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_nmi_nesting = DYNTICK_IRQ_NONIDLE,
270 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
274 * There's a few places, currently just in the tracing infrastructure,
275 * that uses rcu_irq_enter() to make sure RCU is watching. But there's
276 * a small location where that will not even work. In those cases
277 * rcu_irq_enter_disabled() needs to be checked to make sure rcu_irq_enter()
280 static DEFINE_PER_CPU(bool, disable_rcu_irq_enter);
282 bool rcu_irq_enter_disabled(void)
284 return this_cpu_read(disable_rcu_irq_enter);
288 * Record entry into an extended quiescent state. This is only to be
289 * called when not already in an extended quiescent state.
291 static void rcu_dynticks_eqs_enter(void)
293 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
297 * CPUs seeing atomic_add_return() must see prior RCU read-side
298 * critical sections, and we also must force ordering with the
301 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
302 /* Better be in an extended quiescent state! */
303 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
304 (seq & RCU_DYNTICK_CTRL_CTR));
305 /* Better not have special action (TLB flush) pending! */
306 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
307 (seq & RCU_DYNTICK_CTRL_MASK));
311 * Record exit from an extended quiescent state. This is only to be
312 * called from an extended quiescent state.
314 static void rcu_dynticks_eqs_exit(void)
316 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
320 * CPUs seeing atomic_add_return() must see prior idle sojourns,
321 * and we also must force ordering with the next RCU read-side
324 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
325 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
326 !(seq & RCU_DYNTICK_CTRL_CTR));
327 if (seq & RCU_DYNTICK_CTRL_MASK) {
328 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
329 smp_mb__after_atomic(); /* _exit after clearing mask. */
330 /* Prefer duplicate flushes to losing a flush. */
331 rcu_eqs_special_exit();
336 * Reset the current CPU's ->dynticks counter to indicate that the
337 * newly onlined CPU is no longer in an extended quiescent state.
338 * This will either leave the counter unchanged, or increment it
339 * to the next non-quiescent value.
341 * The non-atomic test/increment sequence works because the upper bits
342 * of the ->dynticks counter are manipulated only by the corresponding CPU,
343 * or when the corresponding CPU is offline.
345 static void rcu_dynticks_eqs_online(void)
347 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
349 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
351 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
355 * Is the current CPU in an extended quiescent state?
357 * No ordering, as we are sampling CPU-local information.
359 bool rcu_dynticks_curr_cpu_in_eqs(void)
361 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
363 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
367 * Snapshot the ->dynticks counter with full ordering so as to allow
368 * stable comparison of this counter with past and future snapshots.
370 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
372 int snap = atomic_add_return(0, &rdtp->dynticks);
374 return snap & ~RCU_DYNTICK_CTRL_MASK;
378 * Return true if the snapshot returned from rcu_dynticks_snap()
379 * indicates that RCU is in an extended quiescent state.
381 static bool rcu_dynticks_in_eqs(int snap)
383 return !(snap & RCU_DYNTICK_CTRL_CTR);
387 * Return true if the CPU corresponding to the specified rcu_dynticks
388 * structure has spent some time in an extended quiescent state since
389 * rcu_dynticks_snap() returned the specified snapshot.
391 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
393 return snap != rcu_dynticks_snap(rdtp);
397 * Do a double-increment of the ->dynticks counter to emulate a
398 * momentary idle-CPU quiescent state.
400 static void rcu_dynticks_momentary_idle(void)
402 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
403 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
406 /* It is illegal to call this from idle state. */
407 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
411 * Set the special (bottom) bit of the specified CPU so that it
412 * will take special action (such as flushing its TLB) on the
413 * next exit from an extended quiescent state. Returns true if
414 * the bit was successfully set, or false if the CPU was not in
415 * an extended quiescent state.
417 bool rcu_eqs_special_set(int cpu)
421 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
424 old = atomic_read(&rdtp->dynticks);
425 if (old & RCU_DYNTICK_CTRL_CTR)
427 new = old | RCU_DYNTICK_CTRL_MASK;
428 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
433 * Let the RCU core know that this CPU has gone through the scheduler,
434 * which is a quiescent state. This is called when the need for a
435 * quiescent state is urgent, so we burn an atomic operation and full
436 * memory barriers to let the RCU core know about it, regardless of what
437 * this CPU might (or might not) do in the near future.
439 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
441 * The caller must have disabled interrupts.
443 static void rcu_momentary_dyntick_idle(void)
445 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
446 rcu_dynticks_momentary_idle();
450 * Note a context switch. This is a quiescent state for RCU-sched,
451 * and requires special handling for preemptible RCU.
452 * The caller must have disabled interrupts.
454 void rcu_note_context_switch(bool preempt)
456 barrier(); /* Avoid RCU read-side critical sections leaking down. */
457 trace_rcu_utilization(TPS("Start context switch"));
459 rcu_preempt_note_context_switch(preempt);
460 /* Load rcu_urgent_qs before other flags. */
461 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
463 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
464 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
465 rcu_momentary_dyntick_idle();
466 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
468 rcu_note_voluntary_context_switch_lite(current);
470 trace_rcu_utilization(TPS("End context switch"));
471 barrier(); /* Avoid RCU read-side critical sections leaking up. */
473 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
476 * Register a quiescent state for all RCU flavors. If there is an
477 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
478 * dyntick-idle quiescent state visible to other CPUs (but only for those
479 * RCU flavors in desperate need of a quiescent state, which will normally
480 * be none of them). Either way, do a lightweight quiescent state for
483 * The barrier() calls are redundant in the common case when this is
484 * called externally, but just in case this is called from within this
488 void rcu_all_qs(void)
492 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
495 /* Load rcu_urgent_qs before other flags. */
496 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
500 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
501 barrier(); /* Avoid RCU read-side critical sections leaking down. */
502 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
503 local_irq_save(flags);
504 rcu_momentary_dyntick_idle();
505 local_irq_restore(flags);
507 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
509 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
510 barrier(); /* Avoid RCU read-side critical sections leaking up. */
513 EXPORT_SYMBOL_GPL(rcu_all_qs);
515 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
516 static long blimit = DEFAULT_RCU_BLIMIT;
517 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
518 static long qhimark = DEFAULT_RCU_QHIMARK;
519 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
520 static long qlowmark = DEFAULT_RCU_QLOMARK;
522 module_param(blimit, long, 0444);
523 module_param(qhimark, long, 0444);
524 module_param(qlowmark, long, 0444);
526 static ulong jiffies_till_first_fqs = ULONG_MAX;
527 static ulong jiffies_till_next_fqs = ULONG_MAX;
528 static bool rcu_kick_kthreads;
530 module_param(jiffies_till_first_fqs, ulong, 0644);
531 module_param(jiffies_till_next_fqs, ulong, 0644);
532 module_param(rcu_kick_kthreads, bool, 0644);
535 * How long the grace period must be before we start recruiting
536 * quiescent-state help from rcu_note_context_switch().
538 static ulong jiffies_till_sched_qs = HZ / 10;
539 module_param(jiffies_till_sched_qs, ulong, 0444);
541 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
542 struct rcu_data *rdp);
543 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
544 static void force_quiescent_state(struct rcu_state *rsp);
545 static int rcu_pending(void);
548 * Return the number of RCU batches started thus far for debug & stats.
550 unsigned long rcu_batches_started(void)
552 return rcu_state_p->gpnum;
554 EXPORT_SYMBOL_GPL(rcu_batches_started);
557 * Return the number of RCU-sched batches started thus far for debug & stats.
559 unsigned long rcu_batches_started_sched(void)
561 return rcu_sched_state.gpnum;
563 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
566 * Return the number of RCU BH batches started thus far for debug & stats.
568 unsigned long rcu_batches_started_bh(void)
570 return rcu_bh_state.gpnum;
572 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
575 * Return the number of RCU batches completed thus far for debug & stats.
577 unsigned long rcu_batches_completed(void)
579 return rcu_state_p->completed;
581 EXPORT_SYMBOL_GPL(rcu_batches_completed);
584 * Return the number of RCU-sched batches completed thus far for debug & stats.
586 unsigned long rcu_batches_completed_sched(void)
588 return rcu_sched_state.completed;
590 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
593 * Return the number of RCU BH batches completed thus far for debug & stats.
595 unsigned long rcu_batches_completed_bh(void)
597 return rcu_bh_state.completed;
599 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
602 * Return the number of RCU expedited batches completed thus far for
603 * debug & stats. Odd numbers mean that a batch is in progress, even
604 * numbers mean idle. The value returned will thus be roughly double
605 * the cumulative batches since boot.
607 unsigned long rcu_exp_batches_completed(void)
609 return rcu_state_p->expedited_sequence;
611 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
614 * Return the number of RCU-sched expedited batches completed thus far
615 * for debug & stats. Similar to rcu_exp_batches_completed().
617 unsigned long rcu_exp_batches_completed_sched(void)
619 return rcu_sched_state.expedited_sequence;
621 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
624 * Force a quiescent state.
626 void rcu_force_quiescent_state(void)
628 force_quiescent_state(rcu_state_p);
630 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
633 * Force a quiescent state for RCU BH.
635 void rcu_bh_force_quiescent_state(void)
637 force_quiescent_state(&rcu_bh_state);
639 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
642 * Force a quiescent state for RCU-sched.
644 void rcu_sched_force_quiescent_state(void)
646 force_quiescent_state(&rcu_sched_state);
648 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
651 * Show the state of the grace-period kthreads.
653 void show_rcu_gp_kthreads(void)
655 struct rcu_state *rsp;
657 for_each_rcu_flavor(rsp) {
658 pr_info("%s: wait state: %d ->state: %#lx\n",
659 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
660 /* sched_show_task(rsp->gp_kthread); */
663 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
666 * Record the number of times rcutorture tests have been initiated and
667 * terminated. This information allows the debugfs tracing stats to be
668 * correlated to the rcutorture messages, even when the rcutorture module
669 * is being repeatedly loaded and unloaded. In other words, we cannot
670 * store this state in rcutorture itself.
672 void rcutorture_record_test_transition(void)
674 rcutorture_testseq++;
675 rcutorture_vernum = 0;
677 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
680 * Send along grace-period-related data for rcutorture diagnostics.
682 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
683 unsigned long *gpnum, unsigned long *completed)
685 struct rcu_state *rsp = NULL;
694 case RCU_SCHED_FLAVOR:
695 rsp = &rcu_sched_state;
702 *flags = READ_ONCE(rsp->gp_flags);
703 *gpnum = READ_ONCE(rsp->gpnum);
704 *completed = READ_ONCE(rsp->completed);
706 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
709 * Record the number of writer passes through the current rcutorture test.
710 * This is also used to correlate debugfs tracing stats with the rcutorture
713 void rcutorture_record_progress(unsigned long vernum)
717 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
720 * Return the root node of the specified rcu_state structure.
722 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
724 return &rsp->node[0];
728 * Is there any need for future grace periods?
729 * Interrupts must be disabled. If the caller does not hold the root
730 * rnp_node structure's ->lock, the results are advisory only.
732 static int rcu_future_needs_gp(struct rcu_state *rsp)
734 struct rcu_node *rnp = rcu_get_root(rsp);
735 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
736 int *fp = &rnp->need_future_gp[idx];
738 lockdep_assert_irqs_disabled();
739 return READ_ONCE(*fp);
743 * Does the current CPU require a not-yet-started grace period?
744 * The caller must have disabled interrupts to prevent races with
745 * normal callback registry.
748 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
750 lockdep_assert_irqs_disabled();
751 if (rcu_gp_in_progress(rsp))
752 return false; /* No, a grace period is already in progress. */
753 if (rcu_future_needs_gp(rsp))
754 return true; /* Yes, a no-CBs CPU needs one. */
755 if (!rcu_segcblist_is_enabled(&rdp->cblist))
756 return false; /* No, this is a no-CBs (or offline) CPU. */
757 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
758 return true; /* Yes, CPU has newly registered callbacks. */
759 if (rcu_segcblist_future_gp_needed(&rdp->cblist,
760 READ_ONCE(rsp->completed)))
761 return true; /* Yes, CBs for future grace period. */
762 return false; /* No grace period needed. */
766 * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
768 * Enter idle, doing appropriate accounting. The caller must have
769 * disabled interrupts.
771 static void rcu_eqs_enter_common(bool user)
773 struct rcu_state *rsp;
774 struct rcu_data *rdp;
775 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
777 lockdep_assert_irqs_disabled();
778 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0);
779 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
780 !user && !is_idle_task(current)) {
781 struct task_struct *idle __maybe_unused =
782 idle_task(smp_processor_id());
784 trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0);
785 rcu_ftrace_dump(DUMP_ORIG);
786 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
787 current->pid, current->comm,
788 idle->pid, idle->comm); /* must be idle task! */
790 for_each_rcu_flavor(rsp) {
791 rdp = this_cpu_ptr(rsp->rda);
792 do_nocb_deferred_wakeup(rdp);
794 rcu_prepare_for_idle();
795 __this_cpu_inc(disable_rcu_irq_enter);
796 rdtp->dynticks_nesting = 0; /* Breaks tracing momentarily. */
797 rcu_dynticks_eqs_enter(); /* After this, tracing works again. */
798 __this_cpu_dec(disable_rcu_irq_enter);
799 rcu_dynticks_task_enter();
802 * It is illegal to enter an extended quiescent state while
803 * in an RCU read-side critical section.
805 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
806 "Illegal idle entry in RCU read-side critical section.");
807 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
808 "Illegal idle entry in RCU-bh read-side critical section.");
809 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
810 "Illegal idle entry in RCU-sched read-side critical section.");
814 * Enter an RCU extended quiescent state, which can be either the
815 * idle loop or adaptive-tickless usermode execution.
817 static void rcu_eqs_enter(bool user)
819 struct rcu_dynticks *rdtp;
821 rdtp = this_cpu_ptr(&rcu_dynticks);
822 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
823 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
824 (rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
825 if ((rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
826 rcu_eqs_enter_common(user);
828 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
832 * rcu_idle_enter - inform RCU that current CPU is entering idle
834 * Enter idle mode, in other words, -leave- the mode in which RCU
835 * read-side critical sections can occur. (Though RCU read-side
836 * critical sections can occur in irq handlers in idle, a possibility
837 * handled by irq_enter() and irq_exit().)
839 * We crowbar the ->dynticks_nesting field to zero to allow for
840 * the possibility of usermode upcalls having messed up our count
841 * of interrupt nesting level during the prior busy period.
843 * If you add or remove a call to rcu_idle_enter(), be sure to test with
844 * CONFIG_RCU_EQS_DEBUG=y.
846 void rcu_idle_enter(void)
848 lockdep_assert_irqs_disabled();
849 rcu_eqs_enter(false);
852 #ifdef CONFIG_NO_HZ_FULL
854 * rcu_user_enter - inform RCU that we are resuming userspace.
856 * Enter RCU idle mode right before resuming userspace. No use of RCU
857 * is permitted between this call and rcu_user_exit(). This way the
858 * CPU doesn't need to maintain the tick for RCU maintenance purposes
859 * when the CPU runs in userspace.
861 * If you add or remove a call to rcu_user_enter(), be sure to test with
862 * CONFIG_RCU_EQS_DEBUG=y.
864 void rcu_user_enter(void)
866 lockdep_assert_irqs_disabled();
869 #endif /* CONFIG_NO_HZ_FULL */
872 * rcu_nmi_exit - inform RCU of exit from NMI context
874 * If we are returning from the outermost NMI handler that interrupted an
875 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
876 * to let the RCU grace-period handling know that the CPU is back to
879 * If you add or remove a call to rcu_nmi_exit(), be sure to test
880 * with CONFIG_RCU_EQS_DEBUG=y.
882 void rcu_nmi_exit(void)
884 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
887 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
888 * (We are exiting an NMI handler, so RCU better be paying attention
891 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
892 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
895 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
896 * leave it in non-RCU-idle state.
898 if (rdtp->dynticks_nmi_nesting != 1) {
899 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
900 rdtp->dynticks_nmi_nesting - 2);
904 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
905 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
906 rcu_dynticks_eqs_enter();
910 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
912 * Exit from an interrupt handler, which might possibly result in entering
913 * idle mode, in other words, leaving the mode in which read-side critical
914 * sections can occur. The caller must have disabled interrupts.
916 * This code assumes that the idle loop never does anything that might
917 * result in unbalanced calls to irq_enter() and irq_exit(). If your
918 * architecture's idle loop violates this assumption, RCU will give you what
919 * you deserve, good and hard. But very infrequently and irreproducibly.
921 * Use things like work queues to work around this limitation.
923 * You have been warned.
925 * If you add or remove a call to rcu_irq_exit(), be sure to test with
926 * CONFIG_RCU_EQS_DEBUG=y.
928 void rcu_irq_exit(void)
930 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
932 lockdep_assert_irqs_disabled();
933 if (rdtp->dynticks_nmi_nesting == 1)
934 rcu_prepare_for_idle();
936 if (rdtp->dynticks_nmi_nesting == 0)
937 rcu_dynticks_task_enter();
941 * Wrapper for rcu_irq_exit() where interrupts are enabled.
943 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
944 * with CONFIG_RCU_EQS_DEBUG=y.
946 void rcu_irq_exit_irqson(void)
950 local_irq_save(flags);
952 local_irq_restore(flags);
956 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
958 * If the new value of the ->dynticks_nesting counter was previously zero,
959 * we really have exited idle, and must do the appropriate accounting.
960 * The caller must have disabled interrupts.
962 static void rcu_eqs_exit_common(long long newval, int user)
964 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
966 rcu_dynticks_task_exit();
967 rcu_dynticks_eqs_exit();
968 rcu_cleanup_after_idle();
969 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, newval);
970 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
971 !user && !is_idle_task(current)) {
972 struct task_struct *idle __maybe_unused =
973 idle_task(smp_processor_id());
975 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
976 rdtp->dynticks_nesting, newval);
977 rcu_ftrace_dump(DUMP_ORIG);
978 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
979 current->pid, current->comm,
980 idle->pid, idle->comm); /* must be idle task! */
985 * Exit an RCU extended quiescent state, which can be either the
986 * idle loop or adaptive-tickless usermode execution.
988 static void rcu_eqs_exit(bool user)
990 struct rcu_dynticks *rdtp;
993 lockdep_assert_irqs_disabled();
994 rdtp = this_cpu_ptr(&rcu_dynticks);
995 oldval = rdtp->dynticks_nesting;
996 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
997 if (oldval & DYNTICK_TASK_NEST_MASK) {
998 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
1000 __this_cpu_inc(disable_rcu_irq_enter);
1001 rcu_eqs_exit_common(DYNTICK_TASK_EXIT_IDLE, user);
1002 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
1003 __this_cpu_dec(disable_rcu_irq_enter);
1004 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
1009 * rcu_idle_exit - inform RCU that current CPU is leaving idle
1011 * Exit idle mode, in other words, -enter- the mode in which RCU
1012 * read-side critical sections can occur.
1014 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
1015 * allow for the possibility of usermode upcalls messing up our count
1016 * of interrupt nesting level during the busy period that is just
1019 * If you add or remove a call to rcu_idle_exit(), be sure to test with
1020 * CONFIG_RCU_EQS_DEBUG=y.
1022 void rcu_idle_exit(void)
1024 unsigned long flags;
1026 local_irq_save(flags);
1027 rcu_eqs_exit(false);
1028 local_irq_restore(flags);
1031 #ifdef CONFIG_NO_HZ_FULL
1033 * rcu_user_exit - inform RCU that we are exiting userspace.
1035 * Exit RCU idle mode while entering the kernel because it can
1036 * run a RCU read side critical section anytime.
1038 * If you add or remove a call to rcu_user_exit(), be sure to test with
1039 * CONFIG_RCU_EQS_DEBUG=y.
1041 void rcu_user_exit(void)
1045 #endif /* CONFIG_NO_HZ_FULL */
1048 * rcu_nmi_enter - inform RCU of entry to NMI context
1050 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1051 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1052 * that the CPU is active. This implementation permits nested NMIs, as
1053 * long as the nesting level does not overflow an int. (You will probably
1054 * run out of stack space first.)
1056 * If you add or remove a call to rcu_nmi_enter(), be sure to test
1057 * with CONFIG_RCU_EQS_DEBUG=y.
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 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
1080 rdtp->dynticks_nmi_nesting + incby);
1085 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1087 * Enter an interrupt handler, which might possibly result in exiting
1088 * idle mode, in other words, entering the mode in which read-side critical
1089 * sections can occur. The caller must have disabled interrupts.
1091 * Note that the Linux kernel is fully capable of entering an interrupt
1092 * handler that it never exits, for example when doing upcalls to user mode!
1093 * This code assumes that the idle loop never does upcalls to user mode.
1094 * If your architecture's idle loop does do upcalls to user mode (or does
1095 * anything else that results in unbalanced calls to the irq_enter() and
1096 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1097 * But very infrequently and irreproducibly.
1099 * Use things like work queues to work around this limitation.
1101 * You have been warned.
1103 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1104 * CONFIG_RCU_EQS_DEBUG=y.
1106 void rcu_irq_enter(void)
1108 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1110 lockdep_assert_irqs_disabled();
1111 if (rdtp->dynticks_nmi_nesting == 0)
1112 rcu_dynticks_task_exit();
1114 if (rdtp->dynticks_nmi_nesting == 1)
1115 rcu_cleanup_after_idle();
1119 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1121 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1122 * with CONFIG_RCU_EQS_DEBUG=y.
1124 void rcu_irq_enter_irqson(void)
1126 unsigned long flags;
1128 local_irq_save(flags);
1130 local_irq_restore(flags);
1134 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1136 * Return true if RCU is watching the running CPU, which means that this
1137 * CPU can safely enter RCU read-side critical sections. In other words,
1138 * if the current CPU is in its idle loop and is neither in an interrupt
1139 * or NMI handler, return true.
1141 bool notrace rcu_is_watching(void)
1145 preempt_disable_notrace();
1146 ret = !rcu_dynticks_curr_cpu_in_eqs();
1147 preempt_enable_notrace();
1150 EXPORT_SYMBOL_GPL(rcu_is_watching);
1153 * If a holdout task is actually running, request an urgent quiescent
1154 * state from its CPU. This is unsynchronized, so migrations can cause
1155 * the request to go to the wrong CPU. Which is OK, all that will happen
1156 * is that the CPU's next context switch will be a bit slower and next
1157 * time around this task will generate another request.
1159 void rcu_request_urgent_qs_task(struct task_struct *t)
1166 return; /* This task is not running on that CPU. */
1167 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1170 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1173 * Is the current CPU online? Disable preemption to avoid false positives
1174 * that could otherwise happen due to the current CPU number being sampled,
1175 * this task being preempted, its old CPU being taken offline, resuming
1176 * on some other CPU, then determining that its old CPU is now offline.
1177 * It is OK to use RCU on an offline processor during initial boot, hence
1178 * the check for rcu_scheduler_fully_active. Note also that it is OK
1179 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1180 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1181 * offline to continue to use RCU for one jiffy after marking itself
1182 * offline in the cpu_online_mask. This leniency is necessary given the
1183 * non-atomic nature of the online and offline processing, for example,
1184 * the fact that a CPU enters the scheduler after completing the teardown
1187 * This is also why RCU internally marks CPUs online during in the
1188 * preparation phase and offline after the CPU has been taken down.
1190 * Disable checking if in an NMI handler because we cannot safely report
1191 * errors from NMI handlers anyway.
1193 bool rcu_lockdep_current_cpu_online(void)
1195 struct rcu_data *rdp;
1196 struct rcu_node *rnp;
1202 rdp = this_cpu_ptr(&rcu_sched_data);
1204 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1205 !rcu_scheduler_fully_active;
1209 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1211 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1214 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1216 * If the current CPU is idle or running at a first-level (not nested)
1217 * interrupt from idle, return true. The caller must have at least
1218 * disabled preemption.
1220 static int rcu_is_cpu_rrupt_from_idle(void)
1222 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1226 * We are reporting a quiescent state on behalf of some other CPU, so
1227 * it is our responsibility to check for and handle potential overflow
1228 * of the rcu_node ->gpnum counter with respect to the rcu_data counters.
1229 * After all, the CPU might be in deep idle state, and thus executing no
1232 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1234 lockdep_assert_held(&rnp->lock);
1235 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, rnp->gpnum))
1236 WRITE_ONCE(rdp->gpwrap, true);
1237 if (ULONG_CMP_LT(rdp->rcu_iw_gpnum + ULONG_MAX / 4, rnp->gpnum))
1238 rdp->rcu_iw_gpnum = rnp->gpnum + ULONG_MAX / 4;
1242 * Snapshot the specified CPU's dynticks counter so that we can later
1243 * credit them with an implicit quiescent state. Return 1 if this CPU
1244 * is in dynticks idle mode, which is an extended quiescent state.
1246 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1248 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1249 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1250 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1251 rcu_gpnum_ovf(rdp->mynode, rdp);
1258 * Handler for the irq_work request posted when a grace period has
1259 * gone on for too long, but not yet long enough for an RCU CPU
1260 * stall warning. Set state appropriately, but just complain if
1261 * there is unexpected state on entry.
1263 static void rcu_iw_handler(struct irq_work *iwp)
1265 struct rcu_data *rdp;
1266 struct rcu_node *rnp;
1268 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1270 raw_spin_lock_rcu_node(rnp);
1271 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1272 rdp->rcu_iw_gpnum = rnp->gpnum;
1273 rdp->rcu_iw_pending = false;
1275 raw_spin_unlock_rcu_node(rnp);
1279 * Return true if the specified CPU has passed through a quiescent
1280 * state by virtue of being in or having passed through an dynticks
1281 * idle state since the last call to dyntick_save_progress_counter()
1282 * for this same CPU, or by virtue of having been offline.
1284 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1289 struct rcu_node *rnp = rdp->mynode;
1292 * If the CPU passed through or entered a dynticks idle phase with
1293 * no active irq/NMI handlers, then we can safely pretend that the CPU
1294 * already acknowledged the request to pass through a quiescent
1295 * state. Either way, that CPU cannot possibly be in an RCU
1296 * read-side critical section that started before the beginning
1297 * of the current RCU grace period.
1299 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1300 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1301 rdp->dynticks_fqs++;
1302 rcu_gpnum_ovf(rnp, rdp);
1307 * Has this CPU encountered a cond_resched_rcu_qs() since the
1308 * beginning of the grace period? For this to be the case,
1309 * the CPU has to have noticed the current grace period. This
1310 * might not be the case for nohz_full CPUs looping in the kernel.
1312 jtsq = jiffies_till_sched_qs;
1313 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1314 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1315 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1316 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1317 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1318 rcu_gpnum_ovf(rnp, rdp);
1320 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1321 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1322 smp_store_release(ruqp, true);
1325 /* Check for the CPU being offline. */
1326 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1327 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1329 rcu_gpnum_ovf(rnp, rdp);
1334 * A CPU running for an extended time within the kernel can
1335 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1336 * even context-switching back and forth between a pair of
1337 * in-kernel CPU-bound tasks cannot advance grace periods.
1338 * So if the grace period is old enough, make the CPU pay attention.
1339 * Note that the unsynchronized assignments to the per-CPU
1340 * rcu_need_heavy_qs variable are safe. Yes, setting of
1341 * bits can be lost, but they will be set again on the next
1342 * force-quiescent-state pass. So lost bit sets do not result
1343 * in incorrect behavior, merely in a grace period lasting
1344 * a few jiffies longer than it might otherwise. Because
1345 * there are at most four threads involved, and because the
1346 * updates are only once every few jiffies, the probability of
1347 * lossage (and thus of slight grace-period extension) is
1350 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1351 if (!READ_ONCE(*rnhqp) &&
1352 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1353 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1354 WRITE_ONCE(*rnhqp, true);
1355 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1356 smp_store_release(ruqp, true);
1357 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1361 * If more than halfway to RCU CPU stall-warning time, do a
1362 * resched_cpu() to try to loosen things up a bit. Also check to
1363 * see if the CPU is getting hammered with interrupts, but only
1364 * once per grace period, just to keep the IPIs down to a dull roar.
1366 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1367 resched_cpu(rdp->cpu);
1368 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1369 !rdp->rcu_iw_pending && rdp->rcu_iw_gpnum != rnp->gpnum &&
1370 (rnp->ffmask & rdp->grpmask)) {
1371 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1372 rdp->rcu_iw_pending = true;
1373 rdp->rcu_iw_gpnum = rnp->gpnum;
1374 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1381 static void record_gp_stall_check_time(struct rcu_state *rsp)
1383 unsigned long j = jiffies;
1387 smp_wmb(); /* Record start time before stall time. */
1388 j1 = rcu_jiffies_till_stall_check();
1389 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1390 rsp->jiffies_resched = j + j1 / 2;
1391 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1395 * Convert a ->gp_state value to a character string.
1397 static const char *gp_state_getname(short gs)
1399 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1401 return gp_state_names[gs];
1405 * Complain about starvation of grace-period kthread.
1407 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1413 gpa = READ_ONCE(rsp->gp_activity);
1414 if (j - gpa > 2 * HZ) {
1415 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1417 rsp->gpnum, rsp->completed,
1419 gp_state_getname(rsp->gp_state), rsp->gp_state,
1420 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1421 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1422 if (rsp->gp_kthread) {
1423 sched_show_task(rsp->gp_kthread);
1424 wake_up_process(rsp->gp_kthread);
1430 * Dump stacks of all tasks running on stalled CPUs. First try using
1431 * NMIs, but fall back to manual remote stack tracing on architectures
1432 * that don't support NMI-based stack dumps. The NMI-triggered stack
1433 * traces are more accurate because they are printed by the target CPU.
1435 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1438 unsigned long flags;
1439 struct rcu_node *rnp;
1441 rcu_for_each_leaf_node(rsp, rnp) {
1442 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1443 for_each_leaf_node_possible_cpu(rnp, cpu)
1444 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1445 if (!trigger_single_cpu_backtrace(cpu))
1447 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1452 * If too much time has passed in the current grace period, and if
1453 * so configured, go kick the relevant kthreads.
1455 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1459 if (!rcu_kick_kthreads)
1461 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1462 if (time_after(jiffies, j) && rsp->gp_kthread &&
1463 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1464 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1465 rcu_ftrace_dump(DUMP_ALL);
1466 wake_up_process(rsp->gp_kthread);
1467 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1471 static inline void panic_on_rcu_stall(void)
1473 if (sysctl_panic_on_rcu_stall)
1474 panic("RCU Stall\n");
1477 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1481 unsigned long flags;
1485 struct rcu_node *rnp = rcu_get_root(rsp);
1488 /* Kick and suppress, if so configured. */
1489 rcu_stall_kick_kthreads(rsp);
1490 if (rcu_cpu_stall_suppress)
1493 /* Only let one CPU complain about others per time interval. */
1495 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1496 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1497 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1498 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1501 WRITE_ONCE(rsp->jiffies_stall,
1502 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1503 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1506 * OK, time to rat on our buddy...
1507 * See Documentation/RCU/stallwarn.txt for info on how to debug
1508 * RCU CPU stall warnings.
1510 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1512 print_cpu_stall_info_begin();
1513 rcu_for_each_leaf_node(rsp, rnp) {
1514 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1515 ndetected += rcu_print_task_stall(rnp);
1516 if (rnp->qsmask != 0) {
1517 for_each_leaf_node_possible_cpu(rnp, cpu)
1518 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1519 print_cpu_stall_info(rsp, cpu);
1523 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1526 print_cpu_stall_info_end();
1527 for_each_possible_cpu(cpu)
1528 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1530 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1531 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1532 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1534 rcu_dump_cpu_stacks(rsp);
1536 /* Complain about tasks blocking the grace period. */
1537 rcu_print_detail_task_stall(rsp);
1539 if (READ_ONCE(rsp->gpnum) != gpnum ||
1540 READ_ONCE(rsp->completed) == gpnum) {
1541 pr_err("INFO: Stall ended before state dump start\n");
1544 gpa = READ_ONCE(rsp->gp_activity);
1545 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1546 rsp->name, j - gpa, j, gpa,
1547 jiffies_till_next_fqs,
1548 rcu_get_root(rsp)->qsmask);
1549 /* In this case, the current CPU might be at fault. */
1550 sched_show_task(current);
1554 rcu_check_gp_kthread_starvation(rsp);
1556 panic_on_rcu_stall();
1558 force_quiescent_state(rsp); /* Kick them all. */
1561 static void print_cpu_stall(struct rcu_state *rsp)
1564 unsigned long flags;
1565 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1566 struct rcu_node *rnp = rcu_get_root(rsp);
1569 /* Kick and suppress, if so configured. */
1570 rcu_stall_kick_kthreads(rsp);
1571 if (rcu_cpu_stall_suppress)
1575 * OK, time to rat on ourselves...
1576 * See Documentation/RCU/stallwarn.txt for info on how to debug
1577 * RCU CPU stall warnings.
1579 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1580 print_cpu_stall_info_begin();
1581 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1582 print_cpu_stall_info(rsp, smp_processor_id());
1583 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1584 print_cpu_stall_info_end();
1585 for_each_possible_cpu(cpu)
1586 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1588 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1589 jiffies - rsp->gp_start,
1590 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1592 rcu_check_gp_kthread_starvation(rsp);
1594 rcu_dump_cpu_stacks(rsp);
1596 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1597 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1598 WRITE_ONCE(rsp->jiffies_stall,
1599 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1600 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1602 panic_on_rcu_stall();
1605 * Attempt to revive the RCU machinery by forcing a context switch.
1607 * A context switch would normally allow the RCU state machine to make
1608 * progress and it could be we're stuck in kernel space without context
1609 * switches for an entirely unreasonable amount of time.
1611 resched_cpu(smp_processor_id());
1614 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1616 unsigned long completed;
1617 unsigned long gpnum;
1621 struct rcu_node *rnp;
1623 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1624 !rcu_gp_in_progress(rsp))
1626 rcu_stall_kick_kthreads(rsp);
1630 * Lots of memory barriers to reject false positives.
1632 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1633 * then rsp->gp_start, and finally rsp->completed. These values
1634 * are updated in the opposite order with memory barriers (or
1635 * equivalent) during grace-period initialization and cleanup.
1636 * Now, a false positive can occur if we get an new value of
1637 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1638 * the memory barriers, the only way that this can happen is if one
1639 * grace period ends and another starts between these two fetches.
1640 * Detect this by comparing rsp->completed with the previous fetch
1643 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1644 * and rsp->gp_start suffice to forestall false positives.
1646 gpnum = READ_ONCE(rsp->gpnum);
1647 smp_rmb(); /* Pick up ->gpnum first... */
1648 js = READ_ONCE(rsp->jiffies_stall);
1649 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1650 gps = READ_ONCE(rsp->gp_start);
1651 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1652 completed = READ_ONCE(rsp->completed);
1653 if (ULONG_CMP_GE(completed, gpnum) ||
1654 ULONG_CMP_LT(j, js) ||
1655 ULONG_CMP_GE(gps, js))
1656 return; /* No stall or GP completed since entering function. */
1658 if (rcu_gp_in_progress(rsp) &&
1659 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1661 /* We haven't checked in, so go dump stack. */
1662 print_cpu_stall(rsp);
1664 } else if (rcu_gp_in_progress(rsp) &&
1665 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1667 /* They had a few time units to dump stack, so complain. */
1668 print_other_cpu_stall(rsp, gpnum);
1673 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1675 * Set the stall-warning timeout way off into the future, thus preventing
1676 * any RCU CPU stall-warning messages from appearing in the current set of
1677 * RCU grace periods.
1679 * The caller must disable hard irqs.
1681 void rcu_cpu_stall_reset(void)
1683 struct rcu_state *rsp;
1685 for_each_rcu_flavor(rsp)
1686 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1690 * Determine the value that ->completed will have at the end of the
1691 * next subsequent grace period. This is used to tag callbacks so that
1692 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1693 * been dyntick-idle for an extended period with callbacks under the
1694 * influence of RCU_FAST_NO_HZ.
1696 * The caller must hold rnp->lock with interrupts disabled.
1698 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1699 struct rcu_node *rnp)
1701 lockdep_assert_held(&rnp->lock);
1704 * If RCU is idle, we just wait for the next grace period.
1705 * But we can only be sure that RCU is idle if we are looking
1706 * at the root rcu_node structure -- otherwise, a new grace
1707 * period might have started, but just not yet gotten around
1708 * to initializing the current non-root rcu_node structure.
1710 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1711 return rnp->completed + 1;
1714 * Otherwise, wait for a possible partial grace period and
1715 * then the subsequent full grace period.
1717 return rnp->completed + 2;
1721 * Trace-event helper function for rcu_start_future_gp() and
1722 * rcu_nocb_wait_gp().
1724 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1725 unsigned long c, const char *s)
1727 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1728 rnp->completed, c, rnp->level,
1729 rnp->grplo, rnp->grphi, s);
1733 * Start some future grace period, as needed to handle newly arrived
1734 * callbacks. The required future grace periods are recorded in each
1735 * rcu_node structure's ->need_future_gp field. Returns true if there
1736 * is reason to awaken the grace-period kthread.
1738 * The caller must hold the specified rcu_node structure's ->lock.
1740 static bool __maybe_unused
1741 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1742 unsigned long *c_out)
1746 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1748 lockdep_assert_held(&rnp->lock);
1751 * Pick up grace-period number for new callbacks. If this
1752 * grace period is already marked as needed, return to the caller.
1754 c = rcu_cbs_completed(rdp->rsp, rnp);
1755 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1756 if (rnp->need_future_gp[c & 0x1]) {
1757 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1762 * If either this rcu_node structure or the root rcu_node structure
1763 * believe that a grace period is in progress, then we must wait
1764 * for the one following, which is in "c". Because our request
1765 * will be noticed at the end of the current grace period, we don't
1766 * need to explicitly start one. We only do the lockless check
1767 * of rnp_root's fields if the current rcu_node structure thinks
1768 * there is no grace period in flight, and because we hold rnp->lock,
1769 * the only possible change is when rnp_root's two fields are
1770 * equal, in which case rnp_root->gpnum might be concurrently
1771 * incremented. But that is OK, as it will just result in our
1772 * doing some extra useless work.
1774 if (rnp->gpnum != rnp->completed ||
1775 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1776 rnp->need_future_gp[c & 0x1]++;
1777 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1782 * There might be no grace period in progress. If we don't already
1783 * hold it, acquire the root rcu_node structure's lock in order to
1784 * start one (if needed).
1786 if (rnp != rnp_root)
1787 raw_spin_lock_rcu_node(rnp_root);
1790 * Get a new grace-period number. If there really is no grace
1791 * period in progress, it will be smaller than the one we obtained
1792 * earlier. Adjust callbacks as needed.
1794 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1795 if (!rcu_is_nocb_cpu(rdp->cpu))
1796 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1799 * If the needed for the required grace period is already
1800 * recorded, trace and leave.
1802 if (rnp_root->need_future_gp[c & 0x1]) {
1803 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1807 /* Record the need for the future grace period. */
1808 rnp_root->need_future_gp[c & 0x1]++;
1810 /* If a grace period is not already in progress, start one. */
1811 if (rnp_root->gpnum != rnp_root->completed) {
1812 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1814 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1815 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1818 if (rnp != rnp_root)
1819 raw_spin_unlock_rcu_node(rnp_root);
1827 * Clean up any old requests for the just-ended grace period. Also return
1828 * whether any additional grace periods have been requested.
1830 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1832 int c = rnp->completed;
1834 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1836 rnp->need_future_gp[c & 0x1] = 0;
1837 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1838 trace_rcu_future_gp(rnp, rdp, c,
1839 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1844 * Awaken the grace-period kthread for the specified flavor of RCU.
1845 * Don't do a self-awaken, and don't bother awakening when there is
1846 * nothing for the grace-period kthread to do (as in several CPUs
1847 * raced to awaken, and we lost), and finally don't try to awaken
1848 * a kthread that has not yet been created.
1850 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1852 if (current == rsp->gp_kthread ||
1853 !READ_ONCE(rsp->gp_flags) ||
1856 swake_up(&rsp->gp_wq);
1860 * If there is room, assign a ->completed number to any callbacks on
1861 * this CPU that have not already been assigned. Also accelerate any
1862 * callbacks that were previously assigned a ->completed number that has
1863 * since proven to be too conservative, which can happen if callbacks get
1864 * assigned a ->completed number while RCU is idle, but with reference to
1865 * a non-root rcu_node structure. This function is idempotent, so it does
1866 * not hurt to call it repeatedly. Returns an flag saying that we should
1867 * awaken the RCU grace-period kthread.
1869 * The caller must hold rnp->lock with interrupts disabled.
1871 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1872 struct rcu_data *rdp)
1876 lockdep_assert_held(&rnp->lock);
1878 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1879 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1883 * Callbacks are often registered with incomplete grace-period
1884 * information. Something about the fact that getting exact
1885 * information requires acquiring a global lock... RCU therefore
1886 * makes a conservative estimate of the grace period number at which
1887 * a given callback will become ready to invoke. The following
1888 * code checks this estimate and improves it when possible, thus
1889 * accelerating callback invocation to an earlier grace-period
1892 if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
1893 ret = rcu_start_future_gp(rnp, rdp, NULL);
1895 /* Trace depending on how much we were able to accelerate. */
1896 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1897 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1899 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1904 * Move any callbacks whose grace period has completed to the
1905 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1906 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1907 * sublist. This function is idempotent, so it does not hurt to
1908 * invoke it repeatedly. As long as it is not invoked -too- often...
1909 * Returns true if the RCU grace-period kthread needs to be awakened.
1911 * The caller must hold rnp->lock with interrupts disabled.
1913 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1914 struct rcu_data *rdp)
1916 lockdep_assert_held(&rnp->lock);
1918 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1919 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1923 * Find all callbacks whose ->completed numbers indicate that they
1924 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1926 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1928 /* Classify any remaining callbacks. */
1929 return rcu_accelerate_cbs(rsp, rnp, rdp);
1933 * Update CPU-local rcu_data state to record the beginnings and ends of
1934 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1935 * structure corresponding to the current CPU, and must have irqs disabled.
1936 * Returns true if the grace-period kthread needs to be awakened.
1938 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1939 struct rcu_data *rdp)
1944 lockdep_assert_held(&rnp->lock);
1946 /* Handle the ends of any preceding grace periods first. */
1947 if (rdp->completed == rnp->completed &&
1948 !unlikely(READ_ONCE(rdp->gpwrap))) {
1950 /* No grace period end, so just accelerate recent callbacks. */
1951 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1955 /* Advance callbacks. */
1956 ret = rcu_advance_cbs(rsp, rnp, rdp);
1958 /* Remember that we saw this grace-period completion. */
1959 rdp->completed = rnp->completed;
1960 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1963 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1965 * If the current grace period is waiting for this CPU,
1966 * set up to detect a quiescent state, otherwise don't
1967 * go looking for one.
1969 rdp->gpnum = rnp->gpnum;
1970 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1971 need_gp = !!(rnp->qsmask & rdp->grpmask);
1972 rdp->cpu_no_qs.b.norm = need_gp;
1973 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1974 rdp->core_needs_qs = need_gp;
1975 zero_cpu_stall_ticks(rdp);
1976 WRITE_ONCE(rdp->gpwrap, false);
1977 rcu_gpnum_ovf(rnp, rdp);
1982 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1984 unsigned long flags;
1986 struct rcu_node *rnp;
1988 local_irq_save(flags);
1990 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1991 rdp->completed == READ_ONCE(rnp->completed) &&
1992 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1993 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1994 local_irq_restore(flags);
1997 needwake = __note_gp_changes(rsp, rnp, rdp);
1998 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2000 rcu_gp_kthread_wake(rsp);
2003 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
2006 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
2007 schedule_timeout_uninterruptible(delay);
2011 * Initialize a new grace period. Return false if no grace period required.
2013 static bool rcu_gp_init(struct rcu_state *rsp)
2015 unsigned long oldmask;
2016 struct rcu_data *rdp;
2017 struct rcu_node *rnp = rcu_get_root(rsp);
2019 WRITE_ONCE(rsp->gp_activity, jiffies);
2020 raw_spin_lock_irq_rcu_node(rnp);
2021 if (!READ_ONCE(rsp->gp_flags)) {
2022 /* Spurious wakeup, tell caller to go back to sleep. */
2023 raw_spin_unlock_irq_rcu_node(rnp);
2026 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
2028 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
2030 * Grace period already in progress, don't start another.
2031 * Not supposed to be able to happen.
2033 raw_spin_unlock_irq_rcu_node(rnp);
2037 /* Advance to a new grace period and initialize state. */
2038 record_gp_stall_check_time(rsp);
2039 /* Record GP times before starting GP, hence smp_store_release(). */
2040 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
2041 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
2042 raw_spin_unlock_irq_rcu_node(rnp);
2045 * Apply per-leaf buffered online and offline operations to the
2046 * rcu_node tree. Note that this new grace period need not wait
2047 * for subsequent online CPUs, and that quiescent-state forcing
2048 * will handle subsequent offline CPUs.
2050 rcu_for_each_leaf_node(rsp, rnp) {
2051 rcu_gp_slow(rsp, gp_preinit_delay);
2052 raw_spin_lock_irq_rcu_node(rnp);
2053 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2054 !rnp->wait_blkd_tasks) {
2055 /* Nothing to do on this leaf rcu_node structure. */
2056 raw_spin_unlock_irq_rcu_node(rnp);
2060 /* Record old state, apply changes to ->qsmaskinit field. */
2061 oldmask = rnp->qsmaskinit;
2062 rnp->qsmaskinit = rnp->qsmaskinitnext;
2064 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2065 if (!oldmask != !rnp->qsmaskinit) {
2066 if (!oldmask) /* First online CPU for this rcu_node. */
2067 rcu_init_new_rnp(rnp);
2068 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2069 rnp->wait_blkd_tasks = true;
2070 else /* Last offline CPU and can propagate. */
2071 rcu_cleanup_dead_rnp(rnp);
2075 * If all waited-on tasks from prior grace period are
2076 * done, and if all this rcu_node structure's CPUs are
2077 * still offline, propagate up the rcu_node tree and
2078 * clear ->wait_blkd_tasks. Otherwise, if one of this
2079 * rcu_node structure's CPUs has since come back online,
2080 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2081 * checks for this, so just call it unconditionally).
2083 if (rnp->wait_blkd_tasks &&
2084 (!rcu_preempt_has_tasks(rnp) ||
2086 rnp->wait_blkd_tasks = false;
2087 rcu_cleanup_dead_rnp(rnp);
2090 raw_spin_unlock_irq_rcu_node(rnp);
2094 * Set the quiescent-state-needed bits in all the rcu_node
2095 * structures for all currently online CPUs in breadth-first order,
2096 * starting from the root rcu_node structure, relying on the layout
2097 * of the tree within the rsp->node[] array. Note that other CPUs
2098 * will access only the leaves of the hierarchy, thus seeing that no
2099 * grace period is in progress, at least until the corresponding
2100 * leaf node has been initialized.
2102 * The grace period cannot complete until the initialization
2103 * process finishes, because this kthread handles both.
2105 rcu_for_each_node_breadth_first(rsp, rnp) {
2106 rcu_gp_slow(rsp, gp_init_delay);
2107 raw_spin_lock_irq_rcu_node(rnp);
2108 rdp = this_cpu_ptr(rsp->rda);
2109 rcu_preempt_check_blocked_tasks(rnp);
2110 rnp->qsmask = rnp->qsmaskinit;
2111 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2112 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2113 WRITE_ONCE(rnp->completed, rsp->completed);
2114 if (rnp == rdp->mynode)
2115 (void)__note_gp_changes(rsp, rnp, rdp);
2116 rcu_preempt_boost_start_gp(rnp);
2117 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2118 rnp->level, rnp->grplo,
2119 rnp->grphi, rnp->qsmask);
2120 raw_spin_unlock_irq_rcu_node(rnp);
2121 cond_resched_rcu_qs();
2122 WRITE_ONCE(rsp->gp_activity, jiffies);
2129 * Helper function for swait_event_idle() wakeup at force-quiescent-state
2132 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2134 struct rcu_node *rnp = rcu_get_root(rsp);
2136 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2137 *gfp = READ_ONCE(rsp->gp_flags);
2138 if (*gfp & RCU_GP_FLAG_FQS)
2141 /* The current grace period has completed. */
2142 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2149 * Do one round of quiescent-state forcing.
2151 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2153 struct rcu_node *rnp = rcu_get_root(rsp);
2155 WRITE_ONCE(rsp->gp_activity, jiffies);
2158 /* Collect dyntick-idle snapshots. */
2159 force_qs_rnp(rsp, dyntick_save_progress_counter);
2161 /* Handle dyntick-idle and offline CPUs. */
2162 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2164 /* Clear flag to prevent immediate re-entry. */
2165 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2166 raw_spin_lock_irq_rcu_node(rnp);
2167 WRITE_ONCE(rsp->gp_flags,
2168 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2169 raw_spin_unlock_irq_rcu_node(rnp);
2174 * Clean up after the old grace period.
2176 static void rcu_gp_cleanup(struct rcu_state *rsp)
2178 unsigned long gp_duration;
2179 bool needgp = false;
2181 struct rcu_data *rdp;
2182 struct rcu_node *rnp = rcu_get_root(rsp);
2183 struct swait_queue_head *sq;
2185 WRITE_ONCE(rsp->gp_activity, jiffies);
2186 raw_spin_lock_irq_rcu_node(rnp);
2187 gp_duration = jiffies - rsp->gp_start;
2188 if (gp_duration > rsp->gp_max)
2189 rsp->gp_max = gp_duration;
2192 * We know the grace period is complete, but to everyone else
2193 * it appears to still be ongoing. But it is also the case
2194 * that to everyone else it looks like there is nothing that
2195 * they can do to advance the grace period. It is therefore
2196 * safe for us to drop the lock in order to mark the grace
2197 * period as completed in all of the rcu_node structures.
2199 raw_spin_unlock_irq_rcu_node(rnp);
2202 * Propagate new ->completed value to rcu_node structures so
2203 * that other CPUs don't have to wait until the start of the next
2204 * grace period to process their callbacks. This also avoids
2205 * some nasty RCU grace-period initialization races by forcing
2206 * the end of the current grace period to be completely recorded in
2207 * all of the rcu_node structures before the beginning of the next
2208 * grace period is recorded in any of the rcu_node structures.
2210 rcu_for_each_node_breadth_first(rsp, rnp) {
2211 raw_spin_lock_irq_rcu_node(rnp);
2212 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2213 WARN_ON_ONCE(rnp->qsmask);
2214 WRITE_ONCE(rnp->completed, rsp->gpnum);
2215 rdp = this_cpu_ptr(rsp->rda);
2216 if (rnp == rdp->mynode)
2217 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2218 /* smp_mb() provided by prior unlock-lock pair. */
2219 nocb += rcu_future_gp_cleanup(rsp, rnp);
2220 sq = rcu_nocb_gp_get(rnp);
2221 raw_spin_unlock_irq_rcu_node(rnp);
2222 rcu_nocb_gp_cleanup(sq);
2223 cond_resched_rcu_qs();
2224 WRITE_ONCE(rsp->gp_activity, jiffies);
2225 rcu_gp_slow(rsp, gp_cleanup_delay);
2227 rnp = rcu_get_root(rsp);
2228 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2229 rcu_nocb_gp_set(rnp, nocb);
2231 /* Declare grace period done. */
2232 WRITE_ONCE(rsp->completed, rsp->gpnum);
2233 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2234 rsp->gp_state = RCU_GP_IDLE;
2235 rdp = this_cpu_ptr(rsp->rda);
2236 /* Advance CBs to reduce false positives below. */
2237 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2238 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2239 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2240 trace_rcu_grace_period(rsp->name,
2241 READ_ONCE(rsp->gpnum),
2244 raw_spin_unlock_irq_rcu_node(rnp);
2248 * Body of kthread that handles grace periods.
2250 static int __noreturn rcu_gp_kthread(void *arg)
2256 struct rcu_state *rsp = arg;
2257 struct rcu_node *rnp = rcu_get_root(rsp);
2259 rcu_bind_gp_kthread();
2262 /* Handle grace-period start. */
2264 trace_rcu_grace_period(rsp->name,
2265 READ_ONCE(rsp->gpnum),
2267 rsp->gp_state = RCU_GP_WAIT_GPS;
2268 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2270 rsp->gp_state = RCU_GP_DONE_GPS;
2271 /* Locking provides needed memory barrier. */
2272 if (rcu_gp_init(rsp))
2274 cond_resched_rcu_qs();
2275 WRITE_ONCE(rsp->gp_activity, jiffies);
2276 WARN_ON(signal_pending(current));
2277 trace_rcu_grace_period(rsp->name,
2278 READ_ONCE(rsp->gpnum),
2282 /* Handle quiescent-state forcing. */
2283 first_gp_fqs = true;
2284 j = jiffies_till_first_fqs;
2287 jiffies_till_first_fqs = HZ;
2292 rsp->jiffies_force_qs = jiffies + j;
2293 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2296 trace_rcu_grace_period(rsp->name,
2297 READ_ONCE(rsp->gpnum),
2299 rsp->gp_state = RCU_GP_WAIT_FQS;
2300 ret = swait_event_idle_timeout(rsp->gp_wq,
2301 rcu_gp_fqs_check_wake(rsp, &gf), j);
2302 rsp->gp_state = RCU_GP_DOING_FQS;
2303 /* Locking provides needed memory barriers. */
2304 /* If grace period done, leave loop. */
2305 if (!READ_ONCE(rnp->qsmask) &&
2306 !rcu_preempt_blocked_readers_cgp(rnp))
2308 /* If time for quiescent-state forcing, do it. */
2309 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2310 (gf & RCU_GP_FLAG_FQS)) {
2311 trace_rcu_grace_period(rsp->name,
2312 READ_ONCE(rsp->gpnum),
2314 rcu_gp_fqs(rsp, first_gp_fqs);
2315 first_gp_fqs = false;
2316 trace_rcu_grace_period(rsp->name,
2317 READ_ONCE(rsp->gpnum),
2319 cond_resched_rcu_qs();
2320 WRITE_ONCE(rsp->gp_activity, jiffies);
2321 ret = 0; /* Force full wait till next FQS. */
2322 j = jiffies_till_next_fqs;
2325 jiffies_till_next_fqs = HZ;
2328 jiffies_till_next_fqs = 1;
2331 /* Deal with stray signal. */
2332 cond_resched_rcu_qs();
2333 WRITE_ONCE(rsp->gp_activity, jiffies);
2334 WARN_ON(signal_pending(current));
2335 trace_rcu_grace_period(rsp->name,
2336 READ_ONCE(rsp->gpnum),
2338 ret = 1; /* Keep old FQS timing. */
2340 if (time_after(jiffies, rsp->jiffies_force_qs))
2343 j = rsp->jiffies_force_qs - j;
2347 /* Handle grace-period end. */
2348 rsp->gp_state = RCU_GP_CLEANUP;
2349 rcu_gp_cleanup(rsp);
2350 rsp->gp_state = RCU_GP_CLEANED;
2355 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2356 * in preparation for detecting the next grace period. The caller must hold
2357 * the root node's ->lock and hard irqs must be disabled.
2359 * Note that it is legal for a dying CPU (which is marked as offline) to
2360 * invoke this function. This can happen when the dying CPU reports its
2363 * Returns true if the grace-period kthread must be awakened.
2366 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2367 struct rcu_data *rdp)
2369 lockdep_assert_held(&rnp->lock);
2370 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2372 * Either we have not yet spawned the grace-period
2373 * task, this CPU does not need another grace period,
2374 * or a grace period is already in progress.
2375 * Either way, don't start a new grace period.
2379 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2380 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2384 * We can't do wakeups while holding the rnp->lock, as that
2385 * could cause possible deadlocks with the rq->lock. Defer
2386 * the wakeup to our caller.
2392 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2393 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2394 * is invoked indirectly from rcu_advance_cbs(), which would result in
2395 * endless recursion -- or would do so if it wasn't for the self-deadlock
2396 * that is encountered beforehand.
2398 * Returns true if the grace-period kthread needs to be awakened.
2400 static bool rcu_start_gp(struct rcu_state *rsp)
2402 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2403 struct rcu_node *rnp = rcu_get_root(rsp);
2407 * If there is no grace period in progress right now, any
2408 * callbacks we have up to this point will be satisfied by the
2409 * next grace period. Also, advancing the callbacks reduces the
2410 * probability of false positives from cpu_needs_another_gp()
2411 * resulting in pointless grace periods. So, advance callbacks
2412 * then start the grace period!
2414 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2415 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2420 * Report a full set of quiescent states to the specified rcu_state data
2421 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2422 * kthread if another grace period is required. Whether we wake
2423 * the grace-period kthread or it awakens itself for the next round
2424 * of quiescent-state forcing, that kthread will clean up after the
2425 * just-completed grace period. Note that the caller must hold rnp->lock,
2426 * which is released before return.
2428 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2429 __releases(rcu_get_root(rsp)->lock)
2431 lockdep_assert_held(&rcu_get_root(rsp)->lock);
2432 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2433 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2434 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2435 rcu_gp_kthread_wake(rsp);
2439 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2440 * Allows quiescent states for a group of CPUs to be reported at one go
2441 * to the specified rcu_node structure, though all the CPUs in the group
2442 * must be represented by the same rcu_node structure (which need not be a
2443 * leaf rcu_node structure, though it often will be). The gps parameter
2444 * is the grace-period snapshot, which means that the quiescent states
2445 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2446 * must be held upon entry, and it is released before return.
2449 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2450 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2451 __releases(rnp->lock)
2453 unsigned long oldmask = 0;
2454 struct rcu_node *rnp_c;
2456 lockdep_assert_held(&rnp->lock);
2458 /* Walk up the rcu_node hierarchy. */
2460 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2463 * Our bit has already been cleared, or the
2464 * relevant grace period is already over, so done.
2466 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2469 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2470 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1 &&
2471 rcu_preempt_blocked_readers_cgp(rnp));
2472 rnp->qsmask &= ~mask;
2473 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2474 mask, rnp->qsmask, rnp->level,
2475 rnp->grplo, rnp->grphi,
2477 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2479 /* Other bits still set at this level, so done. */
2480 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2483 mask = rnp->grpmask;
2484 if (rnp->parent == NULL) {
2486 /* No more levels. Exit loop holding root lock. */
2490 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2493 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2494 oldmask = rnp_c->qsmask;
2498 * Get here if we are the last CPU to pass through a quiescent
2499 * state for this grace period. Invoke rcu_report_qs_rsp()
2500 * to clean up and start the next grace period if one is needed.
2502 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2506 * Record a quiescent state for all tasks that were previously queued
2507 * on the specified rcu_node structure and that were blocking the current
2508 * RCU grace period. The caller must hold the specified rnp->lock with
2509 * irqs disabled, and this lock is released upon return, but irqs remain
2512 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2513 struct rcu_node *rnp, unsigned long flags)
2514 __releases(rnp->lock)
2518 struct rcu_node *rnp_p;
2520 lockdep_assert_held(&rnp->lock);
2521 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2522 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2523 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2524 return; /* Still need more quiescent states! */
2527 rnp_p = rnp->parent;
2528 if (rnp_p == NULL) {
2530 * Only one rcu_node structure in the tree, so don't
2531 * try to report up to its nonexistent parent!
2533 rcu_report_qs_rsp(rsp, flags);
2537 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2539 mask = rnp->grpmask;
2540 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2541 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2542 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2546 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2547 * structure. This must be called from the specified CPU.
2550 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2552 unsigned long flags;
2555 struct rcu_node *rnp;
2558 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2559 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2560 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2563 * The grace period in which this quiescent state was
2564 * recorded has ended, so don't report it upwards.
2565 * We will instead need a new quiescent state that lies
2566 * within the current grace period.
2568 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2569 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2570 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2573 mask = rdp->grpmask;
2574 if ((rnp->qsmask & mask) == 0) {
2575 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2577 rdp->core_needs_qs = false;
2580 * This GP can't end until cpu checks in, so all of our
2581 * callbacks can be processed during the next GP.
2583 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2585 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2586 /* ^^^ Released rnp->lock */
2588 rcu_gp_kthread_wake(rsp);
2593 * Check to see if there is a new grace period of which this CPU
2594 * is not yet aware, and if so, set up local rcu_data state for it.
2595 * Otherwise, see if this CPU has just passed through its first
2596 * quiescent state for this grace period, and record that fact if so.
2599 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2601 /* Check for grace-period ends and beginnings. */
2602 note_gp_changes(rsp, rdp);
2605 * Does this CPU still need to do its part for current grace period?
2606 * If no, return and let the other CPUs do their part as well.
2608 if (!rdp->core_needs_qs)
2612 * Was there a quiescent state since the beginning of the grace
2613 * period? If no, then exit and wait for the next call.
2615 if (rdp->cpu_no_qs.b.norm)
2619 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2622 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2626 * Trace the fact that this CPU is going offline.
2628 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2630 RCU_TRACE(unsigned long mask;)
2631 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2632 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2634 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2637 RCU_TRACE(mask = rdp->grpmask;)
2638 trace_rcu_grace_period(rsp->name,
2639 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2644 * All CPUs for the specified rcu_node structure have gone offline,
2645 * and all tasks that were preempted within an RCU read-side critical
2646 * section while running on one of those CPUs have since exited their RCU
2647 * read-side critical section. Some other CPU is reporting this fact with
2648 * the specified rcu_node structure's ->lock held and interrupts disabled.
2649 * This function therefore goes up the tree of rcu_node structures,
2650 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2651 * the leaf rcu_node structure's ->qsmaskinit field has already been
2654 * This function does check that the specified rcu_node structure has
2655 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2656 * prematurely. That said, invoking it after the fact will cost you
2657 * a needless lock acquisition. So once it has done its work, don't
2660 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2663 struct rcu_node *rnp = rnp_leaf;
2665 lockdep_assert_held(&rnp->lock);
2666 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2667 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2670 mask = rnp->grpmask;
2674 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2675 rnp->qsmaskinit &= ~mask;
2676 rnp->qsmask &= ~mask;
2677 if (rnp->qsmaskinit) {
2678 raw_spin_unlock_rcu_node(rnp);
2679 /* irqs remain disabled. */
2682 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2687 * The CPU has been completely removed, and some other CPU is reporting
2688 * this fact from process context. Do the remainder of the cleanup.
2689 * There can only be one CPU hotplug operation at a time, so no need for
2692 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2694 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2695 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2697 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2700 /* Adjust any no-longer-needed kthreads. */
2701 rcu_boost_kthread_setaffinity(rnp, -1);
2705 * Invoke any RCU callbacks that have made it to the end of their grace
2706 * period. Thottle as specified by rdp->blimit.
2708 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2710 unsigned long flags;
2711 struct rcu_head *rhp;
2712 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2715 /* If no callbacks are ready, just return. */
2716 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2717 trace_rcu_batch_start(rsp->name,
2718 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2719 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2720 trace_rcu_batch_end(rsp->name, 0,
2721 !rcu_segcblist_empty(&rdp->cblist),
2722 need_resched(), is_idle_task(current),
2723 rcu_is_callbacks_kthread());
2728 * Extract the list of ready callbacks, disabling to prevent
2729 * races with call_rcu() from interrupt handlers. Leave the
2730 * callback counts, as rcu_barrier() needs to be conservative.
2732 local_irq_save(flags);
2733 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2735 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2736 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2737 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2738 local_irq_restore(flags);
2740 /* Invoke callbacks. */
2741 rhp = rcu_cblist_dequeue(&rcl);
2742 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2743 debug_rcu_head_unqueue(rhp);
2744 if (__rcu_reclaim(rsp->name, rhp))
2745 rcu_cblist_dequeued_lazy(&rcl);
2747 * Stop only if limit reached and CPU has something to do.
2748 * Note: The rcl structure counts down from zero.
2750 if (-rcl.len >= bl &&
2752 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2756 local_irq_save(flags);
2758 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2759 is_idle_task(current), rcu_is_callbacks_kthread());
2761 /* Update counts and requeue any remaining callbacks. */
2762 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2763 smp_mb(); /* List handling before counting for rcu_barrier(). */
2764 rdp->n_cbs_invoked += count;
2765 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2767 /* Reinstate batch limit if we have worked down the excess. */
2768 count = rcu_segcblist_n_cbs(&rdp->cblist);
2769 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2770 rdp->blimit = blimit;
2772 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2773 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2774 rdp->qlen_last_fqs_check = 0;
2775 rdp->n_force_qs_snap = rsp->n_force_qs;
2776 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2777 rdp->qlen_last_fqs_check = count;
2778 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2780 local_irq_restore(flags);
2782 /* Re-invoke RCU core processing if there are callbacks remaining. */
2783 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2788 * Check to see if this CPU is in a non-context-switch quiescent state
2789 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2790 * Also schedule RCU core processing.
2792 * This function must be called from hardirq context. It is normally
2793 * invoked from the scheduling-clock interrupt.
2795 void rcu_check_callbacks(int user)
2797 trace_rcu_utilization(TPS("Start scheduler-tick"));
2798 increment_cpu_stall_ticks();
2799 if (user || rcu_is_cpu_rrupt_from_idle()) {
2802 * Get here if this CPU took its interrupt from user
2803 * mode or from the idle loop, and if this is not a
2804 * nested interrupt. In this case, the CPU is in
2805 * a quiescent state, so note it.
2807 * No memory barrier is required here because both
2808 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2809 * variables that other CPUs neither access nor modify,
2810 * at least not while the corresponding CPU is online.
2816 } else if (!in_softirq()) {
2819 * Get here if this CPU did not take its interrupt from
2820 * softirq, in other words, if it is not interrupting
2821 * a rcu_bh read-side critical section. This is an _bh
2822 * critical section, so note it.
2827 rcu_preempt_check_callbacks();
2831 rcu_note_voluntary_context_switch(current);
2832 trace_rcu_utilization(TPS("End scheduler-tick"));
2836 * Scan the leaf rcu_node structures, processing dyntick state for any that
2837 * have not yet encountered a quiescent state, using the function specified.
2838 * Also initiate boosting for any threads blocked on the root rcu_node.
2840 * The caller must have suppressed start of new grace periods.
2842 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2845 unsigned long flags;
2847 struct rcu_node *rnp;
2849 rcu_for_each_leaf_node(rsp, rnp) {
2850 cond_resched_rcu_qs();
2852 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2853 if (rnp->qsmask == 0) {
2854 if (rcu_state_p == &rcu_sched_state ||
2855 rsp != rcu_state_p ||
2856 rcu_preempt_blocked_readers_cgp(rnp)) {
2858 * No point in scanning bits because they
2859 * are all zero. But we might need to
2860 * priority-boost blocked readers.
2862 rcu_initiate_boost(rnp, flags);
2863 /* rcu_initiate_boost() releases rnp->lock */
2867 (rnp->parent->qsmask & rnp->grpmask)) {
2869 * Race between grace-period
2870 * initialization and task exiting RCU
2871 * read-side critical section: Report.
2873 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2874 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2878 for_each_leaf_node_possible_cpu(rnp, cpu) {
2879 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2880 if ((rnp->qsmask & bit) != 0) {
2881 if (f(per_cpu_ptr(rsp->rda, cpu)))
2886 /* Idle/offline CPUs, report (releases rnp->lock. */
2887 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2889 /* Nothing to do here, so just drop the lock. */
2890 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2896 * Force quiescent states on reluctant CPUs, and also detect which
2897 * CPUs are in dyntick-idle mode.
2899 static void force_quiescent_state(struct rcu_state *rsp)
2901 unsigned long flags;
2903 struct rcu_node *rnp;
2904 struct rcu_node *rnp_old = NULL;
2906 /* Funnel through hierarchy to reduce memory contention. */
2907 rnp = __this_cpu_read(rsp->rda->mynode);
2908 for (; rnp != NULL; rnp = rnp->parent) {
2909 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2910 !raw_spin_trylock(&rnp->fqslock);
2911 if (rnp_old != NULL)
2912 raw_spin_unlock(&rnp_old->fqslock);
2914 rsp->n_force_qs_lh++;
2919 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2921 /* Reached the root of the rcu_node tree, acquire lock. */
2922 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2923 raw_spin_unlock(&rnp_old->fqslock);
2924 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2925 rsp->n_force_qs_lh++;
2926 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2927 return; /* Someone beat us to it. */
2929 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2930 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2931 rcu_gp_kthread_wake(rsp);
2935 * This does the RCU core processing work for the specified rcu_state
2936 * and rcu_data structures. This may be called only from the CPU to
2937 * whom the rdp belongs.
2940 __rcu_process_callbacks(struct rcu_state *rsp)
2942 unsigned long flags;
2944 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2946 WARN_ON_ONCE(!rdp->beenonline);
2948 /* Update RCU state based on any recent quiescent states. */
2949 rcu_check_quiescent_state(rsp, rdp);
2951 /* Does this CPU require a not-yet-started grace period? */
2952 local_irq_save(flags);
2953 if (cpu_needs_another_gp(rsp, rdp)) {
2954 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2955 needwake = rcu_start_gp(rsp);
2956 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2958 rcu_gp_kthread_wake(rsp);
2960 local_irq_restore(flags);
2963 /* If there are callbacks ready, invoke them. */
2964 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2965 invoke_rcu_callbacks(rsp, rdp);
2967 /* Do any needed deferred wakeups of rcuo kthreads. */
2968 do_nocb_deferred_wakeup(rdp);
2972 * Do RCU core processing for the current CPU.
2974 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2976 struct rcu_state *rsp;
2978 if (cpu_is_offline(smp_processor_id()))
2980 trace_rcu_utilization(TPS("Start RCU core"));
2981 for_each_rcu_flavor(rsp)
2982 __rcu_process_callbacks(rsp);
2983 trace_rcu_utilization(TPS("End RCU core"));
2987 * Schedule RCU callback invocation. If the specified type of RCU
2988 * does not support RCU priority boosting, just do a direct call,
2989 * otherwise wake up the per-CPU kernel kthread. Note that because we
2990 * are running on the current CPU with softirqs disabled, the
2991 * rcu_cpu_kthread_task cannot disappear out from under us.
2993 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2995 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2997 if (likely(!rsp->boost)) {
2998 rcu_do_batch(rsp, rdp);
3001 invoke_rcu_callbacks_kthread();
3004 static void invoke_rcu_core(void)
3006 if (cpu_online(smp_processor_id()))
3007 raise_softirq(RCU_SOFTIRQ);
3011 * Handle any core-RCU processing required by a call_rcu() invocation.
3013 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
3014 struct rcu_head *head, unsigned long flags)
3019 * If called from an extended quiescent state, invoke the RCU
3020 * core in order to force a re-evaluation of RCU's idleness.
3022 if (!rcu_is_watching())
3025 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3026 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3030 * Force the grace period if too many callbacks or too long waiting.
3031 * Enforce hysteresis, and don't invoke force_quiescent_state()
3032 * if some other CPU has recently done so. Also, don't bother
3033 * invoking force_quiescent_state() if the newly enqueued callback
3034 * is the only one waiting for a grace period to complete.
3036 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
3037 rdp->qlen_last_fqs_check + qhimark)) {
3039 /* Are we ignoring a completed grace period? */
3040 note_gp_changes(rsp, rdp);
3042 /* Start a new grace period if one not already started. */
3043 if (!rcu_gp_in_progress(rsp)) {
3044 struct rcu_node *rnp_root = rcu_get_root(rsp);
3046 raw_spin_lock_rcu_node(rnp_root);
3047 needwake = rcu_start_gp(rsp);
3048 raw_spin_unlock_rcu_node(rnp_root);
3050 rcu_gp_kthread_wake(rsp);
3052 /* Give the grace period a kick. */
3053 rdp->blimit = LONG_MAX;
3054 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3055 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
3056 force_quiescent_state(rsp);
3057 rdp->n_force_qs_snap = rsp->n_force_qs;
3058 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
3064 * RCU callback function to leak a callback.
3066 static void rcu_leak_callback(struct rcu_head *rhp)
3071 * Helper function for call_rcu() and friends. The cpu argument will
3072 * normally be -1, indicating "currently running CPU". It may specify
3073 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3074 * is expected to specify a CPU.
3077 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3078 struct rcu_state *rsp, int cpu, bool lazy)
3080 unsigned long flags;
3081 struct rcu_data *rdp;
3083 /* Misaligned rcu_head! */
3084 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3086 if (debug_rcu_head_queue(head)) {
3088 * Probable double call_rcu(), so leak the callback.
3089 * Use rcu:rcu_callback trace event to find the previous
3090 * time callback was passed to __call_rcu().
3092 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
3094 WRITE_ONCE(head->func, rcu_leak_callback);
3099 local_irq_save(flags);
3100 rdp = this_cpu_ptr(rsp->rda);
3102 /* Add the callback to our list. */
3103 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
3107 rdp = per_cpu_ptr(rsp->rda, cpu);
3108 if (likely(rdp->mynode)) {
3109 /* Post-boot, so this should be for a no-CBs CPU. */
3110 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3111 WARN_ON_ONCE(offline);
3112 /* Offline CPU, _call_rcu() illegal, leak callback. */
3113 local_irq_restore(flags);
3117 * Very early boot, before rcu_init(). Initialize if needed
3118 * and then drop through to queue the callback.
3121 WARN_ON_ONCE(!rcu_is_watching());
3122 if (rcu_segcblist_empty(&rdp->cblist))
3123 rcu_segcblist_init(&rdp->cblist);
3125 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3127 rcu_idle_count_callbacks_posted();
3129 if (__is_kfree_rcu_offset((unsigned long)func))
3130 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3131 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3132 rcu_segcblist_n_cbs(&rdp->cblist));
3134 trace_rcu_callback(rsp->name, head,
3135 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3136 rcu_segcblist_n_cbs(&rdp->cblist));
3138 /* Go handle any RCU core processing required. */
3139 __call_rcu_core(rsp, rdp, head, flags);
3140 local_irq_restore(flags);
3144 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3145 * @head: structure to be used for queueing the RCU updates.
3146 * @func: actual callback function to be invoked after the grace period
3148 * The callback function will be invoked some time after a full grace
3149 * period elapses, in other words after all currently executing RCU
3150 * read-side critical sections have completed. call_rcu_sched() assumes
3151 * that the read-side critical sections end on enabling of preemption
3152 * or on voluntary preemption.
3153 * RCU read-side critical sections are delimited by:
3155 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3156 * - anything that disables preemption.
3158 * These may be nested.
3160 * See the description of call_rcu() for more detailed information on
3161 * memory ordering guarantees.
3163 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3165 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3167 EXPORT_SYMBOL_GPL(call_rcu_sched);
3170 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3171 * @head: structure to be used for queueing the RCU updates.
3172 * @func: actual callback function to be invoked after the grace period
3174 * The callback function will be invoked some time after a full grace
3175 * period elapses, in other words after all currently executing RCU
3176 * read-side critical sections have completed. call_rcu_bh() assumes
3177 * that the read-side critical sections end on completion of a softirq
3178 * handler. This means that read-side critical sections in process
3179 * context must not be interrupted by softirqs. This interface is to be
3180 * used when most of the read-side critical sections are in softirq context.
3181 * RCU read-side critical sections are delimited by:
3183 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3184 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3186 * These may be nested.
3188 * See the description of call_rcu() for more detailed information on
3189 * memory ordering guarantees.
3191 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3193 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3195 EXPORT_SYMBOL_GPL(call_rcu_bh);
3198 * Queue an RCU callback for lazy invocation after a grace period.
3199 * This will likely be later named something like "call_rcu_lazy()",
3200 * but this change will require some way of tagging the lazy RCU
3201 * callbacks in the list of pending callbacks. Until then, this
3202 * function may only be called from __kfree_rcu().
3204 void kfree_call_rcu(struct rcu_head *head,
3205 rcu_callback_t func)
3207 __call_rcu(head, func, rcu_state_p, -1, 1);
3209 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3212 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3213 * any blocking grace-period wait automatically implies a grace period
3214 * if there is only one CPU online at any point time during execution
3215 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3216 * occasionally incorrectly indicate that there are multiple CPUs online
3217 * when there was in fact only one the whole time, as this just adds
3218 * some overhead: RCU still operates correctly.
3220 static inline int rcu_blocking_is_gp(void)
3224 might_sleep(); /* Check for RCU read-side critical section. */
3226 ret = num_online_cpus() <= 1;
3232 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3234 * Control will return to the caller some time after a full rcu-sched
3235 * grace period has elapsed, in other words after all currently executing
3236 * rcu-sched read-side critical sections have completed. These read-side
3237 * critical sections are delimited by rcu_read_lock_sched() and
3238 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3239 * local_irq_disable(), and so on may be used in place of
3240 * rcu_read_lock_sched().
3242 * This means that all preempt_disable code sequences, including NMI and
3243 * non-threaded hardware-interrupt handlers, in progress on entry will
3244 * have completed before this primitive returns. However, this does not
3245 * guarantee that softirq handlers will have completed, since in some
3246 * kernels, these handlers can run in process context, and can block.
3248 * Note that this guarantee implies further memory-ordering guarantees.
3249 * On systems with more than one CPU, when synchronize_sched() returns,
3250 * each CPU is guaranteed to have executed a full memory barrier since the
3251 * end of its last RCU-sched read-side critical section whose beginning
3252 * preceded the call to synchronize_sched(). In addition, each CPU having
3253 * an RCU read-side critical section that extends beyond the return from
3254 * synchronize_sched() is guaranteed to have executed a full memory barrier
3255 * after the beginning of synchronize_sched() and before the beginning of
3256 * that RCU read-side critical section. Note that these guarantees include
3257 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3258 * that are executing in the kernel.
3260 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3261 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3262 * to have executed a full memory barrier during the execution of
3263 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3264 * again only if the system has more than one CPU).
3266 void synchronize_sched(void)
3268 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3269 lock_is_held(&rcu_lock_map) ||
3270 lock_is_held(&rcu_sched_lock_map),
3271 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3272 if (rcu_blocking_is_gp())
3274 if (rcu_gp_is_expedited())
3275 synchronize_sched_expedited();
3277 wait_rcu_gp(call_rcu_sched);
3279 EXPORT_SYMBOL_GPL(synchronize_sched);
3282 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3284 * Control will return to the caller some time after a full rcu_bh grace
3285 * period has elapsed, in other words after all currently executing rcu_bh
3286 * read-side critical sections have completed. RCU read-side critical
3287 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3288 * and may be nested.
3290 * See the description of synchronize_sched() for more detailed information
3291 * on memory ordering guarantees.
3293 void synchronize_rcu_bh(void)
3295 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3296 lock_is_held(&rcu_lock_map) ||
3297 lock_is_held(&rcu_sched_lock_map),
3298 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3299 if (rcu_blocking_is_gp())
3301 if (rcu_gp_is_expedited())
3302 synchronize_rcu_bh_expedited();
3304 wait_rcu_gp(call_rcu_bh);
3306 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3309 * get_state_synchronize_rcu - Snapshot current RCU state
3311 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3312 * to determine whether or not a full grace period has elapsed in the
3315 unsigned long get_state_synchronize_rcu(void)
3318 * Any prior manipulation of RCU-protected data must happen
3319 * before the load from ->gpnum.
3324 * Make sure this load happens before the purportedly
3325 * time-consuming work between get_state_synchronize_rcu()
3326 * and cond_synchronize_rcu().
3328 return smp_load_acquire(&rcu_state_p->gpnum);
3330 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3333 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3335 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3337 * If a full RCU grace period has elapsed since the earlier call to
3338 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3339 * synchronize_rcu() to wait for a full grace period.
3341 * Yes, this function does not take counter wrap into account. But
3342 * counter wrap is harmless. If the counter wraps, we have waited for
3343 * more than 2 billion grace periods (and way more on a 64-bit system!),
3344 * so waiting for one additional grace period should be just fine.
3346 void cond_synchronize_rcu(unsigned long oldstate)
3348 unsigned long newstate;
3351 * Ensure that this load happens before any RCU-destructive
3352 * actions the caller might carry out after we return.
3354 newstate = smp_load_acquire(&rcu_state_p->completed);
3355 if (ULONG_CMP_GE(oldstate, newstate))
3358 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3361 * get_state_synchronize_sched - Snapshot current RCU-sched state
3363 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3364 * to determine whether or not a full grace period has elapsed in the
3367 unsigned long get_state_synchronize_sched(void)
3370 * Any prior manipulation of RCU-protected data must happen
3371 * before the load from ->gpnum.
3376 * Make sure this load happens before the purportedly
3377 * time-consuming work between get_state_synchronize_sched()
3378 * and cond_synchronize_sched().
3380 return smp_load_acquire(&rcu_sched_state.gpnum);
3382 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3385 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3387 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3389 * If a full RCU-sched grace period has elapsed since the earlier call to
3390 * get_state_synchronize_sched(), just return. Otherwise, invoke
3391 * synchronize_sched() to wait for a full grace period.
3393 * Yes, this function does not take counter wrap into account. But
3394 * counter wrap is harmless. If the counter wraps, we have waited for
3395 * more than 2 billion grace periods (and way more on a 64-bit system!),
3396 * so waiting for one additional grace period should be just fine.
3398 void cond_synchronize_sched(unsigned long oldstate)
3400 unsigned long newstate;
3403 * Ensure that this load happens before any RCU-destructive
3404 * actions the caller might carry out after we return.
3406 newstate = smp_load_acquire(&rcu_sched_state.completed);
3407 if (ULONG_CMP_GE(oldstate, newstate))
3408 synchronize_sched();
3410 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3413 * Check to see if there is any immediate RCU-related work to be done
3414 * by the current CPU, for the specified type of RCU, returning 1 if so.
3415 * The checks are in order of increasing expense: checks that can be
3416 * carried out against CPU-local state are performed first. However,
3417 * we must check for CPU stalls first, else we might not get a chance.
3419 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3421 struct rcu_node *rnp = rdp->mynode;
3423 rdp->n_rcu_pending++;
3425 /* Check for CPU stalls, if enabled. */
3426 check_cpu_stall(rsp, rdp);
3428 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3429 if (rcu_nohz_full_cpu(rsp))
3432 /* Is the RCU core waiting for a quiescent state from this CPU? */
3433 if (rcu_scheduler_fully_active &&
3434 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3435 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
3436 rdp->n_rp_core_needs_qs++;
3437 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3438 rdp->n_rp_report_qs++;
3442 /* Does this CPU have callbacks ready to invoke? */
3443 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
3444 rdp->n_rp_cb_ready++;
3448 /* Has RCU gone idle with this CPU needing another grace period? */
3449 if (cpu_needs_another_gp(rsp, rdp)) {
3450 rdp->n_rp_cpu_needs_gp++;
3454 /* Has another RCU grace period completed? */
3455 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3456 rdp->n_rp_gp_completed++;
3460 /* Has a new RCU grace period started? */
3461 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3462 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3463 rdp->n_rp_gp_started++;
3467 /* Does this CPU need a deferred NOCB wakeup? */
3468 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3469 rdp->n_rp_nocb_defer_wakeup++;
3474 rdp->n_rp_need_nothing++;
3479 * Check to see if there is any immediate RCU-related work to be done
3480 * by the current CPU, returning 1 if so. This function is part of the
3481 * RCU implementation; it is -not- an exported member of the RCU API.
3483 static int rcu_pending(void)
3485 struct rcu_state *rsp;
3487 for_each_rcu_flavor(rsp)
3488 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3494 * Return true if the specified CPU has any callback. If all_lazy is
3495 * non-NULL, store an indication of whether all callbacks are lazy.
3496 * (If there are no callbacks, all of them are deemed to be lazy.)
3498 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3502 struct rcu_data *rdp;
3503 struct rcu_state *rsp;
3505 for_each_rcu_flavor(rsp) {
3506 rdp = this_cpu_ptr(rsp->rda);
3507 if (rcu_segcblist_empty(&rdp->cblist))
3510 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3521 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3522 * the compiler is expected to optimize this away.
3524 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3525 int cpu, unsigned long done)
3527 trace_rcu_barrier(rsp->name, s, cpu,
3528 atomic_read(&rsp->barrier_cpu_count), done);
3532 * RCU callback function for _rcu_barrier(). If we are last, wake
3533 * up the task executing _rcu_barrier().
3535 static void rcu_barrier_callback(struct rcu_head *rhp)
3537 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3538 struct rcu_state *rsp = rdp->rsp;
3540 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3541 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3542 rsp->barrier_sequence);
3543 complete(&rsp->barrier_completion);
3545 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3550 * Called with preemption disabled, and from cross-cpu IRQ context.
3552 static void rcu_barrier_func(void *type)
3554 struct rcu_state *rsp = type;
3555 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3557 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3558 rdp->barrier_head.func = rcu_barrier_callback;
3559 debug_rcu_head_queue(&rdp->barrier_head);
3560 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3561 atomic_inc(&rsp->barrier_cpu_count);
3563 debug_rcu_head_unqueue(&rdp->barrier_head);
3564 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3565 rsp->barrier_sequence);
3570 * Orchestrate the specified type of RCU barrier, waiting for all
3571 * RCU callbacks of the specified type to complete.
3573 static void _rcu_barrier(struct rcu_state *rsp)
3576 struct rcu_data *rdp;
3577 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3579 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3581 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3582 mutex_lock(&rsp->barrier_mutex);
3584 /* Did someone else do our work for us? */
3585 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3586 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3587 rsp->barrier_sequence);
3588 smp_mb(); /* caller's subsequent code after above check. */
3589 mutex_unlock(&rsp->barrier_mutex);
3593 /* Mark the start of the barrier operation. */
3594 rcu_seq_start(&rsp->barrier_sequence);
3595 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3598 * Initialize the count to one rather than to zero in order to
3599 * avoid a too-soon return to zero in case of a short grace period
3600 * (or preemption of this task). Exclude CPU-hotplug operations
3601 * to ensure that no offline CPU has callbacks queued.
3603 init_completion(&rsp->barrier_completion);
3604 atomic_set(&rsp->barrier_cpu_count, 1);
3608 * Force each CPU with callbacks to register a new callback.
3609 * When that callback is invoked, we will know that all of the
3610 * corresponding CPU's preceding callbacks have been invoked.
3612 for_each_possible_cpu(cpu) {
3613 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3615 rdp = per_cpu_ptr(rsp->rda, cpu);
3616 if (rcu_is_nocb_cpu(cpu)) {
3617 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3618 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3619 rsp->barrier_sequence);
3621 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3622 rsp->barrier_sequence);
3623 smp_mb__before_atomic();
3624 atomic_inc(&rsp->barrier_cpu_count);
3625 __call_rcu(&rdp->barrier_head,
3626 rcu_barrier_callback, rsp, cpu, 0);
3628 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3629 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3630 rsp->barrier_sequence);
3631 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3633 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3634 rsp->barrier_sequence);
3640 * Now that we have an rcu_barrier_callback() callback on each
3641 * CPU, and thus each counted, remove the initial count.
3643 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3644 complete(&rsp->barrier_completion);
3646 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3647 wait_for_completion(&rsp->barrier_completion);
3649 /* Mark the end of the barrier operation. */
3650 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3651 rcu_seq_end(&rsp->barrier_sequence);
3653 /* Other rcu_barrier() invocations can now safely proceed. */
3654 mutex_unlock(&rsp->barrier_mutex);
3658 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3660 void rcu_barrier_bh(void)
3662 _rcu_barrier(&rcu_bh_state);
3664 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3667 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3669 void rcu_barrier_sched(void)
3671 _rcu_barrier(&rcu_sched_state);
3673 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3676 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3677 * first CPU in a given leaf rcu_node structure coming online. The caller
3678 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3681 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3684 struct rcu_node *rnp = rnp_leaf;
3686 lockdep_assert_held(&rnp->lock);
3688 mask = rnp->grpmask;
3692 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3693 rnp->qsmaskinit |= mask;
3694 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3699 * Do boot-time initialization of a CPU's per-CPU RCU data.
3702 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3704 unsigned long flags;
3705 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3706 struct rcu_node *rnp = rcu_get_root(rsp);
3708 /* Set up local state, ensuring consistent view of global state. */
3709 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3710 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3711 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3712 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3713 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3716 rcu_boot_init_nocb_percpu_data(rdp);
3717 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3721 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3722 * offline event can be happening at a given time. Note also that we
3723 * can accept some slop in the rsp->completed access due to the fact
3724 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3727 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3729 unsigned long flags;
3730 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3731 struct rcu_node *rnp = rcu_get_root(rsp);
3733 /* Set up local state, ensuring consistent view of global state. */
3734 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3735 rdp->qlen_last_fqs_check = 0;
3736 rdp->n_force_qs_snap = rsp->n_force_qs;
3737 rdp->blimit = blimit;
3738 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3739 !init_nocb_callback_list(rdp))
3740 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3741 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3742 rcu_dynticks_eqs_online();
3743 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3746 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3747 * propagation up the rcu_node tree will happen at the beginning
3748 * of the next grace period.
3751 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3752 rdp->beenonline = true; /* We have now been online. */
3753 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3754 rdp->completed = rnp->completed;
3755 rdp->cpu_no_qs.b.norm = true;
3756 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3757 rdp->core_needs_qs = false;
3758 rdp->rcu_iw_pending = false;
3759 rdp->rcu_iw_gpnum = rnp->gpnum - 1;
3760 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3761 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3765 * Invoked early in the CPU-online process, when pretty much all
3766 * services are available. The incoming CPU is not present.
3768 int rcutree_prepare_cpu(unsigned int cpu)
3770 struct rcu_state *rsp;
3772 for_each_rcu_flavor(rsp)
3773 rcu_init_percpu_data(cpu, rsp);
3775 rcu_prepare_kthreads(cpu);
3776 rcu_spawn_all_nocb_kthreads(cpu);
3782 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3784 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3786 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3788 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3792 * Near the end of the CPU-online process. Pretty much all services
3793 * enabled, and the CPU is now very much alive.
3795 int rcutree_online_cpu(unsigned int cpu)
3797 unsigned long flags;
3798 struct rcu_data *rdp;
3799 struct rcu_node *rnp;
3800 struct rcu_state *rsp;
3802 for_each_rcu_flavor(rsp) {
3803 rdp = per_cpu_ptr(rsp->rda, cpu);
3805 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3806 rnp->ffmask |= rdp->grpmask;
3807 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3809 if (IS_ENABLED(CONFIG_TREE_SRCU))
3810 srcu_online_cpu(cpu);
3811 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3812 return 0; /* Too early in boot for scheduler work. */
3813 sync_sched_exp_online_cleanup(cpu);
3814 rcutree_affinity_setting(cpu, -1);
3819 * Near the beginning of the process. The CPU is still very much alive
3820 * with pretty much all services enabled.
3822 int rcutree_offline_cpu(unsigned int cpu)
3824 unsigned long flags;
3825 struct rcu_data *rdp;
3826 struct rcu_node *rnp;
3827 struct rcu_state *rsp;
3829 for_each_rcu_flavor(rsp) {
3830 rdp = per_cpu_ptr(rsp->rda, cpu);
3832 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3833 rnp->ffmask &= ~rdp->grpmask;
3834 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3837 rcutree_affinity_setting(cpu, cpu);
3838 if (IS_ENABLED(CONFIG_TREE_SRCU))
3839 srcu_offline_cpu(cpu);
3844 * Near the end of the offline process. We do only tracing here.
3846 int rcutree_dying_cpu(unsigned int cpu)
3848 struct rcu_state *rsp;
3850 for_each_rcu_flavor(rsp)
3851 rcu_cleanup_dying_cpu(rsp);
3856 * The outgoing CPU is gone and we are running elsewhere.
3858 int rcutree_dead_cpu(unsigned int cpu)
3860 struct rcu_state *rsp;
3862 for_each_rcu_flavor(rsp) {
3863 rcu_cleanup_dead_cpu(cpu, rsp);
3864 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3870 * Mark the specified CPU as being online so that subsequent grace periods
3871 * (both expedited and normal) will wait on it. Note that this means that
3872 * incoming CPUs are not allowed to use RCU read-side critical sections
3873 * until this function is called. Failing to observe this restriction
3874 * will result in lockdep splats.
3876 * Note that this function is special in that it is invoked directly
3877 * from the incoming CPU rather than from the cpuhp_step mechanism.
3878 * This is because this function must be invoked at a precise location.
3880 void rcu_cpu_starting(unsigned int cpu)
3882 unsigned long flags;
3885 unsigned long oldmask;
3886 struct rcu_data *rdp;
3887 struct rcu_node *rnp;
3888 struct rcu_state *rsp;
3890 for_each_rcu_flavor(rsp) {
3891 rdp = per_cpu_ptr(rsp->rda, cpu);
3893 mask = rdp->grpmask;
3894 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3895 rnp->qsmaskinitnext |= mask;
3896 oldmask = rnp->expmaskinitnext;
3897 rnp->expmaskinitnext |= mask;
3898 oldmask ^= rnp->expmaskinitnext;
3899 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3900 /* Allow lockless access for expedited grace periods. */
3901 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3902 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3904 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3907 #ifdef CONFIG_HOTPLUG_CPU
3909 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3910 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3913 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3915 unsigned long flags;
3917 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3918 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3920 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3921 mask = rdp->grpmask;
3922 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3923 rnp->qsmaskinitnext &= ~mask;
3924 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3928 * The outgoing function has no further need of RCU, so remove it from
3929 * the list of CPUs that RCU must track.
3931 * Note that this function is special in that it is invoked directly
3932 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3933 * This is because this function must be invoked at a precise location.
3935 void rcu_report_dead(unsigned int cpu)
3937 struct rcu_state *rsp;
3939 /* QS for any half-done expedited RCU-sched GP. */
3941 rcu_report_exp_rdp(&rcu_sched_state,
3942 this_cpu_ptr(rcu_sched_state.rda), true);
3944 for_each_rcu_flavor(rsp)
3945 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3948 /* Migrate the dead CPU's callbacks to the current CPU. */
3949 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3951 unsigned long flags;
3952 struct rcu_data *my_rdp;
3953 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3954 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3956 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3957 return; /* No callbacks to migrate. */
3959 local_irq_save(flags);
3960 my_rdp = this_cpu_ptr(rsp->rda);
3961 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3962 local_irq_restore(flags);
3965 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3966 rcu_advance_cbs(rsp, rnp_root, rdp); /* Leverage recent GPs. */
3967 rcu_advance_cbs(rsp, rnp_root, my_rdp); /* Assign GP to pending CBs. */
3968 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3969 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3970 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3971 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3972 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3973 !rcu_segcblist_empty(&rdp->cblist),
3974 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3975 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3976 rcu_segcblist_first_cb(&rdp->cblist));
3980 * The outgoing CPU has just passed through the dying-idle state,
3981 * and we are being invoked from the CPU that was IPIed to continue the
3982 * offline operation. We need to migrate the outgoing CPU's callbacks.
3984 void rcutree_migrate_callbacks(int cpu)
3986 struct rcu_state *rsp;
3988 for_each_rcu_flavor(rsp)
3989 rcu_migrate_callbacks(cpu, rsp);
3994 * On non-huge systems, use expedited RCU grace periods to make suspend
3995 * and hibernation run faster.
3997 static int rcu_pm_notify(struct notifier_block *self,
3998 unsigned long action, void *hcpu)
4001 case PM_HIBERNATION_PREPARE:
4002 case PM_SUSPEND_PREPARE:
4003 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4006 case PM_POST_HIBERNATION:
4007 case PM_POST_SUSPEND:
4008 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4009 rcu_unexpedite_gp();
4018 * Spawn the kthreads that handle each RCU flavor's grace periods.
4020 static int __init rcu_spawn_gp_kthread(void)
4022 unsigned long flags;
4023 int kthread_prio_in = kthread_prio;
4024 struct rcu_node *rnp;
4025 struct rcu_state *rsp;
4026 struct sched_param sp;
4027 struct task_struct *t;
4029 /* Force priority into range. */
4030 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4032 else if (kthread_prio < 0)
4034 else if (kthread_prio > 99)
4036 if (kthread_prio != kthread_prio_in)
4037 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4038 kthread_prio, kthread_prio_in);
4040 rcu_scheduler_fully_active = 1;
4041 for_each_rcu_flavor(rsp) {
4042 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4044 rnp = rcu_get_root(rsp);
4045 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4046 rsp->gp_kthread = t;
4048 sp.sched_priority = kthread_prio;
4049 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4051 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4054 rcu_spawn_nocb_kthreads();
4055 rcu_spawn_boost_kthreads();
4058 early_initcall(rcu_spawn_gp_kthread);
4061 * This function is invoked towards the end of the scheduler's
4062 * initialization process. Before this is called, the idle task might
4063 * contain synchronous grace-period primitives (during which time, this idle
4064 * task is booting the system, and such primitives are no-ops). After this
4065 * function is called, any synchronous grace-period primitives are run as
4066 * expedited, with the requesting task driving the grace period forward.
4067 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4068 * runtime RCU functionality.
4070 void rcu_scheduler_starting(void)
4072 WARN_ON(num_online_cpus() != 1);
4073 WARN_ON(nr_context_switches() > 0);
4074 rcu_test_sync_prims();
4075 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4076 rcu_test_sync_prims();
4080 * Helper function for rcu_init() that initializes one rcu_state structure.
4082 static void __init rcu_init_one(struct rcu_state *rsp)
4084 static const char * const buf[] = RCU_NODE_NAME_INIT;
4085 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4086 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4087 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4089 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4093 struct rcu_node *rnp;
4095 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4097 /* Silence gcc 4.8 false positive about array index out of range. */
4098 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4099 panic("rcu_init_one: rcu_num_lvls out of range");
4101 /* Initialize the level-tracking arrays. */
4103 for (i = 1; i < rcu_num_lvls; i++)
4104 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
4105 rcu_init_levelspread(levelspread, num_rcu_lvl);
4107 /* Initialize the elements themselves, starting from the leaves. */
4109 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4110 cpustride *= levelspread[i];
4111 rnp = rsp->level[i];
4112 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4113 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4114 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4115 &rcu_node_class[i], buf[i]);
4116 raw_spin_lock_init(&rnp->fqslock);
4117 lockdep_set_class_and_name(&rnp->fqslock,
4118 &rcu_fqs_class[i], fqs[i]);
4119 rnp->gpnum = rsp->gpnum;
4120 rnp->completed = rsp->completed;
4122 rnp->qsmaskinit = 0;
4123 rnp->grplo = j * cpustride;
4124 rnp->grphi = (j + 1) * cpustride - 1;
4125 if (rnp->grphi >= nr_cpu_ids)
4126 rnp->grphi = nr_cpu_ids - 1;
4132 rnp->grpnum = j % levelspread[i - 1];
4133 rnp->grpmask = 1UL << rnp->grpnum;
4134 rnp->parent = rsp->level[i - 1] +
4135 j / levelspread[i - 1];
4138 INIT_LIST_HEAD(&rnp->blkd_tasks);
4139 rcu_init_one_nocb(rnp);
4140 init_waitqueue_head(&rnp->exp_wq[0]);
4141 init_waitqueue_head(&rnp->exp_wq[1]);
4142 init_waitqueue_head(&rnp->exp_wq[2]);
4143 init_waitqueue_head(&rnp->exp_wq[3]);
4144 spin_lock_init(&rnp->exp_lock);
4148 init_swait_queue_head(&rsp->gp_wq);
4149 init_swait_queue_head(&rsp->expedited_wq);
4150 rnp = rsp->level[rcu_num_lvls - 1];
4151 for_each_possible_cpu(i) {
4152 while (i > rnp->grphi)
4154 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4155 rcu_boot_init_percpu_data(i, rsp);
4157 list_add(&rsp->flavors, &rcu_struct_flavors);
4161 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4162 * replace the definitions in tree.h because those are needed to size
4163 * the ->node array in the rcu_state structure.
4165 static void __init rcu_init_geometry(void)
4169 int rcu_capacity[RCU_NUM_LVLS];
4172 * Initialize any unspecified boot parameters.
4173 * The default values of jiffies_till_first_fqs and
4174 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4175 * value, which is a function of HZ, then adding one for each
4176 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4178 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4179 if (jiffies_till_first_fqs == ULONG_MAX)
4180 jiffies_till_first_fqs = d;
4181 if (jiffies_till_next_fqs == ULONG_MAX)
4182 jiffies_till_next_fqs = d;
4184 /* If the compile-time values are accurate, just leave. */
4185 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4186 nr_cpu_ids == NR_CPUS)
4188 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4189 rcu_fanout_leaf, nr_cpu_ids);
4192 * The boot-time rcu_fanout_leaf parameter must be at least two
4193 * and cannot exceed the number of bits in the rcu_node masks.
4194 * Complain and fall back to the compile-time values if this
4195 * limit is exceeded.
4197 if (rcu_fanout_leaf < 2 ||
4198 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4199 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4205 * Compute number of nodes that can be handled an rcu_node tree
4206 * with the given number of levels.
4208 rcu_capacity[0] = rcu_fanout_leaf;
4209 for (i = 1; i < RCU_NUM_LVLS; i++)
4210 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4213 * The tree must be able to accommodate the configured number of CPUs.
4214 * If this limit is exceeded, fall back to the compile-time values.
4216 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4217 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4222 /* Calculate the number of levels in the tree. */
4223 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4225 rcu_num_lvls = i + 1;
4227 /* Calculate the number of rcu_nodes at each level of the tree. */
4228 for (i = 0; i < rcu_num_lvls; i++) {
4229 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4230 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4233 /* Calculate the total number of rcu_node structures. */
4235 for (i = 0; i < rcu_num_lvls; i++)
4236 rcu_num_nodes += num_rcu_lvl[i];
4240 * Dump out the structure of the rcu_node combining tree associated
4241 * with the rcu_state structure referenced by rsp.
4243 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4246 struct rcu_node *rnp;
4248 pr_info("rcu_node tree layout dump\n");
4250 rcu_for_each_node_breadth_first(rsp, rnp) {
4251 if (rnp->level != level) {
4256 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4261 void __init rcu_init(void)
4265 rcu_early_boot_tests();
4267 rcu_bootup_announce();
4268 rcu_init_geometry();
4269 rcu_init_one(&rcu_bh_state);
4270 rcu_init_one(&rcu_sched_state);
4272 rcu_dump_rcu_node_tree(&rcu_sched_state);
4273 __rcu_init_preempt();
4274 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4277 * We don't need protection against CPU-hotplug here because
4278 * this is called early in boot, before either interrupts
4279 * or the scheduler are operational.
4281 pm_notifier(rcu_pm_notify, 0);
4282 for_each_online_cpu(cpu) {
4283 rcutree_prepare_cpu(cpu);
4284 rcu_cpu_starting(cpu);
4285 rcutree_online_cpu(cpu);
4289 #include "tree_exp.h"
4290 #include "tree_plugin.h"