1 // SPDX-License-Identifier: GPL-2.0+
3 * Read-Copy Update mechanism for mutual exclusion
5 * Copyright IBM Corporation, 2008
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com> Hierarchical version
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
14 * For detailed explanation of Read-Copy Update mechanism see -
18 #define pr_fmt(fmt) "rcu: " fmt
20 #include <linux/types.h>
21 #include <linux/kernel.h>
22 #include <linux/init.h>
23 #include <linux/spinlock.h>
24 #include <linux/smp.h>
25 #include <linux/rcupdate_wait.h>
26 #include <linux/interrupt.h>
27 #include <linux/sched.h>
28 #include <linux/sched/debug.h>
29 #include <linux/nmi.h>
30 #include <linux/atomic.h>
31 #include <linux/bitops.h>
32 #include <linux/export.h>
33 #include <linux/completion.h>
34 #include <linux/moduleparam.h>
35 #include <linux/percpu.h>
36 #include <linux/notifier.h>
37 #include <linux/cpu.h>
38 #include <linux/mutex.h>
39 #include <linux/time.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/wait.h>
42 #include <linux/kthread.h>
43 #include <uapi/linux/sched/types.h>
44 #include <linux/prefetch.h>
45 #include <linux/delay.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/trace_events.h>
49 #include <linux/suspend.h>
50 #include <linux/ftrace.h>
51 #include <linux/tick.h>
52 #include <linux/sysrq.h>
53 #include <linux/kprobes.h>
54 #include <linux/gfp.h>
55 #include <linux/oom.h>
56 #include <linux/smpboot.h>
57 #include <linux/jiffies.h>
58 #include <linux/sched/isolation.h>
59 #include <linux/sched/clock.h>
60 #include "../time/tick-internal.h"
65 #ifdef MODULE_PARAM_PREFIX
66 #undef MODULE_PARAM_PREFIX
68 #define MODULE_PARAM_PREFIX "rcutree."
70 /* Data structures. */
73 * Steal a bit from the bottom of ->dynticks for idle entry/exit
74 * control. Initially this is for TLB flushing.
76 #define RCU_DYNTICK_CTRL_MASK 0x1
77 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
78 #ifndef rcu_eqs_special_exit
79 #define rcu_eqs_special_exit() do { } while (0)
82 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
83 .dynticks_nesting = 1,
84 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
85 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
87 struct rcu_state rcu_state = {
88 .level = { &rcu_state.node[0] },
89 .gp_state = RCU_GP_IDLE,
90 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
91 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
94 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
95 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
96 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
99 /* Dump rcu_node combining tree at boot to verify correct setup. */
100 static bool dump_tree;
101 module_param(dump_tree, bool, 0444);
102 /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
103 static bool use_softirq = 1;
104 module_param(use_softirq, bool, 0444);
105 /* Control rcu_node-tree auto-balancing at boot time. */
106 static bool rcu_fanout_exact;
107 module_param(rcu_fanout_exact, bool, 0444);
108 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
109 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
110 module_param(rcu_fanout_leaf, int, 0444);
111 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
112 /* Number of rcu_nodes at specified level. */
113 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
114 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
117 * The rcu_scheduler_active variable is initialized to the value
118 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
119 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
120 * RCU can assume that there is but one task, allowing RCU to (for example)
121 * optimize synchronize_rcu() to a simple barrier(). When this variable
122 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
123 * to detect real grace periods. This variable is also used to suppress
124 * boot-time false positives from lockdep-RCU error checking. Finally, it
125 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
126 * is fully initialized, including all of its kthreads having been spawned.
128 int rcu_scheduler_active __read_mostly;
129 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
132 * The rcu_scheduler_fully_active variable transitions from zero to one
133 * during the early_initcall() processing, which is after the scheduler
134 * is capable of creating new tasks. So RCU processing (for example,
135 * creating tasks for RCU priority boosting) must be delayed until after
136 * rcu_scheduler_fully_active transitions from zero to one. We also
137 * currently delay invocation of any RCU callbacks until after this point.
139 * It might later prove better for people registering RCU callbacks during
140 * early boot to take responsibility for these callbacks, but one step at
143 static int rcu_scheduler_fully_active __read_mostly;
145 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
146 unsigned long gps, unsigned long flags);
147 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
148 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
149 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
150 static void invoke_rcu_core(void);
151 static void rcu_report_exp_rdp(struct rcu_data *rdp);
152 static void sync_sched_exp_online_cleanup(int cpu);
154 /* rcuc/rcub kthread realtime priority */
155 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
156 module_param(kthread_prio, int, 0444);
158 /* Delay in jiffies for grace-period initialization delays, debug only. */
160 static int gp_preinit_delay;
161 module_param(gp_preinit_delay, int, 0444);
162 static int gp_init_delay;
163 module_param(gp_init_delay, int, 0444);
164 static int gp_cleanup_delay;
165 module_param(gp_cleanup_delay, int, 0444);
167 /* Retrieve RCU kthreads priority for rcutorture */
168 int rcu_get_gp_kthreads_prio(void)
172 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
175 * Number of grace periods between delays, normalized by the duration of
176 * the delay. The longer the delay, the more the grace periods between
177 * each delay. The reason for this normalization is that it means that,
178 * for non-zero delays, the overall slowdown of grace periods is constant
179 * regardless of the duration of the delay. This arrangement balances
180 * the need for long delays to increase some race probabilities with the
181 * need for fast grace periods to increase other race probabilities.
183 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
186 * Compute the mask of online CPUs for the specified rcu_node structure.
187 * This will not be stable unless the rcu_node structure's ->lock is
188 * held, but the bit corresponding to the current CPU will be stable
191 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
193 return READ_ONCE(rnp->qsmaskinitnext);
197 * Return true if an RCU grace period is in progress. The READ_ONCE()s
198 * permit this function to be invoked without holding the root rcu_node
199 * structure's ->lock, but of course results can be subject to change.
201 static int rcu_gp_in_progress(void)
203 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
207 * Return the number of callbacks queued on the specified CPU.
208 * Handles both the nocbs and normal cases.
210 static long rcu_get_n_cbs_cpu(int cpu)
212 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
214 if (rcu_segcblist_is_enabled(&rdp->cblist))
215 return rcu_segcblist_n_cbs(&rdp->cblist);
219 void rcu_softirq_qs(void)
222 rcu_preempt_deferred_qs(current);
226 * Record entry into an extended quiescent state. This is only to be
227 * called when not already in an extended quiescent state.
229 static void rcu_dynticks_eqs_enter(void)
231 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
235 * CPUs seeing atomic_add_return() must see prior RCU read-side
236 * critical sections, and we also must force ordering with the
239 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
240 /* Better be in an extended quiescent state! */
241 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
242 (seq & RCU_DYNTICK_CTRL_CTR));
243 /* Better not have special action (TLB flush) pending! */
244 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
245 (seq & RCU_DYNTICK_CTRL_MASK));
249 * Record exit from an extended quiescent state. This is only to be
250 * called from an extended quiescent state.
252 static void rcu_dynticks_eqs_exit(void)
254 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
258 * CPUs seeing atomic_add_return() must see prior idle sojourns,
259 * and we also must force ordering with the next RCU read-side
262 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
263 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
264 !(seq & RCU_DYNTICK_CTRL_CTR));
265 if (seq & RCU_DYNTICK_CTRL_MASK) {
266 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
267 smp_mb__after_atomic(); /* _exit after clearing mask. */
268 /* Prefer duplicate flushes to losing a flush. */
269 rcu_eqs_special_exit();
274 * Reset the current CPU's ->dynticks counter to indicate that the
275 * newly onlined CPU is no longer in an extended quiescent state.
276 * This will either leave the counter unchanged, or increment it
277 * to the next non-quiescent value.
279 * The non-atomic test/increment sequence works because the upper bits
280 * of the ->dynticks counter are manipulated only by the corresponding CPU,
281 * or when the corresponding CPU is offline.
283 static void rcu_dynticks_eqs_online(void)
285 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
287 if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
289 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
293 * Is the current CPU in an extended quiescent state?
295 * No ordering, as we are sampling CPU-local information.
297 bool rcu_dynticks_curr_cpu_in_eqs(void)
299 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
301 return !(atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
305 * Snapshot the ->dynticks counter with full ordering so as to allow
306 * stable comparison of this counter with past and future snapshots.
308 int rcu_dynticks_snap(struct rcu_data *rdp)
310 int snap = atomic_add_return(0, &rdp->dynticks);
312 return snap & ~RCU_DYNTICK_CTRL_MASK;
316 * Return true if the snapshot returned from rcu_dynticks_snap()
317 * indicates that RCU is in an extended quiescent state.
319 static bool rcu_dynticks_in_eqs(int snap)
321 return !(snap & RCU_DYNTICK_CTRL_CTR);
325 * Return true if the CPU corresponding to the specified rcu_data
326 * structure has spent some time in an extended quiescent state since
327 * rcu_dynticks_snap() returned the specified snapshot.
329 static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
331 return snap != rcu_dynticks_snap(rdp);
335 * Set the special (bottom) bit of the specified CPU so that it
336 * will take special action (such as flushing its TLB) on the
337 * next exit from an extended quiescent state. Returns true if
338 * the bit was successfully set, or false if the CPU was not in
339 * an extended quiescent state.
341 bool rcu_eqs_special_set(int cpu)
345 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
348 old = atomic_read(&rdp->dynticks);
349 if (old & RCU_DYNTICK_CTRL_CTR)
351 new = old | RCU_DYNTICK_CTRL_MASK;
352 } while (atomic_cmpxchg(&rdp->dynticks, old, new) != old);
357 * Let the RCU core know that this CPU has gone through the scheduler,
358 * which is a quiescent state. This is called when the need for a
359 * quiescent state is urgent, so we burn an atomic operation and full
360 * memory barriers to let the RCU core know about it, regardless of what
361 * this CPU might (or might not) do in the near future.
363 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
365 * The caller must have disabled interrupts and must not be idle.
367 static void __maybe_unused rcu_momentary_dyntick_idle(void)
371 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
372 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
373 &this_cpu_ptr(&rcu_data)->dynticks);
374 /* It is illegal to call this from idle state. */
375 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
376 rcu_preempt_deferred_qs(current);
380 * rcu_is_cpu_rrupt_from_idle - see if interrupted from idle
382 * If the current CPU is idle and running at a first-level (not nested)
383 * interrupt from idle, return true. The caller must have at least
384 * disabled preemption.
386 static int rcu_is_cpu_rrupt_from_idle(void)
388 /* Called only from within the scheduling-clock interrupt */
389 lockdep_assert_in_irq();
391 /* Check for counter underflows */
392 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
393 "RCU dynticks_nesting counter underflow!");
394 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
395 "RCU dynticks_nmi_nesting counter underflow/zero!");
397 /* Are we at first interrupt nesting level? */
398 if (__this_cpu_read(rcu_data.dynticks_nmi_nesting) != 1)
401 /* Does CPU appear to be idle from an RCU standpoint? */
402 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
405 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch ... */
406 #define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */
407 static long blimit = DEFAULT_RCU_BLIMIT;
408 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
409 static long qhimark = DEFAULT_RCU_QHIMARK;
410 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
411 static long qlowmark = DEFAULT_RCU_QLOMARK;
413 module_param(blimit, long, 0444);
414 module_param(qhimark, long, 0444);
415 module_param(qlowmark, long, 0444);
417 static ulong jiffies_till_first_fqs = ULONG_MAX;
418 static ulong jiffies_till_next_fqs = ULONG_MAX;
419 static bool rcu_kick_kthreads;
420 static int rcu_divisor = 7;
421 module_param(rcu_divisor, int, 0644);
423 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
424 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
425 module_param(rcu_resched_ns, long, 0644);
428 * How long the grace period must be before we start recruiting
429 * quiescent-state help from rcu_note_context_switch().
431 static ulong jiffies_till_sched_qs = ULONG_MAX;
432 module_param(jiffies_till_sched_qs, ulong, 0444);
433 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
434 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
437 * Make sure that we give the grace-period kthread time to detect any
438 * idle CPUs before taking active measures to force quiescent states.
439 * However, don't go below 100 milliseconds, adjusted upwards for really
442 static void adjust_jiffies_till_sched_qs(void)
446 /* If jiffies_till_sched_qs was specified, respect the request. */
447 if (jiffies_till_sched_qs != ULONG_MAX) {
448 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
451 /* Otherwise, set to third fqs scan, but bound below on large system. */
452 j = READ_ONCE(jiffies_till_first_fqs) +
453 2 * READ_ONCE(jiffies_till_next_fqs);
454 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
455 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
456 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
457 WRITE_ONCE(jiffies_to_sched_qs, j);
460 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
463 int ret = kstrtoul(val, 0, &j);
466 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
467 adjust_jiffies_till_sched_qs();
472 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
475 int ret = kstrtoul(val, 0, &j);
478 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
479 adjust_jiffies_till_sched_qs();
484 static struct kernel_param_ops first_fqs_jiffies_ops = {
485 .set = param_set_first_fqs_jiffies,
486 .get = param_get_ulong,
489 static struct kernel_param_ops next_fqs_jiffies_ops = {
490 .set = param_set_next_fqs_jiffies,
491 .get = param_get_ulong,
494 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
495 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
496 module_param(rcu_kick_kthreads, bool, 0644);
498 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
499 static int rcu_pending(void);
502 * Return the number of RCU GPs completed thus far for debug & stats.
504 unsigned long rcu_get_gp_seq(void)
506 return READ_ONCE(rcu_state.gp_seq);
508 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
511 * Return the number of RCU expedited batches completed thus far for
512 * debug & stats. Odd numbers mean that a batch is in progress, even
513 * numbers mean idle. The value returned will thus be roughly double
514 * the cumulative batches since boot.
516 unsigned long rcu_exp_batches_completed(void)
518 return rcu_state.expedited_sequence;
520 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
523 * Return the root node of the rcu_state structure.
525 static struct rcu_node *rcu_get_root(void)
527 return &rcu_state.node[0];
531 * Convert a ->gp_state value to a character string.
533 static const char *gp_state_getname(short gs)
535 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
537 return gp_state_names[gs];
541 * Send along grace-period-related data for rcutorture diagnostics.
543 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
544 unsigned long *gp_seq)
548 *flags = READ_ONCE(rcu_state.gp_flags);
549 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
555 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
558 * Enter an RCU extended quiescent state, which can be either the
559 * idle loop or adaptive-tickless usermode execution.
561 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
562 * the possibility of usermode upcalls having messed up our count
563 * of interrupt nesting level during the prior busy period.
565 static void rcu_eqs_enter(bool user)
567 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
569 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
570 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
571 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
572 rdp->dynticks_nesting == 0);
573 if (rdp->dynticks_nesting != 1) {
574 rdp->dynticks_nesting--;
578 lockdep_assert_irqs_disabled();
579 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
580 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
581 rdp = this_cpu_ptr(&rcu_data);
582 do_nocb_deferred_wakeup(rdp);
583 rcu_prepare_for_idle();
584 rcu_preempt_deferred_qs(current);
585 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
586 rcu_dynticks_eqs_enter();
587 rcu_dynticks_task_enter();
591 * rcu_idle_enter - inform RCU that current CPU is entering idle
593 * Enter idle mode, in other words, -leave- the mode in which RCU
594 * read-side critical sections can occur. (Though RCU read-side
595 * critical sections can occur in irq handlers in idle, a possibility
596 * handled by irq_enter() and irq_exit().)
598 * If you add or remove a call to rcu_idle_enter(), be sure to test with
599 * CONFIG_RCU_EQS_DEBUG=y.
601 void rcu_idle_enter(void)
603 lockdep_assert_irqs_disabled();
604 rcu_eqs_enter(false);
607 #ifdef CONFIG_NO_HZ_FULL
609 * rcu_user_enter - inform RCU that we are resuming userspace.
611 * Enter RCU idle mode right before resuming userspace. No use of RCU
612 * is permitted between this call and rcu_user_exit(). This way the
613 * CPU doesn't need to maintain the tick for RCU maintenance purposes
614 * when the CPU runs in userspace.
616 * If you add or remove a call to rcu_user_enter(), be sure to test with
617 * CONFIG_RCU_EQS_DEBUG=y.
619 void rcu_user_enter(void)
621 lockdep_assert_irqs_disabled();
624 #endif /* CONFIG_NO_HZ_FULL */
627 * If we are returning from the outermost NMI handler that interrupted an
628 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
629 * to let the RCU grace-period handling know that the CPU is back to
632 * If you add or remove a call to rcu_nmi_exit_common(), be sure to test
633 * with CONFIG_RCU_EQS_DEBUG=y.
635 static __always_inline void rcu_nmi_exit_common(bool irq)
637 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
640 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
641 * (We are exiting an NMI handler, so RCU better be paying attention
644 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
645 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
648 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
649 * leave it in non-RCU-idle state.
651 if (rdp->dynticks_nmi_nesting != 1) {
652 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
653 atomic_read(&rdp->dynticks));
654 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
655 rdp->dynticks_nmi_nesting - 2);
659 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
660 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
661 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
664 rcu_prepare_for_idle();
666 rcu_dynticks_eqs_enter();
669 rcu_dynticks_task_enter();
673 * rcu_nmi_exit - inform RCU of exit from NMI context
675 * If you add or remove a call to rcu_nmi_exit(), be sure to test
676 * with CONFIG_RCU_EQS_DEBUG=y.
678 void rcu_nmi_exit(void)
680 rcu_nmi_exit_common(false);
684 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
686 * Exit from an interrupt handler, which might possibly result in entering
687 * idle mode, in other words, leaving the mode in which read-side critical
688 * sections can occur. The caller must have disabled interrupts.
690 * This code assumes that the idle loop never does anything that might
691 * result in unbalanced calls to irq_enter() and irq_exit(). If your
692 * architecture's idle loop violates this assumption, RCU will give you what
693 * you deserve, good and hard. But very infrequently and irreproducibly.
695 * Use things like work queues to work around this limitation.
697 * You have been warned.
699 * If you add or remove a call to rcu_irq_exit(), be sure to test with
700 * CONFIG_RCU_EQS_DEBUG=y.
702 void rcu_irq_exit(void)
704 lockdep_assert_irqs_disabled();
705 rcu_nmi_exit_common(true);
709 * Wrapper for rcu_irq_exit() where interrupts are enabled.
711 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
712 * with CONFIG_RCU_EQS_DEBUG=y.
714 void rcu_irq_exit_irqson(void)
718 local_irq_save(flags);
720 local_irq_restore(flags);
724 * Exit an RCU extended quiescent state, which can be either the
725 * idle loop or adaptive-tickless usermode execution.
727 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
728 * allow for the possibility of usermode upcalls messing up our count of
729 * interrupt nesting level during the busy period that is just now starting.
731 static void rcu_eqs_exit(bool user)
733 struct rcu_data *rdp;
736 lockdep_assert_irqs_disabled();
737 rdp = this_cpu_ptr(&rcu_data);
738 oldval = rdp->dynticks_nesting;
739 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
741 rdp->dynticks_nesting++;
744 rcu_dynticks_task_exit();
745 rcu_dynticks_eqs_exit();
746 rcu_cleanup_after_idle();
747 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
748 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
749 WRITE_ONCE(rdp->dynticks_nesting, 1);
750 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
751 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
755 * rcu_idle_exit - inform RCU that current CPU is leaving idle
757 * Exit idle mode, in other words, -enter- the mode in which RCU
758 * read-side critical sections can occur.
760 * If you add or remove a call to rcu_idle_exit(), be sure to test with
761 * CONFIG_RCU_EQS_DEBUG=y.
763 void rcu_idle_exit(void)
767 local_irq_save(flags);
769 local_irq_restore(flags);
772 #ifdef CONFIG_NO_HZ_FULL
774 * rcu_user_exit - inform RCU that we are exiting userspace.
776 * Exit RCU idle mode while entering the kernel because it can
777 * run a RCU read side critical section anytime.
779 * If you add or remove a call to rcu_user_exit(), be sure to test with
780 * CONFIG_RCU_EQS_DEBUG=y.
782 void rcu_user_exit(void)
786 #endif /* CONFIG_NO_HZ_FULL */
789 * rcu_nmi_enter_common - inform RCU of entry to NMI context
790 * @irq: Is this call from rcu_irq_enter?
792 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
793 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
794 * that the CPU is active. This implementation permits nested NMIs, as
795 * long as the nesting level does not overflow an int. (You will probably
796 * run out of stack space first.)
798 * If you add or remove a call to rcu_nmi_enter_common(), be sure to test
799 * with CONFIG_RCU_EQS_DEBUG=y.
801 static __always_inline void rcu_nmi_enter_common(bool irq)
803 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
806 /* Complain about underflow. */
807 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
810 * If idle from RCU viewpoint, atomically increment ->dynticks
811 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
812 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
813 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
814 * to be in the outermost NMI handler that interrupted an RCU-idle
815 * period (observation due to Andy Lutomirski).
817 if (rcu_dynticks_curr_cpu_in_eqs()) {
820 rcu_dynticks_task_exit();
822 rcu_dynticks_eqs_exit();
825 rcu_cleanup_after_idle();
829 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
830 rdp->dynticks_nmi_nesting,
831 rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
832 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
833 rdp->dynticks_nmi_nesting + incby);
838 * rcu_nmi_enter - inform RCU of entry to NMI context
840 void rcu_nmi_enter(void)
842 rcu_nmi_enter_common(false);
844 NOKPROBE_SYMBOL(rcu_nmi_enter);
847 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
849 * Enter an interrupt handler, which might possibly result in exiting
850 * idle mode, in other words, entering the mode in which read-side critical
851 * sections can occur. The caller must have disabled interrupts.
853 * Note that the Linux kernel is fully capable of entering an interrupt
854 * handler that it never exits, for example when doing upcalls to user mode!
855 * This code assumes that the idle loop never does upcalls to user mode.
856 * If your architecture's idle loop does do upcalls to user mode (or does
857 * anything else that results in unbalanced calls to the irq_enter() and
858 * irq_exit() functions), RCU will give you what you deserve, good and hard.
859 * But very infrequently and irreproducibly.
861 * Use things like work queues to work around this limitation.
863 * You have been warned.
865 * If you add or remove a call to rcu_irq_enter(), be sure to test with
866 * CONFIG_RCU_EQS_DEBUG=y.
868 void rcu_irq_enter(void)
870 lockdep_assert_irqs_disabled();
871 rcu_nmi_enter_common(true);
875 * Wrapper for rcu_irq_enter() where interrupts are enabled.
877 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
878 * with CONFIG_RCU_EQS_DEBUG=y.
880 void rcu_irq_enter_irqson(void)
884 local_irq_save(flags);
886 local_irq_restore(flags);
890 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
892 * Return true if RCU is watching the running CPU, which means that this
893 * CPU can safely enter RCU read-side critical sections. In other words,
894 * if the current CPU is not in its idle loop or is in an interrupt or
895 * NMI handler, return true.
897 bool notrace rcu_is_watching(void)
901 preempt_disable_notrace();
902 ret = !rcu_dynticks_curr_cpu_in_eqs();
903 preempt_enable_notrace();
906 EXPORT_SYMBOL_GPL(rcu_is_watching);
909 * If a holdout task is actually running, request an urgent quiescent
910 * state from its CPU. This is unsynchronized, so migrations can cause
911 * the request to go to the wrong CPU. Which is OK, all that will happen
912 * is that the CPU's next context switch will be a bit slower and next
913 * time around this task will generate another request.
915 void rcu_request_urgent_qs_task(struct task_struct *t)
922 return; /* This task is not running on that CPU. */
923 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
926 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
929 * Is the current CPU online as far as RCU is concerned?
931 * Disable preemption to avoid false positives that could otherwise
932 * happen due to the current CPU number being sampled, this task being
933 * preempted, its old CPU being taken offline, resuming on some other CPU,
934 * then determining that its old CPU is now offline.
936 * Disable checking if in an NMI handler because we cannot safely
937 * report errors from NMI handlers anyway. In addition, it is OK to use
938 * RCU on an offline processor during initial boot, hence the check for
939 * rcu_scheduler_fully_active.
941 bool rcu_lockdep_current_cpu_online(void)
943 struct rcu_data *rdp;
944 struct rcu_node *rnp;
947 if (in_nmi() || !rcu_scheduler_fully_active)
950 rdp = this_cpu_ptr(&rcu_data);
952 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
957 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
959 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
962 * We are reporting a quiescent state on behalf of some other CPU, so
963 * it is our responsibility to check for and handle potential overflow
964 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
965 * After all, the CPU might be in deep idle state, and thus executing no
968 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
970 raw_lockdep_assert_held_rcu_node(rnp);
971 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
973 WRITE_ONCE(rdp->gpwrap, true);
974 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
975 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
979 * Snapshot the specified CPU's dynticks counter so that we can later
980 * credit them with an implicit quiescent state. Return 1 if this CPU
981 * is in dynticks idle mode, which is an extended quiescent state.
983 static int dyntick_save_progress_counter(struct rcu_data *rdp)
985 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
986 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
987 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
988 rcu_gpnum_ovf(rdp->mynode, rdp);
995 * Return true if the specified CPU has passed through a quiescent
996 * state by virtue of being in or having passed through an dynticks
997 * idle state since the last call to dyntick_save_progress_counter()
998 * for this same CPU, or by virtue of having been offline.
1000 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1005 struct rcu_node *rnp = rdp->mynode;
1008 * If the CPU passed through or entered a dynticks idle phase with
1009 * no active irq/NMI handlers, then we can safely pretend that the CPU
1010 * already acknowledged the request to pass through a quiescent
1011 * state. Either way, that CPU cannot possibly be in an RCU
1012 * read-side critical section that started before the beginning
1013 * of the current RCU grace period.
1015 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1016 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1017 rcu_gpnum_ovf(rnp, rdp);
1021 /* If waiting too long on an offline CPU, complain. */
1022 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1023 time_after(jiffies, rcu_state.gp_start + HZ)) {
1025 struct rcu_node *rnp1;
1027 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1028 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1029 __func__, rnp->grplo, rnp->grphi, rnp->level,
1030 (long)rnp->gp_seq, (long)rnp->completedqs);
1031 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1032 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1033 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1034 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1035 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1036 __func__, rdp->cpu, ".o"[onl],
1037 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1038 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1039 return 1; /* Break things loose after complaining. */
1043 * A CPU running for an extended time within the kernel can
1044 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1045 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1046 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1047 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1048 * variable are safe because the assignments are repeated if this
1049 * CPU failed to pass through a quiescent state. This code
1050 * also checks .jiffies_resched in case jiffies_to_sched_qs
1053 jtsq = READ_ONCE(jiffies_to_sched_qs);
1054 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1055 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1056 if (!READ_ONCE(*rnhqp) &&
1057 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1058 time_after(jiffies, rcu_state.jiffies_resched))) {
1059 WRITE_ONCE(*rnhqp, true);
1060 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1061 smp_store_release(ruqp, true);
1062 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1063 WRITE_ONCE(*ruqp, true);
1067 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1068 * The above code handles this, but only for straight cond_resched().
1069 * And some in-kernel loops check need_resched() before calling
1070 * cond_resched(), which defeats the above code for CPUs that are
1071 * running in-kernel with scheduling-clock interrupts disabled.
1072 * So hit them over the head with the resched_cpu() hammer!
1074 if (tick_nohz_full_cpu(rdp->cpu) &&
1076 READ_ONCE(rdp->last_fqs_resched) + jtsq * 3)) {
1077 resched_cpu(rdp->cpu);
1078 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1082 * If more than halfway to RCU CPU stall-warning time, invoke
1083 * resched_cpu() more frequently to try to loosen things up a bit.
1084 * Also check to see if the CPU is getting hammered with interrupts,
1085 * but only once per grace period, just to keep the IPIs down to
1088 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1089 if (time_after(jiffies,
1090 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1091 resched_cpu(rdp->cpu);
1092 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1094 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1095 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1096 (rnp->ffmask & rdp->grpmask)) {
1097 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1098 rdp->rcu_iw_pending = true;
1099 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1100 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1107 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1108 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1109 unsigned long gp_seq_req, const char *s)
1111 trace_rcu_future_grace_period(rcu_state.name, rnp->gp_seq, gp_seq_req,
1112 rnp->level, rnp->grplo, rnp->grphi, s);
1116 * rcu_start_this_gp - Request the start of a particular grace period
1117 * @rnp_start: The leaf node of the CPU from which to start.
1118 * @rdp: The rcu_data corresponding to the CPU from which to start.
1119 * @gp_seq_req: The gp_seq of the grace period to start.
1121 * Start the specified grace period, as needed to handle newly arrived
1122 * callbacks. The required future grace periods are recorded in each
1123 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1124 * is reason to awaken the grace-period kthread.
1126 * The caller must hold the specified rcu_node structure's ->lock, which
1127 * is why the caller is responsible for waking the grace-period kthread.
1129 * Returns true if the GP thread needs to be awakened else false.
1131 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1132 unsigned long gp_seq_req)
1135 struct rcu_node *rnp;
1138 * Use funnel locking to either acquire the root rcu_node
1139 * structure's lock or bail out if the need for this grace period
1140 * has already been recorded -- or if that grace period has in
1141 * fact already started. If there is already a grace period in
1142 * progress in a non-leaf node, no recording is needed because the
1143 * end of the grace period will scan the leaf rcu_node structures.
1144 * Note that rnp_start->lock must not be released.
1146 raw_lockdep_assert_held_rcu_node(rnp_start);
1147 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1148 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1149 if (rnp != rnp_start)
1150 raw_spin_lock_rcu_node(rnp);
1151 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1152 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1153 (rnp != rnp_start &&
1154 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1155 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1159 rnp->gp_seq_needed = gp_seq_req;
1160 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1162 * We just marked the leaf or internal node, and a
1163 * grace period is in progress, which means that
1164 * rcu_gp_cleanup() will see the marking. Bail to
1165 * reduce contention.
1167 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1168 TPS("Startedleaf"));
1171 if (rnp != rnp_start && rnp->parent != NULL)
1172 raw_spin_unlock_rcu_node(rnp);
1174 break; /* At root, and perhaps also leaf. */
1177 /* If GP already in progress, just leave, otherwise start one. */
1178 if (rcu_gp_in_progress()) {
1179 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1182 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1183 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1184 rcu_state.gp_req_activity = jiffies;
1185 if (!rcu_state.gp_kthread) {
1186 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1189 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rcu_state.gp_seq), TPS("newreq"));
1190 ret = true; /* Caller must wake GP kthread. */
1192 /* Push furthest requested GP to leaf node and rcu_data structure. */
1193 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1194 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1195 rdp->gp_seq_needed = rnp->gp_seq_needed;
1197 if (rnp != rnp_start)
1198 raw_spin_unlock_rcu_node(rnp);
1203 * Clean up any old requests for the just-ended grace period. Also return
1204 * whether any additional grace periods have been requested.
1206 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1209 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1211 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1213 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1214 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1215 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1220 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1221 * an interrupt or softirq handler), and don't bother awakening when there
1222 * is nothing for the grace-period kthread to do (as in several CPUs raced
1223 * to awaken, and we lost), and finally don't try to awaken a kthread that
1224 * has not yet been created. If all those checks are passed, track some
1225 * debug information and awaken.
1227 * So why do the self-wakeup when in an interrupt or softirq handler
1228 * in the grace-period kthread's context? Because the kthread might have
1229 * been interrupted just as it was going to sleep, and just after the final
1230 * pre-sleep check of the awaken condition. In this case, a wakeup really
1231 * is required, and is therefore supplied.
1233 static void rcu_gp_kthread_wake(void)
1235 if ((current == rcu_state.gp_kthread &&
1236 !in_irq() && !in_serving_softirq()) ||
1237 !READ_ONCE(rcu_state.gp_flags) ||
1238 !rcu_state.gp_kthread)
1240 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1241 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1242 swake_up_one(&rcu_state.gp_wq);
1246 * If there is room, assign a ->gp_seq number to any callbacks on this
1247 * CPU that have not already been assigned. Also accelerate any callbacks
1248 * that were previously assigned a ->gp_seq number that has since proven
1249 * to be too conservative, which can happen if callbacks get assigned a
1250 * ->gp_seq number while RCU is idle, but with reference to a non-root
1251 * rcu_node structure. This function is idempotent, so it does not hurt
1252 * to call it repeatedly. Returns an flag saying that we should awaken
1253 * the RCU grace-period kthread.
1255 * The caller must hold rnp->lock with interrupts disabled.
1257 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1259 unsigned long gp_seq_req;
1262 rcu_lockdep_assert_cblist_protected(rdp);
1263 raw_lockdep_assert_held_rcu_node(rnp);
1265 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1266 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1270 * Callbacks are often registered with incomplete grace-period
1271 * information. Something about the fact that getting exact
1272 * information requires acquiring a global lock... RCU therefore
1273 * makes a conservative estimate of the grace period number at which
1274 * a given callback will become ready to invoke. The following
1275 * code checks this estimate and improves it when possible, thus
1276 * accelerating callback invocation to an earlier grace-period
1279 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1280 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1281 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1283 /* Trace depending on how much we were able to accelerate. */
1284 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1285 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1287 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1292 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1293 * rcu_node structure's ->lock be held. It consults the cached value
1294 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1295 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1296 * while holding the leaf rcu_node structure's ->lock.
1298 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1299 struct rcu_data *rdp)
1304 rcu_lockdep_assert_cblist_protected(rdp);
1305 c = rcu_seq_snap(&rcu_state.gp_seq);
1306 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1307 /* Old request still live, so mark recent callbacks. */
1308 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1311 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1312 needwake = rcu_accelerate_cbs(rnp, rdp);
1313 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1315 rcu_gp_kthread_wake();
1319 * Move any callbacks whose grace period has completed to the
1320 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1321 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1322 * sublist. This function is idempotent, so it does not hurt to
1323 * invoke it repeatedly. As long as it is not invoked -too- often...
1324 * Returns true if the RCU grace-period kthread needs to be awakened.
1326 * The caller must hold rnp->lock with interrupts disabled.
1328 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1330 rcu_lockdep_assert_cblist_protected(rdp);
1331 raw_lockdep_assert_held_rcu_node(rnp);
1333 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1334 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1338 * Find all callbacks whose ->gp_seq numbers indicate that they
1339 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1341 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1343 /* Classify any remaining callbacks. */
1344 return rcu_accelerate_cbs(rnp, rdp);
1348 * Move and classify callbacks, but only if doing so won't require
1349 * that the RCU grace-period kthread be awakened.
1351 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1352 struct rcu_data *rdp)
1354 rcu_lockdep_assert_cblist_protected(rdp);
1355 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1356 !raw_spin_trylock_rcu_node(rnp))
1358 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1359 raw_spin_unlock_rcu_node(rnp);
1363 * Update CPU-local rcu_data state to record the beginnings and ends of
1364 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1365 * structure corresponding to the current CPU, and must have irqs disabled.
1366 * Returns true if the grace-period kthread needs to be awakened.
1368 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1372 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1373 rcu_segcblist_is_offloaded(&rdp->cblist);
1375 raw_lockdep_assert_held_rcu_node(rnp);
1377 if (rdp->gp_seq == rnp->gp_seq)
1378 return false; /* Nothing to do. */
1380 /* Handle the ends of any preceding grace periods first. */
1381 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1382 unlikely(READ_ONCE(rdp->gpwrap))) {
1384 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1385 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1388 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1391 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1392 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1393 unlikely(READ_ONCE(rdp->gpwrap))) {
1395 * If the current grace period is waiting for this CPU,
1396 * set up to detect a quiescent state, otherwise don't
1397 * go looking for one.
1399 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1400 need_gp = !!(rnp->qsmask & rdp->grpmask);
1401 rdp->cpu_no_qs.b.norm = need_gp;
1402 rdp->core_needs_qs = need_gp;
1403 zero_cpu_stall_ticks(rdp);
1405 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1406 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1407 rdp->gp_seq_needed = rnp->gp_seq_needed;
1408 WRITE_ONCE(rdp->gpwrap, false);
1409 rcu_gpnum_ovf(rnp, rdp);
1413 static void note_gp_changes(struct rcu_data *rdp)
1415 unsigned long flags;
1417 struct rcu_node *rnp;
1419 local_irq_save(flags);
1421 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1422 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1423 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1424 local_irq_restore(flags);
1427 needwake = __note_gp_changes(rnp, rdp);
1428 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1430 rcu_gp_kthread_wake();
1433 static void rcu_gp_slow(int delay)
1436 !(rcu_seq_ctr(rcu_state.gp_seq) %
1437 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1438 schedule_timeout_uninterruptible(delay);
1442 * Initialize a new grace period. Return false if no grace period required.
1444 static bool rcu_gp_init(void)
1446 unsigned long flags;
1447 unsigned long oldmask;
1449 struct rcu_data *rdp;
1450 struct rcu_node *rnp = rcu_get_root();
1452 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1453 raw_spin_lock_irq_rcu_node(rnp);
1454 if (!READ_ONCE(rcu_state.gp_flags)) {
1455 /* Spurious wakeup, tell caller to go back to sleep. */
1456 raw_spin_unlock_irq_rcu_node(rnp);
1459 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1461 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1463 * Grace period already in progress, don't start another.
1464 * Not supposed to be able to happen.
1466 raw_spin_unlock_irq_rcu_node(rnp);
1470 /* Advance to a new grace period and initialize state. */
1471 record_gp_stall_check_time();
1472 /* Record GP times before starting GP, hence rcu_seq_start(). */
1473 rcu_seq_start(&rcu_state.gp_seq);
1474 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1475 raw_spin_unlock_irq_rcu_node(rnp);
1478 * Apply per-leaf buffered online and offline operations to the
1479 * rcu_node tree. Note that this new grace period need not wait
1480 * for subsequent online CPUs, and that quiescent-state forcing
1481 * will handle subsequent offline CPUs.
1483 rcu_state.gp_state = RCU_GP_ONOFF;
1484 rcu_for_each_leaf_node(rnp) {
1485 raw_spin_lock(&rcu_state.ofl_lock);
1486 raw_spin_lock_irq_rcu_node(rnp);
1487 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1488 !rnp->wait_blkd_tasks) {
1489 /* Nothing to do on this leaf rcu_node structure. */
1490 raw_spin_unlock_irq_rcu_node(rnp);
1491 raw_spin_unlock(&rcu_state.ofl_lock);
1495 /* Record old state, apply changes to ->qsmaskinit field. */
1496 oldmask = rnp->qsmaskinit;
1497 rnp->qsmaskinit = rnp->qsmaskinitnext;
1499 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1500 if (!oldmask != !rnp->qsmaskinit) {
1501 if (!oldmask) { /* First online CPU for rcu_node. */
1502 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1503 rcu_init_new_rnp(rnp);
1504 } else if (rcu_preempt_has_tasks(rnp)) {
1505 rnp->wait_blkd_tasks = true; /* blocked tasks */
1506 } else { /* Last offline CPU and can propagate. */
1507 rcu_cleanup_dead_rnp(rnp);
1512 * If all waited-on tasks from prior grace period are
1513 * done, and if all this rcu_node structure's CPUs are
1514 * still offline, propagate up the rcu_node tree and
1515 * clear ->wait_blkd_tasks. Otherwise, if one of this
1516 * rcu_node structure's CPUs has since come back online,
1517 * simply clear ->wait_blkd_tasks.
1519 if (rnp->wait_blkd_tasks &&
1520 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1521 rnp->wait_blkd_tasks = false;
1522 if (!rnp->qsmaskinit)
1523 rcu_cleanup_dead_rnp(rnp);
1526 raw_spin_unlock_irq_rcu_node(rnp);
1527 raw_spin_unlock(&rcu_state.ofl_lock);
1529 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1532 * Set the quiescent-state-needed bits in all the rcu_node
1533 * structures for all currently online CPUs in breadth-first
1534 * order, starting from the root rcu_node structure, relying on the
1535 * layout of the tree within the rcu_state.node[] array. Note that
1536 * other CPUs will access only the leaves of the hierarchy, thus
1537 * seeing that no grace period is in progress, at least until the
1538 * corresponding leaf node has been initialized.
1540 * The grace period cannot complete until the initialization
1541 * process finishes, because this kthread handles both.
1543 rcu_state.gp_state = RCU_GP_INIT;
1544 rcu_for_each_node_breadth_first(rnp) {
1545 rcu_gp_slow(gp_init_delay);
1546 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1547 rdp = this_cpu_ptr(&rcu_data);
1548 rcu_preempt_check_blocked_tasks(rnp);
1549 rnp->qsmask = rnp->qsmaskinit;
1550 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1551 if (rnp == rdp->mynode)
1552 (void)__note_gp_changes(rnp, rdp);
1553 rcu_preempt_boost_start_gp(rnp);
1554 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1555 rnp->level, rnp->grplo,
1556 rnp->grphi, rnp->qsmask);
1557 /* Quiescent states for tasks on any now-offline CPUs. */
1558 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1559 rnp->rcu_gp_init_mask = mask;
1560 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1561 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1563 raw_spin_unlock_irq_rcu_node(rnp);
1564 cond_resched_tasks_rcu_qs();
1565 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1572 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1575 static bool rcu_gp_fqs_check_wake(int *gfp)
1577 struct rcu_node *rnp = rcu_get_root();
1579 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1580 *gfp = READ_ONCE(rcu_state.gp_flags);
1581 if (*gfp & RCU_GP_FLAG_FQS)
1584 /* The current grace period has completed. */
1585 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1592 * Do one round of quiescent-state forcing.
1594 static void rcu_gp_fqs(bool first_time)
1596 struct rcu_node *rnp = rcu_get_root();
1598 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1599 rcu_state.n_force_qs++;
1601 /* Collect dyntick-idle snapshots. */
1602 force_qs_rnp(dyntick_save_progress_counter);
1604 /* Handle dyntick-idle and offline CPUs. */
1605 force_qs_rnp(rcu_implicit_dynticks_qs);
1607 /* Clear flag to prevent immediate re-entry. */
1608 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1609 raw_spin_lock_irq_rcu_node(rnp);
1610 WRITE_ONCE(rcu_state.gp_flags,
1611 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1612 raw_spin_unlock_irq_rcu_node(rnp);
1617 * Loop doing repeated quiescent-state forcing until the grace period ends.
1619 static void rcu_gp_fqs_loop(void)
1625 struct rcu_node *rnp = rcu_get_root();
1627 first_gp_fqs = true;
1628 j = READ_ONCE(jiffies_till_first_fqs);
1632 rcu_state.jiffies_force_qs = jiffies + j;
1633 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1634 jiffies + (j ? 3 * j : 2));
1636 trace_rcu_grace_period(rcu_state.name,
1637 READ_ONCE(rcu_state.gp_seq),
1639 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1640 ret = swait_event_idle_timeout_exclusive(
1641 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1642 rcu_state.gp_state = RCU_GP_DOING_FQS;
1643 /* Locking provides needed memory barriers. */
1644 /* If grace period done, leave loop. */
1645 if (!READ_ONCE(rnp->qsmask) &&
1646 !rcu_preempt_blocked_readers_cgp(rnp))
1648 /* If time for quiescent-state forcing, do it. */
1649 if (ULONG_CMP_GE(jiffies, rcu_state.jiffies_force_qs) ||
1650 (gf & RCU_GP_FLAG_FQS)) {
1651 trace_rcu_grace_period(rcu_state.name,
1652 READ_ONCE(rcu_state.gp_seq),
1654 rcu_gp_fqs(first_gp_fqs);
1655 first_gp_fqs = false;
1656 trace_rcu_grace_period(rcu_state.name,
1657 READ_ONCE(rcu_state.gp_seq),
1659 cond_resched_tasks_rcu_qs();
1660 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1661 ret = 0; /* Force full wait till next FQS. */
1662 j = READ_ONCE(jiffies_till_next_fqs);
1664 /* Deal with stray signal. */
1665 cond_resched_tasks_rcu_qs();
1666 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1667 WARN_ON(signal_pending(current));
1668 trace_rcu_grace_period(rcu_state.name,
1669 READ_ONCE(rcu_state.gp_seq),
1671 ret = 1; /* Keep old FQS timing. */
1673 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1676 j = rcu_state.jiffies_force_qs - j;
1682 * Clean up after the old grace period.
1684 static void rcu_gp_cleanup(void)
1686 unsigned long gp_duration;
1687 bool needgp = false;
1688 unsigned long new_gp_seq;
1690 struct rcu_data *rdp;
1691 struct rcu_node *rnp = rcu_get_root();
1692 struct swait_queue_head *sq;
1694 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1695 raw_spin_lock_irq_rcu_node(rnp);
1696 rcu_state.gp_end = jiffies;
1697 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1698 if (gp_duration > rcu_state.gp_max)
1699 rcu_state.gp_max = gp_duration;
1702 * We know the grace period is complete, but to everyone else
1703 * it appears to still be ongoing. But it is also the case
1704 * that to everyone else it looks like there is nothing that
1705 * they can do to advance the grace period. It is therefore
1706 * safe for us to drop the lock in order to mark the grace
1707 * period as completed in all of the rcu_node structures.
1709 raw_spin_unlock_irq_rcu_node(rnp);
1712 * Propagate new ->gp_seq value to rcu_node structures so that
1713 * other CPUs don't have to wait until the start of the next grace
1714 * period to process their callbacks. This also avoids some nasty
1715 * RCU grace-period initialization races by forcing the end of
1716 * the current grace period to be completely recorded in all of
1717 * the rcu_node structures before the beginning of the next grace
1718 * period is recorded in any of the rcu_node structures.
1720 new_gp_seq = rcu_state.gp_seq;
1721 rcu_seq_end(&new_gp_seq);
1722 rcu_for_each_node_breadth_first(rnp) {
1723 raw_spin_lock_irq_rcu_node(rnp);
1724 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1725 dump_blkd_tasks(rnp, 10);
1726 WARN_ON_ONCE(rnp->qsmask);
1727 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1728 rdp = this_cpu_ptr(&rcu_data);
1729 if (rnp == rdp->mynode)
1730 needgp = __note_gp_changes(rnp, rdp) || needgp;
1731 /* smp_mb() provided by prior unlock-lock pair. */
1732 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1733 sq = rcu_nocb_gp_get(rnp);
1734 raw_spin_unlock_irq_rcu_node(rnp);
1735 rcu_nocb_gp_cleanup(sq);
1736 cond_resched_tasks_rcu_qs();
1737 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1738 rcu_gp_slow(gp_cleanup_delay);
1740 rnp = rcu_get_root();
1741 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1743 /* Declare grace period done, trace first to use old GP number. */
1744 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1745 rcu_seq_end(&rcu_state.gp_seq);
1746 rcu_state.gp_state = RCU_GP_IDLE;
1747 /* Check for GP requests since above loop. */
1748 rdp = this_cpu_ptr(&rcu_data);
1749 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1750 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1751 TPS("CleanupMore"));
1754 /* Advance CBs to reduce false positives below. */
1755 offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1756 rcu_segcblist_is_offloaded(&rdp->cblist);
1757 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1758 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1759 rcu_state.gp_req_activity = jiffies;
1760 trace_rcu_grace_period(rcu_state.name,
1761 READ_ONCE(rcu_state.gp_seq),
1764 WRITE_ONCE(rcu_state.gp_flags,
1765 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1767 raw_spin_unlock_irq_rcu_node(rnp);
1771 * Body of kthread that handles grace periods.
1773 static int __noreturn rcu_gp_kthread(void *unused)
1775 rcu_bind_gp_kthread();
1778 /* Handle grace-period start. */
1780 trace_rcu_grace_period(rcu_state.name,
1781 READ_ONCE(rcu_state.gp_seq),
1783 rcu_state.gp_state = RCU_GP_WAIT_GPS;
1784 swait_event_idle_exclusive(rcu_state.gp_wq,
1785 READ_ONCE(rcu_state.gp_flags) &
1787 rcu_state.gp_state = RCU_GP_DONE_GPS;
1788 /* Locking provides needed memory barrier. */
1791 cond_resched_tasks_rcu_qs();
1792 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1793 WARN_ON(signal_pending(current));
1794 trace_rcu_grace_period(rcu_state.name,
1795 READ_ONCE(rcu_state.gp_seq),
1799 /* Handle quiescent-state forcing. */
1802 /* Handle grace-period end. */
1803 rcu_state.gp_state = RCU_GP_CLEANUP;
1805 rcu_state.gp_state = RCU_GP_CLEANED;
1810 * Report a full set of quiescent states to the rcu_state data structure.
1811 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1812 * another grace period is required. Whether we wake the grace-period
1813 * kthread or it awakens itself for the next round of quiescent-state
1814 * forcing, that kthread will clean up after the just-completed grace
1815 * period. Note that the caller must hold rnp->lock, which is released
1818 static void rcu_report_qs_rsp(unsigned long flags)
1819 __releases(rcu_get_root()->lock)
1821 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1822 WARN_ON_ONCE(!rcu_gp_in_progress());
1823 WRITE_ONCE(rcu_state.gp_flags,
1824 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1825 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1826 rcu_gp_kthread_wake();
1830 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1831 * Allows quiescent states for a group of CPUs to be reported at one go
1832 * to the specified rcu_node structure, though all the CPUs in the group
1833 * must be represented by the same rcu_node structure (which need not be a
1834 * leaf rcu_node structure, though it often will be). The gps parameter
1835 * is the grace-period snapshot, which means that the quiescent states
1836 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1837 * must be held upon entry, and it is released before return.
1839 * As a special case, if mask is zero, the bit-already-cleared check is
1840 * disabled. This allows propagating quiescent state due to resumed tasks
1841 * during grace-period initialization.
1843 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1844 unsigned long gps, unsigned long flags)
1845 __releases(rnp->lock)
1847 unsigned long oldmask = 0;
1848 struct rcu_node *rnp_c;
1850 raw_lockdep_assert_held_rcu_node(rnp);
1852 /* Walk up the rcu_node hierarchy. */
1854 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1857 * Our bit has already been cleared, or the
1858 * relevant grace period is already over, so done.
1860 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1863 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1864 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1865 rcu_preempt_blocked_readers_cgp(rnp));
1866 rnp->qsmask &= ~mask;
1867 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1868 mask, rnp->qsmask, rnp->level,
1869 rnp->grplo, rnp->grphi,
1871 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1873 /* Other bits still set at this level, so done. */
1874 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1877 rnp->completedqs = rnp->gp_seq;
1878 mask = rnp->grpmask;
1879 if (rnp->parent == NULL) {
1881 /* No more levels. Exit loop holding root lock. */
1885 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1888 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1889 oldmask = rnp_c->qsmask;
1893 * Get here if we are the last CPU to pass through a quiescent
1894 * state for this grace period. Invoke rcu_report_qs_rsp()
1895 * to clean up and start the next grace period if one is needed.
1897 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1901 * Record a quiescent state for all tasks that were previously queued
1902 * on the specified rcu_node structure and that were blocking the current
1903 * RCU grace period. The caller must hold the corresponding rnp->lock with
1904 * irqs disabled, and this lock is released upon return, but irqs remain
1907 static void __maybe_unused
1908 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1909 __releases(rnp->lock)
1913 struct rcu_node *rnp_p;
1915 raw_lockdep_assert_held_rcu_node(rnp);
1916 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPTION)) ||
1917 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1919 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1920 return; /* Still need more quiescent states! */
1923 rnp->completedqs = rnp->gp_seq;
1924 rnp_p = rnp->parent;
1925 if (rnp_p == NULL) {
1927 * Only one rcu_node structure in the tree, so don't
1928 * try to report up to its nonexistent parent!
1930 rcu_report_qs_rsp(flags);
1934 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1936 mask = rnp->grpmask;
1937 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1938 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1939 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1943 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1944 * structure. This must be called from the specified CPU.
1947 rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
1949 unsigned long flags;
1951 bool needwake = false;
1952 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1953 rcu_segcblist_is_offloaded(&rdp->cblist);
1954 struct rcu_node *rnp;
1957 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1958 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
1962 * The grace period in which this quiescent state was
1963 * recorded has ended, so don't report it upwards.
1964 * We will instead need a new quiescent state that lies
1965 * within the current grace period.
1967 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
1968 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1971 mask = rdp->grpmask;
1972 rdp->core_needs_qs = false;
1973 if ((rnp->qsmask & mask) == 0) {
1974 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1977 * This GP can't end until cpu checks in, so all of our
1978 * callbacks can be processed during the next GP.
1981 needwake = rcu_accelerate_cbs(rnp, rdp);
1983 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1984 /* ^^^ Released rnp->lock */
1986 rcu_gp_kthread_wake();
1991 * Check to see if there is a new grace period of which this CPU
1992 * is not yet aware, and if so, set up local rcu_data state for it.
1993 * Otherwise, see if this CPU has just passed through its first
1994 * quiescent state for this grace period, and record that fact if so.
1997 rcu_check_quiescent_state(struct rcu_data *rdp)
1999 /* Check for grace-period ends and beginnings. */
2000 note_gp_changes(rdp);
2003 * Does this CPU still need to do its part for current grace period?
2004 * If no, return and let the other CPUs do their part as well.
2006 if (!rdp->core_needs_qs)
2010 * Was there a quiescent state since the beginning of the grace
2011 * period? If no, then exit and wait for the next call.
2013 if (rdp->cpu_no_qs.b.norm)
2017 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2020 rcu_report_qs_rdp(rdp->cpu, rdp);
2024 * Near the end of the offline process. Trace the fact that this CPU
2027 int rcutree_dying_cpu(unsigned int cpu)
2030 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2031 struct rcu_node *rnp = rdp->mynode;
2033 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2036 blkd = !!(rnp->qsmask & rdp->grpmask);
2037 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq,
2038 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2043 * All CPUs for the specified rcu_node structure have gone offline,
2044 * and all tasks that were preempted within an RCU read-side critical
2045 * section while running on one of those CPUs have since exited their RCU
2046 * read-side critical section. Some other CPU is reporting this fact with
2047 * the specified rcu_node structure's ->lock held and interrupts disabled.
2048 * This function therefore goes up the tree of rcu_node structures,
2049 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2050 * the leaf rcu_node structure's ->qsmaskinit field has already been
2053 * This function does check that the specified rcu_node structure has
2054 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2055 * prematurely. That said, invoking it after the fact will cost you
2056 * a needless lock acquisition. So once it has done its work, don't
2059 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2062 struct rcu_node *rnp = rnp_leaf;
2064 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2065 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2066 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2067 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2070 mask = rnp->grpmask;
2074 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2075 rnp->qsmaskinit &= ~mask;
2076 /* Between grace periods, so better already be zero! */
2077 WARN_ON_ONCE(rnp->qsmask);
2078 if (rnp->qsmaskinit) {
2079 raw_spin_unlock_rcu_node(rnp);
2080 /* irqs remain disabled. */
2083 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2088 * The CPU has been completely removed, and some other CPU is reporting
2089 * this fact from process context. Do the remainder of the cleanup.
2090 * There can only be one CPU hotplug operation at a time, so no need for
2093 int rcutree_dead_cpu(unsigned int cpu)
2095 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2096 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2098 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2101 /* Adjust any no-longer-needed kthreads. */
2102 rcu_boost_kthread_setaffinity(rnp, -1);
2103 /* Do any needed no-CB deferred wakeups from this CPU. */
2104 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2109 * Invoke any RCU callbacks that have made it to the end of their grace
2110 * period. Thottle as specified by rdp->blimit.
2112 static void rcu_do_batch(struct rcu_data *rdp)
2114 unsigned long flags;
2115 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2116 rcu_segcblist_is_offloaded(&rdp->cblist);
2117 struct rcu_head *rhp;
2118 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2120 long pending, tlimit = 0;
2122 /* If no callbacks are ready, just return. */
2123 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2124 trace_rcu_batch_start(rcu_state.name,
2125 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2126 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2127 trace_rcu_batch_end(rcu_state.name, 0,
2128 !rcu_segcblist_empty(&rdp->cblist),
2129 need_resched(), is_idle_task(current),
2130 rcu_is_callbacks_kthread());
2135 * Extract the list of ready callbacks, disabling to prevent
2136 * races with call_rcu() from interrupt handlers. Leave the
2137 * callback counts, as rcu_barrier() needs to be conservative.
2139 local_irq_save(flags);
2141 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2142 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2143 bl = max(rdp->blimit, pending >> rcu_divisor);
2144 if (unlikely(bl > 100))
2145 tlimit = local_clock() + rcu_resched_ns;
2146 trace_rcu_batch_start(rcu_state.name,
2147 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2148 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2149 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2151 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2152 rcu_nocb_unlock_irqrestore(rdp, flags);
2154 /* Invoke callbacks. */
2155 rhp = rcu_cblist_dequeue(&rcl);
2156 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2157 debug_rcu_head_unqueue(rhp);
2158 if (__rcu_reclaim(rcu_state.name, rhp))
2159 rcu_cblist_dequeued_lazy(&rcl);
2161 * Stop only if limit reached and CPU has something to do.
2162 * Note: The rcl structure counts down from zero.
2164 if (-rcl.len >= bl && !offloaded &&
2166 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2168 if (unlikely(tlimit)) {
2169 /* only call local_clock() every 32 callbacks */
2170 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2172 /* Exceeded the time limit, so leave. */
2176 WARN_ON_ONCE(in_serving_softirq());
2178 lockdep_assert_irqs_enabled();
2179 cond_resched_tasks_rcu_qs();
2180 lockdep_assert_irqs_enabled();
2185 local_irq_save(flags);
2188 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2189 is_idle_task(current), rcu_is_callbacks_kthread());
2191 /* Update counts and requeue any remaining callbacks. */
2192 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2193 smp_mb(); /* List handling before counting for rcu_barrier(). */
2194 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2196 /* Reinstate batch limit if we have worked down the excess. */
2197 count = rcu_segcblist_n_cbs(&rdp->cblist);
2198 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2199 rdp->blimit = blimit;
2201 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2202 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2203 rdp->qlen_last_fqs_check = 0;
2204 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2205 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2206 rdp->qlen_last_fqs_check = count;
2209 * The following usually indicates a double call_rcu(). To track
2210 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2212 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2213 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2214 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2216 rcu_nocb_unlock_irqrestore(rdp, flags);
2218 /* Re-invoke RCU core processing if there are callbacks remaining. */
2219 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2224 * This function is invoked from each scheduling-clock interrupt,
2225 * and checks to see if this CPU is in a non-context-switch quiescent
2226 * state, for example, user mode or idle loop. It also schedules RCU
2227 * core processing. If the current grace period has gone on too long,
2228 * it will ask the scheduler to manufacture a context switch for the sole
2229 * purpose of providing a providing the needed quiescent state.
2231 void rcu_sched_clock_irq(int user)
2233 trace_rcu_utilization(TPS("Start scheduler-tick"));
2234 raw_cpu_inc(rcu_data.ticks_this_gp);
2235 /* The load-acquire pairs with the store-release setting to true. */
2236 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2237 /* Idle and userspace execution already are quiescent states. */
2238 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2239 set_tsk_need_resched(current);
2240 set_preempt_need_resched();
2242 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2244 rcu_flavor_sched_clock_irq(user);
2248 trace_rcu_utilization(TPS("End scheduler-tick"));
2252 * Scan the leaf rcu_node structures. For each structure on which all
2253 * CPUs have reported a quiescent state and on which there are tasks
2254 * blocking the current grace period, initiate RCU priority boosting.
2255 * Otherwise, invoke the specified function to check dyntick state for
2256 * each CPU that has not yet reported a quiescent state.
2258 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2261 unsigned long flags;
2263 struct rcu_node *rnp;
2265 rcu_for_each_leaf_node(rnp) {
2266 cond_resched_tasks_rcu_qs();
2268 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2269 if (rnp->qsmask == 0) {
2270 if (!IS_ENABLED(CONFIG_PREEMPTION) ||
2271 rcu_preempt_blocked_readers_cgp(rnp)) {
2273 * No point in scanning bits because they
2274 * are all zero. But we might need to
2275 * priority-boost blocked readers.
2277 rcu_initiate_boost(rnp, flags);
2278 /* rcu_initiate_boost() releases rnp->lock */
2281 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2284 for_each_leaf_node_possible_cpu(rnp, cpu) {
2285 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2286 if ((rnp->qsmask & bit) != 0) {
2287 if (f(per_cpu_ptr(&rcu_data, cpu)))
2292 /* Idle/offline CPUs, report (releases rnp->lock). */
2293 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2295 /* Nothing to do here, so just drop the lock. */
2296 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2302 * Force quiescent states on reluctant CPUs, and also detect which
2303 * CPUs are in dyntick-idle mode.
2305 void rcu_force_quiescent_state(void)
2307 unsigned long flags;
2309 struct rcu_node *rnp;
2310 struct rcu_node *rnp_old = NULL;
2312 /* Funnel through hierarchy to reduce memory contention. */
2313 rnp = __this_cpu_read(rcu_data.mynode);
2314 for (; rnp != NULL; rnp = rnp->parent) {
2315 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2316 !raw_spin_trylock(&rnp->fqslock);
2317 if (rnp_old != NULL)
2318 raw_spin_unlock(&rnp_old->fqslock);
2323 /* rnp_old == rcu_get_root(), rnp == NULL. */
2325 /* Reached the root of the rcu_node tree, acquire lock. */
2326 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2327 raw_spin_unlock(&rnp_old->fqslock);
2328 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2329 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2330 return; /* Someone beat us to it. */
2332 WRITE_ONCE(rcu_state.gp_flags,
2333 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2334 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2335 rcu_gp_kthread_wake();
2337 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2339 /* Perform RCU core processing work for the current CPU. */
2340 static __latent_entropy void rcu_core(void)
2342 unsigned long flags;
2343 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2344 struct rcu_node *rnp = rdp->mynode;
2345 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2346 rcu_segcblist_is_offloaded(&rdp->cblist);
2348 if (cpu_is_offline(smp_processor_id()))
2350 trace_rcu_utilization(TPS("Start RCU core"));
2351 WARN_ON_ONCE(!rdp->beenonline);
2353 /* Report any deferred quiescent states if preemption enabled. */
2354 if (!(preempt_count() & PREEMPT_MASK)) {
2355 rcu_preempt_deferred_qs(current);
2356 } else if (rcu_preempt_need_deferred_qs(current)) {
2357 set_tsk_need_resched(current);
2358 set_preempt_need_resched();
2361 /* Update RCU state based on any recent quiescent states. */
2362 rcu_check_quiescent_state(rdp);
2364 /* No grace period and unregistered callbacks? */
2365 if (!rcu_gp_in_progress() &&
2366 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2367 local_irq_save(flags);
2368 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2369 rcu_accelerate_cbs_unlocked(rnp, rdp);
2370 local_irq_restore(flags);
2373 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2375 /* If there are callbacks ready, invoke them. */
2376 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2377 likely(READ_ONCE(rcu_scheduler_fully_active)))
2380 /* Do any needed deferred wakeups of rcuo kthreads. */
2381 do_nocb_deferred_wakeup(rdp);
2382 trace_rcu_utilization(TPS("End RCU core"));
2385 static void rcu_core_si(struct softirq_action *h)
2390 static void rcu_wake_cond(struct task_struct *t, int status)
2393 * If the thread is yielding, only wake it when this
2394 * is invoked from idle
2396 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2400 static void invoke_rcu_core_kthread(void)
2402 struct task_struct *t;
2403 unsigned long flags;
2405 local_irq_save(flags);
2406 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2407 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2408 if (t != NULL && t != current)
2409 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2410 local_irq_restore(flags);
2414 * Wake up this CPU's rcuc kthread to do RCU core processing.
2416 static void invoke_rcu_core(void)
2418 if (!cpu_online(smp_processor_id()))
2421 raise_softirq(RCU_SOFTIRQ);
2423 invoke_rcu_core_kthread();
2426 static void rcu_cpu_kthread_park(unsigned int cpu)
2428 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2431 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2433 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2437 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2438 * the RCU softirq used in configurations of RCU that do not support RCU
2439 * priority boosting.
2441 static void rcu_cpu_kthread(unsigned int cpu)
2443 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2444 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2447 for (spincnt = 0; spincnt < 10; spincnt++) {
2448 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
2450 *statusp = RCU_KTHREAD_RUNNING;
2451 local_irq_disable();
2459 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2460 *statusp = RCU_KTHREAD_WAITING;
2464 *statusp = RCU_KTHREAD_YIELDING;
2465 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2466 schedule_timeout_interruptible(2);
2467 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2468 *statusp = RCU_KTHREAD_WAITING;
2471 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2472 .store = &rcu_data.rcu_cpu_kthread_task,
2473 .thread_should_run = rcu_cpu_kthread_should_run,
2474 .thread_fn = rcu_cpu_kthread,
2475 .thread_comm = "rcuc/%u",
2476 .setup = rcu_cpu_kthread_setup,
2477 .park = rcu_cpu_kthread_park,
2481 * Spawn per-CPU RCU core processing kthreads.
2483 static int __init rcu_spawn_core_kthreads(void)
2487 for_each_possible_cpu(cpu)
2488 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2489 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2491 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2492 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2495 early_initcall(rcu_spawn_core_kthreads);
2498 * Handle any core-RCU processing required by a call_rcu() invocation.
2500 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2501 unsigned long flags)
2504 * If called from an extended quiescent state, invoke the RCU
2505 * core in order to force a re-evaluation of RCU's idleness.
2507 if (!rcu_is_watching())
2510 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2511 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2515 * Force the grace period if too many callbacks or too long waiting.
2516 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2517 * if some other CPU has recently done so. Also, don't bother
2518 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2519 * is the only one waiting for a grace period to complete.
2521 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2522 rdp->qlen_last_fqs_check + qhimark)) {
2524 /* Are we ignoring a completed grace period? */
2525 note_gp_changes(rdp);
2527 /* Start a new grace period if one not already started. */
2528 if (!rcu_gp_in_progress()) {
2529 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2531 /* Give the grace period a kick. */
2532 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2533 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2534 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2535 rcu_force_quiescent_state();
2536 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2537 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2543 * RCU callback function to leak a callback.
2545 static void rcu_leak_callback(struct rcu_head *rhp)
2550 * Helper function for call_rcu() and friends. The cpu argument will
2551 * normally be -1, indicating "currently running CPU". It may specify
2552 * a CPU only if that CPU is a no-CBs CPU. Currently, only rcu_barrier()
2553 * is expected to specify a CPU.
2556 __call_rcu(struct rcu_head *head, rcu_callback_t func, bool lazy)
2558 unsigned long flags;
2559 struct rcu_data *rdp;
2562 /* Misaligned rcu_head! */
2563 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2565 if (debug_rcu_head_queue(head)) {
2567 * Probable double call_rcu(), so leak the callback.
2568 * Use rcu:rcu_callback trace event to find the previous
2569 * time callback was passed to __call_rcu().
2571 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2573 WRITE_ONCE(head->func, rcu_leak_callback);
2578 local_irq_save(flags);
2579 rdp = this_cpu_ptr(&rcu_data);
2581 /* Add the callback to our list. */
2582 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2583 // This can trigger due to call_rcu() from offline CPU:
2584 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2585 WARN_ON_ONCE(!rcu_is_watching());
2586 // Very early boot, before rcu_init(). Initialize if needed
2587 // and then drop through to queue the callback.
2588 if (rcu_segcblist_empty(&rdp->cblist))
2589 rcu_segcblist_init(&rdp->cblist);
2591 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2592 return; // Enqueued onto ->nocb_bypass, so just leave.
2593 /* If we get here, rcu_nocb_try_bypass() acquired ->nocb_lock. */
2594 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2595 if (__is_kfree_rcu_offset((unsigned long)func))
2596 trace_rcu_kfree_callback(rcu_state.name, head,
2597 (unsigned long)func,
2598 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2599 rcu_segcblist_n_cbs(&rdp->cblist));
2601 trace_rcu_callback(rcu_state.name, head,
2602 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2603 rcu_segcblist_n_cbs(&rdp->cblist));
2605 /* Go handle any RCU core processing required. */
2606 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2607 unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2608 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2610 __call_rcu_core(rdp, head, flags);
2611 local_irq_restore(flags);
2616 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2617 * @head: structure to be used for queueing the RCU updates.
2618 * @func: actual callback function to be invoked after the grace period
2620 * The callback function will be invoked some time after a full grace
2621 * period elapses, in other words after all pre-existing RCU read-side
2622 * critical sections have completed. However, the callback function
2623 * might well execute concurrently with RCU read-side critical sections
2624 * that started after call_rcu() was invoked. RCU read-side critical
2625 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2626 * may be nested. In addition, regions of code across which interrupts,
2627 * preemption, or softirqs have been disabled also serve as RCU read-side
2628 * critical sections. This includes hardware interrupt handlers, softirq
2629 * handlers, and NMI handlers.
2631 * Note that all CPUs must agree that the grace period extended beyond
2632 * all pre-existing RCU read-side critical section. On systems with more
2633 * than one CPU, this means that when "func()" is invoked, each CPU is
2634 * guaranteed to have executed a full memory barrier since the end of its
2635 * last RCU read-side critical section whose beginning preceded the call
2636 * to call_rcu(). It also means that each CPU executing an RCU read-side
2637 * critical section that continues beyond the start of "func()" must have
2638 * executed a memory barrier after the call_rcu() but before the beginning
2639 * of that RCU read-side critical section. Note that these guarantees
2640 * include CPUs that are offline, idle, or executing in user mode, as
2641 * well as CPUs that are executing in the kernel.
2643 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2644 * resulting RCU callback function "func()", then both CPU A and CPU B are
2645 * guaranteed to execute a full memory barrier during the time interval
2646 * between the call to call_rcu() and the invocation of "func()" -- even
2647 * if CPU A and CPU B are the same CPU (but again only if the system has
2648 * more than one CPU).
2650 void call_rcu(struct rcu_head *head, rcu_callback_t func)
2652 __call_rcu(head, func, 0);
2654 EXPORT_SYMBOL_GPL(call_rcu);
2657 * Queue an RCU callback for lazy invocation after a grace period.
2658 * This will likely be later named something like "call_rcu_lazy()",
2659 * but this change will require some way of tagging the lazy RCU
2660 * callbacks in the list of pending callbacks. Until then, this
2661 * function may only be called from __kfree_rcu().
2663 void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
2665 __call_rcu(head, func, 1);
2667 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2670 * During early boot, any blocking grace-period wait automatically
2671 * implies a grace period. Later on, this is never the case for PREEMPT.
2673 * Howevr, because a context switch is a grace period for !PREEMPT, any
2674 * blocking grace-period wait automatically implies a grace period if
2675 * there is only one CPU online at any point time during execution of
2676 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
2677 * occasionally incorrectly indicate that there are multiple CPUs online
2678 * when there was in fact only one the whole time, as this just adds some
2679 * overhead: RCU still operates correctly.
2681 static int rcu_blocking_is_gp(void)
2685 if (IS_ENABLED(CONFIG_PREEMPTION))
2686 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
2687 might_sleep(); /* Check for RCU read-side critical section. */
2689 ret = num_online_cpus() <= 1;
2695 * synchronize_rcu - wait until a grace period has elapsed.
2697 * Control will return to the caller some time after a full grace
2698 * period has elapsed, in other words after all currently executing RCU
2699 * read-side critical sections have completed. Note, however, that
2700 * upon return from synchronize_rcu(), the caller might well be executing
2701 * concurrently with new RCU read-side critical sections that began while
2702 * synchronize_rcu() was waiting. RCU read-side critical sections are
2703 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
2704 * In addition, regions of code across which interrupts, preemption, or
2705 * softirqs have been disabled also serve as RCU read-side critical
2706 * sections. This includes hardware interrupt handlers, softirq handlers,
2709 * Note that this guarantee implies further memory-ordering guarantees.
2710 * On systems with more than one CPU, when synchronize_rcu() returns,
2711 * each CPU is guaranteed to have executed a full memory barrier since
2712 * the end of its last RCU read-side critical section whose beginning
2713 * preceded the call to synchronize_rcu(). In addition, each CPU having
2714 * an RCU read-side critical section that extends beyond the return from
2715 * synchronize_rcu() is guaranteed to have executed a full memory barrier
2716 * after the beginning of synchronize_rcu() and before the beginning of
2717 * that RCU read-side critical section. Note that these guarantees include
2718 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2719 * that are executing in the kernel.
2721 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
2722 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2723 * to have executed a full memory barrier during the execution of
2724 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
2725 * again only if the system has more than one CPU).
2727 void synchronize_rcu(void)
2729 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
2730 lock_is_held(&rcu_lock_map) ||
2731 lock_is_held(&rcu_sched_lock_map),
2732 "Illegal synchronize_rcu() in RCU read-side critical section");
2733 if (rcu_blocking_is_gp())
2735 if (rcu_gp_is_expedited())
2736 synchronize_rcu_expedited();
2738 wait_rcu_gp(call_rcu);
2740 EXPORT_SYMBOL_GPL(synchronize_rcu);
2743 * get_state_synchronize_rcu - Snapshot current RCU state
2745 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2746 * to determine whether or not a full grace period has elapsed in the
2749 unsigned long get_state_synchronize_rcu(void)
2752 * Any prior manipulation of RCU-protected data must happen
2753 * before the load from ->gp_seq.
2756 return rcu_seq_snap(&rcu_state.gp_seq);
2758 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2761 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2763 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2765 * If a full RCU grace period has elapsed since the earlier call to
2766 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2767 * synchronize_rcu() to wait for a full grace period.
2769 * Yes, this function does not take counter wrap into account. But
2770 * counter wrap is harmless. If the counter wraps, we have waited for
2771 * more than 2 billion grace periods (and way more on a 64-bit system!),
2772 * so waiting for one additional grace period should be just fine.
2774 void cond_synchronize_rcu(unsigned long oldstate)
2776 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
2779 smp_mb(); /* Ensure GP ends before subsequent accesses. */
2781 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2784 * Check to see if there is any immediate RCU-related work to be done by
2785 * the current CPU, returning 1 if so and zero otherwise. The checks are
2786 * in order of increasing expense: checks that can be carried out against
2787 * CPU-local state are performed first. However, we must check for CPU
2788 * stalls first, else we might not get a chance.
2790 static int rcu_pending(void)
2792 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2793 struct rcu_node *rnp = rdp->mynode;
2795 /* Check for CPU stalls, if enabled. */
2796 check_cpu_stall(rdp);
2798 /* Does this CPU need a deferred NOCB wakeup? */
2799 if (rcu_nocb_need_deferred_wakeup(rdp))
2802 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2803 if (rcu_nohz_full_cpu())
2806 /* Is the RCU core waiting for a quiescent state from this CPU? */
2807 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
2810 /* Does this CPU have callbacks ready to invoke? */
2811 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2814 /* Has RCU gone idle with this CPU needing another grace period? */
2815 if (!rcu_gp_in_progress() &&
2816 rcu_segcblist_is_enabled(&rdp->cblist) &&
2817 (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
2818 !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
2819 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2822 /* Have RCU grace period completed or started? */
2823 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
2824 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
2832 * Helper function for rcu_barrier() tracing. If tracing is disabled,
2833 * the compiler is expected to optimize this away.
2835 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
2837 trace_rcu_barrier(rcu_state.name, s, cpu,
2838 atomic_read(&rcu_state.barrier_cpu_count), done);
2842 * RCU callback function for rcu_barrier(). If we are last, wake
2843 * up the task executing rcu_barrier().
2845 static void rcu_barrier_callback(struct rcu_head *rhp)
2847 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
2848 rcu_barrier_trace(TPS("LastCB"), -1,
2849 rcu_state.barrier_sequence);
2850 complete(&rcu_state.barrier_completion);
2852 rcu_barrier_trace(TPS("CB"), -1, rcu_state.barrier_sequence);
2857 * Called with preemption disabled, and from cross-cpu IRQ context.
2859 static void rcu_barrier_func(void *unused)
2861 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2863 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
2864 rdp->barrier_head.func = rcu_barrier_callback;
2865 debug_rcu_head_queue(&rdp->barrier_head);
2867 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
2868 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
2869 atomic_inc(&rcu_state.barrier_cpu_count);
2871 debug_rcu_head_unqueue(&rdp->barrier_head);
2872 rcu_barrier_trace(TPS("IRQNQ"), -1,
2873 rcu_state.barrier_sequence);
2875 rcu_nocb_unlock(rdp);
2879 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
2881 * Note that this primitive does not necessarily wait for an RCU grace period
2882 * to complete. For example, if there are no RCU callbacks queued anywhere
2883 * in the system, then rcu_barrier() is within its rights to return
2884 * immediately, without waiting for anything, much less an RCU grace period.
2886 void rcu_barrier(void)
2889 struct rcu_data *rdp;
2890 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
2892 rcu_barrier_trace(TPS("Begin"), -1, s);
2894 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2895 mutex_lock(&rcu_state.barrier_mutex);
2897 /* Did someone else do our work for us? */
2898 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
2899 rcu_barrier_trace(TPS("EarlyExit"), -1,
2900 rcu_state.barrier_sequence);
2901 smp_mb(); /* caller's subsequent code after above check. */
2902 mutex_unlock(&rcu_state.barrier_mutex);
2906 /* Mark the start of the barrier operation. */
2907 rcu_seq_start(&rcu_state.barrier_sequence);
2908 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
2911 * Initialize the count to one rather than to zero in order to
2912 * avoid a too-soon return to zero in case of a short grace period
2913 * (or preemption of this task). Exclude CPU-hotplug operations
2914 * to ensure that no offline CPU has callbacks queued.
2916 init_completion(&rcu_state.barrier_completion);
2917 atomic_set(&rcu_state.barrier_cpu_count, 1);
2921 * Force each CPU with callbacks to register a new callback.
2922 * When that callback is invoked, we will know that all of the
2923 * corresponding CPU's preceding callbacks have been invoked.
2925 for_each_possible_cpu(cpu) {
2926 rdp = per_cpu_ptr(&rcu_data, cpu);
2927 if (!cpu_online(cpu) &&
2928 !rcu_segcblist_is_offloaded(&rdp->cblist))
2930 if (rcu_segcblist_n_cbs(&rdp->cblist)) {
2931 rcu_barrier_trace(TPS("OnlineQ"), cpu,
2932 rcu_state.barrier_sequence);
2933 smp_call_function_single(cpu, rcu_barrier_func, NULL, 1);
2935 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
2936 rcu_state.barrier_sequence);
2942 * Now that we have an rcu_barrier_callback() callback on each
2943 * CPU, and thus each counted, remove the initial count.
2945 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count))
2946 complete(&rcu_state.barrier_completion);
2948 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2949 wait_for_completion(&rcu_state.barrier_completion);
2951 /* Mark the end of the barrier operation. */
2952 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
2953 rcu_seq_end(&rcu_state.barrier_sequence);
2955 /* Other rcu_barrier() invocations can now safely proceed. */
2956 mutex_unlock(&rcu_state.barrier_mutex);
2958 EXPORT_SYMBOL_GPL(rcu_barrier);
2961 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
2962 * first CPU in a given leaf rcu_node structure coming online. The caller
2963 * must hold the corresponding leaf rcu_node ->lock with interrrupts
2966 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
2970 struct rcu_node *rnp = rnp_leaf;
2972 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2973 WARN_ON_ONCE(rnp->wait_blkd_tasks);
2975 mask = rnp->grpmask;
2979 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
2980 oldmask = rnp->qsmaskinit;
2981 rnp->qsmaskinit |= mask;
2982 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
2989 * Do boot-time initialization of a CPU's per-CPU RCU data.
2992 rcu_boot_init_percpu_data(int cpu)
2994 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2996 /* Set up local state, ensuring consistent view of global state. */
2997 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
2998 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
2999 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
3000 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3001 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3002 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3003 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3005 rcu_boot_init_nocb_percpu_data(rdp);
3009 * Invoked early in the CPU-online process, when pretty much all services
3010 * are available. The incoming CPU is not present.
3012 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3013 * offline event can be happening at a given time. Note also that we can
3014 * accept some slop in the rsp->gp_seq access due to the fact that this
3015 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3016 * And any offloaded callbacks are being numbered elsewhere.
3018 int rcutree_prepare_cpu(unsigned int cpu)
3020 unsigned long flags;
3021 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3022 struct rcu_node *rnp = rcu_get_root();
3024 /* Set up local state, ensuring consistent view of global state. */
3025 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3026 rdp->qlen_last_fqs_check = 0;
3027 rdp->n_force_qs_snap = rcu_state.n_force_qs;
3028 rdp->blimit = blimit;
3029 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3030 !rcu_segcblist_is_offloaded(&rdp->cblist))
3031 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3032 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
3033 rcu_dynticks_eqs_online();
3034 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3037 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3038 * propagation up the rcu_node tree will happen at the beginning
3039 * of the next grace period.
3042 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3043 rdp->beenonline = true; /* We have now been online. */
3044 rdp->gp_seq = rnp->gp_seq;
3045 rdp->gp_seq_needed = rnp->gp_seq;
3046 rdp->cpu_no_qs.b.norm = true;
3047 rdp->core_needs_qs = false;
3048 rdp->rcu_iw_pending = false;
3049 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3050 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3051 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3052 rcu_prepare_kthreads(cpu);
3053 rcu_spawn_cpu_nocb_kthread(cpu);
3059 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3061 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3063 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3065 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3069 * Near the end of the CPU-online process. Pretty much all services
3070 * enabled, and the CPU is now very much alive.
3072 int rcutree_online_cpu(unsigned int cpu)
3074 unsigned long flags;
3075 struct rcu_data *rdp;
3076 struct rcu_node *rnp;
3078 rdp = per_cpu_ptr(&rcu_data, cpu);
3080 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3081 rnp->ffmask |= rdp->grpmask;
3082 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3083 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3084 return 0; /* Too early in boot for scheduler work. */
3085 sync_sched_exp_online_cleanup(cpu);
3086 rcutree_affinity_setting(cpu, -1);
3091 * Near the beginning of the process. The CPU is still very much alive
3092 * with pretty much all services enabled.
3094 int rcutree_offline_cpu(unsigned int cpu)
3096 unsigned long flags;
3097 struct rcu_data *rdp;
3098 struct rcu_node *rnp;
3100 rdp = per_cpu_ptr(&rcu_data, cpu);
3102 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3103 rnp->ffmask &= ~rdp->grpmask;
3104 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3106 rcutree_affinity_setting(cpu, cpu);
3110 static DEFINE_PER_CPU(int, rcu_cpu_started);
3113 * Mark the specified CPU as being online so that subsequent grace periods
3114 * (both expedited and normal) will wait on it. Note that this means that
3115 * incoming CPUs are not allowed to use RCU read-side critical sections
3116 * until this function is called. Failing to observe this restriction
3117 * will result in lockdep splats.
3119 * Note that this function is special in that it is invoked directly
3120 * from the incoming CPU rather than from the cpuhp_step mechanism.
3121 * This is because this function must be invoked at a precise location.
3123 void rcu_cpu_starting(unsigned int cpu)
3125 unsigned long flags;
3128 unsigned long oldmask;
3129 struct rcu_data *rdp;
3130 struct rcu_node *rnp;
3132 if (per_cpu(rcu_cpu_started, cpu))
3135 per_cpu(rcu_cpu_started, cpu) = 1;
3137 rdp = per_cpu_ptr(&rcu_data, cpu);
3139 mask = rdp->grpmask;
3140 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3141 rnp->qsmaskinitnext |= mask;
3142 oldmask = rnp->expmaskinitnext;
3143 rnp->expmaskinitnext |= mask;
3144 oldmask ^= rnp->expmaskinitnext;
3145 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3146 /* Allow lockless access for expedited grace periods. */
3147 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */
3148 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3149 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3150 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3151 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3152 /* Report QS -after- changing ->qsmaskinitnext! */
3153 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3155 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3157 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3160 #ifdef CONFIG_HOTPLUG_CPU
3162 * The outgoing function has no further need of RCU, so remove it from
3163 * the rcu_node tree's ->qsmaskinitnext bit masks.
3165 * Note that this function is special in that it is invoked directly
3166 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3167 * This is because this function must be invoked at a precise location.
3169 void rcu_report_dead(unsigned int cpu)
3171 unsigned long flags;
3173 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3174 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3176 /* QS for any half-done expedited grace period. */
3178 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
3180 rcu_preempt_deferred_qs(current);
3182 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3183 mask = rdp->grpmask;
3184 raw_spin_lock(&rcu_state.ofl_lock);
3185 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3186 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3187 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3188 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3189 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3190 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3191 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3193 rnp->qsmaskinitnext &= ~mask;
3194 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3195 raw_spin_unlock(&rcu_state.ofl_lock);
3197 per_cpu(rcu_cpu_started, cpu) = 0;
3201 * The outgoing CPU has just passed through the dying-idle state, and we
3202 * are being invoked from the CPU that was IPIed to continue the offline
3203 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
3205 void rcutree_migrate_callbacks(int cpu)
3207 unsigned long flags;
3208 struct rcu_data *my_rdp;
3209 struct rcu_node *my_rnp;
3210 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3213 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
3214 rcu_segcblist_empty(&rdp->cblist))
3215 return; /* No callbacks to migrate. */
3217 local_irq_save(flags);
3218 my_rdp = this_cpu_ptr(&rcu_data);
3219 my_rnp = my_rdp->mynode;
3220 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
3221 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
3222 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
3223 /* Leverage recent GPs and set GP for new callbacks. */
3224 needwake = rcu_advance_cbs(my_rnp, rdp) ||
3225 rcu_advance_cbs(my_rnp, my_rdp);
3226 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3227 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
3228 rcu_segcblist_disable(&rdp->cblist);
3229 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3230 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3231 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
3232 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
3233 __call_rcu_nocb_wake(my_rdp, true, flags);
3235 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
3236 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
3239 rcu_gp_kthread_wake();
3240 lockdep_assert_irqs_enabled();
3241 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3242 !rcu_segcblist_empty(&rdp->cblist),
3243 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3244 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3245 rcu_segcblist_first_cb(&rdp->cblist));
3250 * On non-huge systems, use expedited RCU grace periods to make suspend
3251 * and hibernation run faster.
3253 static int rcu_pm_notify(struct notifier_block *self,
3254 unsigned long action, void *hcpu)
3257 case PM_HIBERNATION_PREPARE:
3258 case PM_SUSPEND_PREPARE:
3261 case PM_POST_HIBERNATION:
3262 case PM_POST_SUSPEND:
3263 rcu_unexpedite_gp();
3272 * Spawn the kthreads that handle RCU's grace periods.
3274 static int __init rcu_spawn_gp_kthread(void)
3276 unsigned long flags;
3277 int kthread_prio_in = kthread_prio;
3278 struct rcu_node *rnp;
3279 struct sched_param sp;
3280 struct task_struct *t;
3282 /* Force priority into range. */
3283 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
3284 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
3286 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3288 else if (kthread_prio < 0)
3290 else if (kthread_prio > 99)
3293 if (kthread_prio != kthread_prio_in)
3294 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3295 kthread_prio, kthread_prio_in);
3297 rcu_scheduler_fully_active = 1;
3298 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
3299 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
3302 sp.sched_priority = kthread_prio;
3303 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3305 rnp = rcu_get_root();
3306 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3307 rcu_state.gp_kthread = t;
3308 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3310 rcu_spawn_nocb_kthreads();
3311 rcu_spawn_boost_kthreads();
3314 early_initcall(rcu_spawn_gp_kthread);
3317 * This function is invoked towards the end of the scheduler's
3318 * initialization process. Before this is called, the idle task might
3319 * contain synchronous grace-period primitives (during which time, this idle
3320 * task is booting the system, and such primitives are no-ops). After this
3321 * function is called, any synchronous grace-period primitives are run as
3322 * expedited, with the requesting task driving the grace period forward.
3323 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3324 * runtime RCU functionality.
3326 void rcu_scheduler_starting(void)
3328 WARN_ON(num_online_cpus() != 1);
3329 WARN_ON(nr_context_switches() > 0);
3330 rcu_test_sync_prims();
3331 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3332 rcu_test_sync_prims();
3336 * Helper function for rcu_init() that initializes the rcu_state structure.
3338 static void __init rcu_init_one(void)
3340 static const char * const buf[] = RCU_NODE_NAME_INIT;
3341 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3342 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3343 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3345 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3349 struct rcu_node *rnp;
3351 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3353 /* Silence gcc 4.8 false positive about array index out of range. */
3354 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3355 panic("rcu_init_one: rcu_num_lvls out of range");
3357 /* Initialize the level-tracking arrays. */
3359 for (i = 1; i < rcu_num_lvls; i++)
3360 rcu_state.level[i] =
3361 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
3362 rcu_init_levelspread(levelspread, num_rcu_lvl);
3364 /* Initialize the elements themselves, starting from the leaves. */
3366 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3367 cpustride *= levelspread[i];
3368 rnp = rcu_state.level[i];
3369 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3370 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3371 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3372 &rcu_node_class[i], buf[i]);
3373 raw_spin_lock_init(&rnp->fqslock);
3374 lockdep_set_class_and_name(&rnp->fqslock,
3375 &rcu_fqs_class[i], fqs[i]);
3376 rnp->gp_seq = rcu_state.gp_seq;
3377 rnp->gp_seq_needed = rcu_state.gp_seq;
3378 rnp->completedqs = rcu_state.gp_seq;
3380 rnp->qsmaskinit = 0;
3381 rnp->grplo = j * cpustride;
3382 rnp->grphi = (j + 1) * cpustride - 1;
3383 if (rnp->grphi >= nr_cpu_ids)
3384 rnp->grphi = nr_cpu_ids - 1;
3390 rnp->grpnum = j % levelspread[i - 1];
3391 rnp->grpmask = BIT(rnp->grpnum);
3392 rnp->parent = rcu_state.level[i - 1] +
3393 j / levelspread[i - 1];
3396 INIT_LIST_HEAD(&rnp->blkd_tasks);
3397 rcu_init_one_nocb(rnp);
3398 init_waitqueue_head(&rnp->exp_wq[0]);
3399 init_waitqueue_head(&rnp->exp_wq[1]);
3400 init_waitqueue_head(&rnp->exp_wq[2]);
3401 init_waitqueue_head(&rnp->exp_wq[3]);
3402 spin_lock_init(&rnp->exp_lock);
3406 init_swait_queue_head(&rcu_state.gp_wq);
3407 init_swait_queue_head(&rcu_state.expedited_wq);
3408 rnp = rcu_first_leaf_node();
3409 for_each_possible_cpu(i) {
3410 while (i > rnp->grphi)
3412 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
3413 rcu_boot_init_percpu_data(i);
3418 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3419 * replace the definitions in tree.h because those are needed to size
3420 * the ->node array in the rcu_state structure.
3422 static void __init rcu_init_geometry(void)
3426 int rcu_capacity[RCU_NUM_LVLS];
3429 * Initialize any unspecified boot parameters.
3430 * The default values of jiffies_till_first_fqs and
3431 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3432 * value, which is a function of HZ, then adding one for each
3433 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3435 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3436 if (jiffies_till_first_fqs == ULONG_MAX)
3437 jiffies_till_first_fqs = d;
3438 if (jiffies_till_next_fqs == ULONG_MAX)
3439 jiffies_till_next_fqs = d;
3440 adjust_jiffies_till_sched_qs();
3442 /* If the compile-time values are accurate, just leave. */
3443 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3444 nr_cpu_ids == NR_CPUS)
3446 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3447 rcu_fanout_leaf, nr_cpu_ids);
3450 * The boot-time rcu_fanout_leaf parameter must be at least two
3451 * and cannot exceed the number of bits in the rcu_node masks.
3452 * Complain and fall back to the compile-time values if this
3453 * limit is exceeded.
3455 if (rcu_fanout_leaf < 2 ||
3456 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3457 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3463 * Compute number of nodes that can be handled an rcu_node tree
3464 * with the given number of levels.
3466 rcu_capacity[0] = rcu_fanout_leaf;
3467 for (i = 1; i < RCU_NUM_LVLS; i++)
3468 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
3471 * The tree must be able to accommodate the configured number of CPUs.
3472 * If this limit is exceeded, fall back to the compile-time values.
3474 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
3475 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3480 /* Calculate the number of levels in the tree. */
3481 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
3483 rcu_num_lvls = i + 1;
3485 /* Calculate the number of rcu_nodes at each level of the tree. */
3486 for (i = 0; i < rcu_num_lvls; i++) {
3487 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
3488 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
3491 /* Calculate the total number of rcu_node structures. */
3493 for (i = 0; i < rcu_num_lvls; i++)
3494 rcu_num_nodes += num_rcu_lvl[i];
3498 * Dump out the structure of the rcu_node combining tree associated
3499 * with the rcu_state structure.
3501 static void __init rcu_dump_rcu_node_tree(void)
3504 struct rcu_node *rnp;
3506 pr_info("rcu_node tree layout dump\n");
3508 rcu_for_each_node_breadth_first(rnp) {
3509 if (rnp->level != level) {
3514 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
3519 struct workqueue_struct *rcu_gp_wq;
3520 struct workqueue_struct *rcu_par_gp_wq;
3522 void __init rcu_init(void)
3526 rcu_early_boot_tests();
3528 rcu_bootup_announce();
3529 rcu_init_geometry();
3532 rcu_dump_rcu_node_tree();
3534 open_softirq(RCU_SOFTIRQ, rcu_core_si);
3537 * We don't need protection against CPU-hotplug here because
3538 * this is called early in boot, before either interrupts
3539 * or the scheduler are operational.
3541 pm_notifier(rcu_pm_notify, 0);
3542 for_each_online_cpu(cpu) {
3543 rcutree_prepare_cpu(cpu);
3544 rcu_cpu_starting(cpu);
3545 rcutree_online_cpu(cpu);
3548 /* Create workqueue for expedited GPs and for Tree SRCU. */
3549 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
3550 WARN_ON(!rcu_gp_wq);
3551 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
3552 WARN_ON(!rcu_par_gp_wq);
3556 #include "tree_stall.h"
3557 #include "tree_exp.h"
3558 #include "tree_plugin.h"