define DEFINE_RCU_TPS(sname) \
82 static char sname##_varname[] = #sname; \
83 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
84 # define RCU_STATE_NAME(sname) sname##_varname
86 # define DEFINE_RCU_TPS(sname)
87 # define RCU_STATE_NAME(sname) __stringify(sname)
90 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
91 DEFINE_RCU_TPS(sname) \
92 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
93 struct rcu_state sname##_state = { \
94 .level = { &sname##_state.node[0] }, \
95 .rda = &sname##_data, \
97 .gp_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
101 .name = RCU_STATE_NAME(sname), \
103 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
104 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
107 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
110 static struct rcu_state *const rcu_state_p;
111 LIST_HEAD(rcu_struct_flavors);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree;
115 module_param(dump_tree, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact;
118 module_param(rcu_fanout_exact, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121 module_param(rcu_fanout_leaf, int, 0444);
122 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123 /* Number of rcu_nodes at specified level. */
124 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126 /* panic() on RCU Stall sysctl. */
127 int sysctl_panic_on_rcu_stall __read_mostly;
130 * The rcu_scheduler_active variable is initialized to the value
131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
132 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
133 * RCU can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_rcu() to a simple barrier(). When this variable
135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
136 * to detect real grace periods. This variable is also used to suppress
137 * boot-time false positives from lockdep-RCU error checking. Finally, it
138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
139 * is fully initialized, including all of its kthreads having been spawned.
141 int rcu_scheduler_active __read_mostly;
142 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
156 static int rcu_scheduler_fully_active __read_mostly;
158 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
159 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163 static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
165 static void sync_sched_exp_online_cleanup(int cpu);
167 /* rcuc/rcub kthread realtime priority */
168 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
169 module_param(kthread_prio, int, 0644);
171 /* Delay in jiffies for grace-period initialization delays, debug only. */
173 static int gp_preinit_delay;
174 module_param(gp_preinit_delay, int, 0444);
175 static int gp_init_delay;
176 module_param(gp_init_delay, int, 0444);
177 static int gp_cleanup_delay;
178 module_param(gp_cleanup_delay, int, 0444);
181 * Number of grace periods between delays, normalized by the duration of
182 * the delay. The longer the delay, the more the grace periods between
183 * each delay. The reason for this normalization is that it means that,
184 * for non-zero delays, the overall slowdown of grace periods is constant
185 * regardless of the duration of the delay. This arrangement balances
186 * the need for long delays to increase some race probabilities with the
187 * need for fast grace periods to increase other race probabilities.
189 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
192 * Track the rcutorture test sequence number and the update version
193 * number within a given test. The rcutorture_testseq is incremented
194 * on every rcutorture module load and unload, so has an odd value
195 * when a test is running. The rcutorture_vernum is set to zero
196 * when rcutorture starts and is incremented on each rcutorture update.
197 * These variables enable correlating rcutorture output with the
198 * RCU tracing information.
200 unsigned long rcutorture_testseq;
201 unsigned long rcutorture_vernum;
204 * Compute the mask of online CPUs for the specified rcu_node structure.
205 * This will not be stable unless the rcu_node structure's ->lock is
206 * held, but the bit corresponding to the current CPU will be stable
209 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
211 return READ_ONCE(rnp->qsmaskinitnext);
215 * Return true if an RCU grace period is in progress. The READ_ONCE()s
216 * permit this function to be invoked without holding the root rcu_node
217 * structure's ->lock, but of course results can be subject to change.
219 static int rcu_gp_in_progress(struct rcu_state *rsp)
221 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
225 * Note a quiescent state. Because we do not need to know
226 * how many quiescent states passed, just if there was at least
227 * one since the start of the grace period, this just sets a flag.
228 * The caller must have disabled preemption.
230 void rcu_sched_qs(void)
232 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
233 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
235 trace_rcu_grace_period(TPS("rcu_sched"),
236 __this_cpu_read(rcu_sched_data.gpnum),
238 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
239 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
241 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
242 rcu_report_exp_rdp(&rcu_sched_state,
243 this_cpu_ptr(&rcu_sched_data), true);
248 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
249 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
250 trace_rcu_grace_period(TPS("rcu_bh"),
251 __this_cpu_read(rcu_bh_data.gpnum),
253 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
258 * Steal a bit from the bottom of ->dynticks for idle entry/exit
259 * control. Initially this is for TLB flushing.
261 #define RCU_DYNTICK_CTRL_MASK 0x1
262 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
263 #ifndef rcu_eqs_special_exit
264 #define rcu_eqs_special_exit() do { } while (0)
267 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
268 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
269 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
273 * There's a few places, currently just in the tracing infrastructure,
274 * that uses rcu_irq_enter() to make sure RCU is watching. But there's
275 * a small location where that will not even work. In those cases
276 * rcu_irq_enter_disabled() needs to be checked to make sure rcu_irq_enter()
279 static DEFINE_PER_CPU(bool, disable_rcu_irq_enter);
281 bool rcu_irq_enter_disabled(void)
283 return this_cpu_read(disable_rcu_irq_enter);
287 * Record entry into an extended quiescent state. This is only to be
288 * called when not already in an extended quiescent state.
290 static void rcu_dynticks_eqs_enter(void)
292 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
296 * CPUs seeing atomic_add_return() must see prior RCU read-side
297 * critical sections, and we also must force ordering with the
300 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
301 /* Better be in an extended quiescent state! */
302 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
303 (seq & RCU_DYNTICK_CTRL_CTR));
304 /* Better not have special action (TLB flush) pending! */
305 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
306 (seq & RCU_DYNTICK_CTRL_MASK));
310 * Record exit from an extended quiescent state. This is only to be
311 * called from an extended quiescent state.
313 static void rcu_dynticks_eqs_exit(void)
315 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
319 * CPUs seeing atomic_add_return() must see prior idle sojourns,
320 * and we also must force ordering with the next RCU read-side
323 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
324 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
325 !(seq & RCU_DYNTICK_CTRL_CTR));
326 if (seq & RCU_DYNTICK_CTRL_MASK) {
327 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
328 smp_mb__after_atomic(); /* _exit after clearing mask. */
329 /* Prefer duplicate flushes to losing a flush. */
330 rcu_eqs_special_exit();
335 * Reset the current CPU's ->dynticks counter to indicate that the
336 * newly onlined CPU is no longer in an extended quiescent state.
337 * This will either leave the counter unchanged, or increment it
338 * to the next non-quiescent value.
340 * The non-atomic test/increment sequence works because the upper bits
341 * of the ->dynticks counter are manipulated only by the corresponding CPU,
342 * or when the corresponding CPU is offline.
344 static void rcu_dynticks_eqs_online(void)
346 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
348 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
350 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
354 * Is the current CPU in an extended quiescent state?
356 * No ordering, as we are sampling CPU-local information.
358 bool rcu_dynticks_curr_cpu_in_eqs(void)
360 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
362 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
366 * Snapshot the ->dynticks counter with full ordering so as to allow
367 * stable comparison of this counter with past and future snapshots.
369 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
371 int snap = atomic_add_return(0, &rdtp->dynticks);
373 return snap & ~RCU_DYNTICK_CTRL_MASK;
377 * Return true if the snapshot returned from rcu_dynticks_snap()
378 * indicates that RCU is in an extended quiescent state.
380 static bool rcu_dynticks_in_eqs(int snap)
382 return !(snap & RCU_DYNTICK_CTRL_CTR);
386 * Return true if the CPU corresponding to the specified rcu_dynticks
387 * structure has spent some time in an extended quiescent state since
388 * rcu_dynticks_snap() returned the specified snapshot.
390 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
392 return snap != rcu_dynticks_snap(rdtp);
396 * Do a double-increment of the ->dynticks counter to emulate a
397 * momentary idle-CPU quiescent state.
399 static void rcu_dynticks_momentary_idle(void)
401 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
402 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
405 /* It is illegal to call this from idle state. */
406 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
410 * Set the special (bottom) bit of the specified CPU so that it
411 * will take special action (such as flushing its TLB) on the
412 * next exit from an extended quiescent state. Returns true if
413 * the bit was successfully set, or false if the CPU was not in
414 * an extended quiescent state.
416 bool rcu_eqs_special_set(int cpu)
420 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
423 old = atomic_read(&rdtp->dynticks);
424 if (old & RCU_DYNTICK_CTRL_CTR)
426 new = old | RCU_DYNTICK_CTRL_MASK;
427 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
432 * Let the RCU core know that this CPU has gone through the scheduler,
433 * which is a quiescent state. This is called when the need for a
434 * quiescent state is urgent, so we burn an atomic operation and full
435 * memory barriers to let the RCU core know about it, regardless of what
436 * this CPU might (or might not) do in the near future.
438 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
440 * The caller must have disabled interrupts.
442 static void rcu_momentary_dyntick_idle(void)
444 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
445 rcu_dynticks_momentary_idle();
449 * Note a context switch. This is a quiescent state for RCU-sched,
450 * and requires special handling for preemptible RCU.
451 * The caller must have disabled interrupts.
453 void rcu_note_context_switch(bool preempt)
455 barrier(); /* Avoid RCU read-side critical sections leaking down. */
456 trace_rcu_utilization(TPS("Start context switch"));
458 rcu_preempt_note_context_switch(preempt);
459 /* Load rcu_urgent_qs before other flags. */
460 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
462 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
463 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
464 rcu_momentary_dyntick_idle();
465 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
467 rcu_note_voluntary_context_switch_lite(current);
469 trace_rcu_utilization(TPS("End context switch"));
470 barrier(); /* Avoid RCU read-side critical sections leaking up. */
472 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
475 * Register a quiescent state for all RCU flavors. If there is an
476 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
477 * dyntick-idle quiescent state visible to other CPUs (but only for those
478 * RCU flavors in desperate need of a quiescent state, which will normally
479 * be none of them). Either way, do a lightweight quiescent state for
482 * The barrier() calls are redundant in the common case when this is
483 * called externally, but just in case this is called from within this
487 void rcu_all_qs(void)
491 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
494 /* Load rcu_urgent_qs before other flags. */
495 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
499 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
500 barrier(); /* Avoid RCU read-side critical sections leaking down. */
501 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
502 local_irq_save(flags);
503 rcu_momentary_dyntick_idle();
504 local_irq_restore(flags);
506 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
508 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
509 barrier(); /* Avoid RCU read-side critical sections leaking up. */
512 EXPORT_SYMBOL_GPL(rcu_all_qs);
514 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
515 static long blimit = DEFAULT_RCU_BLIMIT;
516 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
517 static long qhimark = DEFAULT_RCU_QHIMARK;
518 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
519 static long qlowmark = DEFAULT_RCU_QLOMARK;
521 module_param(blimit, long, 0444);
522 module_param(qhimark, long, 0444);
523 module_param(qlowmark, long, 0444);
525 static ulong jiffies_till_first_fqs = ULONG_MAX;
526 static ulong jiffies_till_next_fqs = ULONG_MAX;
527 static bool rcu_kick_kthreads;
529 module_param(jiffies_till_first_fqs, ulong, 0644);
530 module_param(jiffies_till_next_fqs, ulong, 0644);
531 module_param(rcu_kick_kthreads, bool, 0644);
534 * How long the grace period must be before we start recruiting
535 * quiescent-state help from rcu_note_context_switch().
537 static ulong jiffies_till_sched_qs = HZ / 20;
538 module_param(jiffies_till_sched_qs, ulong, 0644);
540 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
541 struct rcu_data *rdp);
542 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
543 static void force_quiescent_state(struct rcu_state *rsp);
544 static int rcu_pending(void);
547 * Return the number of RCU batches started thus far for debug & stats.
549 unsigned long rcu_batches_started(void)
551 return rcu_state_p->gpnum;
553 EXPORT_SYMBOL_GPL(rcu_batches_started);
556 * Return the number of RCU-sched batches started thus far for debug & stats.
558 unsigned long rcu_batches_started_sched(void)
560 return rcu_sched_state.gpnum;
562 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
565 * Return the number of RCU BH batches started thus far for debug & stats.
567 unsigned long rcu_batches_started_bh(void)
569 return rcu_bh_state.gpnum;
571 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
574 * Return the number of RCU batches completed thus far for debug & stats.
576 unsigned long rcu_batches_completed(void)
578 return rcu_state_p->completed;
580 EXPORT_SYMBOL_GPL(rcu_batches_completed);
583 * Return the number of RCU-sched batches completed thus far for debug & stats.
585 unsigned long rcu_batches_completed_sched(void)
587 return rcu_sched_state.completed;
589 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
592 * Return the number of RCU BH batches completed thus far for debug & stats.
594 unsigned long rcu_batches_completed_bh(void)
596 return rcu_bh_state.completed;
598 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
601 * Return the number of RCU expedited batches completed thus far for
602 * debug & stats. Odd numbers mean that a batch is in progress, even
603 * numbers mean idle. The value returned will thus be roughly double
604 * the cumulative batches since boot.
606 unsigned long rcu_exp_batches_completed(void)
608 return rcu_state_p->expedited_sequence;
610 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
613 * Return the number of RCU-sched expedited batches completed thus far
614 * for debug & stats. Similar to rcu_exp_batches_completed().
616 unsigned long rcu_exp_batches_completed_sched(void)
618 return rcu_sched_state.expedited_sequence;
620 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
623 * Force a quiescent state.
625 void rcu_force_quiescent_state(void)
627 force_quiescent_state(rcu_state_p);
629 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
632 * Force a quiescent state for RCU BH.
634 void rcu_bh_force_quiescent_state(void)
636 force_quiescent_state(&rcu_bh_state);
638 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
641 * Force a quiescent state for RCU-sched.
643 void rcu_sched_force_quiescent_state(void)
645 force_quiescent_state(&rcu_sched_state);
647 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
650 * Show the state of the grace-period kthreads.
652 void show_rcu_gp_kthreads(void)
654 struct rcu_state *rsp;
656 for_each_rcu_flavor(rsp) {
657 pr_info("%s: wait state: %d ->state: %#lx\n",
658 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
659 /* sched_show_task(rsp->gp_kthread); */
662 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
665 * Record the number of times rcutorture tests have been initiated and
666 * terminated. This information allows the debugfs tracing stats to be
667 * correlated to the rcutorture messages, even when the rcutorture module
668 * is being repeatedly loaded and unloaded. In other words, we cannot
669 * store this state in rcutorture itself.
671 void rcutorture_record_test_transition(void)
673 rcutorture_testseq++;
674 rcutorture_vernum = 0;
676 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
679 * Send along grace-period-related data for rcutorture diagnostics.
681 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
682 unsigned long *gpnum, unsigned long *completed)
684 struct rcu_state *rsp = NULL;
693 case RCU_SCHED_FLAVOR:
694 rsp = &rcu_sched_state;
701 *flags = READ_ONCE(rsp->gp_flags);
702 *gpnum = READ_ONCE(rsp->gpnum);
703 *completed = READ_ONCE(rsp->completed);
705 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
708 * Record the number of writer passes through the current rcutorture test.
709 * This is also used to correlate debugfs tracing stats with the rcutorture
712 void rcutorture_record_progress(unsigned long vernum)
716 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
719 * Return the root node of the specified rcu_state structure.
721 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
723 return &rsp->node[0];
727 * Is there any need for future grace periods?
728 * Interrupts must be disabled. If the caller does not hold the root
729 * rnp_node structure's ->lock, the results are advisory only.
731 static int rcu_future_needs_gp(struct rcu_state *rsp)
733 struct rcu_node *rnp = rcu_get_root(rsp);
734 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
735 int *fp = &rnp->need_future_gp[idx];
737 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_future_needs_gp() invoked with irqs enabled!!!");
738 return READ_ONCE(*fp);
742 * Does the current CPU require a not-yet-started grace period?
743 * The caller must have disabled interrupts to prevent races with
744 * normal callback registry.
747 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
749 RCU_LOCKDEP_WARN(!irqs_disabled(), "cpu_needs_another_gp() invoked with irqs enabled!!!");
750 if (rcu_gp_in_progress(rsp))
751 return false; /* No, a grace period is already in progress. */
752 if (rcu_future_needs_gp(rsp))
753 return true; /* Yes, a no-CBs CPU needs one. */
754 if (!rcu_segcblist_is_enabled(&rdp->cblist))
755 return false; /* No, this is a no-CBs (or offline) CPU. */
756 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
757 return true; /* Yes, CPU has newly registered callbacks. */
758 if (rcu_segcblist_future_gp_needed(&rdp->cblist,
759 READ_ONCE(rsp->completed)))
760 return true; /* Yes, CBs for future grace period. */
761 return false; /* No grace period needed. */
765 * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
767 * Enter idle, doing appropriate accounting. The caller must have
768 * disabled interrupts.
770 static void rcu_eqs_enter_common(bool user)
772 struct rcu_state *rsp;
773 struct rcu_data *rdp;
774 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
776 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_eqs_enter_common() invoked with irqs enabled!!!");
777 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0);
778 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
779 !user && !is_idle_task(current)) {
780 struct task_struct *idle __maybe_unused =
781 idle_task(smp_processor_id());
783 trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0);
784 rcu_ftrace_dump(DUMP_ORIG);
785 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
786 current->pid, current->comm,
787 idle->pid, idle->comm); /* must be idle task! */
789 for_each_rcu_flavor(rsp) {
790 rdp = this_cpu_ptr(rsp->rda);
791 do_nocb_deferred_wakeup(rdp);
793 rcu_prepare_for_idle();
794 __this_cpu_inc(disable_rcu_irq_enter);
795 rdtp->dynticks_nesting = 0; /* Breaks tracing momentarily. */
796 rcu_dynticks_eqs_enter(); /* After this, tracing works again. */
797 __this_cpu_dec(disable_rcu_irq_enter);
798 rcu_dynticks_task_enter();
801 * It is illegal to enter an extended quiescent state while
802 * in an RCU read-side critical section.
804 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
805 "Illegal idle entry in RCU read-side critical section.");
806 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
807 "Illegal idle entry in RCU-bh read-side critical section.");
808 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
809 "Illegal idle entry in RCU-sched read-side critical section.");
813 * Enter an RCU extended quiescent state, which can be either the
814 * idle loop or adaptive-tickless usermode execution.
816 static void rcu_eqs_enter(bool user)
818 struct rcu_dynticks *rdtp;
820 rdtp = this_cpu_ptr(&rcu_dynticks);
821 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
822 (rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
823 if ((rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
824 rcu_eqs_enter_common(user);
826 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
830 * rcu_idle_enter - inform RCU that current CPU is entering idle
832 * Enter idle mode, in other words, -leave- the mode in which RCU
833 * read-side critical sections can occur. (Though RCU read-side
834 * critical sections can occur in irq handlers in idle, a possibility
835 * handled by irq_enter() and irq_exit().)
837 * We crowbar the ->dynticks_nesting field to zero to allow for
838 * the possibility of usermode upcalls having messed up our count
839 * of interrupt nesting level during the prior busy period.
841 void rcu_idle_enter(void)
843 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_idle_enter() invoked with irqs enabled!!!");
844 rcu_eqs_enter(false);
847 #ifdef CONFIG_NO_HZ_FULL
849 * rcu_user_enter - inform RCU that we are resuming userspace.
851 * Enter RCU idle mode right before resuming userspace. No use of RCU
852 * is permitted between this call and rcu_user_exit(). This way the
853 * CPU doesn't need to maintain the tick for RCU maintenance purposes
854 * when the CPU runs in userspace.
856 void rcu_user_enter(void)
858 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_user_enter() invoked with irqs enabled!!!");
861 #endif /* CONFIG_NO_HZ_FULL */
864 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
866 * Exit from an interrupt handler, which might possibly result in entering
867 * idle mode, in other words, leaving the mode in which read-side critical
868 * sections can occur. The caller must have disabled interrupts.
870 * This code assumes that the idle loop never does anything that might
871 * result in unbalanced calls to irq_enter() and irq_exit(). If your
872 * architecture violates this assumption, RCU will give you what you
873 * deserve, good and hard. But very infrequently and irreproducibly.
875 * Use things like work queues to work around this limitation.
877 * You have been warned.
879 void rcu_irq_exit(void)
881 struct rcu_dynticks *rdtp;
883 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
884 rdtp = this_cpu_ptr(&rcu_dynticks);
886 /* Page faults can happen in NMI handlers, so check... */
887 if (rdtp->dynticks_nmi_nesting)
890 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
891 rdtp->dynticks_nesting < 1);
892 if (rdtp->dynticks_nesting <= 1) {
893 rcu_eqs_enter_common(true);
895 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nesting, rdtp->dynticks_nesting - 1);
896 rdtp->dynticks_nesting--;
901 * Wrapper for rcu_irq_exit() where interrupts are enabled.
903 void rcu_irq_exit_irqson(void)
907 local_irq_save(flags);
909 local_irq_restore(flags);
913 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
915 * If the new value of the ->dynticks_nesting counter was previously zero,
916 * we really have exited idle, and must do the appropriate accounting.
917 * The caller must have disabled interrupts.
919 static void rcu_eqs_exit_common(long long oldval, int user)
921 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
923 rcu_dynticks_task_exit();
924 rcu_dynticks_eqs_exit();
925 rcu_cleanup_after_idle();
926 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
927 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
928 !user && !is_idle_task(current)) {
929 struct task_struct *idle __maybe_unused =
930 idle_task(smp_processor_id());
932 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
933 oldval, rdtp->dynticks_nesting);
934 rcu_ftrace_dump(DUMP_ORIG);
935 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
936 current->pid, current->comm,
937 idle->pid, idle->comm); /* must be idle task! */
942 * Exit an RCU extended quiescent state, which can be either the
943 * idle loop or adaptive-tickless usermode execution.
945 static void rcu_eqs_exit(bool user)
947 struct rcu_dynticks *rdtp;
950 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_eqs_exit() invoked with irqs enabled!!!");
951 rdtp = this_cpu_ptr(&rcu_dynticks);
952 oldval = rdtp->dynticks_nesting;
953 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
954 if (oldval & DYNTICK_TASK_NEST_MASK) {
955 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
957 __this_cpu_inc(disable_rcu_irq_enter);
958 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
959 rcu_eqs_exit_common(oldval, user);
960 __this_cpu_dec(disable_rcu_irq_enter);
965 * rcu_idle_exit - inform RCU that current CPU is leaving idle
967 * Exit idle mode, in other words, -enter- the mode in which RCU
968 * read-side critical sections can occur.
970 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
971 * allow for the possibility of usermode upcalls messing up our count
972 * of interrupt nesting level during the busy period that is just
975 void rcu_idle_exit(void)
979 local_irq_save(flags);
981 local_irq_restore(flags);
984 #ifdef CONFIG_NO_HZ_FULL
986 * rcu_user_exit - inform RCU that we are exiting userspace.
988 * Exit RCU idle mode while entering the kernel because it can
989 * run a RCU read side critical section anytime.
991 void rcu_user_exit(void)
995 #endif /* CONFIG_NO_HZ_FULL */
998 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1000 * Enter an interrupt handler, which might possibly result in exiting
1001 * idle mode, in other words, entering the mode in which read-side critical
1002 * sections can occur. The caller must have disabled interrupts.
1004 * Note that the Linux kernel is fully capable of entering an interrupt
1005 * handler that it never exits, for example when doing upcalls to
1006 * user mode! This code assumes that the idle loop never does upcalls to
1007 * user mode. If your architecture does do upcalls from the idle loop (or
1008 * does anything else that results in unbalanced calls to the irq_enter()
1009 * and irq_exit() functions), RCU will give you what you deserve, good
1010 * and hard. But very infrequently and irreproducibly.
1012 * Use things like work queues to work around this limitation.
1014 * You have been warned.
1016 void rcu_irq_enter(void)
1018 struct rcu_dynticks *rdtp;
1021 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
1022 rdtp = this_cpu_ptr(&rcu_dynticks);
1024 /* Page faults can happen in NMI handlers, so check... */
1025 if (rdtp->dynticks_nmi_nesting)
1028 oldval = rdtp->dynticks_nesting;
1029 rdtp->dynticks_nesting++;
1030 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
1031 rdtp->dynticks_nesting == 0);
1033 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
1035 rcu_eqs_exit_common(oldval, true);
1039 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1041 void rcu_irq_enter_irqson(void)
1043 unsigned long flags;
1045 local_irq_save(flags);
1047 local_irq_restore(flags);
1051 * rcu_nmi_enter - inform RCU of entry to NMI context
1053 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1054 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1055 * that the CPU is active. This implementation permits nested NMIs, as
1056 * long as the nesting level does not overflow an int. (You will probably
1057 * run out of stack space first.)
1059 void rcu_nmi_enter(void)
1061 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1064 /* Complain about underflow. */
1065 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
1068 * If idle from RCU viewpoint, atomically increment ->dynticks
1069 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1070 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1071 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1072 * to be in the outermost NMI handler that interrupted an RCU-idle
1073 * period (observation due to Andy Lutomirski).
1075 if (rcu_dynticks_curr_cpu_in_eqs()) {
1076 rcu_dynticks_eqs_exit();
1079 rdtp->dynticks_nmi_nesting += incby;
1084 * rcu_nmi_exit - inform RCU of exit from NMI context
1086 * If we are returning from the outermost NMI handler that interrupted an
1087 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
1088 * to let the RCU grace-period handling know that the CPU is back to
1091 void rcu_nmi_exit(void)
1093 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1096 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1097 * (We are exiting an NMI handler, so RCU better be paying attention
1100 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
1101 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
1104 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1105 * leave it in non-RCU-idle state.
1107 if (rdtp->dynticks_nmi_nesting != 1) {
1108 rdtp->dynticks_nmi_nesting -= 2;
1112 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1113 rdtp->dynticks_nmi_nesting = 0;
1114 rcu_dynticks_eqs_enter();
1118 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1120 * Return true if RCU is watching the running CPU, which means that this
1121 * CPU can safely enter RCU read-side critical sections. In other words,
1122 * if the current CPU is in its idle loop and is neither in an interrupt
1123 * or NMI handler, return true.
1125 bool notrace rcu_is_watching(void)
1129 preempt_disable_notrace();
1130 ret = !rcu_dynticks_curr_cpu_in_eqs();
1131 preempt_enable_notrace();
1134 EXPORT_SYMBOL_GPL(rcu_is_watching);
1137 * If a holdout task is actually running, request an urgent quiescent
1138 * state from its CPU. This is unsynchronized, so migrations can cause
1139 * the request to go to the wrong CPU. Which is OK, all that will happen
1140 * is that the CPU's next context switch will be a bit slower and next
1141 * time around this task will generate another request.
1143 void rcu_request_urgent_qs_task(struct task_struct *t)
1150 return; /* This task is not running on that CPU. */
1151 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1154 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1157 * Is the current CPU online? Disable preemption to avoid false positives
1158 * that could otherwise happen due to the current CPU number being sampled,
1159 * this task being preempted, its old CPU being taken offline, resuming
1160 * on some other CPU, then determining that its old CPU is now offline.
1161 * It is OK to use RCU on an offline processor during initial boot, hence
1162 * the check for rcu_scheduler_fully_active. Note also that it is OK
1163 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1164 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1165 * offline to continue to use RCU for one jiffy after marking itself
1166 * offline in the cpu_online_mask. This leniency is necessary given the
1167 * non-atomic nature of the online and offline processing, for example,
1168 * the fact that a CPU enters the scheduler after completing the teardown
1171 * This is also why RCU internally marks CPUs online during in the
1172 * preparation phase and offline after the CPU has been taken down.
1174 * Disable checking if in an NMI handler because we cannot safely report
1175 * errors from NMI handlers anyway.
1177 bool rcu_lockdep_current_cpu_online(void)
1179 struct rcu_data *rdp;
1180 struct rcu_node *rnp;
1186 rdp = this_cpu_ptr(&rcu_sched_data);
1188 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1189 !rcu_scheduler_fully_active;
1193 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1195 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1198 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1200 * If the current CPU is idle or running at a first-level (not nested)
1201 * interrupt from idle, return true. The caller must have at least
1202 * disabled preemption.
1204 static int rcu_is_cpu_rrupt_from_idle(void)
1206 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1210 * Snapshot the specified CPU's dynticks counter so that we can later
1211 * credit them with an implicit quiescent state. Return 1 if this CPU
1212 * is in dynticks idle mode, which is an extended quiescent state.
1214 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1216 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1217 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1218 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1219 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1220 rdp->mynode->gpnum))
1221 WRITE_ONCE(rdp->gpwrap, true);
1228 * Return true if the specified CPU has passed through a quiescent
1229 * state by virtue of being in or having passed through an dynticks
1230 * idle state since the last call to dyntick_save_progress_counter()
1231 * for this same CPU, or by virtue of having been offline.
1233 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1238 unsigned long rjtsc;
1239 struct rcu_node *rnp;
1242 * If the CPU passed through or entered a dynticks idle phase with
1243 * no active irq/NMI handlers, then we can safely pretend that the CPU
1244 * already acknowledged the request to pass through a quiescent
1245 * state. Either way, that CPU cannot possibly be in an RCU
1246 * read-side critical section that started before the beginning
1247 * of the current RCU grace period.
1249 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1250 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1251 rdp->dynticks_fqs++;
1255 /* Compute and saturate jiffies_till_sched_qs. */
1256 jtsq = jiffies_till_sched_qs;
1257 rjtsc = rcu_jiffies_till_stall_check();
1258 if (jtsq > rjtsc / 2) {
1259 WRITE_ONCE(jiffies_till_sched_qs, rjtsc);
1261 } else if (jtsq < 1) {
1262 WRITE_ONCE(jiffies_till_sched_qs, 1);
1267 * Has this CPU encountered a cond_resched_rcu_qs() since the
1268 * beginning of the grace period? For this to be the case,
1269 * the CPU has to have noticed the current grace period. This
1270 * might not be the case for nohz_full CPUs looping in the kernel.
1273 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1274 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1275 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1276 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1277 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1280 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1281 smp_store_release(ruqp, true);
1284 /* Check for the CPU being offline. */
1285 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1286 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1292 * A CPU running for an extended time within the kernel can
1293 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1294 * even context-switching back and forth between a pair of
1295 * in-kernel CPU-bound tasks cannot advance grace periods.
1296 * So if the grace period is old enough, make the CPU pay attention.
1297 * Note that the unsynchronized assignments to the per-CPU
1298 * rcu_need_heavy_qs variable are safe. Yes, setting of
1299 * bits can be lost, but they will be set again on the next
1300 * force-quiescent-state pass. So lost bit sets do not result
1301 * in incorrect behavior, merely in a grace period lasting
1302 * a few jiffies longer than it might otherwise. Because
1303 * there are at most four threads involved, and because the
1304 * updates are only once every few jiffies, the probability of
1305 * lossage (and thus of slight grace-period extension) is
1308 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1309 * is set too high, we override with half of the RCU CPU stall
1312 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1313 if (!READ_ONCE(*rnhqp) &&
1314 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1315 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1316 WRITE_ONCE(*rnhqp, true);
1317 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1318 smp_store_release(ruqp, true);
1319 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1323 * If more than halfway to RCU CPU stall-warning time, do
1324 * a resched_cpu() to try to loosen things up a bit.
1326 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2)
1327 resched_cpu(rdp->cpu);
1332 static void record_gp_stall_check_time(struct rcu_state *rsp)
1334 unsigned long j = jiffies;
1338 smp_wmb(); /* Record start time before stall time. */
1339 j1 = rcu_jiffies_till_stall_check();
1340 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1341 rsp->jiffies_resched = j + j1 / 2;
1342 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1346 * Convert a ->gp_state value to a character string.
1348 static const char *gp_state_getname(short gs)
1350 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1352 return gp_state_names[gs];
1356 * Complain about starvation of grace-period kthread.
1358 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1364 gpa = READ_ONCE(rsp->gp_activity);
1365 if (j - gpa > 2 * HZ) {
1366 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1368 rsp->gpnum, rsp->completed,
1370 gp_state_getname(rsp->gp_state), rsp->gp_state,
1371 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1372 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1373 if (rsp->gp_kthread) {
1374 sched_show_task(rsp->gp_kthread);
1375 wake_up_process(rsp->gp_kthread);
1381 * Dump stacks of all tasks running on stalled CPUs. First try using
1382 * NMIs, but fall back to manual remote stack tracing on architectures
1383 * that don't support NMI-based stack dumps. The NMI-triggered stack
1384 * traces are more accurate because they are printed by the target CPU.
1386 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1389 unsigned long flags;
1390 struct rcu_node *rnp;
1392 rcu_for_each_leaf_node(rsp, rnp) {
1393 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1394 for_each_leaf_node_possible_cpu(rnp, cpu)
1395 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1396 if (!trigger_single_cpu_backtrace(cpu))
1398 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1403 * If too much time has passed in the current grace period, and if
1404 * so configured, go kick the relevant kthreads.
1406 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1410 if (!rcu_kick_kthreads)
1412 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1413 if (time_after(jiffies, j) && rsp->gp_kthread &&
1414 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1415 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1416 rcu_ftrace_dump(DUMP_ALL);
1417 wake_up_process(rsp->gp_kthread);
1418 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1422 static inline void panic_on_rcu_stall(void)
1424 if (sysctl_panic_on_rcu_stall)
1425 panic("RCU Stall\n");
1428 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1432 unsigned long flags;
1436 struct rcu_node *rnp = rcu_get_root(rsp);
1439 /* Kick and suppress, if so configured. */
1440 rcu_stall_kick_kthreads(rsp);
1441 if (rcu_cpu_stall_suppress)
1444 /* Only let one CPU complain about others per time interval. */
1446 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1447 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1448 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1449 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1452 WRITE_ONCE(rsp->jiffies_stall,
1453 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1454 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1457 * OK, time to rat on our buddy...
1458 * See Documentation/RCU/stallwarn.txt for info on how to debug
1459 * RCU CPU stall warnings.
1461 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1463 print_cpu_stall_info_begin();
1464 rcu_for_each_leaf_node(rsp, rnp) {
1465 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1466 ndetected += rcu_print_task_stall(rnp);
1467 if (rnp->qsmask != 0) {
1468 for_each_leaf_node_possible_cpu(rnp, cpu)
1469 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1470 print_cpu_stall_info(rsp, cpu);
1474 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1477 print_cpu_stall_info_end();
1478 for_each_possible_cpu(cpu)
1479 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1481 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1482 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1483 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1485 rcu_dump_cpu_stacks(rsp);
1487 /* Complain about tasks blocking the grace period. */
1488 rcu_print_detail_task_stall(rsp);
1490 if (READ_ONCE(rsp->gpnum) != gpnum ||
1491 READ_ONCE(rsp->completed) == gpnum) {
1492 pr_err("INFO: Stall ended before state dump start\n");
1495 gpa = READ_ONCE(rsp->gp_activity);
1496 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1497 rsp->name, j - gpa, j, gpa,
1498 jiffies_till_next_fqs,
1499 rcu_get_root(rsp)->qsmask);
1500 /* In this case, the current CPU might be at fault. */
1501 sched_show_task(current);
1505 rcu_check_gp_kthread_starvation(rsp);
1507 panic_on_rcu_stall();
1509 force_quiescent_state(rsp); /* Kick them all. */
1512 static void print_cpu_stall(struct rcu_state *rsp)
1515 unsigned long flags;
1516 struct rcu_node *rnp = rcu_get_root(rsp);
1519 /* Kick and suppress, if so configured. */
1520 rcu_stall_kick_kthreads(rsp);
1521 if (rcu_cpu_stall_suppress)
1525 * OK, time to rat on ourselves...
1526 * See Documentation/RCU/stallwarn.txt for info on how to debug
1527 * RCU CPU stall warnings.
1529 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1530 print_cpu_stall_info_begin();
1531 print_cpu_stall_info(rsp, smp_processor_id());
1532 print_cpu_stall_info_end();
1533 for_each_possible_cpu(cpu)
1534 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1536 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1537 jiffies - rsp->gp_start,
1538 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1540 rcu_check_gp_kthread_starvation(rsp);
1542 rcu_dump_cpu_stacks(rsp);
1544 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1545 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1546 WRITE_ONCE(rsp->jiffies_stall,
1547 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1548 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1550 panic_on_rcu_stall();
1553 * Attempt to revive the RCU machinery by forcing a context switch.
1555 * A context switch would normally allow the RCU state machine to make
1556 * progress and it could be we're stuck in kernel space without context
1557 * switches for an entirely unreasonable amount of time.
1559 resched_cpu(smp_processor_id());
1562 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1564 unsigned long completed;
1565 unsigned long gpnum;
1569 struct rcu_node *rnp;
1571 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1572 !rcu_gp_in_progress(rsp))
1574 rcu_stall_kick_kthreads(rsp);
1578 * Lots of memory barriers to reject false positives.
1580 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1581 * then rsp->gp_start, and finally rsp->completed. These values
1582 * are updated in the opposite order with memory barriers (or
1583 * equivalent) during grace-period initialization and cleanup.
1584 * Now, a false positive can occur if we get an new value of
1585 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1586 * the memory barriers, the only way that this can happen is if one
1587 * grace period ends and another starts between these two fetches.
1588 * Detect this by comparing rsp->completed with the previous fetch
1591 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1592 * and rsp->gp_start suffice to forestall false positives.
1594 gpnum = READ_ONCE(rsp->gpnum);
1595 smp_rmb(); /* Pick up ->gpnum first... */
1596 js = READ_ONCE(rsp->jiffies_stall);
1597 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1598 gps = READ_ONCE(rsp->gp_start);
1599 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1600 completed = READ_ONCE(rsp->completed);
1601 if (ULONG_CMP_GE(completed, gpnum) ||
1602 ULONG_CMP_LT(j, js) ||
1603 ULONG_CMP_GE(gps, js))
1604 return; /* No stall or GP completed since entering function. */
1606 if (rcu_gp_in_progress(rsp) &&
1607 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1609 /* We haven't checked in, so go dump stack. */
1610 print_cpu_stall(rsp);
1612 } else if (rcu_gp_in_progress(rsp) &&
1613 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1615 /* They had a few time units to dump stack, so complain. */
1616 print_other_cpu_stall(rsp, gpnum);
1621 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1623 * Set the stall-warning timeout way off into the future, thus preventing
1624 * any RCU CPU stall-warning messages from appearing in the current set of
1625 * RCU grace periods.
1627 * The caller must disable hard irqs.
1629 void rcu_cpu_stall_reset(void)
1631 struct rcu_state *rsp;
1633 for_each_rcu_flavor(rsp)
1634 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1638 * Determine the value that ->completed will have at the end of the
1639 * next subsequent grace period. This is used to tag callbacks so that
1640 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1641 * been dyntick-idle for an extended period with callbacks under the
1642 * influence of RCU_FAST_NO_HZ.
1644 * The caller must hold rnp->lock with interrupts disabled.
1646 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1647 struct rcu_node *rnp)
1649 lockdep_assert_held(&rnp->lock);
1652 * If RCU is idle, we just wait for the next grace period.
1653 * But we can only be sure that RCU is idle if we are looking
1654 * at the root rcu_node structure -- otherwise, a new grace
1655 * period might have started, but just not yet gotten around
1656 * to initializing the current non-root rcu_node structure.
1658 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1659 return rnp->completed + 1;
1662 * Otherwise, wait for a possible partial grace period and
1663 * then the subsequent full grace period.
1665 return rnp->completed + 2;
1669 * Trace-event helper function for rcu_start_future_gp() and
1670 * rcu_nocb_wait_gp().
1672 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1673 unsigned long c, const char *s)
1675 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1676 rnp->completed, c, rnp->level,
1677 rnp->grplo, rnp->grphi, s);
1681 * Start some future grace period, as needed to handle newly arrived
1682 * callbacks. The required future grace periods are recorded in each
1683 * rcu_node structure's ->need_future_gp field. Returns true if there
1684 * is reason to awaken the grace-period kthread.
1686 * The caller must hold the specified rcu_node structure's ->lock.
1688 static bool __maybe_unused
1689 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1690 unsigned long *c_out)
1694 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1696 lockdep_assert_held(&rnp->lock);
1699 * Pick up grace-period number for new callbacks. If this
1700 * grace period is already marked as needed, return to the caller.
1702 c = rcu_cbs_completed(rdp->rsp, rnp);
1703 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1704 if (rnp->need_future_gp[c & 0x1]) {
1705 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1710 * If either this rcu_node structure or the root rcu_node structure
1711 * believe that a grace period is in progress, then we must wait
1712 * for the one following, which is in "c". Because our request
1713 * will be noticed at the end of the current grace period, we don't
1714 * need to explicitly start one. We only do the lockless check
1715 * of rnp_root's fields if the current rcu_node structure thinks
1716 * there is no grace period in flight, and because we hold rnp->lock,
1717 * the only possible change is when rnp_root's two fields are
1718 * equal, in which case rnp_root->gpnum might be concurrently
1719 * incremented. But that is OK, as it will just result in our
1720 * doing some extra useless work.
1722 if (rnp->gpnum != rnp->completed ||
1723 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1724 rnp->need_future_gp[c & 0x1]++;
1725 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1730 * There might be no grace period in progress. If we don't already
1731 * hold it, acquire the root rcu_node structure's lock in order to
1732 * start one (if needed).
1734 if (rnp != rnp_root)
1735 raw_spin_lock_rcu_node(rnp_root);
1738 * Get a new grace-period number. If there really is no grace
1739 * period in progress, it will be smaller than the one we obtained
1740 * earlier. Adjust callbacks as needed.
1742 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1743 if (!rcu_is_nocb_cpu(rdp->cpu))
1744 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1747 * If the needed for the required grace period is already
1748 * recorded, trace and leave.
1750 if (rnp_root->need_future_gp[c & 0x1]) {
1751 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1755 /* Record the need for the future grace period. */
1756 rnp_root->need_future_gp[c & 0x1]++;
1758 /* If a grace period is not already in progress, start one. */
1759 if (rnp_root->gpnum != rnp_root->completed) {
1760 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1762 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1763 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1766 if (rnp != rnp_root)
1767 raw_spin_unlock_rcu_node(rnp_root);
1775 * Clean up any old requests for the just-ended grace period. Also return
1776 * whether any additional grace periods have been requested.
1778 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1780 int c = rnp->completed;
1782 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1784 rnp->need_future_gp[c & 0x1] = 0;
1785 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1786 trace_rcu_future_gp(rnp, rdp, c,
1787 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1792 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1793 * an interrupt or softirq handler), and don't bother awakening when there
1794 * is nothing for the grace-period kthread to do (as in several CPUs raced
1795 * to awaken, and we lost), and finally don't try to awaken a kthread that
1796 * has not yet been created. If all those checks are passed, track some
1797 * debug information and awaken.
1799 * So why do the self-wakeup when in an interrupt or softirq handler
1800 * in the grace-period kthread's context? Because the kthread might have
1801 * been interrupted just as it was going to sleep, and just after the final
1802 * pre-sleep check of the awaken condition. In this case, a wakeup really
1803 * is required, and is therefore supplied.
1805 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1807 if ((current == rsp->gp_kthread &&
1808 !in_interrupt() && !in_serving_softirq()) ||
1809 !READ_ONCE(rsp->gp_flags) ||
1812 swake_up(&rsp->gp_wq);
1816 * If there is room, assign a ->completed number to any callbacks on
1817 * this CPU that have not already been assigned. Also accelerate any
1818 * callbacks that were previously assigned a ->completed number that has
1819 * since proven to be too conservative, which can happen if callbacks get
1820 * assigned a ->completed number while RCU is idle, but with reference to
1821 * a non-root rcu_node structure. This function is idempotent, so it does
1822 * not hurt to call it repeatedly. Returns an flag saying that we should
1823 * awaken the RCU grace-period kthread.
1825 * The caller must hold rnp->lock with interrupts disabled.
1827 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1828 struct rcu_data *rdp)
1832 lockdep_assert_held(&rnp->lock);
1834 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1835 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1839 * Callbacks are often registered with incomplete grace-period
1840 * information. Something about the fact that getting exact
1841 * information requires acquiring a global lock... RCU therefore
1842 * makes a conservative estimate of the grace period number at which
1843 * a given callback will become ready to invoke. The following
1844 * code checks this estimate and improves it when possible, thus
1845 * accelerating callback invocation to an earlier grace-period
1848 if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
1849 ret = rcu_start_future_gp(rnp, rdp, NULL);
1851 /* Trace depending on how much we were able to accelerate. */
1852 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1853 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1855 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1860 * Move any callbacks whose grace period has completed to the
1861 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1862 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1863 * sublist. This function is idempotent, so it does not hurt to
1864 * invoke it repeatedly. As long as it is not invoked -too- often...
1865 * Returns true if the RCU grace-period kthread needs to be awakened.
1867 * The caller must hold rnp->lock with interrupts disabled.
1869 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1870 struct rcu_data *rdp)
1872 lockdep_assert_held(&rnp->lock);
1874 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1875 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1879 * Find all callbacks whose ->completed numbers indicate that they
1880 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1882 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1884 /* Classify any remaining callbacks. */
1885 return rcu_accelerate_cbs(rsp, rnp, rdp);
1889 * Update CPU-local rcu_data state to record the beginnings and ends of
1890 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1891 * structure corresponding to the current CPU, and must have irqs disabled.
1892 * Returns true if the grace-period kthread needs to be awakened.
1894 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1895 struct rcu_data *rdp)
1900 lockdep_assert_held(&rnp->lock);
1902 /* Handle the ends of any preceding grace periods first. */
1903 if (rdp->completed == rnp->completed &&
1904 !unlikely(READ_ONCE(rdp->gpwrap))) {
1906 /* No grace period end, so just accelerate recent callbacks. */
1907 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1911 /* Advance callbacks. */
1912 ret = rcu_advance_cbs(rsp, rnp, rdp);
1914 /* Remember that we saw this grace-period completion. */
1915 rdp->completed = rnp->completed;
1916 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1919 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1921 * If the current grace period is waiting for this CPU,
1922 * set up to detect a quiescent state, otherwise don't
1923 * go looking for one.
1925 rdp->gpnum = rnp->gpnum;
1926 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1927 need_gp = !!(rnp->qsmask & rdp->grpmask);
1928 rdp->cpu_no_qs.b.norm = need_gp;
1929 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1930 rdp->core_needs_qs = need_gp;
1931 zero_cpu_stall_ticks(rdp);
1932 WRITE_ONCE(rdp->gpwrap, false);
1937 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1939 unsigned long flags;
1941 struct rcu_node *rnp;
1943 local_irq_save(flags);
1945 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1946 rdp->completed == READ_ONCE(rnp->completed) &&
1947 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1948 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1949 local_irq_restore(flags);
1952 needwake = __note_gp_changes(rsp, rnp, rdp);
1953 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1955 rcu_gp_kthread_wake(rsp);
1958 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1961 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1962 schedule_timeout_uninterruptible(delay);
1966 * Initialize a new grace period. Return false if no grace period required.
1968 static bool rcu_gp_init(struct rcu_state *rsp)
1970 unsigned long oldmask;
1971 struct rcu_data *rdp;
1972 struct rcu_node *rnp = rcu_get_root(rsp);
1974 WRITE_ONCE(rsp->gp_activity, jiffies);
1975 raw_spin_lock_irq_rcu_node(rnp);
1976 if (!READ_ONCE(rsp->gp_flags)) {
1977 /* Spurious wakeup, tell caller to go back to sleep. */
1978 raw_spin_unlock_irq_rcu_node(rnp);
1981 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1983 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1985 * Grace period already in progress, don't start another.
1986 * Not supposed to be able to happen.
1988 raw_spin_unlock_irq_rcu_node(rnp);
1992 /* Advance to a new grace period and initialize state. */
1993 record_gp_stall_check_time(rsp);
1994 /* Record GP times before starting GP, hence smp_store_release(). */
1995 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1996 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1997 raw_spin_unlock_irq_rcu_node(rnp);
2000 * Apply per-leaf buffered online and offline operations to the
2001 * rcu_node tree. Note that this new grace period need not wait
2002 * for subsequent online CPUs, and that quiescent-state forcing
2003 * will handle subsequent offline CPUs.
2005 rcu_for_each_leaf_node(rsp, rnp) {
2006 rcu_gp_slow(rsp, gp_preinit_delay);
2007 raw_spin_lock_irq_rcu_node(rnp);
2008 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2009 !rnp->wait_blkd_tasks) {
2010 /* Nothing to do on this leaf rcu_node structure. */
2011 raw_spin_unlock_irq_rcu_node(rnp);
2015 /* Record old state, apply changes to ->qsmaskinit field. */
2016 oldmask = rnp->qsmaskinit;
2017 rnp->qsmaskinit = rnp->qsmaskinitnext;
2019 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2020 if (!oldmask != !rnp->qsmaskinit) {
2021 if (!oldmask) /* First online CPU for this rcu_node. */
2022 rcu_init_new_rnp(rnp);
2023 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2024 rnp->wait_blkd_tasks = true;
2025 else /* Last offline CPU and can propagate. */
2026 rcu_cleanup_dead_rnp(rnp);
2030 * If all waited-on tasks from prior grace period are
2031 * done, and if all this rcu_node structure's CPUs are
2032 * still offline, propagate up the rcu_node tree and
2033 * clear ->wait_blkd_tasks. Otherwise, if one of this
2034 * rcu_node structure's CPUs has since come back online,
2035 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2036 * checks for this, so just call it unconditionally).
2038 if (rnp->wait_blkd_tasks &&
2039 (!rcu_preempt_has_tasks(rnp) ||
2041 rnp->wait_blkd_tasks = false;
2042 rcu_cleanup_dead_rnp(rnp);
2045 raw_spin_unlock_irq_rcu_node(rnp);
2049 * Set the quiescent-state-needed bits in all the rcu_node
2050 * structures for all currently online CPUs in breadth-first order,
2051 * starting from the root rcu_node structure, relying on the layout
2052 * of the tree within the rsp->node[] array. Note that other CPUs
2053 * will access only the leaves of the hierarchy, thus seeing that no
2054 * grace period is in progress, at least until the corresponding
2055 * leaf node has been initialized.
2057 * The grace period cannot complete until the initialization
2058 * process finishes, because this kthread handles both.
2060 rcu_for_each_node_breadth_first(rsp, rnp) {
2061 rcu_gp_slow(rsp, gp_init_delay);
2062 raw_spin_lock_irq_rcu_node(rnp);
2063 rdp = this_cpu_ptr(rsp->rda);
2064 rcu_preempt_check_blocked_tasks(rnp);
2065 rnp->qsmask = rnp->qsmaskinit;
2066 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2067 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2068 WRITE_ONCE(rnp->completed, rsp->completed);
2069 if (rnp == rdp->mynode)
2070 (void)__note_gp_changes(rsp, rnp, rdp);
2071 rcu_preempt_boost_start_gp(rnp);
2072 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2073 rnp->level, rnp->grplo,
2074 rnp->grphi, rnp->qsmask);
2075 raw_spin_unlock_irq_rcu_node(rnp);
2076 cond_resched_rcu_qs();
2077 WRITE_ONCE(rsp->gp_activity, jiffies);
2084 * Helper function for swait_event_idle() wakeup at force-quiescent-state
2087 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2089 struct rcu_node *rnp = rcu_get_root(rsp);
2091 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2092 *gfp = READ_ONCE(rsp->gp_flags);
2093 if (*gfp & RCU_GP_FLAG_FQS)
2096 /* The current grace period has completed. */
2097 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2104 * Do one round of quiescent-state forcing.
2106 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2108 struct rcu_node *rnp = rcu_get_root(rsp);
2110 WRITE_ONCE(rsp->gp_activity, jiffies);
2113 /* Collect dyntick-idle snapshots. */
2114 force_qs_rnp(rsp, dyntick_save_progress_counter);
2116 /* Handle dyntick-idle and offline CPUs. */
2117 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2119 /* Clear flag to prevent immediate re-entry. */
2120 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2121 raw_spin_lock_irq_rcu_node(rnp);
2122 WRITE_ONCE(rsp->gp_flags,
2123 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2124 raw_spin_unlock_irq_rcu_node(rnp);
2129 * Clean up after the old grace period.
2131 static void rcu_gp_cleanup(struct rcu_state *rsp)
2133 unsigned long gp_duration;
2134 bool needgp = false;
2136 struct rcu_data *rdp;
2137 struct rcu_node *rnp = rcu_get_root(rsp);
2138 struct swait_queue_head *sq;
2140 WRITE_ONCE(rsp->gp_activity, jiffies);
2141 raw_spin_lock_irq_rcu_node(rnp);
2142 gp_duration = jiffies - rsp->gp_start;
2143 if (gp_duration > rsp->gp_max)
2144 rsp->gp_max = gp_duration;
2147 * We know the grace period is complete, but to everyone else
2148 * it appears to still be ongoing. But it is also the case
2149 * that to everyone else it looks like there is nothing that
2150 * they can do to advance the grace period. It is therefore
2151 * safe for us to drop the lock in order to mark the grace
2152 * period as completed in all of the rcu_node structures.
2154 raw_spin_unlock_irq_rcu_node(rnp);
2157 * Propagate new ->completed value to rcu_node structures so
2158 * that other CPUs don't have to wait until the start of the next
2159 * grace period to process their callbacks. This also avoids
2160 * some nasty RCU grace-period initialization races by forcing
2161 * the end of the current grace period to be completely recorded in
2162 * all of the rcu_node structures before the beginning of the next
2163 * grace period is recorded in any of the rcu_node structures.
2165 rcu_for_each_node_breadth_first(rsp, rnp) {
2166 raw_spin_lock_irq_rcu_node(rnp);
2167 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2168 WARN_ON_ONCE(rnp->qsmask);
2169 WRITE_ONCE(rnp->completed, rsp->gpnum);
2170 rdp = this_cpu_ptr(rsp->rda);
2171 if (rnp == rdp->mynode)
2172 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2173 /* smp_mb() provided by prior unlock-lock pair. */
2174 nocb += rcu_future_gp_cleanup(rsp, rnp);
2175 sq = rcu_nocb_gp_get(rnp);
2176 raw_spin_unlock_irq_rcu_node(rnp);
2177 rcu_nocb_gp_cleanup(sq);
2178 cond_resched_rcu_qs();
2179 WRITE_ONCE(rsp->gp_activity, jiffies);
2180 rcu_gp_slow(rsp, gp_cleanup_delay);
2182 rnp = rcu_get_root(rsp);
2183 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2184 rcu_nocb_gp_set(rnp, nocb);
2186 /* Declare grace period done. */
2187 WRITE_ONCE(rsp->completed, rsp->gpnum);
2188 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2189 rsp->gp_state = RCU_GP_IDLE;
2190 rdp = this_cpu_ptr(rsp->rda);
2191 /* Advance CBs to reduce false positives below. */
2192 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2193 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2194 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2195 trace_rcu_grace_period(rsp->name,
2196 READ_ONCE(rsp->gpnum),
2199 raw_spin_unlock_irq_rcu_node(rnp);
2203 * Body of kthread that handles grace periods.
2205 static int __noreturn rcu_gp_kthread(void *arg)
2211 struct rcu_state *rsp = arg;
2212 struct rcu_node *rnp = rcu_get_root(rsp);
2214 rcu_bind_gp_kthread();
2217 /* Handle grace-period start. */
2219 trace_rcu_grace_period(rsp->name,
2220 READ_ONCE(rsp->gpnum),
2222 rsp->gp_state = RCU_GP_WAIT_GPS;
2223 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2225 rsp->gp_state = RCU_GP_DONE_GPS;
2226 /* Locking provides needed memory barrier. */
2227 if (rcu_gp_init(rsp))
2229 cond_resched_rcu_qs();
2230 WRITE_ONCE(rsp->gp_activity, jiffies);
2231 WARN_ON(signal_pending(current));
2232 trace_rcu_grace_period(rsp->name,
2233 READ_ONCE(rsp->gpnum),
2237 /* Handle quiescent-state forcing. */
2238 first_gp_fqs = true;
2239 j = jiffies_till_first_fqs;
2242 jiffies_till_first_fqs = HZ;
2247 rsp->jiffies_force_qs = jiffies + j;
2248 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2251 trace_rcu_grace_period(rsp->name,
2252 READ_ONCE(rsp->gpnum),
2254 rsp->gp_state = RCU_GP_WAIT_FQS;
2255 ret = swait_event_idle_timeout(rsp->gp_wq,
2256 rcu_gp_fqs_check_wake(rsp, &gf), j);
2257 rsp->gp_state = RCU_GP_DOING_FQS;
2258 /* Locking provides needed memory barriers. */
2259 /* If grace period done, leave loop. */
2260 if (!READ_ONCE(rnp->qsmask) &&
2261 !rcu_preempt_blocked_readers_cgp(rnp))
2263 /* If time for quiescent-state forcing, do it. */
2264 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2265 (gf & RCU_GP_FLAG_FQS)) {
2266 trace_rcu_grace_period(rsp->name,
2267 READ_ONCE(rsp->gpnum),
2269 rcu_gp_fqs(rsp, first_gp_fqs);
2270 first_gp_fqs = false;
2271 trace_rcu_grace_period(rsp->name,
2272 READ_ONCE(rsp->gpnum),
2274 cond_resched_rcu_qs();
2275 WRITE_ONCE(rsp->gp_activity, jiffies);
2276 ret = 0; /* Force full wait till next FQS. */
2277 j = jiffies_till_next_fqs;
2280 jiffies_till_next_fqs = HZ;
2283 jiffies_till_next_fqs = 1;
2286 /* Deal with stray signal. */
2287 cond_resched_rcu_qs();
2288 WRITE_ONCE(rsp->gp_activity, jiffies);
2289 WARN_ON(signal_pending(current));
2290 trace_rcu_grace_period(rsp->name,
2291 READ_ONCE(rsp->gpnum),
2293 ret = 1; /* Keep old FQS timing. */
2295 if (time_after(jiffies, rsp->jiffies_force_qs))
2298 j = rsp->jiffies_force_qs - j;
2302 /* Handle grace-period end. */
2303 rsp->gp_state = RCU_GP_CLEANUP;
2304 rcu_gp_cleanup(rsp);
2305 rsp->gp_state = RCU_GP_CLEANED;
2310 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2311 * in preparation for detecting the next grace period. The caller must hold
2312 * the root node's ->lock and hard irqs must be disabled.
2314 * Note that it is legal for a dying CPU (which is marked as offline) to
2315 * invoke this function. This can happen when the dying CPU reports its
2318 * Returns true if the grace-period kthread must be awakened.
2321 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2322 struct rcu_data *rdp)
2324 lockdep_assert_held(&rnp->lock);
2325 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2327 * Either we have not yet spawned the grace-period
2328 * task, this CPU does not need another grace period,
2329 * or a grace period is already in progress.
2330 * Either way, don't start a new grace period.
2334 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2335 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2339 * We can't do wakeups while holding the rnp->lock, as that
2340 * could cause possible deadlocks with the rq->lock. Defer
2341 * the wakeup to our caller.
2347 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2348 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2349 * is invoked indirectly from rcu_advance_cbs(), which would result in
2350 * endless recursion -- or would do so if it wasn't for the self-deadlock
2351 * that is encountered beforehand.
2353 * Returns true if the grace-period kthread needs to be awakened.
2355 static bool rcu_start_gp(struct rcu_state *rsp)
2357 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2358 struct rcu_node *rnp = rcu_get_root(rsp);
2362 * If there is no grace period in progress right now, any
2363 * callbacks we have up to this point will be satisfied by the
2364 * next grace period. Also, advancing the callbacks reduces the
2365 * probability of false positives from cpu_needs_another_gp()
2366 * resulting in pointless grace periods. So, advance callbacks
2367 * then start the grace period!
2369 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2370 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2375 * Report a full set of quiescent states to the specified rcu_state data
2376 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2377 * kthread if another grace period is required. Whether we wake
2378 * the grace-period kthread or it awakens itself for the next round
2379 * of quiescent-state forcing, that kthread will clean up after the
2380 * just-completed grace period. Note that the caller must hold rnp->lock,
2381 * which is released before return.
2383 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2384 __releases(rcu_get_root(rsp)->lock)
2386 lockdep_assert_held(&rcu_get_root(rsp)->lock);
2387 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2388 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2389 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2390 rcu_gp_kthread_wake(rsp);
2394 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2395 * Allows quiescent states for a group of CPUs to be reported at one go
2396 * to the specified rcu_node structure, though all the CPUs in the group
2397 * must be represented by the same rcu_node structure (which need not be a
2398 * leaf rcu_node structure, though it often will be). The gps parameter
2399 * is the grace-period snapshot, which means that the quiescent states
2400 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2401 * must be held upon entry, and it is released before return.
2404 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2405 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2406 __releases(rnp->lock)
2408 unsigned long oldmask = 0;
2409 struct rcu_node *rnp_c;
2411 lockdep_assert_held(&rnp->lock);
2413 /* Walk up the rcu_node hierarchy. */
2415 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2418 * Our bit has already been cleared, or the
2419 * relevant grace period is already over, so done.
2421 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2424 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2425 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1 &&
2426 rcu_preempt_blocked_readers_cgp(rnp));
2427 rnp->qsmask &= ~mask;
2428 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2429 mask, rnp->qsmask, rnp->level,
2430 rnp->grplo, rnp->grphi,
2432 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2434 /* Other bits still set at this level, so done. */
2435 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2438 mask = rnp->grpmask;
2439 if (rnp->parent == NULL) {
2441 /* No more levels. Exit loop holding root lock. */
2445 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2448 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2449 oldmask = rnp_c->qsmask;
2453 * Get here if we are the last CPU to pass through a quiescent
2454 * state for this grace period. Invoke rcu_report_qs_rsp()
2455 * to clean up and start the next grace period if one is needed.
2457 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2461 * Record a quiescent state for all tasks that were previously queued
2462 * on the specified rcu_node structure and that were blocking the current
2463 * RCU grace period. The caller must hold the specified rnp->lock with
2464 * irqs disabled, and this lock is released upon return, but irqs remain
2467 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2468 struct rcu_node *rnp, unsigned long flags)
2469 __releases(rnp->lock)
2473 struct rcu_node *rnp_p;
2475 lockdep_assert_held(&rnp->lock);
2476 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2477 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2478 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2479 return; /* Still need more quiescent states! */
2482 rnp_p = rnp->parent;
2483 if (rnp_p == NULL) {
2485 * Only one rcu_node structure in the tree, so don't
2486 * try to report up to its nonexistent parent!
2488 rcu_report_qs_rsp(rsp, flags);
2492 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2494 mask = rnp->grpmask;
2495 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2496 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2497 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2501 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2502 * structure. This must be called from the specified CPU.
2505 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2507 unsigned long flags;
2510 struct rcu_node *rnp;
2513 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2514 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2515 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2518 * The grace period in which this quiescent state was
2519 * recorded has ended, so don't report it upwards.
2520 * We will instead need a new quiescent state that lies
2521 * within the current grace period.
2523 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2524 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2525 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2528 mask = rdp->grpmask;
2529 if ((rnp->qsmask & mask) == 0) {
2530 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2532 rdp->core_needs_qs = false;
2535 * This GP can't end until cpu checks in, so all of our
2536 * callbacks can be processed during the next GP.
2538 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2540 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2541 /* ^^^ Released rnp->lock */
2543 rcu_gp_kthread_wake(rsp);
2548 * Check to see if there is a new grace period of which this CPU
2549 * is not yet aware, and if so, set up local rcu_data state for it.
2550 * Otherwise, see if this CPU has just passed through its first
2551 * quiescent state for this grace period, and record that fact if so.
2554 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2556 /* Check for grace-period ends and beginnings. */
2557 note_gp_changes(rsp, rdp);
2560 * Does this CPU still need to do its part for current grace period?
2561 * If no, return and let the other CPUs do their part as well.
2563 if (!rdp->core_needs_qs)
2567 * Was there a quiescent state since the beginning of the grace
2568 * period? If no, then exit and wait for the next call.
2570 if (rdp->cpu_no_qs.b.norm)
2574 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2577 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2581 * Trace the fact that this CPU is going offline.
2583 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2585 RCU_TRACE(unsigned long mask;)
2586 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2587 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2589 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2592 RCU_TRACE(mask = rdp->grpmask;)
2593 trace_rcu_grace_period(rsp->name,
2594 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2599 * All CPUs for the specified rcu_node structure have gone offline,
2600 * and all tasks that were preempted within an RCU read-side critical
2601 * section while running on one of those CPUs have since exited their RCU
2602 * read-side critical section. Some other CPU is reporting this fact with
2603 * the specified rcu_node structure's ->lock held and interrupts disabled.
2604 * This function therefore goes up the tree of rcu_node structures,
2605 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2606 * the leaf rcu_node structure's ->qsmaskinit field has already been
2609 * This function does check that the specified rcu_node structure has
2610 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2611 * prematurely. That said, invoking it after the fact will cost you
2612 * a needless lock acquisition. So once it has done its work, don't
2615 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2618 struct rcu_node *rnp = rnp_leaf;
2620 lockdep_assert_held(&rnp->lock);
2621 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2622 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2625 mask = rnp->grpmask;
2629 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2630 rnp->qsmaskinit &= ~mask;
2631 rnp->qsmask &= ~mask;
2632 if (rnp->qsmaskinit) {
2633 raw_spin_unlock_rcu_node(rnp);
2634 /* irqs remain disabled. */
2637 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2642 * The CPU has been completely removed, and some other CPU is reporting
2643 * this fact from process context. Do the remainder of the cleanup.
2644 * There can only be one CPU hotplug operation at a time, so no need for
2647 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2649 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2650 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2652 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2655 /* Adjust any no-longer-needed kthreads. */
2656 rcu_boost_kthread_setaffinity(rnp, -1);
2660 * Invoke any RCU callbacks that have made it to the end of their grace
2661 * period. Thottle as specified by rdp->blimit.
2663 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2665 unsigned long flags;
2666 struct rcu_head *rhp;
2667 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2670 /* If no callbacks are ready, just return. */
2671 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2672 trace_rcu_batch_start(rsp->name,
2673 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2674 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2675 trace_rcu_batch_end(rsp->name, 0,
2676 !rcu_segcblist_empty(&rdp->cblist),
2677 need_resched(), is_idle_task(current),
2678 rcu_is_callbacks_kthread());
2683 * Extract the list of ready callbacks, disabling to prevent
2684 * races with call_rcu() from interrupt handlers. Leave the
2685 * callback counts, as rcu_barrier() needs to be conservative.
2687 local_irq_save(flags);
2688 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2690 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2691 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2692 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2693 local_irq_restore(flags);
2695 /* Invoke callbacks. */
2696 rhp = rcu_cblist_dequeue(&rcl);
2697 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2698 debug_rcu_head_unqueue(rhp);
2699 if (__rcu_reclaim(rsp->name, rhp))
2700 rcu_cblist_dequeued_lazy(&rcl);
2702 * Stop only if limit reached and CPU has something to do.
2703 * Note: The rcl structure counts down from zero.
2705 if (-rcl.len >= bl &&
2707 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2711 local_irq_save(flags);
2713 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2714 is_idle_task(current), rcu_is_callbacks_kthread());
2716 /* Update counts and requeue any remaining callbacks. */
2717 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2718 smp_mb(); /* List handling before counting for rcu_barrier(). */
2719 rdp->n_cbs_invoked += count;
2720 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2722 /* Reinstate batch limit if we have worked down the excess. */
2723 count = rcu_segcblist_n_cbs(&rdp->cblist);
2724 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2725 rdp->blimit = blimit;
2727 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2728 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2729 rdp->qlen_last_fqs_check = 0;
2730 rdp->n_force_qs_snap = rsp->n_force_qs;
2731 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2732 rdp->qlen_last_fqs_check = count;
2733 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2735 local_irq_restore(flags);
2737 /* Re-invoke RCU core processing if there are callbacks remaining. */
2738 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2743 * Check to see if this CPU is in a non-context-switch quiescent state
2744 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2745 * Also schedule RCU core processing.
2747 * This function must be called from hardirq context. It is normally
2748 * invoked from the scheduling-clock interrupt.
2750 void rcu_check_callbacks(int user)
2752 trace_rcu_utilization(TPS("Start scheduler-tick"));
2753 increment_cpu_stall_ticks();
2754 if (user || rcu_is_cpu_rrupt_from_idle()) {
2757 * Get here if this CPU took its interrupt from user
2758 * mode or from the idle loop, and if this is not a
2759 * nested interrupt. In this case, the CPU is in
2760 * a quiescent state, so note it.
2762 * No memory barrier is required here because both
2763 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2764 * variables that other CPUs neither access nor modify,
2765 * at least not while the corresponding CPU is online.
2771 } else if (!in_softirq()) {
2774 * Get here if this CPU did not take its interrupt from
2775 * softirq, in other words, if it is not interrupting
2776 * a rcu_bh read-side critical section. This is an _bh
2777 * critical section, so note it.
2782 rcu_preempt_check_callbacks();
2783 /* The load-acquire pairs with the store-release setting to true. */
2784 if (smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
2785 /* Idle and userspace execution already are quiescent states. */
2786 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2787 set_tsk_need_resched(current);
2788 set_preempt_need_resched();
2790 __this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
2795 rcu_note_voluntary_context_switch(current);
2796 trace_rcu_utilization(TPS("End scheduler-tick"));
2800 * Scan the leaf rcu_node structures, processing dyntick state for any that
2801 * have not yet encountered a quiescent state, using the function specified.
2802 * Also initiate boosting for any threads blocked on the root rcu_node.
2804 * The caller must have suppressed start of new grace periods.
2806 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2809 unsigned long flags;
2811 struct rcu_node *rnp;
2813 rcu_for_each_leaf_node(rsp, rnp) {
2814 cond_resched_rcu_qs();
2816 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2817 if (rnp->qsmask == 0) {
2818 if (rcu_state_p == &rcu_sched_state ||
2819 rsp != rcu_state_p ||
2820 rcu_preempt_blocked_readers_cgp(rnp)) {
2822 * No point in scanning bits because they
2823 * are all zero. But we might need to
2824 * priority-boost blocked readers.
2826 rcu_initiate_boost(rnp, flags);
2827 /* rcu_initiate_boost() releases rnp->lock */
2831 (rnp->parent->qsmask & rnp->grpmask)) {
2833 * Race between grace-period
2834 * initialization and task exiting RCU
2835 * read-side critical section: Report.
2837 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2838 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2842 for_each_leaf_node_possible_cpu(rnp, cpu) {
2843 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2844 if ((rnp->qsmask & bit) != 0) {
2845 if (f(per_cpu_ptr(rsp->rda, cpu)))
2850 /* Idle/offline CPUs, report (releases rnp->lock. */
2851 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2853 /* Nothing to do here, so just drop the lock. */
2854 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2860 * Force quiescent states on reluctant CPUs, and also detect which
2861 * CPUs are in dyntick-idle mode.
2863 static void force_quiescent_state(struct rcu_state *rsp)
2865 unsigned long flags;
2867 struct rcu_node *rnp;
2868 struct rcu_node *rnp_old = NULL;
2870 /* Funnel through hierarchy to reduce memory contention. */
2871 rnp = __this_cpu_read(rsp->rda->mynode);
2872 for (; rnp != NULL; rnp = rnp->parent) {
2873 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2874 !raw_spin_trylock(&rnp->fqslock);
2875 if (rnp_old != NULL)
2876 raw_spin_unlock(&rnp_old->fqslock);
2878 rsp->n_force_qs_lh++;
2883 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2885 /* Reached the root of the rcu_node tree, acquire lock. */
2886 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2887 raw_spin_unlock(&rnp_old->fqslock);
2888 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2889 rsp->n_force_qs_lh++;
2890 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2891 return; /* Someone beat us to it. */
2893 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2894 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2895 rcu_gp_kthread_wake(rsp);
2899 * This does the RCU core processing work for the specified rcu_state
2900 * and rcu_data structures. This may be called only from the CPU to
2901 * whom the rdp belongs.
2904 __rcu_process_callbacks(struct rcu_state *rsp)
2906 unsigned long flags;
2908 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2910 WARN_ON_ONCE(!rdp->beenonline);
2912 /* Update RCU state based on any recent quiescent states. */
2913 rcu_check_quiescent_state(rsp, rdp);
2915 /* Does this CPU require a not-yet-started grace period? */
2916 local_irq_save(flags);
2917 if (cpu_needs_another_gp(rsp, rdp)) {
2918 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2919 needwake = rcu_start_gp(rsp);
2920 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2922 rcu_gp_kthread_wake(rsp);
2924 local_irq_restore(flags);
2927 /* If there are callbacks ready, invoke them. */
2928 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2929 invoke_rcu_callbacks(rsp, rdp);
2931 /* Do any needed deferred wakeups of rcuo kthreads. */
2932 do_nocb_deferred_wakeup(rdp);
2936 * Do RCU core processing for the current CPU.
2938 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2940 struct rcu_state *rsp;
2942 if (cpu_is_offline(smp_processor_id()))
2944 trace_rcu_utilization(TPS("Start RCU core"));
2945 for_each_rcu_flavor(rsp)
2946 __rcu_process_callbacks(rsp);
2947 trace_rcu_utilization(TPS("End RCU core"));
2951 * Schedule RCU callback invocation. If the specified type of RCU
2952 * does not support RCU priority boosting, just do a direct call,
2953 * otherwise wake up the per-CPU kernel kthread. Note that because we
2954 * are running on the current CPU with softirqs disabled, the
2955 * rcu_cpu_kthread_task cannot disappear out from under us.
2957 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2959 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2961 if (likely(!rsp->boost)) {
2962 rcu_do_batch(rsp, rdp);
2965 invoke_rcu_callbacks_kthread();
2968 static void invoke_rcu_core(void)
2970 if (cpu_online(smp_processor_id()))
2971 raise_softirq(RCU_SOFTIRQ);
2975 * Handle any core-RCU processing required by a call_rcu() invocation.
2977 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2978 struct rcu_head *head, unsigned long flags)
2983 * If called from an extended quiescent state, invoke the RCU
2984 * core in order to force a re-evaluation of RCU's idleness.
2986 if (!rcu_is_watching())
2989 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2990 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2994 * Force the grace period if too many callbacks or too long waiting.
2995 * Enforce hysteresis, and don't invoke force_quiescent_state()
2996 * if some other CPU has recently done so. Also, don't bother
2997 * invoking force_quiescent_state() if the newly enqueued callback
2998 * is the only one waiting for a grace period to complete.
3000 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
3001 rdp->qlen_last_fqs_check + qhimark)) {
3003 /* Are we ignoring a completed grace period? */
3004 note_gp_changes(rsp, rdp);
3006 /* Start a new grace period if one not already started. */
3007 if (!rcu_gp_in_progress(rsp)) {
3008 struct rcu_node *rnp_root = rcu_get_root(rsp);
3010 raw_spin_lock_rcu_node(rnp_root);
3011 needwake = rcu_start_gp(rsp);
3012 raw_spin_unlock_rcu_node(rnp_root);
3014 rcu_gp_kthread_wake(rsp);
3016 /* Give the grace period a kick. */
3017 rdp->blimit = LONG_MAX;
3018 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3019 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
3020 force_quiescent_state(rsp);
3021 rdp->n_force_qs_snap = rsp->n_force_qs;
3022 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
3028 * RCU callback function to leak a callback.
3030 static void rcu_leak_callback(struct rcu_head *rhp)
3035 * Helper function for call_rcu() and friends. The cpu argument will
3036 * normally be -1, indicating "currently running CPU". It may specify
3037 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3038 * is expected to specify a CPU.
3041 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3042 struct rcu_state *rsp, int cpu, bool lazy)
3044 unsigned long flags;
3045 struct rcu_data *rdp;
3047 /* Misaligned rcu_head! */
3048 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3050 if (debug_rcu_head_queue(head)) {
3052 * Probable double call_rcu(), so leak the callback.
3053 * Use rcu:rcu_callback trace event to find the previous
3054 * time callback was passed to __call_rcu().
3056 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
3058 WRITE_ONCE(head->func, rcu_leak_callback);
3063 local_irq_save(flags);
3064 rdp = this_cpu_ptr(rsp->rda);
3066 /* Add the callback to our list. */
3067 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
3071 rdp = per_cpu_ptr(rsp->rda, cpu);
3072 if (likely(rdp->mynode)) {
3073 /* Post-boot, so this should be for a no-CBs CPU. */
3074 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3075 WARN_ON_ONCE(offline);
3076 /* Offline CPU, _call_rcu() illegal, leak callback. */
3077 local_irq_restore(flags);
3081 * Very early boot, before rcu_init(). Initialize if needed
3082 * and then drop through to queue the callback.
3085 WARN_ON_ONCE(!rcu_is_watching());
3086 if (rcu_segcblist_empty(&rdp->cblist))
3087 rcu_segcblist_init(&rdp->cblist);
3089 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3091 rcu_idle_count_callbacks_posted();
3093 if (__is_kfree_rcu_offset((unsigned long)func))
3094 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3095 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3096 rcu_segcblist_n_cbs(&rdp->cblist));
3098 trace_rcu_callback(rsp->name, head,
3099 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3100 rcu_segcblist_n_cbs(&rdp->cblist));
3102 /* Go handle any RCU core processing required. */
3103 __call_rcu_core(rsp, rdp, head, flags);
3104 local_irq_restore(flags);
3108 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3109 * @head: structure to be used for queueing the RCU updates.
3110 * @func: actual callback function to be invoked after the grace period
3112 * The callback function will be invoked some time after a full grace
3113 * period elapses, in other words after all currently executing RCU
3114 * read-side critical sections have completed. call_rcu_sched() assumes
3115 * that the read-side critical sections end on enabling of preemption
3116 * or on voluntary preemption.
3117 * RCU read-side critical sections are delimited by:
3119 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3120 * - anything that disables preemption.
3122 * These may be nested.
3124 * See the description of call_rcu() for more detailed information on
3125 * memory ordering guarantees.
3127 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3129 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3131 EXPORT_SYMBOL_GPL(call_rcu_sched);
3134 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3135 * @head: structure to be used for queueing the RCU updates.
3136 * @func: actual callback function to be invoked after the grace period
3138 * The callback function will be invoked some time after a full grace
3139 * period elapses, in other words after all currently executing RCU
3140 * read-side critical sections have completed. call_rcu_bh() assumes
3141 * that the read-side critical sections end on completion of a softirq
3142 * handler. This means that read-side critical sections in process
3143 * context must not be interrupted by softirqs. This interface is to be
3144 * used when most of the read-side critical sections are in softirq context.
3145 * RCU read-side critical sections are delimited by:
3147 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3148 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3150 * These may be nested.
3152 * See the description of call_rcu() for more detailed information on
3153 * memory ordering guarantees.
3155 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3157 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3159 EXPORT_SYMBOL_GPL(call_rcu_bh);
3162 * Queue an RCU callback for lazy invocation after a grace period.
3163 * This will likely be later named something like "call_rcu_lazy()",
3164 * but this change will require some way of tagging the lazy RCU
3165 * callbacks in the list of pending callbacks. Until then, this
3166 * function may only be called from __kfree_rcu().
3168 void kfree_call_rcu(struct rcu_head *head,
3169 rcu_callback_t func)
3171 __call_rcu(head, func, rcu_state_p, -1, 1);
3173 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3176 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3177 * any blocking grace-period wait automatically implies a grace period
3178 * if there is only one CPU online at any point time during execution
3179 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3180 * occasionally incorrectly indicate that there are multiple CPUs online
3181 * when there was in fact only one the whole time, as this just adds
3182 * some overhead: RCU still operates correctly.
3184 static inline int rcu_blocking_is_gp(void)
3188 might_sleep(); /* Check for RCU read-side critical section. */
3190 ret = num_online_cpus() <= 1;
3196 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3198 * Control will return to the caller some time after a full rcu-sched
3199 * grace period has elapsed, in other words after all currently executing
3200 * rcu-sched read-side critical sections have completed. These read-side
3201 * critical sections are delimited by rcu_read_lock_sched() and
3202 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3203 * local_irq_disable(), and so on may be used in place of
3204 * rcu_read_lock_sched().
3206 * This means that all preempt_disable code sequences, including NMI and
3207 * non-threaded hardware-interrupt handlers, in progress on entry will
3208 * have completed before this primitive returns. However, this does not
3209 * guarantee that softirq handlers will have completed, since in some
3210 * kernels, these handlers can run in process context, and can block.
3212 * Note that this guarantee implies further memory-ordering guarantees.
3213 * On systems with more than one CPU, when synchronize_sched() returns,
3214 * each CPU is guaranteed to have executed a full memory barrier since the
3215 * end of its last RCU-sched read-side critical section whose beginning
3216 * preceded the call to synchronize_sched(). In addition, each CPU having
3217 * an RCU read-side critical section that extends beyond the return from
3218 * synchronize_sched() is guaranteed to have executed a full memory barrier
3219 * after the beginning of synchronize_sched() and before the beginning of
3220 * that RCU read-side critical section. Note that these guarantees include
3221 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3222 * that are executing in the kernel.
3224 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3225 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3226 * to have executed a full memory barrier during the execution of
3227 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3228 * again only if the system has more than one CPU).
3230 void synchronize_sched(void)
3232 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3233 lock_is_held(&rcu_lock_map) ||
3234 lock_is_held(&rcu_sched_lock_map),
3235 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3236 if (rcu_blocking_is_gp())
3238 if (rcu_gp_is_expedited())
3239 synchronize_sched_expedited();
3241 wait_rcu_gp(call_rcu_sched);
3243 EXPORT_SYMBOL_GPL(synchronize_sched);
3246 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3248 * Control will return to the caller some time after a full rcu_bh grace
3249 * period has elapsed, in other words after all currently executing rcu_bh
3250 * read-side critical sections have completed. RCU read-side critical
3251 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3252 * and may be nested.
3254 * See the description of synchronize_sched() for more detailed information
3255 * on memory ordering guarantees.
3257 void synchronize_rcu_bh(void)
3259 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3260 lock_is_held(&rcu_lock_map) ||
3261 lock_is_held(&rcu_sched_lock_map),
3262 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3263 if (rcu_blocking_is_gp())
3265 if (rcu_gp_is_expedited())
3266 synchronize_rcu_bh_expedited();
3268 wait_rcu_gp(call_rcu_bh);
3270 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3273 * get_state_synchronize_rcu - Snapshot current RCU state
3275 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3276 * to determine whether or not a full grace period has elapsed in the
3279 unsigned long get_state_synchronize_rcu(void)
3282 * Any prior manipulation of RCU-protected data must happen
3283 * before the load from ->gpnum.
3288 * Make sure this load happens before the purportedly
3289 * time-consuming work between get_state_synchronize_rcu()
3290 * and cond_synchronize_rcu().
3292 return smp_load_acquire(&rcu_state_p->gpnum);
3294 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3297 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3299 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3301 * If a full RCU grace period has elapsed since the earlier call to
3302 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3303 * synchronize_rcu() to wait for a full grace period.
3305 * Yes, this function does not take counter wrap into account. But
3306 * counter wrap is harmless. If the counter wraps, we have waited for
3307 * more than 2 billion grace periods (and way more on a 64-bit system!),
3308 * so waiting for one additional grace period should be just fine.
3310 void cond_synchronize_rcu(unsigned long oldstate)
3312 unsigned long newstate;
3315 * Ensure that this load happens before any RCU-destructive
3316 * actions the caller might carry out after we return.
3318 newstate = smp_load_acquire(&rcu_state_p->completed);
3319 if (ULONG_CMP_GE(oldstate, newstate))
3322 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3325 * get_state_synchronize_sched - Snapshot current RCU-sched state
3327 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3328 * to determine whether or not a full grace period has elapsed in the
3331 unsigned long get_state_synchronize_sched(void)
3334 * Any prior manipulation of RCU-protected data must happen
3335 * before the load from ->gpnum.
3340 * Make sure this load happens before the purportedly
3341 * time-consuming work between get_state_synchronize_sched()
3342 * and cond_synchronize_sched().
3344 return smp_load_acquire(&rcu_sched_state.gpnum);
3346 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3349 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3351 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3353 * If a full RCU-sched grace period has elapsed since the earlier call to
3354 * get_state_synchronize_sched(), just return. Otherwise, invoke
3355 * synchronize_sched() to wait for a full grace period.
3357 * Yes, this function does not take counter wrap into account. But
3358 * counter wrap is harmless. If the counter wraps, we have waited for
3359 * more than 2 billion grace periods (and way more on a 64-bit system!),
3360 * so waiting for one additional grace period should be just fine.
3362 void cond_synchronize_sched(unsigned long oldstate)
3364 unsigned long newstate;
3367 * Ensure that this load happens before any RCU-destructive
3368 * actions the caller might carry out after we return.
3370 newstate = smp_load_acquire(&rcu_sched_state.completed);
3371 if (ULONG_CMP_GE(oldstate, newstate))
3372 synchronize_sched();
3374 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3377 * Check to see if there is any immediate RCU-related work to be done
3378 * by the current CPU, for the specified type of RCU, returning 1 if so.
3379 * The checks are in order of increasing expense: checks that can be
3380 * carried out against CPU-local state are performed first. However,
3381 * we must check for CPU stalls first, else we might not get a chance.
3383 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3385 struct rcu_node *rnp = rdp->mynode;
3387 rdp->n_rcu_pending++;
3389 /* Check for CPU stalls, if enabled. */
3390 check_cpu_stall(rsp, rdp);
3392 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3393 if (rcu_nohz_full_cpu(rsp))
3396 /* Is the RCU core waiting for a quiescent state from this CPU? */
3397 if (rcu_scheduler_fully_active &&
3398 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3399 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
3400 rdp->n_rp_core_needs_qs++;
3401 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3402 rdp->n_rp_report_qs++;
3406 /* Does this CPU have callbacks ready to invoke? */
3407 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
3408 rdp->n_rp_cb_ready++;
3412 /* Has RCU gone idle with this CPU needing another grace period? */
3413 if (cpu_needs_another_gp(rsp, rdp)) {
3414 rdp->n_rp_cpu_needs_gp++;
3418 /* Has another RCU grace period completed? */
3419 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3420 rdp->n_rp_gp_completed++;
3424 /* Has a new RCU grace period started? */
3425 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3426 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3427 rdp->n_rp_gp_started++;
3431 /* Does this CPU need a deferred NOCB wakeup? */
3432 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3433 rdp->n_rp_nocb_defer_wakeup++;
3438 rdp->n_rp_need_nothing++;
3443 * Check to see if there is any immediate RCU-related work to be done
3444 * by the current CPU, returning 1 if so. This function is part of the
3445 * RCU implementation; it is -not- an exported member of the RCU API.
3447 static int rcu_pending(void)
3449 struct rcu_state *rsp;
3451 for_each_rcu_flavor(rsp)
3452 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3458 * Return true if the specified CPU has any callback. If all_lazy is
3459 * non-NULL, store an indication of whether all callbacks are lazy.
3460 * (If there are no callbacks, all of them are deemed to be lazy.)
3462 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3466 struct rcu_data *rdp;
3467 struct rcu_state *rsp;
3469 for_each_rcu_flavor(rsp) {
3470 rdp = this_cpu_ptr(rsp->rda);
3471 if (rcu_segcblist_empty(&rdp->cblist))
3474 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3485 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3486 * the compiler is expected to optimize this away.
3488 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3489 int cpu, unsigned long done)
3491 trace_rcu_barrier(rsp->name, s, cpu,
3492 atomic_read(&rsp->barrier_cpu_count), done);
3496 * RCU callback function for _rcu_barrier(). If we are last, wake
3497 * up the task executing _rcu_barrier().
3499 static void rcu_barrier_callback(struct rcu_head *rhp)
3501 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3502 struct rcu_state *rsp = rdp->rsp;
3504 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3505 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3506 rsp->barrier_sequence);
3507 complete(&rsp->barrier_completion);
3509 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3514 * Called with preemption disabled, and from cross-cpu IRQ context.
3516 static void rcu_barrier_func(void *type)
3518 struct rcu_state *rsp = type;
3519 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3521 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3522 rdp->barrier_head.func = rcu_barrier_callback;
3523 debug_rcu_head_queue(&rdp->barrier_head);
3524 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3525 atomic_inc(&rsp->barrier_cpu_count);
3527 debug_rcu_head_unqueue(&rdp->barrier_head);
3528 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3529 rsp->barrier_sequence);
3534 * Orchestrate the specified type of RCU barrier, waiting for all
3535 * RCU callbacks of the specified type to complete.
3537 static void _rcu_barrier(struct rcu_state *rsp)
3540 struct rcu_data *rdp;
3541 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3543 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3545 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3546 mutex_lock(&rsp->barrier_mutex);
3548 /* Did someone else do our work for us? */
3549 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3550 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3551 rsp->barrier_sequence);
3552 smp_mb(); /* caller's subsequent code after above check. */
3553 mutex_unlock(&rsp->barrier_mutex);
3557 /* Mark the start of the barrier operation. */
3558 rcu_seq_start(&rsp->barrier_sequence);
3559 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3562 * Initialize the count to one rather than to zero in order to
3563 * avoid a too-soon return to zero in case of a short grace period
3564 * (or preemption of this task). Exclude CPU-hotplug operations
3565 * to ensure that no offline CPU has callbacks queued.
3567 init_completion(&rsp->barrier_completion);
3568 atomic_set(&rsp->barrier_cpu_count, 1);
3572 * Force each CPU with callbacks to register a new callback.
3573 * When that callback is invoked, we will know that all of the
3574 * corresponding CPU's preceding callbacks have been invoked.
3576 for_each_possible_cpu(cpu) {
3577 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3579 rdp = per_cpu_ptr(rsp->rda, cpu);
3580 if (rcu_is_nocb_cpu(cpu)) {
3581 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3582 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3583 rsp->barrier_sequence);
3585 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3586 rsp->barrier_sequence);
3587 smp_mb__before_atomic();
3588 atomic_inc(&rsp->barrier_cpu_count);
3589 __call_rcu(&rdp->barrier_head,
3590 rcu_barrier_callback, rsp, cpu, 0);
3592 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3593 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3594 rsp->barrier_sequence);
3595 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3597 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3598 rsp->barrier_sequence);
3604 * Now that we have an rcu_barrier_callback() callback on each
3605 * CPU, and thus each counted, remove the initial count.
3607 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3608 complete(&rsp->barrier_completion);
3610 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3611 wait_for_completion(&rsp->barrier_completion);
3613 /* Mark the end of the barrier operation. */
3614 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3615 rcu_seq_end(&rsp->barrier_sequence);
3617 /* Other rcu_barrier() invocations can now safely proceed. */
3618 mutex_unlock(&rsp->barrier_mutex);
3622 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3624 void rcu_barrier_bh(void)
3626 _rcu_barrier(&rcu_bh_state);
3628 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3631 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3633 void rcu_barrier_sched(void)
3635 _rcu_barrier(&rcu_sched_state);
3637 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3640 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3641 * first CPU in a given leaf rcu_node structure coming online. The caller
3642 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3645 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3648 struct rcu_node *rnp = rnp_leaf;
3650 lockdep_assert_held(&rnp->lock);
3652 mask = rnp->grpmask;
3656 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3657 rnp->qsmaskinit |= mask;
3658 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3663 * Do boot-time initialization of a CPU's per-CPU RCU data.
3666 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3668 unsigned long flags;
3669 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3670 struct rcu_node *rnp = rcu_get_root(rsp);
3672 /* Set up local state, ensuring consistent view of global state. */
3673 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3674 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3675 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3676 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3677 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3680 rcu_boot_init_nocb_percpu_data(rdp);
3681 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3685 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3686 * offline event can be happening at a given time. Note also that we
3687 * can accept some slop in the rsp->completed access due to the fact
3688 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3691 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3693 unsigned long flags;
3694 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3695 struct rcu_node *rnp = rcu_get_root(rsp);
3697 /* Set up local state, ensuring consistent view of global state. */
3698 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3699 rdp->qlen_last_fqs_check = 0;
3700 rdp->n_force_qs_snap = rsp->n_force_qs;
3701 rdp->blimit = blimit;
3702 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3703 !init_nocb_callback_list(rdp))
3704 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3705 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3706 rcu_dynticks_eqs_online();
3707 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3710 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3711 * propagation up the rcu_node tree will happen at the beginning
3712 * of the next grace period.
3715 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3716 rdp->beenonline = true; /* We have now been online. */
3717 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3718 rdp->completed = rnp->completed;
3719 rdp->cpu_no_qs.b.norm = true;
3720 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3721 rdp->core_needs_qs = false;
3722 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3723 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3727 * Invoked early in the CPU-online process, when pretty much all
3728 * services are available. The incoming CPU is not present.
3730 int rcutree_prepare_cpu(unsigned int cpu)
3732 struct rcu_state *rsp;
3734 for_each_rcu_flavor(rsp)
3735 rcu_init_percpu_data(cpu, rsp);
3737 rcu_prepare_kthreads(cpu);
3738 rcu_spawn_all_nocb_kthreads(cpu);
3744 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3746 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3748 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3750 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3754 * Near the end of the CPU-online process. Pretty much all services
3755 * enabled, and the CPU is now very much alive.
3757 int rcutree_online_cpu(unsigned int cpu)
3759 sync_sched_exp_online_cleanup(cpu);
3760 rcutree_affinity_setting(cpu, -1);
3761 if (IS_ENABLED(CONFIG_TREE_SRCU))
3762 srcu_online_cpu(cpu);
3767 * Near the beginning of the process. The CPU is still very much alive
3768 * with pretty much all services enabled.
3770 int rcutree_offline_cpu(unsigned int cpu)
3772 rcutree_affinity_setting(cpu, cpu);
3773 if (IS_ENABLED(CONFIG_TREE_SRCU))
3774 srcu_offline_cpu(cpu);
3779 * Near the end of the offline process. We do only tracing here.
3781 int rcutree_dying_cpu(unsigned int cpu)
3783 struct rcu_state *rsp;
3785 for_each_rcu_flavor(rsp)
3786 rcu_cleanup_dying_cpu(rsp);
3791 * The outgoing CPU is gone and we are running elsewhere.
3793 int rcutree_dead_cpu(unsigned int cpu)
3795 struct rcu_state *rsp;
3797 for_each_rcu_flavor(rsp) {
3798 rcu_cleanup_dead_cpu(cpu, rsp);
3799 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3805 * Mark the specified CPU as being online so that subsequent grace periods
3806 * (both expedited and normal) will wait on it. Note that this means that
3807 * incoming CPUs are not allowed to use RCU read-side critical sections
3808 * until this function is called. Failing to observe this restriction
3809 * will result in lockdep splats.
3811 * Note that this function is special in that it is invoked directly
3812 * from the incoming CPU rather than from the cpuhp_step mechanism.
3813 * This is because this function must be invoked at a precise location.
3815 void rcu_cpu_starting(unsigned int cpu)
3817 unsigned long flags;
3820 unsigned long oldmask;
3821 struct rcu_data *rdp;
3822 struct rcu_node *rnp;
3823 struct rcu_state *rsp;
3825 for_each_rcu_flavor(rsp) {
3826 rdp = per_cpu_ptr(rsp->rda, cpu);
3828 mask = rdp->grpmask;
3829 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3830 rnp->qsmaskinitnext |= mask;
3831 oldmask = rnp->expmaskinitnext;
3832 rnp->expmaskinitnext |= mask;
3833 oldmask ^= rnp->expmaskinitnext;
3834 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3835 /* Allow lockless access for expedited grace periods. */
3836 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3837 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3839 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3842 #ifdef CONFIG_HOTPLUG_CPU
3844 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3845 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3848 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3850 unsigned long flags;
3852 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3853 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3855 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3856 mask = rdp->grpmask;
3857 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3858 rnp->qsmaskinitnext &= ~mask;
3859 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3863 * The outgoing function has no further need of RCU, so remove it from
3864 * the list of CPUs that RCU must track.
3866 * Note that this function is special in that it is invoked directly
3867 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3868 * This is because this function must be invoked at a precise location.
3870 void rcu_report_dead(unsigned int cpu)
3872 struct rcu_state *rsp;
3874 /* QS for any half-done expedited RCU-sched GP. */
3876 rcu_report_exp_rdp(&rcu_sched_state,
3877 this_cpu_ptr(rcu_sched_state.rda), true);
3879 for_each_rcu_flavor(rsp)
3880 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3883 /* Migrate the dead CPU's callbacks to the current CPU. */
3884 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3886 unsigned long flags;
3887 struct rcu_data *my_rdp;
3888 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3889 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3891 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3892 return; /* No callbacks to migrate. */
3894 local_irq_save(flags);
3895 my_rdp = this_cpu_ptr(rsp->rda);
3896 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3897 local_irq_restore(flags);
3900 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3901 rcu_advance_cbs(rsp, rnp_root, rdp); /* Leverage recent GPs. */
3902 rcu_advance_cbs(rsp, rnp_root, my_rdp); /* Assign GP to pending CBs. */
3903 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3904 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3905 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3906 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3907 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3908 !rcu_segcblist_empty(&rdp->cblist),
3909 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3910 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3911 rcu_segcblist_first_cb(&rdp->cblist));
3915 * The outgoing CPU has just passed through the dying-idle state,
3916 * and we are being invoked from the CPU that was IPIed to continue the
3917 * offline operation. We need to migrate the outgoing CPU's callbacks.
3919 void rcutree_migrate_callbacks(int cpu)
3921 struct rcu_state *rsp;
3923 for_each_rcu_flavor(rsp)
3924 rcu_migrate_callbacks(cpu, rsp);
3929 * On non-huge systems, use expedited RCU grace periods to make suspend
3930 * and hibernation run faster.
3932 static int rcu_pm_notify(struct notifier_block *self,
3933 unsigned long action, void *hcpu)
3936 case PM_HIBERNATION_PREPARE:
3937 case PM_SUSPEND_PREPARE:
3938 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3941 case PM_POST_HIBERNATION:
3942 case PM_POST_SUSPEND:
3943 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3944 rcu_unexpedite_gp();
3953 * Spawn the kthreads that handle each RCU flavor's grace periods.
3955 static int __init rcu_spawn_gp_kthread(void)
3957 unsigned long flags;
3958 int kthread_prio_in = kthread_prio;
3959 struct rcu_node *rnp;
3960 struct rcu_state *rsp;
3961 struct sched_param sp;
3962 struct task_struct *t;
3964 /* Force priority into range. */
3965 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3967 else if (kthread_prio < 0)
3969 else if (kthread_prio > 99)
3971 if (kthread_prio != kthread_prio_in)
3972 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3973 kthread_prio, kthread_prio_in);
3975 rcu_scheduler_fully_active = 1;
3976 for_each_rcu_flavor(rsp) {
3977 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3979 rnp = rcu_get_root(rsp);
3980 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3981 rsp->gp_kthread = t;
3983 sp.sched_priority = kthread_prio;
3984 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3986 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3989 rcu_spawn_nocb_kthreads();
3990 rcu_spawn_boost_kthreads();
3993 early_initcall(rcu_spawn_gp_kthread);
3996 * This function is invoked towards the end of the scheduler's
3997 * initialization process. Before this is called, the idle task might
3998 * contain synchronous grace-period primitives (during which time, this idle
3999 * task is booting the system, and such primitives are no-ops). After this
4000 * function is called, any synchronous grace-period primitives are run as
4001 * expedited, with the requesting task driving the grace period forward.
4002 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4003 * runtime RCU functionality.
4005 void rcu_scheduler_starting(void)
4007 WARN_ON(num_online_cpus() != 1);
4008 WARN_ON(nr_context_switches() > 0);
4009 rcu_test_sync_prims();
4010 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4011 rcu_test_sync_prims();
4015 * Helper function for rcu_init() that initializes one rcu_state structure.
4017 static void __init rcu_init_one(struct rcu_state *rsp)
4019 static const char * const buf[] = RCU_NODE_NAME_INIT;
4020 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4021 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4022 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4024 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4028 struct rcu_node *rnp;
4030 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4032 /* Silence gcc 4.8 false positive about array index out of range. */
4033 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4034 panic("rcu_init_one: rcu_num_lvls out of range");
4036 /* Initialize the level-tracking arrays. */
4038 for (i = 1; i < rcu_num_lvls; i++)
4039 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
4040 rcu_init_levelspread(levelspread, num_rcu_lvl);
4042 /* Initialize the elements themselves, starting from the leaves. */
4044 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4045 cpustride *= levelspread[i];
4046 rnp = rsp->level[i];
4047 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4048 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4049 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4050 &rcu_node_class[i], buf[i]);
4051 raw_spin_lock_init(&rnp->fqslock);
4052 lockdep_set_class_and_name(&rnp->fqslock,
4053 &rcu_fqs_class[i], fqs[i]);
4054 rnp->gpnum = rsp->gpnum;
4055 rnp->completed = rsp->completed;
4057 rnp->qsmaskinit = 0;
4058 rnp->grplo = j * cpustride;
4059 rnp->grphi = (j + 1) * cpustride - 1;
4060 if (rnp->grphi >= nr_cpu_ids)
4061 rnp->grphi = nr_cpu_ids - 1;
4067 rnp->grpnum = j % levelspread[i - 1];
4068 rnp->grpmask = 1UL << rnp->grpnum;
4069 rnp->parent = rsp->level[i - 1] +
4070 j / levelspread[i - 1];
4073 INIT_LIST_HEAD(&rnp->blkd_tasks);
4074 rcu_init_one_nocb(rnp);
4075 init_waitqueue_head(&rnp->exp_wq[0]);
4076 init_waitqueue_head(&rnp->exp_wq[1]);
4077 init_waitqueue_head(&rnp->exp_wq[2]);
4078 init_waitqueue_head(&rnp->exp_wq[3]);
4079 spin_lock_init(&rnp->exp_lock);
4083 init_swait_queue_head(&rsp->gp_wq);
4084 init_swait_queue_head(&rsp->expedited_wq);
4085 rnp = rsp->level[rcu_num_lvls - 1];
4086 for_each_possible_cpu(i) {
4087 while (i > rnp->grphi)
4089 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4090 rcu_boot_init_percpu_data(i, rsp);
4092 list_add(&rsp->flavors, &rcu_struct_flavors);
4096 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4097 * replace the definitions in tree.h because those are needed to size
4098 * the ->node array in the rcu_state structure.
4100 static void __init rcu_init_geometry(void)
4104 int rcu_capacity[RCU_NUM_LVLS];
4107 * Initialize any unspecified boot parameters.
4108 * The default values of jiffies_till_first_fqs and
4109 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4110 * value, which is a function of HZ, then adding one for each
4111 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4113 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4114 if (jiffies_till_first_fqs == ULONG_MAX)
4115 jiffies_till_first_fqs = d;
4116 if (jiffies_till_next_fqs == ULONG_MAX)
4117 jiffies_till_next_fqs = d;
4119 /* If the compile-time values are accurate, just leave. */
4120 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4121 nr_cpu_ids == NR_CPUS)
4123 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4124 rcu_fanout_leaf, nr_cpu_ids);
4127 * The boot-time rcu_fanout_leaf parameter must be at least two
4128 * and cannot exceed the number of bits in the rcu_node masks.
4129 * Complain and fall back to the compile-time values if this
4130 * limit is exceeded.
4132 if (rcu_fanout_leaf < 2 ||
4133 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4134 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4140 * Compute number of nodes that can be handled an rcu_node tree
4141 * with the given number of levels.
4143 rcu_capacity[0] = rcu_fanout_leaf;
4144 for (i = 1; i < RCU_NUM_LVLS; i++)
4145 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4148 * The tree must be able to accommodate the configured number of CPUs.
4149 * If this limit is exceeded, fall back to the compile-time values.
4151 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4152 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4157 /* Calculate the number of levels in the tree. */
4158 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4160 rcu_num_lvls = i + 1;
4162 /* Calculate the number of rcu_nodes at each level of the tree. */
4163 for (i = 0; i < rcu_num_lvls; i++) {
4164 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4165 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4168 /* Calculate the total number of rcu_node structures. */
4170 for (i = 0; i < rcu_num_lvls; i++)
4171 rcu_num_nodes += num_rcu_lvl[i];
4175 * Dump out the structure of the rcu_node combining tree associated
4176 * with the rcu_state structure referenced by rsp.
4178 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4181 struct rcu_node *rnp;
4183 pr_info("rcu_node tree layout dump\n");
4185 rcu_for_each_node_breadth_first(rsp, rnp) {
4186 if (rnp->level != level) {
4191 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4196 void __init rcu_init(void)
4200 rcu_early_boot_tests();
4202 rcu_bootup_announce();
4203 rcu_init_geometry();
4204 rcu_init_one(&rcu_bh_state);
4205 rcu_init_one(&rcu_sched_state);
4207 rcu_dump_rcu_node_tree(&rcu_sched_state);
4208 __rcu_init_preempt();
4209 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4212 * We don't need protection against CPU-hotplug here because
4213 * this is called early in boot, before either interrupts
4214 * or the scheduler are operational.
4216 pm_notifier(rcu_pm_notify, 0);
4217 for_each_online_cpu(cpu) {
4218 rcutree_prepare_cpu(cpu);
4219 rcu_cpu_starting(cpu);
4220 if (IS_ENABLED(CONFIG_TREE_SRCU))
4221 srcu_online_cpu(cpu);
4225 #include "tree_exp.h"
4226 #include "tree_plugin.h"