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 -
31 #define pr_fmt(fmt) "rcu: " fmt
33 #include <linux/types.h>
34 #include <linux/kernel.h>
35 #include <linux/init.h>
36 #include <linux/spinlock.h>
37 #include <linux/smp.h>
38 #include <linux/rcupdate_wait.h>
39 #include <linux/interrupt.h>
40 #include <linux/sched.h>
41 #include <linux/sched/debug.h>
42 #include <linux/nmi.h>
43 #include <linux/atomic.h>
44 #include <linux/bitops.h>
45 #include <linux/export.h>
46 #include <linux/completion.h>
47 #include <linux/moduleparam.h>
48 #include <linux/percpu.h>
49 #include <linux/notifier.h>
50 #include <linux/cpu.h>
51 #include <linux/mutex.h>
52 #include <linux/time.h>
53 #include <linux/kernel_stat.h>
54 #include <linux/wait.h>
55 #include <linux/kthread.h>
56 #include <uapi/linux/sched/types.h>
57 #include <linux/prefetch.h>
58 #include <linux/delay.h>
59 #include <linux/stop_machine.h>
60 #include <linux/random.h>
61 #include <linux/trace_events.h>
62 #include <linux/suspend.h>
63 #include <linux/ftrace.h>
64 #include <linux/tick.h>
69 #ifdef MODULE_PARAM_PREFIX
70 #undef MODULE_PARAM_PREFIX
72 #define MODULE_PARAM_PREFIX "rcutree."
74 /* Data structures. */
76 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data);
77 struct rcu_state rcu_state = {
78 .level = { &rcu_state.node[0] },
79 .gp_state = RCU_GP_IDLE,
80 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
81 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
84 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
85 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
86 .ofl_lock = __SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
89 /* Dump rcu_node combining tree at boot to verify correct setup. */
90 static bool dump_tree;
91 module_param(dump_tree, bool, 0444);
92 /* Control rcu_node-tree auto-balancing at boot time. */
93 static bool rcu_fanout_exact;
94 module_param(rcu_fanout_exact, bool, 0444);
95 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
96 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
97 module_param(rcu_fanout_leaf, int, 0444);
98 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
99 /* Number of rcu_nodes at specified level. */
100 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
101 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
102 /* panic() on RCU Stall sysctl. */
103 int sysctl_panic_on_rcu_stall __read_mostly;
106 * The rcu_scheduler_active variable is initialized to the value
107 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
108 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
109 * RCU can assume that there is but one task, allowing RCU to (for example)
110 * optimize synchronize_rcu() to a simple barrier(). When this variable
111 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
112 * to detect real grace periods. This variable is also used to suppress
113 * boot-time false positives from lockdep-RCU error checking. Finally, it
114 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
115 * is fully initialized, including all of its kthreads having been spawned.
117 int rcu_scheduler_active __read_mostly;
118 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
121 * The rcu_scheduler_fully_active variable transitions from zero to one
122 * during the early_initcall() processing, which is after the scheduler
123 * is capable of creating new tasks. So RCU processing (for example,
124 * creating tasks for RCU priority boosting) must be delayed until after
125 * rcu_scheduler_fully_active transitions from zero to one. We also
126 * currently delay invocation of any RCU callbacks until after this point.
128 * It might later prove better for people registering RCU callbacks during
129 * early boot to take responsibility for these callbacks, but one step at
132 static int rcu_scheduler_fully_active __read_mostly;
134 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
135 unsigned long gps, unsigned long flags);
136 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
137 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
138 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
139 static void invoke_rcu_core(void);
140 static void invoke_rcu_callbacks(struct rcu_data *rdp);
141 static void rcu_report_exp_rdp(struct rcu_data *rdp);
142 static void sync_sched_exp_online_cleanup(int cpu);
144 /* rcuc/rcub kthread realtime priority */
145 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
146 module_param(kthread_prio, int, 0644);
148 /* Delay in jiffies for grace-period initialization delays, debug only. */
150 static int gp_preinit_delay;
151 module_param(gp_preinit_delay, int, 0444);
152 static int gp_init_delay;
153 module_param(gp_init_delay, int, 0444);
154 static int gp_cleanup_delay;
155 module_param(gp_cleanup_delay, int, 0444);
157 /* Retrieve RCU kthreads priority for rcutorture */
158 int rcu_get_gp_kthreads_prio(void)
162 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
165 * Number of grace periods between delays, normalized by the duration of
166 * the delay. The longer the delay, the more the grace periods between
167 * each delay. The reason for this normalization is that it means that,
168 * for non-zero delays, the overall slowdown of grace periods is constant
169 * regardless of the duration of the delay. This arrangement balances
170 * the need for long delays to increase some race probabilities with the
171 * need for fast grace periods to increase other race probabilities.
173 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
176 * Compute the mask of online CPUs for the specified rcu_node structure.
177 * This will not be stable unless the rcu_node structure's ->lock is
178 * held, but the bit corresponding to the current CPU will be stable
181 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
183 return READ_ONCE(rnp->qsmaskinitnext);
187 * Return true if an RCU grace period is in progress. The READ_ONCE()s
188 * permit this function to be invoked without holding the root rcu_node
189 * structure's ->lock, but of course results can be subject to change.
191 static int rcu_gp_in_progress(void)
193 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
196 void rcu_softirq_qs(void)
199 rcu_preempt_deferred_qs(current);
203 * Steal a bit from the bottom of ->dynticks for idle entry/exit
204 * control. Initially this is for TLB flushing.
206 #define RCU_DYNTICK_CTRL_MASK 0x1
207 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
208 #ifndef rcu_eqs_special_exit
209 #define rcu_eqs_special_exit() do { } while (0)
212 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
213 .dynticks_nesting = 1,
214 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
215 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
219 * Record entry into an extended quiescent state. This is only to be
220 * called when not already in an extended quiescent state.
222 static void rcu_dynticks_eqs_enter(void)
224 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
228 * CPUs seeing atomic_add_return() must see prior RCU read-side
229 * critical sections, and we also must force ordering with the
232 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
233 /* Better be in an extended quiescent state! */
234 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
235 (seq & RCU_DYNTICK_CTRL_CTR));
236 /* Better not have special action (TLB flush) pending! */
237 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
238 (seq & RCU_DYNTICK_CTRL_MASK));
242 * Record exit from an extended quiescent state. This is only to be
243 * called from an extended quiescent state.
245 static void rcu_dynticks_eqs_exit(void)
247 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
251 * CPUs seeing atomic_add_return() must see prior idle sojourns,
252 * and we also must force ordering with the next RCU read-side
255 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
256 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
257 !(seq & RCU_DYNTICK_CTRL_CTR));
258 if (seq & RCU_DYNTICK_CTRL_MASK) {
259 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
260 smp_mb__after_atomic(); /* _exit after clearing mask. */
261 /* Prefer duplicate flushes to losing a flush. */
262 rcu_eqs_special_exit();
267 * Reset the current CPU's ->dynticks counter to indicate that the
268 * newly onlined CPU is no longer in an extended quiescent state.
269 * This will either leave the counter unchanged, or increment it
270 * to the next non-quiescent value.
272 * The non-atomic test/increment sequence works because the upper bits
273 * of the ->dynticks counter are manipulated only by the corresponding CPU,
274 * or when the corresponding CPU is offline.
276 static void rcu_dynticks_eqs_online(void)
278 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
280 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
282 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
286 * Is the current CPU in an extended quiescent state?
288 * No ordering, as we are sampling CPU-local information.
290 bool rcu_dynticks_curr_cpu_in_eqs(void)
292 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
294 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
298 * Snapshot the ->dynticks counter with full ordering so as to allow
299 * stable comparison of this counter with past and future snapshots.
301 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
303 int snap = atomic_add_return(0, &rdtp->dynticks);
305 return snap & ~RCU_DYNTICK_CTRL_MASK;
309 * Return true if the snapshot returned from rcu_dynticks_snap()
310 * indicates that RCU is in an extended quiescent state.
312 static bool rcu_dynticks_in_eqs(int snap)
314 return !(snap & RCU_DYNTICK_CTRL_CTR);
318 * Return true if the CPU corresponding to the specified rcu_dynticks
319 * structure has spent some time in an extended quiescent state since
320 * rcu_dynticks_snap() returned the specified snapshot.
322 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
324 return snap != rcu_dynticks_snap(rdtp);
328 * Set the special (bottom) bit of the specified CPU so that it
329 * will take special action (such as flushing its TLB) on the
330 * next exit from an extended quiescent state. Returns true if
331 * the bit was successfully set, or false if the CPU was not in
332 * an extended quiescent state.
334 bool rcu_eqs_special_set(int cpu)
338 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
341 old = atomic_read(&rdtp->dynticks);
342 if (old & RCU_DYNTICK_CTRL_CTR)
344 new = old | RCU_DYNTICK_CTRL_MASK;
345 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
350 * Let the RCU core know that this CPU has gone through the scheduler,
351 * which is a quiescent state. This is called when the need for a
352 * quiescent state is urgent, so we burn an atomic operation and full
353 * memory barriers to let the RCU core know about it, regardless of what
354 * this CPU might (or might not) do in the near future.
356 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
358 * The caller must have disabled interrupts and must not be idle.
360 static void __maybe_unused rcu_momentary_dyntick_idle(void)
362 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
365 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
366 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
367 /* It is illegal to call this from idle state. */
368 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
369 rcu_preempt_deferred_qs(current);
373 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
375 * If the current CPU is idle or running at a first-level (not nested)
376 * interrupt from idle, return true. The caller must have at least
377 * disabled preemption.
379 static int rcu_is_cpu_rrupt_from_idle(void)
381 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
382 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
385 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
386 static long blimit = DEFAULT_RCU_BLIMIT;
387 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
388 static long qhimark = DEFAULT_RCU_QHIMARK;
389 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
390 static long qlowmark = DEFAULT_RCU_QLOMARK;
392 module_param(blimit, long, 0444);
393 module_param(qhimark, long, 0444);
394 module_param(qlowmark, long, 0444);
396 static ulong jiffies_till_first_fqs = ULONG_MAX;
397 static ulong jiffies_till_next_fqs = ULONG_MAX;
398 static bool rcu_kick_kthreads;
401 * How long the grace period must be before we start recruiting
402 * quiescent-state help from rcu_note_context_switch().
404 static ulong jiffies_till_sched_qs = ULONG_MAX;
405 module_param(jiffies_till_sched_qs, ulong, 0444);
406 static ulong jiffies_to_sched_qs; /* Adjusted version of above if not default */
407 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
410 * Make sure that we give the grace-period kthread time to detect any
411 * idle CPUs before taking active measures to force quiescent states.
412 * However, don't go below 100 milliseconds, adjusted upwards for really
415 static void adjust_jiffies_till_sched_qs(void)
419 /* If jiffies_till_sched_qs was specified, respect the request. */
420 if (jiffies_till_sched_qs != ULONG_MAX) {
421 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
424 j = READ_ONCE(jiffies_till_first_fqs) +
425 2 * READ_ONCE(jiffies_till_next_fqs);
426 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
427 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
428 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
429 WRITE_ONCE(jiffies_to_sched_qs, j);
432 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
435 int ret = kstrtoul(val, 0, &j);
438 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
439 adjust_jiffies_till_sched_qs();
444 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
447 int ret = kstrtoul(val, 0, &j);
450 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
451 adjust_jiffies_till_sched_qs();
456 static struct kernel_param_ops first_fqs_jiffies_ops = {
457 .set = param_set_first_fqs_jiffies,
458 .get = param_get_ulong,
461 static struct kernel_param_ops next_fqs_jiffies_ops = {
462 .set = param_set_next_fqs_jiffies,
463 .get = param_get_ulong,
466 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
467 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
468 module_param(rcu_kick_kthreads, bool, 0644);
470 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
471 static void force_quiescent_state(void);
472 static int rcu_pending(void);
475 * Return the number of RCU GPs completed thus far for debug & stats.
477 unsigned long rcu_get_gp_seq(void)
479 return READ_ONCE(rcu_state.gp_seq);
481 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
484 * Return the number of RCU expedited batches completed thus far for
485 * debug & stats. Odd numbers mean that a batch is in progress, even
486 * numbers mean idle. The value returned will thus be roughly double
487 * the cumulative batches since boot.
489 unsigned long rcu_exp_batches_completed(void)
491 return rcu_state.expedited_sequence;
493 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
496 * Force a quiescent state.
498 void rcu_force_quiescent_state(void)
500 force_quiescent_state();
502 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
505 * Show the state of the grace-period kthreads.
507 void show_rcu_gp_kthreads(void)
510 struct rcu_data *rdp;
511 struct rcu_node *rnp;
513 pr_info("%s: wait state: %d ->state: %#lx\n", rcu_state.name,
514 rcu_state.gp_state, rcu_state.gp_kthread->state);
515 rcu_for_each_node_breadth_first(rnp) {
516 if (ULONG_CMP_GE(rcu_state.gp_seq, rnp->gp_seq_needed))
518 pr_info("\trcu_node %d:%d ->gp_seq %lu ->gp_seq_needed %lu\n",
519 rnp->grplo, rnp->grphi, rnp->gp_seq,
521 if (!rcu_is_leaf_node(rnp))
523 for_each_leaf_node_possible_cpu(rnp, cpu) {
524 rdp = per_cpu_ptr(&rcu_data, cpu);
526 ULONG_CMP_GE(rcu_state.gp_seq,
529 pr_info("\tcpu %d ->gp_seq_needed %lu\n",
530 cpu, rdp->gp_seq_needed);
533 /* sched_show_task(rcu_state.gp_kthread); */
535 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
538 * Send along grace-period-related data for rcutorture diagnostics.
540 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
541 unsigned long *gp_seq)
546 case RCU_SCHED_FLAVOR:
547 *flags = READ_ONCE(rcu_state.gp_flags);
548 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
554 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
557 * Return the root node of the rcu_state structure.
559 static struct rcu_node *rcu_get_root(void)
561 return &rcu_state.node[0];
565 * Enter an RCU extended quiescent state, which can be either the
566 * idle loop or adaptive-tickless usermode execution.
568 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
569 * the possibility of usermode upcalls having messed up our count
570 * of interrupt nesting level during the prior busy period.
572 static void rcu_eqs_enter(bool user)
574 struct rcu_data *rdp;
575 struct rcu_dynticks *rdtp;
577 rdtp = this_cpu_ptr(&rcu_dynticks);
578 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
579 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
580 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
581 rdtp->dynticks_nesting == 0);
582 if (rdtp->dynticks_nesting != 1) {
583 rdtp->dynticks_nesting--;
587 lockdep_assert_irqs_disabled();
588 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
589 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
590 rdp = this_cpu_ptr(&rcu_data);
591 do_nocb_deferred_wakeup(rdp);
592 rcu_prepare_for_idle();
593 rcu_preempt_deferred_qs(current);
594 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
595 rcu_dynticks_eqs_enter();
596 rcu_dynticks_task_enter();
600 * rcu_idle_enter - inform RCU that current CPU is entering idle
602 * Enter idle mode, in other words, -leave- the mode in which RCU
603 * read-side critical sections can occur. (Though RCU read-side
604 * critical sections can occur in irq handlers in idle, a possibility
605 * handled by irq_enter() and irq_exit().)
607 * If you add or remove a call to rcu_idle_enter(), be sure to test with
608 * CONFIG_RCU_EQS_DEBUG=y.
610 void rcu_idle_enter(void)
612 lockdep_assert_irqs_disabled();
613 rcu_eqs_enter(false);
616 #ifdef CONFIG_NO_HZ_FULL
618 * rcu_user_enter - inform RCU that we are resuming userspace.
620 * Enter RCU idle mode right before resuming userspace. No use of RCU
621 * is permitted between this call and rcu_user_exit(). This way the
622 * CPU doesn't need to maintain the tick for RCU maintenance purposes
623 * when the CPU runs in userspace.
625 * If you add or remove a call to rcu_user_enter(), be sure to test with
626 * CONFIG_RCU_EQS_DEBUG=y.
628 void rcu_user_enter(void)
630 lockdep_assert_irqs_disabled();
633 #endif /* CONFIG_NO_HZ_FULL */
636 * If we are returning from the outermost NMI handler that interrupted an
637 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
638 * to let the RCU grace-period handling know that the CPU is back to
641 * If you add or remove a call to rcu_nmi_exit_common(), be sure to test
642 * with CONFIG_RCU_EQS_DEBUG=y.
644 static __always_inline void rcu_nmi_exit_common(bool irq)
646 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
649 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
650 * (We are exiting an NMI handler, so RCU better be paying attention
653 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
654 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
657 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
658 * leave it in non-RCU-idle state.
660 if (rdtp->dynticks_nmi_nesting != 1) {
661 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
662 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
663 rdtp->dynticks_nmi_nesting - 2);
667 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
668 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
669 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
672 rcu_prepare_for_idle();
674 rcu_dynticks_eqs_enter();
677 rcu_dynticks_task_enter();
681 * rcu_nmi_exit - inform RCU of exit from NMI context
682 * @irq: Is this call from rcu_irq_exit?
684 * If you add or remove a call to rcu_nmi_exit(), be sure to test
685 * with CONFIG_RCU_EQS_DEBUG=y.
687 void rcu_nmi_exit(void)
689 rcu_nmi_exit_common(false);
693 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
695 * Exit from an interrupt handler, which might possibly result in entering
696 * idle mode, in other words, leaving the mode in which read-side critical
697 * sections can occur. The caller must have disabled interrupts.
699 * This code assumes that the idle loop never does anything that might
700 * result in unbalanced calls to irq_enter() and irq_exit(). If your
701 * architecture's idle loop violates this assumption, RCU will give you what
702 * you deserve, good and hard. But very infrequently and irreproducibly.
704 * Use things like work queues to work around this limitation.
706 * You have been warned.
708 * If you add or remove a call to rcu_irq_exit(), be sure to test with
709 * CONFIG_RCU_EQS_DEBUG=y.
711 void rcu_irq_exit(void)
713 lockdep_assert_irqs_disabled();
714 rcu_nmi_exit_common(true);
718 * Wrapper for rcu_irq_exit() where interrupts are enabled.
720 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
721 * with CONFIG_RCU_EQS_DEBUG=y.
723 void rcu_irq_exit_irqson(void)
727 local_irq_save(flags);
729 local_irq_restore(flags);
733 * Exit an RCU extended quiescent state, which can be either the
734 * idle loop or adaptive-tickless usermode execution.
736 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
737 * allow for the possibility of usermode upcalls messing up our count of
738 * interrupt nesting level during the busy period that is just now starting.
740 static void rcu_eqs_exit(bool user)
742 struct rcu_dynticks *rdtp;
745 lockdep_assert_irqs_disabled();
746 rdtp = this_cpu_ptr(&rcu_dynticks);
747 oldval = rdtp->dynticks_nesting;
748 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
750 rdtp->dynticks_nesting++;
753 rcu_dynticks_task_exit();
754 rcu_dynticks_eqs_exit();
755 rcu_cleanup_after_idle();
756 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, 1, rdtp->dynticks);
757 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
758 WRITE_ONCE(rdtp->dynticks_nesting, 1);
759 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting);
760 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
764 * rcu_idle_exit - inform RCU that current CPU is leaving idle
766 * Exit idle mode, in other words, -enter- the mode in which RCU
767 * read-side critical sections can occur.
769 * If you add or remove a call to rcu_idle_exit(), be sure to test with
770 * CONFIG_RCU_EQS_DEBUG=y.
772 void rcu_idle_exit(void)
776 local_irq_save(flags);
778 local_irq_restore(flags);
781 #ifdef CONFIG_NO_HZ_FULL
783 * rcu_user_exit - inform RCU that we are exiting userspace.
785 * Exit RCU idle mode while entering the kernel because it can
786 * run a RCU read side critical section anytime.
788 * If you add or remove a call to rcu_user_exit(), be sure to test with
789 * CONFIG_RCU_EQS_DEBUG=y.
791 void rcu_user_exit(void)
795 #endif /* CONFIG_NO_HZ_FULL */
798 * rcu_nmi_enter_common - inform RCU of entry to NMI context
799 * @irq: Is this call from rcu_irq_enter?
801 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
802 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
803 * that the CPU is active. This implementation permits nested NMIs, as
804 * long as the nesting level does not overflow an int. (You will probably
805 * run out of stack space first.)
807 * If you add or remove a call to rcu_nmi_enter_common(), be sure to test
808 * with CONFIG_RCU_EQS_DEBUG=y.
810 static __always_inline void rcu_nmi_enter_common(bool irq)
812 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
815 /* Complain about underflow. */
816 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
819 * If idle from RCU viewpoint, atomically increment ->dynticks
820 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
821 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
822 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
823 * to be in the outermost NMI handler that interrupted an RCU-idle
824 * period (observation due to Andy Lutomirski).
826 if (rcu_dynticks_curr_cpu_in_eqs()) {
829 rcu_dynticks_task_exit();
831 rcu_dynticks_eqs_exit();
834 rcu_cleanup_after_idle();
838 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
839 rdtp->dynticks_nmi_nesting,
840 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
841 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
842 rdtp->dynticks_nmi_nesting + incby);
847 * rcu_nmi_enter - inform RCU of entry to NMI context
849 void rcu_nmi_enter(void)
851 rcu_nmi_enter_common(false);
855 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
857 * Enter an interrupt handler, which might possibly result in exiting
858 * idle mode, in other words, entering the mode in which read-side critical
859 * sections can occur. The caller must have disabled interrupts.
861 * Note that the Linux kernel is fully capable of entering an interrupt
862 * handler that it never exits, for example when doing upcalls to user mode!
863 * This code assumes that the idle loop never does upcalls to user mode.
864 * If your architecture's idle loop does do upcalls to user mode (or does
865 * anything else that results in unbalanced calls to the irq_enter() and
866 * irq_exit() functions), RCU will give you what you deserve, good and hard.
867 * But very infrequently and irreproducibly.
869 * Use things like work queues to work around this limitation.
871 * You have been warned.
873 * If you add or remove a call to rcu_irq_enter(), be sure to test with
874 * CONFIG_RCU_EQS_DEBUG=y.
876 void rcu_irq_enter(void)
878 lockdep_assert_irqs_disabled();
879 rcu_nmi_enter_common(true);
883 * Wrapper for rcu_irq_enter() where interrupts are enabled.
885 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
886 * with CONFIG_RCU_EQS_DEBUG=y.
888 void rcu_irq_enter_irqson(void)
892 local_irq_save(flags);
894 local_irq_restore(flags);
898 * rcu_is_watching - see if RCU thinks that the current CPU is idle
900 * Return true if RCU is watching the running CPU, which means that this
901 * CPU can safely enter RCU read-side critical sections. In other words,
902 * if the current CPU is in its idle loop and is neither in an interrupt
903 * or NMI handler, return true.
905 bool notrace rcu_is_watching(void)
909 preempt_disable_notrace();
910 ret = !rcu_dynticks_curr_cpu_in_eqs();
911 preempt_enable_notrace();
914 EXPORT_SYMBOL_GPL(rcu_is_watching);
917 * If a holdout task is actually running, request an urgent quiescent
918 * state from its CPU. This is unsynchronized, so migrations can cause
919 * the request to go to the wrong CPU. Which is OK, all that will happen
920 * is that the CPU's next context switch will be a bit slower and next
921 * time around this task will generate another request.
923 void rcu_request_urgent_qs_task(struct task_struct *t)
930 return; /* This task is not running on that CPU. */
931 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
934 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
937 * Is the current CPU online as far as RCU is concerned?
939 * Disable preemption to avoid false positives that could otherwise
940 * happen due to the current CPU number being sampled, this task being
941 * preempted, its old CPU being taken offline, resuming on some other CPU,
942 * then determining that its old CPU is now offline.
944 * Disable checking if in an NMI handler because we cannot safely
945 * report errors from NMI handlers anyway. In addition, it is OK to use
946 * RCU on an offline processor during initial boot, hence the check for
947 * rcu_scheduler_fully_active.
949 bool rcu_lockdep_current_cpu_online(void)
951 struct rcu_data *rdp;
952 struct rcu_node *rnp;
955 if (in_nmi() || !rcu_scheduler_fully_active)
958 rdp = this_cpu_ptr(&rcu_data);
960 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
965 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
967 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
970 * We are reporting a quiescent state on behalf of some other CPU, so
971 * it is our responsibility to check for and handle potential overflow
972 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
973 * After all, the CPU might be in deep idle state, and thus executing no
976 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
978 raw_lockdep_assert_held_rcu_node(rnp);
979 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
981 WRITE_ONCE(rdp->gpwrap, true);
982 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
983 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
987 * Snapshot the specified CPU's dynticks counter so that we can later
988 * credit them with an implicit quiescent state. Return 1 if this CPU
989 * is in dynticks idle mode, which is an extended quiescent state.
991 static int dyntick_save_progress_counter(struct rcu_data *rdp)
993 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
994 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
995 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
996 rcu_gpnum_ovf(rdp->mynode, rdp);
1003 * Handler for the irq_work request posted when a grace period has
1004 * gone on for too long, but not yet long enough for an RCU CPU
1005 * stall warning. Set state appropriately, but just complain if
1006 * there is unexpected state on entry.
1008 static void rcu_iw_handler(struct irq_work *iwp)
1010 struct rcu_data *rdp;
1011 struct rcu_node *rnp;
1013 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1015 raw_spin_lock_rcu_node(rnp);
1016 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1017 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1018 rdp->rcu_iw_pending = false;
1020 raw_spin_unlock_rcu_node(rnp);
1024 * Return true if the specified CPU has passed through a quiescent
1025 * state by virtue of being in or having passed through an dynticks
1026 * idle state since the last call to dyntick_save_progress_counter()
1027 * for this same CPU, or by virtue of having been offline.
1029 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1034 struct rcu_node *rnp = rdp->mynode;
1037 * If the CPU passed through or entered a dynticks idle phase with
1038 * no active irq/NMI handlers, then we can safely pretend that the CPU
1039 * already acknowledged the request to pass through a quiescent
1040 * state. Either way, that CPU cannot possibly be in an RCU
1041 * read-side critical section that started before the beginning
1042 * of the current RCU grace period.
1044 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1045 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1046 rdp->dynticks_fqs++;
1047 rcu_gpnum_ovf(rnp, rdp);
1051 /* If waiting too long on an offline CPU, complain. */
1052 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1053 time_after(jiffies, rcu_state.gp_start + HZ)) {
1055 struct rcu_node *rnp1;
1057 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1058 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1059 __func__, rnp->grplo, rnp->grphi, rnp->level,
1060 (long)rnp->gp_seq, (long)rnp->completedqs);
1061 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1062 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1063 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1064 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1065 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1066 __func__, rdp->cpu, ".o"[onl],
1067 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1068 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1069 return 1; /* Break things loose after complaining. */
1073 * A CPU running for an extended time within the kernel can
1074 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1075 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1076 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1077 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1078 * variable are safe because the assignments are repeated if this
1079 * CPU failed to pass through a quiescent state. This code
1080 * also checks .jiffies_resched in case jiffies_to_sched_qs
1083 jtsq = READ_ONCE(jiffies_to_sched_qs);
1084 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1085 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1086 if (!READ_ONCE(*rnhqp) &&
1087 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1088 time_after(jiffies, rcu_state.jiffies_resched))) {
1089 WRITE_ONCE(*rnhqp, true);
1090 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1091 smp_store_release(ruqp, true);
1092 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1093 WRITE_ONCE(*ruqp, true);
1097 * NO_HZ_FULL CPUs can run in-kernel without rcu_check_callbacks!
1098 * The above code handles this, but only for straight cond_resched().
1099 * And some in-kernel loops check need_resched() before calling
1100 * cond_resched(), which defeats the above code for CPUs that are
1101 * running in-kernel with scheduling-clock interrupts disabled.
1102 * So hit them over the head with the resched_cpu() hammer!
1104 if (tick_nohz_full_cpu(rdp->cpu) &&
1106 READ_ONCE(rdp->last_fqs_resched) + jtsq * 3)) {
1107 resched_cpu(rdp->cpu);
1108 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1112 * If more than halfway to RCU CPU stall-warning time, invoke
1113 * resched_cpu() more frequently to try to loosen things up a bit.
1114 * Also check to see if the CPU is getting hammered with interrupts,
1115 * but only once per grace period, just to keep the IPIs down to
1118 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1119 if (time_after(jiffies,
1120 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1121 resched_cpu(rdp->cpu);
1122 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1124 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1125 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1126 (rnp->ffmask & rdp->grpmask)) {
1127 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1128 rdp->rcu_iw_pending = true;
1129 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1130 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1137 static void record_gp_stall_check_time(void)
1139 unsigned long j = jiffies;
1142 rcu_state.gp_start = j;
1143 j1 = rcu_jiffies_till_stall_check();
1144 /* Record ->gp_start before ->jiffies_stall. */
1145 smp_store_release(&rcu_state.jiffies_stall, j + j1); /* ^^^ */
1146 rcu_state.jiffies_resched = j + j1 / 2;
1147 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
1151 * Convert a ->gp_state value to a character string.
1153 static const char *gp_state_getname(short gs)
1155 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1157 return gp_state_names[gs];
1161 * Complain about starvation of grace-period kthread.
1163 static void rcu_check_gp_kthread_starvation(void)
1165 struct task_struct *gpk = rcu_state.gp_kthread;
1168 j = jiffies - READ_ONCE(rcu_state.gp_activity);
1170 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1172 (long)rcu_seq_current(&rcu_state.gp_seq),
1174 gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
1175 gpk ? gpk->state : ~0, gpk ? task_cpu(gpk) : -1);
1177 pr_err("RCU grace-period kthread stack dump:\n");
1178 sched_show_task(gpk);
1179 wake_up_process(gpk);
1185 * Dump stacks of all tasks running on stalled CPUs. First try using
1186 * NMIs, but fall back to manual remote stack tracing on architectures
1187 * that don't support NMI-based stack dumps. The NMI-triggered stack
1188 * traces are more accurate because they are printed by the target CPU.
1190 static void rcu_dump_cpu_stacks(void)
1193 unsigned long flags;
1194 struct rcu_node *rnp;
1196 rcu_for_each_leaf_node(rnp) {
1197 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1198 for_each_leaf_node_possible_cpu(rnp, cpu)
1199 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1200 if (!trigger_single_cpu_backtrace(cpu))
1202 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1207 * If too much time has passed in the current grace period, and if
1208 * so configured, go kick the relevant kthreads.
1210 static void rcu_stall_kick_kthreads(void)
1214 if (!rcu_kick_kthreads)
1216 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
1217 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
1218 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
1219 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
1221 rcu_ftrace_dump(DUMP_ALL);
1222 wake_up_process(rcu_state.gp_kthread);
1223 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
1227 static void panic_on_rcu_stall(void)
1229 if (sysctl_panic_on_rcu_stall)
1230 panic("RCU Stall\n");
1233 static void print_other_cpu_stall(unsigned long gp_seq)
1236 unsigned long flags;
1240 struct rcu_node *rnp = rcu_get_root();
1243 /* Kick and suppress, if so configured. */
1244 rcu_stall_kick_kthreads();
1245 if (rcu_cpu_stall_suppress)
1249 * OK, time to rat on our buddy...
1250 * See Documentation/RCU/stallwarn.txt for info on how to debug
1251 * RCU CPU stall warnings.
1253 pr_err("INFO: %s detected stalls on CPUs/tasks:", rcu_state.name);
1254 print_cpu_stall_info_begin();
1255 rcu_for_each_leaf_node(rnp) {
1256 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1257 ndetected += rcu_print_task_stall(rnp);
1258 if (rnp->qsmask != 0) {
1259 for_each_leaf_node_possible_cpu(rnp, cpu)
1260 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1261 print_cpu_stall_info(cpu);
1265 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1268 print_cpu_stall_info_end();
1269 for_each_possible_cpu(cpu)
1270 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(&rcu_data,
1272 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
1273 smp_processor_id(), (long)(jiffies - rcu_state.gp_start),
1274 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
1276 rcu_dump_cpu_stacks();
1278 /* Complain about tasks blocking the grace period. */
1279 rcu_print_detail_task_stall();
1281 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
1282 pr_err("INFO: Stall ended before state dump start\n");
1285 gpa = READ_ONCE(rcu_state.gp_activity);
1286 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1287 rcu_state.name, j - gpa, j, gpa,
1288 READ_ONCE(jiffies_till_next_fqs),
1289 rcu_get_root()->qsmask);
1290 /* In this case, the current CPU might be at fault. */
1291 sched_show_task(current);
1294 /* Rewrite if needed in case of slow consoles. */
1295 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
1296 WRITE_ONCE(rcu_state.jiffies_stall,
1297 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1299 rcu_check_gp_kthread_starvation();
1301 panic_on_rcu_stall();
1303 force_quiescent_state(); /* Kick them all. */
1306 static void print_cpu_stall(void)
1309 unsigned long flags;
1310 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1311 struct rcu_node *rnp = rcu_get_root();
1314 /* Kick and suppress, if so configured. */
1315 rcu_stall_kick_kthreads();
1316 if (rcu_cpu_stall_suppress)
1320 * OK, time to rat on ourselves...
1321 * See Documentation/RCU/stallwarn.txt for info on how to debug
1322 * RCU CPU stall warnings.
1324 pr_err("INFO: %s self-detected stall on CPU", rcu_state.name);
1325 print_cpu_stall_info_begin();
1326 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1327 print_cpu_stall_info(smp_processor_id());
1328 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1329 print_cpu_stall_info_end();
1330 for_each_possible_cpu(cpu)
1331 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(&rcu_data,
1333 pr_cont(" (t=%lu jiffies g=%ld q=%lu)\n",
1334 jiffies - rcu_state.gp_start,
1335 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
1337 rcu_check_gp_kthread_starvation();
1339 rcu_dump_cpu_stacks();
1341 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1342 /* Rewrite if needed in case of slow consoles. */
1343 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
1344 WRITE_ONCE(rcu_state.jiffies_stall,
1345 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1346 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1348 panic_on_rcu_stall();
1351 * Attempt to revive the RCU machinery by forcing a context switch.
1353 * A context switch would normally allow the RCU state machine to make
1354 * progress and it could be we're stuck in kernel space without context
1355 * switches for an entirely unreasonable amount of time.
1357 set_tsk_need_resched(current);
1358 set_preempt_need_resched();
1361 static void check_cpu_stall(struct rcu_data *rdp)
1369 struct rcu_node *rnp;
1371 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1372 !rcu_gp_in_progress())
1374 rcu_stall_kick_kthreads();
1378 * Lots of memory barriers to reject false positives.
1380 * The idea is to pick up rcu_state.gp_seq, then
1381 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
1382 * another copy of rcu_state.gp_seq. These values are updated in
1383 * the opposite order with memory barriers (or equivalent) during
1384 * grace-period initialization and cleanup. Now, a false positive
1385 * can occur if we get an new value of rcu_state.gp_start and a old
1386 * value of rcu_state.jiffies_stall. But given the memory barriers,
1387 * the only way that this can happen is if one grace period ends
1388 * and another starts between these two fetches. This is detected
1389 * by comparing the second fetch of rcu_state.gp_seq with the
1390 * previous fetch from rcu_state.gp_seq.
1392 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
1393 * and rcu_state.gp_start suffice to forestall false positives.
1395 gs1 = READ_ONCE(rcu_state.gp_seq);
1396 smp_rmb(); /* Pick up ->gp_seq first... */
1397 js = READ_ONCE(rcu_state.jiffies_stall);
1398 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1399 gps = READ_ONCE(rcu_state.gp_start);
1400 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
1401 gs2 = READ_ONCE(rcu_state.gp_seq);
1403 ULONG_CMP_LT(j, js) ||
1404 ULONG_CMP_GE(gps, js))
1405 return; /* No stall or GP completed since entering function. */
1407 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1408 if (rcu_gp_in_progress() &&
1409 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
1410 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
1412 /* We haven't checked in, so go dump stack. */
1415 } else if (rcu_gp_in_progress() &&
1416 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
1417 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
1419 /* They had a few time units to dump stack, so complain. */
1420 print_other_cpu_stall(gs2);
1425 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1427 * Set the stall-warning timeout way off into the future, thus preventing
1428 * any RCU CPU stall-warning messages from appearing in the current set of
1429 * RCU grace periods.
1431 * The caller must disable hard irqs.
1433 void rcu_cpu_stall_reset(void)
1435 WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
1438 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1439 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1440 unsigned long gp_seq_req, const char *s)
1442 trace_rcu_future_grace_period(rcu_state.name, rnp->gp_seq, gp_seq_req,
1443 rnp->level, rnp->grplo, rnp->grphi, s);
1447 * rcu_start_this_gp - Request the start of a particular grace period
1448 * @rnp_start: The leaf node of the CPU from which to start.
1449 * @rdp: The rcu_data corresponding to the CPU from which to start.
1450 * @gp_seq_req: The gp_seq of the grace period to start.
1452 * Start the specified grace period, as needed to handle newly arrived
1453 * callbacks. The required future grace periods are recorded in each
1454 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1455 * is reason to awaken the grace-period kthread.
1457 * The caller must hold the specified rcu_node structure's ->lock, which
1458 * is why the caller is responsible for waking the grace-period kthread.
1460 * Returns true if the GP thread needs to be awakened else false.
1462 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1463 unsigned long gp_seq_req)
1466 struct rcu_node *rnp;
1469 * Use funnel locking to either acquire the root rcu_node
1470 * structure's lock or bail out if the need for this grace period
1471 * has already been recorded -- or if that grace period has in
1472 * fact already started. If there is already a grace period in
1473 * progress in a non-leaf node, no recording is needed because the
1474 * end of the grace period will scan the leaf rcu_node structures.
1475 * Note that rnp_start->lock must not be released.
1477 raw_lockdep_assert_held_rcu_node(rnp_start);
1478 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1479 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1480 if (rnp != rnp_start)
1481 raw_spin_lock_rcu_node(rnp);
1482 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1483 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1484 (rnp != rnp_start &&
1485 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1486 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1490 rnp->gp_seq_needed = gp_seq_req;
1491 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1493 * We just marked the leaf or internal node, and a
1494 * grace period is in progress, which means that
1495 * rcu_gp_cleanup() will see the marking. Bail to
1496 * reduce contention.
1498 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1499 TPS("Startedleaf"));
1502 if (rnp != rnp_start && rnp->parent != NULL)
1503 raw_spin_unlock_rcu_node(rnp);
1505 break; /* At root, and perhaps also leaf. */
1508 /* If GP already in progress, just leave, otherwise start one. */
1509 if (rcu_gp_in_progress()) {
1510 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1513 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1514 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1515 rcu_state.gp_req_activity = jiffies;
1516 if (!rcu_state.gp_kthread) {
1517 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1520 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rcu_state.gp_seq), TPS("newreq"));
1521 ret = true; /* Caller must wake GP kthread. */
1523 /* Push furthest requested GP to leaf node and rcu_data structure. */
1524 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1525 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1526 rdp->gp_seq_needed = rnp->gp_seq_needed;
1528 if (rnp != rnp_start)
1529 raw_spin_unlock_rcu_node(rnp);
1534 * Clean up any old requests for the just-ended grace period. Also return
1535 * whether any additional grace periods have been requested.
1537 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1540 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1542 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1544 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1545 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1546 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1551 * Awaken the grace-period kthread. Don't do a self-awaken, and don't
1552 * bother awakening when there is nothing for the grace-period kthread
1553 * to do (as in several CPUs raced to awaken, and we lost), and finally
1554 * don't try to awaken a kthread that has not yet been created.
1556 static void rcu_gp_kthread_wake(void)
1558 if (current == rcu_state.gp_kthread ||
1559 !READ_ONCE(rcu_state.gp_flags) ||
1560 !rcu_state.gp_kthread)
1562 swake_up_one(&rcu_state.gp_wq);
1566 * If there is room, assign a ->gp_seq number to any callbacks on this
1567 * CPU that have not already been assigned. Also accelerate any callbacks
1568 * that were previously assigned a ->gp_seq number that has since proven
1569 * to be too conservative, which can happen if callbacks get assigned a
1570 * ->gp_seq number while RCU is idle, but with reference to a non-root
1571 * rcu_node structure. This function is idempotent, so it does not hurt
1572 * to call it repeatedly. Returns an flag saying that we should awaken
1573 * the RCU grace-period kthread.
1575 * The caller must hold rnp->lock with interrupts disabled.
1577 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1579 unsigned long gp_seq_req;
1582 raw_lockdep_assert_held_rcu_node(rnp);
1584 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1585 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1589 * Callbacks are often registered with incomplete grace-period
1590 * information. Something about the fact that getting exact
1591 * information requires acquiring a global lock... RCU therefore
1592 * makes a conservative estimate of the grace period number at which
1593 * a given callback will become ready to invoke. The following
1594 * code checks this estimate and improves it when possible, thus
1595 * accelerating callback invocation to an earlier grace-period
1598 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1599 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1600 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1602 /* Trace depending on how much we were able to accelerate. */
1603 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1604 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1606 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1611 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1612 * rcu_node structure's ->lock be held. It consults the cached value
1613 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1614 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1615 * while holding the leaf rcu_node structure's ->lock.
1617 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1618 struct rcu_data *rdp)
1623 lockdep_assert_irqs_disabled();
1624 c = rcu_seq_snap(&rcu_state.gp_seq);
1625 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1626 /* Old request still live, so mark recent callbacks. */
1627 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1630 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1631 needwake = rcu_accelerate_cbs(rnp, rdp);
1632 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1634 rcu_gp_kthread_wake();
1638 * Move any callbacks whose grace period has completed to the
1639 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1640 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1641 * sublist. This function is idempotent, so it does not hurt to
1642 * invoke it repeatedly. As long as it is not invoked -too- often...
1643 * Returns true if the RCU grace-period kthread needs to be awakened.
1645 * The caller must hold rnp->lock with interrupts disabled.
1647 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1649 raw_lockdep_assert_held_rcu_node(rnp);
1651 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1652 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1656 * Find all callbacks whose ->gp_seq numbers indicate that they
1657 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1659 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1661 /* Classify any remaining callbacks. */
1662 return rcu_accelerate_cbs(rnp, rdp);
1666 * Update CPU-local rcu_data state to record the beginnings and ends of
1667 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1668 * structure corresponding to the current CPU, and must have irqs disabled.
1669 * Returns true if the grace-period kthread needs to be awakened.
1671 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1676 raw_lockdep_assert_held_rcu_node(rnp);
1678 if (rdp->gp_seq == rnp->gp_seq)
1679 return false; /* Nothing to do. */
1681 /* Handle the ends of any preceding grace periods first. */
1682 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1683 unlikely(READ_ONCE(rdp->gpwrap))) {
1684 ret = rcu_advance_cbs(rnp, rdp); /* Advance callbacks. */
1685 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1687 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent callbacks. */
1690 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1691 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1692 unlikely(READ_ONCE(rdp->gpwrap))) {
1694 * If the current grace period is waiting for this CPU,
1695 * set up to detect a quiescent state, otherwise don't
1696 * go looking for one.
1698 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1699 need_gp = !!(rnp->qsmask & rdp->grpmask);
1700 rdp->cpu_no_qs.b.norm = need_gp;
1701 rdp->core_needs_qs = need_gp;
1702 zero_cpu_stall_ticks(rdp);
1704 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1705 if (ULONG_CMP_GE(rnp->gp_seq_needed, rdp->gp_seq_needed) || rdp->gpwrap)
1706 rdp->gp_seq_needed = rnp->gp_seq_needed;
1707 WRITE_ONCE(rdp->gpwrap, false);
1708 rcu_gpnum_ovf(rnp, rdp);
1712 static void note_gp_changes(struct rcu_data *rdp)
1714 unsigned long flags;
1716 struct rcu_node *rnp;
1718 local_irq_save(flags);
1720 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1721 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1722 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1723 local_irq_restore(flags);
1726 needwake = __note_gp_changes(rnp, rdp);
1727 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1729 rcu_gp_kthread_wake();
1732 static void rcu_gp_slow(int delay)
1735 !(rcu_seq_ctr(rcu_state.gp_seq) %
1736 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1737 schedule_timeout_uninterruptible(delay);
1741 * Initialize a new grace period. Return false if no grace period required.
1743 static bool rcu_gp_init(void)
1745 unsigned long flags;
1746 unsigned long oldmask;
1748 struct rcu_data *rdp;
1749 struct rcu_node *rnp = rcu_get_root();
1751 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1752 raw_spin_lock_irq_rcu_node(rnp);
1753 if (!READ_ONCE(rcu_state.gp_flags)) {
1754 /* Spurious wakeup, tell caller to go back to sleep. */
1755 raw_spin_unlock_irq_rcu_node(rnp);
1758 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1760 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1762 * Grace period already in progress, don't start another.
1763 * Not supposed to be able to happen.
1765 raw_spin_unlock_irq_rcu_node(rnp);
1769 /* Advance to a new grace period and initialize state. */
1770 record_gp_stall_check_time();
1771 /* Record GP times before starting GP, hence rcu_seq_start(). */
1772 rcu_seq_start(&rcu_state.gp_seq);
1773 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1774 raw_spin_unlock_irq_rcu_node(rnp);
1777 * Apply per-leaf buffered online and offline operations to the
1778 * rcu_node tree. Note that this new grace period need not wait
1779 * for subsequent online CPUs, and that quiescent-state forcing
1780 * will handle subsequent offline CPUs.
1782 rcu_state.gp_state = RCU_GP_ONOFF;
1783 rcu_for_each_leaf_node(rnp) {
1784 spin_lock(&rcu_state.ofl_lock);
1785 raw_spin_lock_irq_rcu_node(rnp);
1786 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1787 !rnp->wait_blkd_tasks) {
1788 /* Nothing to do on this leaf rcu_node structure. */
1789 raw_spin_unlock_irq_rcu_node(rnp);
1790 spin_unlock(&rcu_state.ofl_lock);
1794 /* Record old state, apply changes to ->qsmaskinit field. */
1795 oldmask = rnp->qsmaskinit;
1796 rnp->qsmaskinit = rnp->qsmaskinitnext;
1798 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1799 if (!oldmask != !rnp->qsmaskinit) {
1800 if (!oldmask) { /* First online CPU for rcu_node. */
1801 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1802 rcu_init_new_rnp(rnp);
1803 } else if (rcu_preempt_has_tasks(rnp)) {
1804 rnp->wait_blkd_tasks = true; /* blocked tasks */
1805 } else { /* Last offline CPU and can propagate. */
1806 rcu_cleanup_dead_rnp(rnp);
1811 * If all waited-on tasks from prior grace period are
1812 * done, and if all this rcu_node structure's CPUs are
1813 * still offline, propagate up the rcu_node tree and
1814 * clear ->wait_blkd_tasks. Otherwise, if one of this
1815 * rcu_node structure's CPUs has since come back online,
1816 * simply clear ->wait_blkd_tasks.
1818 if (rnp->wait_blkd_tasks &&
1819 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1820 rnp->wait_blkd_tasks = false;
1821 if (!rnp->qsmaskinit)
1822 rcu_cleanup_dead_rnp(rnp);
1825 raw_spin_unlock_irq_rcu_node(rnp);
1826 spin_unlock(&rcu_state.ofl_lock);
1828 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1831 * Set the quiescent-state-needed bits in all the rcu_node
1832 * structures for all currently online CPUs in breadth-first
1833 * order, starting from the root rcu_node structure, relying on the
1834 * layout of the tree within the rcu_state.node[] array. Note that
1835 * other CPUs will access only the leaves of the hierarchy, thus
1836 * seeing that no grace period is in progress, at least until the
1837 * corresponding leaf node has been initialized.
1839 * The grace period cannot complete until the initialization
1840 * process finishes, because this kthread handles both.
1842 rcu_state.gp_state = RCU_GP_INIT;
1843 rcu_for_each_node_breadth_first(rnp) {
1844 rcu_gp_slow(gp_init_delay);
1845 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1846 rdp = this_cpu_ptr(&rcu_data);
1847 rcu_preempt_check_blocked_tasks(rnp);
1848 rnp->qsmask = rnp->qsmaskinit;
1849 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1850 if (rnp == rdp->mynode)
1851 (void)__note_gp_changes(rnp, rdp);
1852 rcu_preempt_boost_start_gp(rnp);
1853 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1854 rnp->level, rnp->grplo,
1855 rnp->grphi, rnp->qsmask);
1856 /* Quiescent states for tasks on any now-offline CPUs. */
1857 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1858 rnp->rcu_gp_init_mask = mask;
1859 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1860 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1862 raw_spin_unlock_irq_rcu_node(rnp);
1863 cond_resched_tasks_rcu_qs();
1864 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1871 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1874 static bool rcu_gp_fqs_check_wake(int *gfp)
1876 struct rcu_node *rnp = rcu_get_root();
1878 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1879 *gfp = READ_ONCE(rcu_state.gp_flags);
1880 if (*gfp & RCU_GP_FLAG_FQS)
1883 /* The current grace period has completed. */
1884 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1891 * Do one round of quiescent-state forcing.
1893 static void rcu_gp_fqs(bool first_time)
1895 struct rcu_node *rnp = rcu_get_root();
1897 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1898 rcu_state.n_force_qs++;
1900 /* Collect dyntick-idle snapshots. */
1901 force_qs_rnp(dyntick_save_progress_counter);
1903 /* Handle dyntick-idle and offline CPUs. */
1904 force_qs_rnp(rcu_implicit_dynticks_qs);
1906 /* Clear flag to prevent immediate re-entry. */
1907 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1908 raw_spin_lock_irq_rcu_node(rnp);
1909 WRITE_ONCE(rcu_state.gp_flags,
1910 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1911 raw_spin_unlock_irq_rcu_node(rnp);
1916 * Loop doing repeated quiescent-state forcing until the grace period ends.
1918 static void rcu_gp_fqs_loop(void)
1924 struct rcu_node *rnp = rcu_get_root();
1926 first_gp_fqs = true;
1927 j = READ_ONCE(jiffies_till_first_fqs);
1931 rcu_state.jiffies_force_qs = jiffies + j;
1932 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1935 trace_rcu_grace_period(rcu_state.name,
1936 READ_ONCE(rcu_state.gp_seq),
1938 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1939 ret = swait_event_idle_timeout_exclusive(
1940 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1941 rcu_state.gp_state = RCU_GP_DOING_FQS;
1942 /* Locking provides needed memory barriers. */
1943 /* If grace period done, leave loop. */
1944 if (!READ_ONCE(rnp->qsmask) &&
1945 !rcu_preempt_blocked_readers_cgp(rnp))
1947 /* If time for quiescent-state forcing, do it. */
1948 if (ULONG_CMP_GE(jiffies, rcu_state.jiffies_force_qs) ||
1949 (gf & RCU_GP_FLAG_FQS)) {
1950 trace_rcu_grace_period(rcu_state.name,
1951 READ_ONCE(rcu_state.gp_seq),
1953 rcu_gp_fqs(first_gp_fqs);
1954 first_gp_fqs = false;
1955 trace_rcu_grace_period(rcu_state.name,
1956 READ_ONCE(rcu_state.gp_seq),
1958 cond_resched_tasks_rcu_qs();
1959 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1960 ret = 0; /* Force full wait till next FQS. */
1961 j = READ_ONCE(jiffies_till_next_fqs);
1963 /* Deal with stray signal. */
1964 cond_resched_tasks_rcu_qs();
1965 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1966 WARN_ON(signal_pending(current));
1967 trace_rcu_grace_period(rcu_state.name,
1968 READ_ONCE(rcu_state.gp_seq),
1970 ret = 1; /* Keep old FQS timing. */
1972 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1975 j = rcu_state.jiffies_force_qs - j;
1981 * Clean up after the old grace period.
1983 static void rcu_gp_cleanup(void)
1985 unsigned long gp_duration;
1986 bool needgp = false;
1987 unsigned long new_gp_seq;
1988 struct rcu_data *rdp;
1989 struct rcu_node *rnp = rcu_get_root();
1990 struct swait_queue_head *sq;
1992 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1993 raw_spin_lock_irq_rcu_node(rnp);
1994 gp_duration = jiffies - rcu_state.gp_start;
1995 if (gp_duration > rcu_state.gp_max)
1996 rcu_state.gp_max = gp_duration;
1999 * We know the grace period is complete, but to everyone else
2000 * it appears to still be ongoing. But it is also the case
2001 * that to everyone else it looks like there is nothing that
2002 * they can do to advance the grace period. It is therefore
2003 * safe for us to drop the lock in order to mark the grace
2004 * period as completed in all of the rcu_node structures.
2006 raw_spin_unlock_irq_rcu_node(rnp);
2009 * Propagate new ->gp_seq value to rcu_node structures so that
2010 * other CPUs don't have to wait until the start of the next grace
2011 * period to process their callbacks. This also avoids some nasty
2012 * RCU grace-period initialization races by forcing the end of
2013 * the current grace period to be completely recorded in all of
2014 * the rcu_node structures before the beginning of the next grace
2015 * period is recorded in any of the rcu_node structures.
2017 new_gp_seq = rcu_state.gp_seq;
2018 rcu_seq_end(&new_gp_seq);
2019 rcu_for_each_node_breadth_first(rnp) {
2020 raw_spin_lock_irq_rcu_node(rnp);
2021 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2022 dump_blkd_tasks(rnp, 10);
2023 WARN_ON_ONCE(rnp->qsmask);
2024 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2025 rdp = this_cpu_ptr(&rcu_data);
2026 if (rnp == rdp->mynode)
2027 needgp = __note_gp_changes(rnp, rdp) || needgp;
2028 /* smp_mb() provided by prior unlock-lock pair. */
2029 needgp = rcu_future_gp_cleanup(rnp) || needgp;
2030 sq = rcu_nocb_gp_get(rnp);
2031 raw_spin_unlock_irq_rcu_node(rnp);
2032 rcu_nocb_gp_cleanup(sq);
2033 cond_resched_tasks_rcu_qs();
2034 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2035 rcu_gp_slow(gp_cleanup_delay);
2037 rnp = rcu_get_root();
2038 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2040 /* Declare grace period done. */
2041 rcu_seq_end(&rcu_state.gp_seq);
2042 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2043 rcu_state.gp_state = RCU_GP_IDLE;
2044 /* Check for GP requests since above loop. */
2045 rdp = this_cpu_ptr(&rcu_data);
2046 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2047 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2048 TPS("CleanupMore"));
2051 /* Advance CBs to reduce false positives below. */
2052 if (!rcu_accelerate_cbs(rnp, rdp) && needgp) {
2053 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2054 rcu_state.gp_req_activity = jiffies;
2055 trace_rcu_grace_period(rcu_state.name,
2056 READ_ONCE(rcu_state.gp_seq),
2059 WRITE_ONCE(rcu_state.gp_flags,
2060 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2062 raw_spin_unlock_irq_rcu_node(rnp);
2066 * Body of kthread that handles grace periods.
2068 static int __noreturn rcu_gp_kthread(void *unused)
2070 rcu_bind_gp_kthread();
2073 /* Handle grace-period start. */
2075 trace_rcu_grace_period(rcu_state.name,
2076 READ_ONCE(rcu_state.gp_seq),
2078 rcu_state.gp_state = RCU_GP_WAIT_GPS;
2079 swait_event_idle_exclusive(rcu_state.gp_wq,
2080 READ_ONCE(rcu_state.gp_flags) &
2082 rcu_state.gp_state = RCU_GP_DONE_GPS;
2083 /* Locking provides needed memory barrier. */
2086 cond_resched_tasks_rcu_qs();
2087 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2088 WARN_ON(signal_pending(current));
2089 trace_rcu_grace_period(rcu_state.name,
2090 READ_ONCE(rcu_state.gp_seq),
2094 /* Handle quiescent-state forcing. */
2097 /* Handle grace-period end. */
2098 rcu_state.gp_state = RCU_GP_CLEANUP;
2100 rcu_state.gp_state = RCU_GP_CLEANED;
2105 * Report a full set of quiescent states to the rcu_state data structure.
2106 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2107 * another grace period is required. Whether we wake the grace-period
2108 * kthread or it awakens itself for the next round of quiescent-state
2109 * forcing, that kthread will clean up after the just-completed grace
2110 * period. Note that the caller must hold rnp->lock, which is released
2113 static void rcu_report_qs_rsp(unsigned long flags)
2114 __releases(rcu_get_root()->lock)
2116 raw_lockdep_assert_held_rcu_node(rcu_get_root());
2117 WARN_ON_ONCE(!rcu_gp_in_progress());
2118 WRITE_ONCE(rcu_state.gp_flags,
2119 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2120 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2121 rcu_gp_kthread_wake();
2125 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2126 * Allows quiescent states for a group of CPUs to be reported at one go
2127 * to the specified rcu_node structure, though all the CPUs in the group
2128 * must be represented by the same rcu_node structure (which need not be a
2129 * leaf rcu_node structure, though it often will be). The gps parameter
2130 * is the grace-period snapshot, which means that the quiescent states
2131 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2132 * must be held upon entry, and it is released before return.
2134 * As a special case, if mask is zero, the bit-already-cleared check is
2135 * disabled. This allows propagating quiescent state due to resumed tasks
2136 * during grace-period initialization.
2138 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2139 unsigned long gps, unsigned long flags)
2140 __releases(rnp->lock)
2142 unsigned long oldmask = 0;
2143 struct rcu_node *rnp_c;
2145 raw_lockdep_assert_held_rcu_node(rnp);
2147 /* Walk up the rcu_node hierarchy. */
2149 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2152 * Our bit has already been cleared, or the
2153 * relevant grace period is already over, so done.
2155 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2158 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2159 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2160 rcu_preempt_blocked_readers_cgp(rnp));
2161 rnp->qsmask &= ~mask;
2162 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2163 mask, rnp->qsmask, rnp->level,
2164 rnp->grplo, rnp->grphi,
2166 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2168 /* Other bits still set at this level, so done. */
2169 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2172 rnp->completedqs = rnp->gp_seq;
2173 mask = rnp->grpmask;
2174 if (rnp->parent == NULL) {
2176 /* No more levels. Exit loop holding root lock. */
2180 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2183 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2184 oldmask = rnp_c->qsmask;
2188 * Get here if we are the last CPU to pass through a quiescent
2189 * state for this grace period. Invoke rcu_report_qs_rsp()
2190 * to clean up and start the next grace period if one is needed.
2192 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2196 * Record a quiescent state for all tasks that were previously queued
2197 * on the specified rcu_node structure and that were blocking the current
2198 * RCU grace period. The caller must hold the corresponding rnp->lock with
2199 * irqs disabled, and this lock is released upon return, but irqs remain
2202 static void __maybe_unused
2203 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2204 __releases(rnp->lock)
2208 struct rcu_node *rnp_p;
2210 raw_lockdep_assert_held_rcu_node(rnp);
2211 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT)) ||
2212 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2214 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2215 return; /* Still need more quiescent states! */
2218 rnp->completedqs = rnp->gp_seq;
2219 rnp_p = rnp->parent;
2220 if (rnp_p == NULL) {
2222 * Only one rcu_node structure in the tree, so don't
2223 * try to report up to its nonexistent parent!
2225 rcu_report_qs_rsp(flags);
2229 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2231 mask = rnp->grpmask;
2232 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2233 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2234 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2238 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2239 * structure. This must be called from the specified CPU.
2242 rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
2244 unsigned long flags;
2247 struct rcu_node *rnp;
2250 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2251 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2255 * The grace period in which this quiescent state was
2256 * recorded has ended, so don't report it upwards.
2257 * We will instead need a new quiescent state that lies
2258 * within the current grace period.
2260 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2261 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2264 mask = rdp->grpmask;
2265 if ((rnp->qsmask & mask) == 0) {
2266 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2268 rdp->core_needs_qs = false;
2271 * This GP can't end until cpu checks in, so all of our
2272 * callbacks can be processed during the next GP.
2274 needwake = rcu_accelerate_cbs(rnp, rdp);
2276 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2277 /* ^^^ Released rnp->lock */
2279 rcu_gp_kthread_wake();
2284 * Check to see if there is a new grace period of which this CPU
2285 * is not yet aware, and if so, set up local rcu_data state for it.
2286 * Otherwise, see if this CPU has just passed through its first
2287 * quiescent state for this grace period, and record that fact if so.
2290 rcu_check_quiescent_state(struct rcu_data *rdp)
2292 /* Check for grace-period ends and beginnings. */
2293 note_gp_changes(rdp);
2296 * Does this CPU still need to do its part for current grace period?
2297 * If no, return and let the other CPUs do their part as well.
2299 if (!rdp->core_needs_qs)
2303 * Was there a quiescent state since the beginning of the grace
2304 * period? If no, then exit and wait for the next call.
2306 if (rdp->cpu_no_qs.b.norm)
2310 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2313 rcu_report_qs_rdp(rdp->cpu, rdp);
2317 * Near the end of the offline process. Trace the fact that this CPU
2320 int rcutree_dying_cpu(unsigned int cpu)
2322 RCU_TRACE(bool blkd;)
2323 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(&rcu_data);)
2324 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2326 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2329 RCU_TRACE(blkd = !!(rnp->qsmask & rdp->grpmask);)
2330 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq,
2331 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2336 * All CPUs for the specified rcu_node structure have gone offline,
2337 * and all tasks that were preempted within an RCU read-side critical
2338 * section while running on one of those CPUs have since exited their RCU
2339 * read-side critical section. Some other CPU is reporting this fact with
2340 * the specified rcu_node structure's ->lock held and interrupts disabled.
2341 * This function therefore goes up the tree of rcu_node structures,
2342 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2343 * the leaf rcu_node structure's ->qsmaskinit field has already been
2346 * This function does check that the specified rcu_node structure has
2347 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2348 * prematurely. That said, invoking it after the fact will cost you
2349 * a needless lock acquisition. So once it has done its work, don't
2352 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2355 struct rcu_node *rnp = rnp_leaf;
2357 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2358 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2359 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2360 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2363 mask = rnp->grpmask;
2367 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2368 rnp->qsmaskinit &= ~mask;
2369 /* Between grace periods, so better already be zero! */
2370 WARN_ON_ONCE(rnp->qsmask);
2371 if (rnp->qsmaskinit) {
2372 raw_spin_unlock_rcu_node(rnp);
2373 /* irqs remain disabled. */
2376 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2381 * The CPU has been completely removed, and some other CPU is reporting
2382 * this fact from process context. Do the remainder of the cleanup.
2383 * There can only be one CPU hotplug operation at a time, so no need for
2386 int rcutree_dead_cpu(unsigned int cpu)
2388 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2389 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2391 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2394 /* Adjust any no-longer-needed kthreads. */
2395 rcu_boost_kthread_setaffinity(rnp, -1);
2396 /* Do any needed no-CB deferred wakeups from this CPU. */
2397 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2402 * Invoke any RCU callbacks that have made it to the end of their grace
2403 * period. Thottle as specified by rdp->blimit.
2405 static void rcu_do_batch(struct rcu_data *rdp)
2407 unsigned long flags;
2408 struct rcu_head *rhp;
2409 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2412 /* If no callbacks are ready, just return. */
2413 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2414 trace_rcu_batch_start(rcu_state.name,
2415 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2416 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2417 trace_rcu_batch_end(rcu_state.name, 0,
2418 !rcu_segcblist_empty(&rdp->cblist),
2419 need_resched(), is_idle_task(current),
2420 rcu_is_callbacks_kthread());
2425 * Extract the list of ready callbacks, disabling to prevent
2426 * races with call_rcu() from interrupt handlers. Leave the
2427 * callback counts, as rcu_barrier() needs to be conservative.
2429 local_irq_save(flags);
2430 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2432 trace_rcu_batch_start(rcu_state.name,
2433 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2434 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2435 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2436 local_irq_restore(flags);
2438 /* Invoke callbacks. */
2439 rhp = rcu_cblist_dequeue(&rcl);
2440 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2441 debug_rcu_head_unqueue(rhp);
2442 if (__rcu_reclaim(rcu_state.name, rhp))
2443 rcu_cblist_dequeued_lazy(&rcl);
2445 * Stop only if limit reached and CPU has something to do.
2446 * Note: The rcl structure counts down from zero.
2448 if (-rcl.len >= bl &&
2450 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2454 local_irq_save(flags);
2456 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2457 is_idle_task(current), rcu_is_callbacks_kthread());
2459 /* Update counts and requeue any remaining callbacks. */
2460 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2461 smp_mb(); /* List handling before counting for rcu_barrier(). */
2462 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2464 /* Reinstate batch limit if we have worked down the excess. */
2465 count = rcu_segcblist_n_cbs(&rdp->cblist);
2466 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2467 rdp->blimit = blimit;
2469 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2470 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2471 rdp->qlen_last_fqs_check = 0;
2472 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2473 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2474 rdp->qlen_last_fqs_check = count;
2477 * The following usually indicates a double call_rcu(). To track
2478 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2480 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2482 local_irq_restore(flags);
2484 /* Re-invoke RCU core processing if there are callbacks remaining. */
2485 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2490 * Check to see if this CPU is in a non-context-switch quiescent state
2491 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2492 * Also schedule RCU core processing.
2494 * This function must be called from hardirq context. It is normally
2495 * invoked from the scheduling-clock interrupt.
2497 void rcu_check_callbacks(int user)
2499 trace_rcu_utilization(TPS("Start scheduler-tick"));
2500 raw_cpu_inc(rcu_data.ticks_this_gp);
2501 /* The load-acquire pairs with the store-release setting to true. */
2502 if (smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
2503 /* Idle and userspace execution already are quiescent states. */
2504 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2505 set_tsk_need_resched(current);
2506 set_preempt_need_resched();
2508 __this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
2510 rcu_flavor_check_callbacks(user);
2514 trace_rcu_utilization(TPS("End scheduler-tick"));
2518 * Scan the leaf rcu_node structures, processing dyntick state for any that
2519 * have not yet encountered a quiescent state, using the function specified.
2520 * Also initiate boosting for any threads blocked on the root rcu_node.
2522 * The caller must have suppressed start of new grace periods.
2524 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2527 unsigned long flags;
2529 struct rcu_node *rnp;
2531 rcu_for_each_leaf_node(rnp) {
2532 cond_resched_tasks_rcu_qs();
2534 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2535 if (rnp->qsmask == 0) {
2536 if (!IS_ENABLED(CONFIG_PREEMPT) ||
2537 rcu_preempt_blocked_readers_cgp(rnp)) {
2539 * No point in scanning bits because they
2540 * are all zero. But we might need to
2541 * priority-boost blocked readers.
2543 rcu_initiate_boost(rnp, flags);
2544 /* rcu_initiate_boost() releases rnp->lock */
2547 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2550 for_each_leaf_node_possible_cpu(rnp, cpu) {
2551 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2552 if ((rnp->qsmask & bit) != 0) {
2553 if (f(per_cpu_ptr(&rcu_data, cpu)))
2558 /* Idle/offline CPUs, report (releases rnp->lock). */
2559 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2561 /* Nothing to do here, so just drop the lock. */
2562 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2568 * Force quiescent states on reluctant CPUs, and also detect which
2569 * CPUs are in dyntick-idle mode.
2571 static void force_quiescent_state(void)
2573 unsigned long flags;
2575 struct rcu_node *rnp;
2576 struct rcu_node *rnp_old = NULL;
2578 /* Funnel through hierarchy to reduce memory contention. */
2579 rnp = __this_cpu_read(rcu_data.mynode);
2580 for (; rnp != NULL; rnp = rnp->parent) {
2581 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2582 !raw_spin_trylock(&rnp->fqslock);
2583 if (rnp_old != NULL)
2584 raw_spin_unlock(&rnp_old->fqslock);
2589 /* rnp_old == rcu_get_root(), rnp == NULL. */
2591 /* Reached the root of the rcu_node tree, acquire lock. */
2592 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2593 raw_spin_unlock(&rnp_old->fqslock);
2594 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2595 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2596 return; /* Someone beat us to it. */
2598 WRITE_ONCE(rcu_state.gp_flags,
2599 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2600 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2601 rcu_gp_kthread_wake();
2605 * This function checks for grace-period requests that fail to motivate
2606 * RCU to come out of its idle mode.
2609 rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp)
2611 const unsigned long gpssdelay = rcu_jiffies_till_stall_check() * HZ;
2612 unsigned long flags;
2614 struct rcu_node *rnp_root = rcu_get_root();
2615 static atomic_t warned = ATOMIC_INIT(0);
2617 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
2618 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed))
2620 j = jiffies; /* Expensive access, and in common case don't get here. */
2621 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
2622 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
2623 atomic_read(&warned))
2626 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2628 if (rcu_gp_in_progress() ||
2629 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2630 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
2631 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
2632 atomic_read(&warned)) {
2633 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2636 /* Hold onto the leaf lock to make others see warned==1. */
2638 if (rnp_root != rnp)
2639 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
2641 if (rcu_gp_in_progress() ||
2642 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2643 time_before(j, rcu_state.gp_req_activity + gpssdelay) ||
2644 time_before(j, rcu_state.gp_activity + gpssdelay) ||
2645 atomic_xchg(&warned, 1)) {
2646 raw_spin_unlock_rcu_node(rnp_root); /* irqs remain disabled. */
2647 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2650 pr_alert("%s: g%ld->%ld gar:%lu ga:%lu f%#x gs:%d %s->state:%#lx\n",
2651 __func__, (long)READ_ONCE(rcu_state.gp_seq),
2652 (long)READ_ONCE(rnp_root->gp_seq_needed),
2653 j - rcu_state.gp_req_activity, j - rcu_state.gp_activity,
2654 rcu_state.gp_flags, rcu_state.gp_state, rcu_state.name,
2655 rcu_state.gp_kthread ? rcu_state.gp_kthread->state : 0x1ffffL);
2657 if (rnp_root != rnp)
2658 raw_spin_unlock_rcu_node(rnp_root);
2659 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2663 * This does the RCU core processing work for the specified rcu_data
2664 * structures. This may be called only from the CPU to whom the rdp
2667 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2669 unsigned long flags;
2670 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2671 struct rcu_node *rnp = rdp->mynode;
2673 if (cpu_is_offline(smp_processor_id()))
2675 trace_rcu_utilization(TPS("Start RCU core"));
2676 WARN_ON_ONCE(!rdp->beenonline);
2678 /* Report any deferred quiescent states if preemption enabled. */
2679 if (!(preempt_count() & PREEMPT_MASK)) {
2680 rcu_preempt_deferred_qs(current);
2681 } else if (rcu_preempt_need_deferred_qs(current)) {
2682 set_tsk_need_resched(current);
2683 set_preempt_need_resched();
2686 /* Update RCU state based on any recent quiescent states. */
2687 rcu_check_quiescent_state(rdp);
2689 /* No grace period and unregistered callbacks? */
2690 if (!rcu_gp_in_progress() &&
2691 rcu_segcblist_is_enabled(&rdp->cblist)) {
2692 local_irq_save(flags);
2693 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2694 rcu_accelerate_cbs_unlocked(rnp, rdp);
2695 local_irq_restore(flags);
2698 rcu_check_gp_start_stall(rnp, rdp);
2700 /* If there are callbacks ready, invoke them. */
2701 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2702 invoke_rcu_callbacks(rdp);
2704 /* Do any needed deferred wakeups of rcuo kthreads. */
2705 do_nocb_deferred_wakeup(rdp);
2706 trace_rcu_utilization(TPS("End RCU core"));
2710 * Schedule RCU callback invocation. If the running implementation of RCU
2711 * does not support RCU priority boosting, just do a direct call, otherwise
2712 * wake up the per-CPU kernel kthread. Note that because we are running
2713 * on the current CPU with softirqs disabled, the rcu_cpu_kthread_task
2714 * cannot disappear out from under us.
2716 static void invoke_rcu_callbacks(struct rcu_data *rdp)
2718 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2720 if (likely(!rcu_state.boost)) {
2724 invoke_rcu_callbacks_kthread();
2727 static void invoke_rcu_core(void)
2729 if (cpu_online(smp_processor_id()))
2730 raise_softirq(RCU_SOFTIRQ);
2734 * Handle any core-RCU processing required by a call_rcu() invocation.
2736 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2737 unsigned long flags)
2740 * If called from an extended quiescent state, invoke the RCU
2741 * core in order to force a re-evaluation of RCU's idleness.
2743 if (!rcu_is_watching())
2746 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2747 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2751 * Force the grace period if too many callbacks or too long waiting.
2752 * Enforce hysteresis, and don't invoke force_quiescent_state()
2753 * if some other CPU has recently done so. Also, don't bother
2754 * invoking force_quiescent_state() if the newly enqueued callback
2755 * is the only one waiting for a grace period to complete.
2757 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2758 rdp->qlen_last_fqs_check + qhimark)) {
2760 /* Are we ignoring a completed grace period? */
2761 note_gp_changes(rdp);
2763 /* Start a new grace period if one not already started. */
2764 if (!rcu_gp_in_progress()) {
2765 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2767 /* Give the grace period a kick. */
2768 rdp->blimit = LONG_MAX;
2769 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2770 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2771 force_quiescent_state();
2772 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2773 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2779 * RCU callback function to leak a callback.
2781 static void rcu_leak_callback(struct rcu_head *rhp)
2786 * Helper function for call_rcu() and friends. The cpu argument will
2787 * normally be -1, indicating "currently running CPU". It may specify
2788 * a CPU only if that CPU is a no-CBs CPU. Currently, only rcu_barrier()
2789 * is expected to specify a CPU.
2792 __call_rcu(struct rcu_head *head, rcu_callback_t func, int cpu, bool lazy)
2794 unsigned long flags;
2795 struct rcu_data *rdp;
2797 /* Misaligned rcu_head! */
2798 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2800 if (debug_rcu_head_queue(head)) {
2802 * Probable double call_rcu(), so leak the callback.
2803 * Use rcu:rcu_callback trace event to find the previous
2804 * time callback was passed to __call_rcu().
2806 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
2808 WRITE_ONCE(head->func, rcu_leak_callback);
2813 local_irq_save(flags);
2814 rdp = this_cpu_ptr(&rcu_data);
2816 /* Add the callback to our list. */
2817 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
2821 rdp = per_cpu_ptr(&rcu_data, cpu);
2822 if (likely(rdp->mynode)) {
2823 /* Post-boot, so this should be for a no-CBs CPU. */
2824 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2825 WARN_ON_ONCE(offline);
2826 /* Offline CPU, _call_rcu() illegal, leak callback. */
2827 local_irq_restore(flags);
2831 * Very early boot, before rcu_init(). Initialize if needed
2832 * and then drop through to queue the callback.
2835 WARN_ON_ONCE(!rcu_is_watching());
2836 if (rcu_segcblist_empty(&rdp->cblist))
2837 rcu_segcblist_init(&rdp->cblist);
2839 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2841 rcu_idle_count_callbacks_posted();
2843 if (__is_kfree_rcu_offset((unsigned long)func))
2844 trace_rcu_kfree_callback(rcu_state.name, head,
2845 (unsigned long)func,
2846 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2847 rcu_segcblist_n_cbs(&rdp->cblist));
2849 trace_rcu_callback(rcu_state.name, head,
2850 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2851 rcu_segcblist_n_cbs(&rdp->cblist));
2853 /* Go handle any RCU core processing required. */
2854 __call_rcu_core(rdp, head, flags);
2855 local_irq_restore(flags);
2859 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2860 * @head: structure to be used for queueing the RCU updates.
2861 * @func: actual callback function to be invoked after the grace period
2863 * The callback function will be invoked some time after a full grace
2864 * period elapses, in other words after all pre-existing RCU read-side
2865 * critical sections have completed. However, the callback function
2866 * might well execute concurrently with RCU read-side critical sections
2867 * that started after call_rcu() was invoked. RCU read-side critical
2868 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2869 * may be nested. In addition, regions of code across which interrupts,
2870 * preemption, or softirqs have been disabled also serve as RCU read-side
2871 * critical sections. This includes hardware interrupt handlers, softirq
2872 * handlers, and NMI handlers.
2874 * Note that all CPUs must agree that the grace period extended beyond
2875 * all pre-existing RCU read-side critical section. On systems with more
2876 * than one CPU, this means that when "func()" is invoked, each CPU is
2877 * guaranteed to have executed a full memory barrier since the end of its
2878 * last RCU read-side critical section whose beginning preceded the call
2879 * to call_rcu(). It also means that each CPU executing an RCU read-side
2880 * critical section that continues beyond the start of "func()" must have
2881 * executed a memory barrier after the call_rcu() but before the beginning
2882 * of that RCU read-side critical section. Note that these guarantees
2883 * include CPUs that are offline, idle, or executing in user mode, as
2884 * well as CPUs that are executing in the kernel.
2886 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2887 * resulting RCU callback function "func()", then both CPU A and CPU B are
2888 * guaranteed to execute a full memory barrier during the time interval
2889 * between the call to call_rcu() and the invocation of "func()" -- even
2890 * if CPU A and CPU B are the same CPU (but again only if the system has
2891 * more than one CPU).
2893 void call_rcu(struct rcu_head *head, rcu_callback_t func)
2895 __call_rcu(head, func, -1, 0);
2897 EXPORT_SYMBOL_GPL(call_rcu);
2900 * Queue an RCU callback for lazy invocation after a grace period.
2901 * This will likely be later named something like "call_rcu_lazy()",
2902 * but this change will require some way of tagging the lazy RCU
2903 * callbacks in the list of pending callbacks. Until then, this
2904 * function may only be called from __kfree_rcu().
2906 void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
2908 __call_rcu(head, func, -1, 1);
2910 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2913 * get_state_synchronize_rcu - Snapshot current RCU state
2915 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2916 * to determine whether or not a full grace period has elapsed in the
2919 unsigned long get_state_synchronize_rcu(void)
2922 * Any prior manipulation of RCU-protected data must happen
2923 * before the load from ->gp_seq.
2926 return rcu_seq_snap(&rcu_state.gp_seq);
2928 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2931 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2933 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2935 * If a full RCU grace period has elapsed since the earlier call to
2936 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2937 * synchronize_rcu() to wait for a full grace period.
2939 * Yes, this function does not take counter wrap into account. But
2940 * counter wrap is harmless. If the counter wraps, we have waited for
2941 * more than 2 billion grace periods (and way more on a 64-bit system!),
2942 * so waiting for one additional grace period should be just fine.
2944 void cond_synchronize_rcu(unsigned long oldstate)
2946 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
2949 smp_mb(); /* Ensure GP ends before subsequent accesses. */
2951 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2954 * Check to see if there is any immediate RCU-related work to be done by
2955 * the current CPU, returning 1 if so and zero otherwise. The checks are
2956 * in order of increasing expense: checks that can be carried out against
2957 * CPU-local state are performed first. However, we must check for CPU
2958 * stalls first, else we might not get a chance.
2960 static int rcu_pending(void)
2962 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2963 struct rcu_node *rnp = rdp->mynode;
2965 /* Check for CPU stalls, if enabled. */
2966 check_cpu_stall(rdp);
2968 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2969 if (rcu_nohz_full_cpu())
2972 /* Is the RCU core waiting for a quiescent state from this CPU? */
2973 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
2976 /* Does this CPU have callbacks ready to invoke? */
2977 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2980 /* Has RCU gone idle with this CPU needing another grace period? */
2981 if (!rcu_gp_in_progress() &&
2982 rcu_segcblist_is_enabled(&rdp->cblist) &&
2983 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2986 /* Have RCU grace period completed or started? */
2987 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
2988 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
2991 /* Does this CPU need a deferred NOCB wakeup? */
2992 if (rcu_nocb_need_deferred_wakeup(rdp))
3000 * Return true if the specified CPU has any callback. If all_lazy is
3001 * non-NULL, store an indication of whether all callbacks are lazy.
3002 * (If there are no callbacks, all of them are deemed to be lazy.)
3004 static bool rcu_cpu_has_callbacks(bool *all_lazy)
3008 struct rcu_data *rdp;
3010 rdp = this_cpu_ptr(&rcu_data);
3011 if (!rcu_segcblist_empty(&rdp->cblist)) {
3013 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist))
3022 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3023 * the compiler is expected to optimize this away.
3025 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3027 trace_rcu_barrier(rcu_state.name, s, cpu,
3028 atomic_read(&rcu_state.barrier_cpu_count), done);
3032 * RCU callback function for rcu_barrier(). If we are last, wake
3033 * up the task executing rcu_barrier().
3035 static void rcu_barrier_callback(struct rcu_head *rhp)
3037 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3038 rcu_barrier_trace(TPS("LastCB"), -1,
3039 rcu_state.barrier_sequence);
3040 complete(&rcu_state.barrier_completion);
3042 rcu_barrier_trace(TPS("CB"), -1, rcu_state.barrier_sequence);
3047 * Called with preemption disabled, and from cross-cpu IRQ context.
3049 static void rcu_barrier_func(void *unused)
3051 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
3053 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3054 rdp->barrier_head.func = rcu_barrier_callback;
3055 debug_rcu_head_queue(&rdp->barrier_head);
3056 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3057 atomic_inc(&rcu_state.barrier_cpu_count);
3059 debug_rcu_head_unqueue(&rdp->barrier_head);
3060 rcu_barrier_trace(TPS("IRQNQ"), -1,
3061 rcu_state.barrier_sequence);
3066 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3068 * Note that this primitive does not necessarily wait for an RCU grace period
3069 * to complete. For example, if there are no RCU callbacks queued anywhere
3070 * in the system, then rcu_barrier() is within its rights to return
3071 * immediately, without waiting for anything, much less an RCU grace period.
3073 void rcu_barrier(void)
3076 struct rcu_data *rdp;
3077 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3079 rcu_barrier_trace(TPS("Begin"), -1, s);
3081 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3082 mutex_lock(&rcu_state.barrier_mutex);
3084 /* Did someone else do our work for us? */
3085 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
3086 rcu_barrier_trace(TPS("EarlyExit"), -1,
3087 rcu_state.barrier_sequence);
3088 smp_mb(); /* caller's subsequent code after above check. */
3089 mutex_unlock(&rcu_state.barrier_mutex);
3093 /* Mark the start of the barrier operation. */
3094 rcu_seq_start(&rcu_state.barrier_sequence);
3095 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
3098 * Initialize the count to one rather than to zero in order to
3099 * avoid a too-soon return to zero in case of a short grace period
3100 * (or preemption of this task). Exclude CPU-hotplug operations
3101 * to ensure that no offline CPU has callbacks queued.
3103 init_completion(&rcu_state.barrier_completion);
3104 atomic_set(&rcu_state.barrier_cpu_count, 1);
3108 * Force each CPU with callbacks to register a new callback.
3109 * When that callback is invoked, we will know that all of the
3110 * corresponding CPU's preceding callbacks have been invoked.
3112 for_each_possible_cpu(cpu) {
3113 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3115 rdp = per_cpu_ptr(&rcu_data, cpu);
3116 if (rcu_is_nocb_cpu(cpu)) {
3117 if (!rcu_nocb_cpu_needs_barrier(cpu)) {
3118 rcu_barrier_trace(TPS("OfflineNoCB"), cpu,
3119 rcu_state.barrier_sequence);
3121 rcu_barrier_trace(TPS("OnlineNoCB"), cpu,
3122 rcu_state.barrier_sequence);
3123 smp_mb__before_atomic();
3124 atomic_inc(&rcu_state.barrier_cpu_count);
3125 __call_rcu(&rdp->barrier_head,
3126 rcu_barrier_callback, cpu, 0);
3128 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3129 rcu_barrier_trace(TPS("OnlineQ"), cpu,
3130 rcu_state.barrier_sequence);
3131 smp_call_function_single(cpu, rcu_barrier_func, NULL, 1);
3133 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
3134 rcu_state.barrier_sequence);
3140 * Now that we have an rcu_barrier_callback() callback on each
3141 * CPU, and thus each counted, remove the initial count.
3143 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count))
3144 complete(&rcu_state.barrier_completion);
3146 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3147 wait_for_completion(&rcu_state.barrier_completion);
3149 /* Mark the end of the barrier operation. */
3150 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
3151 rcu_seq_end(&rcu_state.barrier_sequence);
3153 /* Other rcu_barrier() invocations can now safely proceed. */
3154 mutex_unlock(&rcu_state.barrier_mutex);
3156 EXPORT_SYMBOL_GPL(rcu_barrier);
3159 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3160 * first CPU in a given leaf rcu_node structure coming online. The caller
3161 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3164 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3168 struct rcu_node *rnp = rnp_leaf;
3170 raw_lockdep_assert_held_rcu_node(rnp_leaf);
3171 WARN_ON_ONCE(rnp->wait_blkd_tasks);
3173 mask = rnp->grpmask;
3177 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3178 oldmask = rnp->qsmaskinit;
3179 rnp->qsmaskinit |= mask;
3180 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3187 * Do boot-time initialization of a CPU's per-CPU RCU data.
3190 rcu_boot_init_percpu_data(int cpu)
3192 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3194 /* Set up local state, ensuring consistent view of global state. */
3195 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3196 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3197 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3198 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3199 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3200 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3201 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3202 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3204 rcu_boot_init_nocb_percpu_data(rdp);
3208 * Invoked early in the CPU-online process, when pretty much all services
3209 * are available. The incoming CPU is not present.
3211 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3212 * offline event can be happening at a given time. Note also that we can
3213 * accept some slop in the rsp->gp_seq access due to the fact that this
3214 * CPU cannot possibly have any RCU callbacks in flight yet.
3216 int rcutree_prepare_cpu(unsigned int cpu)
3218 unsigned long flags;
3219 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3220 struct rcu_node *rnp = rcu_get_root();
3222 /* Set up local state, ensuring consistent view of global state. */
3223 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3224 rdp->qlen_last_fqs_check = 0;
3225 rdp->n_force_qs_snap = rcu_state.n_force_qs;
3226 rdp->blimit = blimit;
3227 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3228 !init_nocb_callback_list(rdp))
3229 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3230 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3231 rcu_dynticks_eqs_online();
3232 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3235 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3236 * propagation up the rcu_node tree will happen at the beginning
3237 * of the next grace period.
3240 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3241 rdp->beenonline = true; /* We have now been online. */
3242 rdp->gp_seq = rnp->gp_seq;
3243 rdp->gp_seq_needed = rnp->gp_seq;
3244 rdp->cpu_no_qs.b.norm = true;
3245 rdp->core_needs_qs = false;
3246 rdp->rcu_iw_pending = false;
3247 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3248 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3249 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3250 rcu_prepare_kthreads(cpu);
3251 rcu_spawn_all_nocb_kthreads(cpu);
3257 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3259 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3261 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3263 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3267 * Near the end of the CPU-online process. Pretty much all services
3268 * enabled, and the CPU is now very much alive.
3270 int rcutree_online_cpu(unsigned int cpu)
3272 unsigned long flags;
3273 struct rcu_data *rdp;
3274 struct rcu_node *rnp;
3276 rdp = per_cpu_ptr(&rcu_data, cpu);
3278 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3279 rnp->ffmask |= rdp->grpmask;
3280 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3281 if (IS_ENABLED(CONFIG_TREE_SRCU))
3282 srcu_online_cpu(cpu);
3283 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3284 return 0; /* Too early in boot for scheduler work. */
3285 sync_sched_exp_online_cleanup(cpu);
3286 rcutree_affinity_setting(cpu, -1);
3291 * Near the beginning of the process. The CPU is still very much alive
3292 * with pretty much all services enabled.
3294 int rcutree_offline_cpu(unsigned int cpu)
3296 unsigned long flags;
3297 struct rcu_data *rdp;
3298 struct rcu_node *rnp;
3300 rdp = per_cpu_ptr(&rcu_data, cpu);
3302 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3303 rnp->ffmask &= ~rdp->grpmask;
3304 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3306 rcutree_affinity_setting(cpu, cpu);
3307 if (IS_ENABLED(CONFIG_TREE_SRCU))
3308 srcu_offline_cpu(cpu);
3312 static DEFINE_PER_CPU(int, rcu_cpu_started);
3315 * Mark the specified CPU as being online so that subsequent grace periods
3316 * (both expedited and normal) will wait on it. Note that this means that
3317 * incoming CPUs are not allowed to use RCU read-side critical sections
3318 * until this function is called. Failing to observe this restriction
3319 * will result in lockdep splats.
3321 * Note that this function is special in that it is invoked directly
3322 * from the incoming CPU rather than from the cpuhp_step mechanism.
3323 * This is because this function must be invoked at a precise location.
3325 void rcu_cpu_starting(unsigned int cpu)
3327 unsigned long flags;
3330 unsigned long oldmask;
3331 struct rcu_data *rdp;
3332 struct rcu_node *rnp;
3334 if (per_cpu(rcu_cpu_started, cpu))
3337 per_cpu(rcu_cpu_started, cpu) = 1;
3339 rdp = per_cpu_ptr(&rcu_data, cpu);
3341 mask = rdp->grpmask;
3342 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3343 rnp->qsmaskinitnext |= mask;
3344 oldmask = rnp->expmaskinitnext;
3345 rnp->expmaskinitnext |= mask;
3346 oldmask ^= rnp->expmaskinitnext;
3347 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3348 /* Allow lockless access for expedited grace periods. */
3349 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */
3350 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3351 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3352 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3353 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3354 /* Report QS -after- changing ->qsmaskinitnext! */
3355 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3357 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3359 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3362 #ifdef CONFIG_HOTPLUG_CPU
3364 * The outgoing function has no further need of RCU, so remove it from
3365 * the rcu_node tree's ->qsmaskinitnext bit masks.
3367 * Note that this function is special in that it is invoked directly
3368 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3369 * This is because this function must be invoked at a precise location.
3371 void rcu_report_dead(unsigned int cpu)
3373 unsigned long flags;
3375 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3376 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3378 /* QS for any half-done expedited grace period. */
3380 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
3382 rcu_preempt_deferred_qs(current);
3384 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3385 mask = rdp->grpmask;
3386 spin_lock(&rcu_state.ofl_lock);
3387 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3388 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3389 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3390 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3391 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3392 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3393 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3395 rnp->qsmaskinitnext &= ~mask;
3396 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3397 spin_unlock(&rcu_state.ofl_lock);
3399 per_cpu(rcu_cpu_started, cpu) = 0;
3403 * The outgoing CPU has just passed through the dying-idle state, and we
3404 * are being invoked from the CPU that was IPIed to continue the offline
3405 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
3407 void rcutree_migrate_callbacks(int cpu)
3409 unsigned long flags;
3410 struct rcu_data *my_rdp;
3411 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3412 struct rcu_node *rnp_root = rcu_get_root();
3415 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3416 return; /* No callbacks to migrate. */
3418 local_irq_save(flags);
3419 my_rdp = this_cpu_ptr(&rcu_data);
3420 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3421 local_irq_restore(flags);
3424 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3425 /* Leverage recent GPs and set GP for new callbacks. */
3426 needwake = rcu_advance_cbs(rnp_root, rdp) ||
3427 rcu_advance_cbs(rnp_root, my_rdp);
3428 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3429 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3430 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3431 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3433 rcu_gp_kthread_wake();
3434 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3435 !rcu_segcblist_empty(&rdp->cblist),
3436 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3437 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3438 rcu_segcblist_first_cb(&rdp->cblist));
3443 * On non-huge systems, use expedited RCU grace periods to make suspend
3444 * and hibernation run faster.
3446 static int rcu_pm_notify(struct notifier_block *self,
3447 unsigned long action, void *hcpu)
3450 case PM_HIBERNATION_PREPARE:
3451 case PM_SUSPEND_PREPARE:
3452 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3455 case PM_POST_HIBERNATION:
3456 case PM_POST_SUSPEND:
3457 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3458 rcu_unexpedite_gp();
3467 * Spawn the kthreads that handle RCU's grace periods.
3469 static int __init rcu_spawn_gp_kthread(void)
3471 unsigned long flags;
3472 int kthread_prio_in = kthread_prio;
3473 struct rcu_node *rnp;
3474 struct sched_param sp;
3475 struct task_struct *t;
3477 /* Force priority into range. */
3478 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
3479 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
3481 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3483 else if (kthread_prio < 0)
3485 else if (kthread_prio > 99)
3488 if (kthread_prio != kthread_prio_in)
3489 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3490 kthread_prio, kthread_prio_in);
3492 rcu_scheduler_fully_active = 1;
3493 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
3495 rnp = rcu_get_root();
3496 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3497 rcu_state.gp_kthread = t;
3499 sp.sched_priority = kthread_prio;
3500 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3502 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3504 rcu_spawn_nocb_kthreads();
3505 rcu_spawn_boost_kthreads();
3508 early_initcall(rcu_spawn_gp_kthread);
3511 * This function is invoked towards the end of the scheduler's
3512 * initialization process. Before this is called, the idle task might
3513 * contain synchronous grace-period primitives (during which time, this idle
3514 * task is booting the system, and such primitives are no-ops). After this
3515 * function is called, any synchronous grace-period primitives are run as
3516 * expedited, with the requesting task driving the grace period forward.
3517 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3518 * runtime RCU functionality.
3520 void rcu_scheduler_starting(void)
3522 WARN_ON(num_online_cpus() != 1);
3523 WARN_ON(nr_context_switches() > 0);
3524 rcu_test_sync_prims();
3525 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3526 rcu_test_sync_prims();
3530 * Helper function for rcu_init() that initializes the rcu_state structure.
3532 static void __init rcu_init_one(void)
3534 static const char * const buf[] = RCU_NODE_NAME_INIT;
3535 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3536 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3537 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3539 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3543 struct rcu_node *rnp;
3545 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3547 /* Silence gcc 4.8 false positive about array index out of range. */
3548 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3549 panic("rcu_init_one: rcu_num_lvls out of range");
3551 /* Initialize the level-tracking arrays. */
3553 for (i = 1; i < rcu_num_lvls; i++)
3554 rcu_state.level[i] =
3555 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
3556 rcu_init_levelspread(levelspread, num_rcu_lvl);
3558 /* Initialize the elements themselves, starting from the leaves. */
3560 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3561 cpustride *= levelspread[i];
3562 rnp = rcu_state.level[i];
3563 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3564 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3565 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3566 &rcu_node_class[i], buf[i]);
3567 raw_spin_lock_init(&rnp->fqslock);
3568 lockdep_set_class_and_name(&rnp->fqslock,
3569 &rcu_fqs_class[i], fqs[i]);
3570 rnp->gp_seq = rcu_state.gp_seq;
3571 rnp->gp_seq_needed = rcu_state.gp_seq;
3572 rnp->completedqs = rcu_state.gp_seq;
3574 rnp->qsmaskinit = 0;
3575 rnp->grplo = j * cpustride;
3576 rnp->grphi = (j + 1) * cpustride - 1;
3577 if (rnp->grphi >= nr_cpu_ids)
3578 rnp->grphi = nr_cpu_ids - 1;
3584 rnp->grpnum = j % levelspread[i - 1];
3585 rnp->grpmask = BIT(rnp->grpnum);
3586 rnp->parent = rcu_state.level[i - 1] +
3587 j / levelspread[i - 1];
3590 INIT_LIST_HEAD(&rnp->blkd_tasks);
3591 rcu_init_one_nocb(rnp);
3592 init_waitqueue_head(&rnp->exp_wq[0]);
3593 init_waitqueue_head(&rnp->exp_wq[1]);
3594 init_waitqueue_head(&rnp->exp_wq[2]);
3595 init_waitqueue_head(&rnp->exp_wq[3]);
3596 spin_lock_init(&rnp->exp_lock);
3600 init_swait_queue_head(&rcu_state.gp_wq);
3601 init_swait_queue_head(&rcu_state.expedited_wq);
3602 rnp = rcu_first_leaf_node();
3603 for_each_possible_cpu(i) {
3604 while (i > rnp->grphi)
3606 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
3607 rcu_boot_init_percpu_data(i);
3612 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3613 * replace the definitions in tree.h because those are needed to size
3614 * the ->node array in the rcu_state structure.
3616 static void __init rcu_init_geometry(void)
3620 int rcu_capacity[RCU_NUM_LVLS];
3623 * Initialize any unspecified boot parameters.
3624 * The default values of jiffies_till_first_fqs and
3625 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3626 * value, which is a function of HZ, then adding one for each
3627 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3629 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3630 if (jiffies_till_first_fqs == ULONG_MAX)
3631 jiffies_till_first_fqs = d;
3632 if (jiffies_till_next_fqs == ULONG_MAX)
3633 jiffies_till_next_fqs = d;
3634 if (jiffies_till_sched_qs == ULONG_MAX)
3635 adjust_jiffies_till_sched_qs();
3637 /* If the compile-time values are accurate, just leave. */
3638 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3639 nr_cpu_ids == NR_CPUS)
3641 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3642 rcu_fanout_leaf, nr_cpu_ids);
3645 * The boot-time rcu_fanout_leaf parameter must be at least two
3646 * and cannot exceed the number of bits in the rcu_node masks.
3647 * Complain and fall back to the compile-time values if this
3648 * limit is exceeded.
3650 if (rcu_fanout_leaf < 2 ||
3651 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3652 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3658 * Compute number of nodes that can be handled an rcu_node tree
3659 * with the given number of levels.
3661 rcu_capacity[0] = rcu_fanout_leaf;
3662 for (i = 1; i < RCU_NUM_LVLS; i++)
3663 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
3666 * The tree must be able to accommodate the configured number of CPUs.
3667 * If this limit is exceeded, fall back to the compile-time values.
3669 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
3670 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3675 /* Calculate the number of levels in the tree. */
3676 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
3678 rcu_num_lvls = i + 1;
3680 /* Calculate the number of rcu_nodes at each level of the tree. */
3681 for (i = 0; i < rcu_num_lvls; i++) {
3682 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
3683 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
3686 /* Calculate the total number of rcu_node structures. */
3688 for (i = 0; i < rcu_num_lvls; i++)
3689 rcu_num_nodes += num_rcu_lvl[i];
3693 * Dump out the structure of the rcu_node combining tree associated
3694 * with the rcu_state structure.
3696 static void __init rcu_dump_rcu_node_tree(void)
3699 struct rcu_node *rnp;
3701 pr_info("rcu_node tree layout dump\n");
3703 rcu_for_each_node_breadth_first(rnp) {
3704 if (rnp->level != level) {
3709 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
3714 struct workqueue_struct *rcu_gp_wq;
3715 struct workqueue_struct *rcu_par_gp_wq;
3717 void __init rcu_init(void)
3721 rcu_early_boot_tests();
3723 rcu_bootup_announce();
3724 rcu_init_geometry();
3727 rcu_dump_rcu_node_tree();
3728 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3731 * We don't need protection against CPU-hotplug here because
3732 * this is called early in boot, before either interrupts
3733 * or the scheduler are operational.
3735 pm_notifier(rcu_pm_notify, 0);
3736 for_each_online_cpu(cpu) {
3737 rcutree_prepare_cpu(cpu);
3738 rcu_cpu_starting(cpu);
3739 rcutree_online_cpu(cpu);
3742 /* Create workqueue for expedited GPs and for Tree SRCU. */
3743 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
3744 WARN_ON(!rcu_gp_wq);
3745 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
3746 WARN_ON(!rcu_par_gp_wq);
3749 #include "tree_exp.h"
3750 #include "tree_plugin.h"