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 rcu_prepare_for_idle();
583 rcu_preempt_deferred_qs(current);
584 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
585 rcu_dynticks_eqs_enter();
586 rcu_dynticks_task_enter();
590 * rcu_idle_enter - inform RCU that current CPU is entering idle
592 * Enter idle mode, in other words, -leave- the mode in which RCU
593 * read-side critical sections can occur. (Though RCU read-side
594 * critical sections can occur in irq handlers in idle, a possibility
595 * handled by irq_enter() and irq_exit().)
597 * If you add or remove a call to rcu_idle_enter(), be sure to test with
598 * CONFIG_RCU_EQS_DEBUG=y.
600 void rcu_idle_enter(void)
602 lockdep_assert_irqs_disabled();
603 rcu_eqs_enter(false);
606 #ifdef CONFIG_NO_HZ_FULL
608 * rcu_user_enter - inform RCU that we are resuming userspace.
610 * Enter RCU idle mode right before resuming userspace. No use of RCU
611 * is permitted between this call and rcu_user_exit(). This way the
612 * CPU doesn't need to maintain the tick for RCU maintenance purposes
613 * when the CPU runs in userspace.
615 * If you add or remove a call to rcu_user_enter(), be sure to test with
616 * CONFIG_RCU_EQS_DEBUG=y.
618 void rcu_user_enter(void)
620 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
622 lockdep_assert_irqs_disabled();
624 instrumentation_begin();
625 do_nocb_deferred_wakeup(rdp);
626 instrumentation_end();
630 #endif /* CONFIG_NO_HZ_FULL */
633 * If we are returning from the outermost NMI handler that interrupted an
634 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
635 * to let the RCU grace-period handling know that the CPU is back to
638 * If you add or remove a call to rcu_nmi_exit_common(), be sure to test
639 * with CONFIG_RCU_EQS_DEBUG=y.
641 static __always_inline void rcu_nmi_exit_common(bool irq)
643 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
646 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
647 * (We are exiting an NMI handler, so RCU better be paying attention
650 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
651 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
654 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
655 * leave it in non-RCU-idle state.
657 if (rdp->dynticks_nmi_nesting != 1) {
658 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
659 atomic_read(&rdp->dynticks));
660 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
661 rdp->dynticks_nmi_nesting - 2);
665 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
666 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
667 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
670 rcu_prepare_for_idle();
672 rcu_dynticks_eqs_enter();
675 rcu_dynticks_task_enter();
679 * rcu_nmi_exit - inform RCU of exit from NMI context
681 * If you add or remove a call to rcu_nmi_exit(), be sure to test
682 * with CONFIG_RCU_EQS_DEBUG=y.
684 void rcu_nmi_exit(void)
686 rcu_nmi_exit_common(false);
690 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
692 * Exit from an interrupt handler, which might possibly result in entering
693 * idle mode, in other words, leaving the mode in which read-side critical
694 * sections can occur. The caller must have disabled interrupts.
696 * This code assumes that the idle loop never does anything that might
697 * result in unbalanced calls to irq_enter() and irq_exit(). If your
698 * architecture's idle loop violates this assumption, RCU will give you what
699 * you deserve, good and hard. But very infrequently and irreproducibly.
701 * Use things like work queues to work around this limitation.
703 * You have been warned.
705 * If you add or remove a call to rcu_irq_exit(), be sure to test with
706 * CONFIG_RCU_EQS_DEBUG=y.
708 void rcu_irq_exit(void)
710 lockdep_assert_irqs_disabled();
711 rcu_nmi_exit_common(true);
715 * Wrapper for rcu_irq_exit() where interrupts are enabled.
717 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
718 * with CONFIG_RCU_EQS_DEBUG=y.
720 void rcu_irq_exit_irqson(void)
724 local_irq_save(flags);
726 local_irq_restore(flags);
730 * Exit an RCU extended quiescent state, which can be either the
731 * idle loop or adaptive-tickless usermode execution.
733 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
734 * allow for the possibility of usermode upcalls messing up our count of
735 * interrupt nesting level during the busy period that is just now starting.
737 static void rcu_eqs_exit(bool user)
739 struct rcu_data *rdp;
742 lockdep_assert_irqs_disabled();
743 rdp = this_cpu_ptr(&rcu_data);
744 oldval = rdp->dynticks_nesting;
745 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
747 rdp->dynticks_nesting++;
750 rcu_dynticks_task_exit();
751 rcu_dynticks_eqs_exit();
752 rcu_cleanup_after_idle();
753 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
754 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
755 WRITE_ONCE(rdp->dynticks_nesting, 1);
756 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
757 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
761 * rcu_idle_exit - inform RCU that current CPU is leaving idle
763 * Exit idle mode, in other words, -enter- the mode in which RCU
764 * read-side critical sections can occur.
766 * If you add or remove a call to rcu_idle_exit(), be sure to test with
767 * CONFIG_RCU_EQS_DEBUG=y.
769 void rcu_idle_exit(void)
773 local_irq_save(flags);
775 local_irq_restore(flags);
778 #ifdef CONFIG_NO_HZ_FULL
780 * rcu_user_exit - inform RCU that we are exiting userspace.
782 * Exit RCU idle mode while entering the kernel because it can
783 * run a RCU read side critical section anytime.
785 * If you add or remove a call to rcu_user_exit(), be sure to test with
786 * CONFIG_RCU_EQS_DEBUG=y.
788 void rcu_user_exit(void)
792 #endif /* CONFIG_NO_HZ_FULL */
795 * rcu_nmi_enter_common - inform RCU of entry to NMI context
796 * @irq: Is this call from rcu_irq_enter?
798 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
799 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
800 * that the CPU is active. This implementation permits nested NMIs, as
801 * long as the nesting level does not overflow an int. (You will probably
802 * run out of stack space first.)
804 * If you add or remove a call to rcu_nmi_enter_common(), be sure to test
805 * with CONFIG_RCU_EQS_DEBUG=y.
807 static __always_inline void rcu_nmi_enter_common(bool irq)
809 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
812 /* Complain about underflow. */
813 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
816 * If idle from RCU viewpoint, atomically increment ->dynticks
817 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
818 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
819 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
820 * to be in the outermost NMI handler that interrupted an RCU-idle
821 * period (observation due to Andy Lutomirski).
823 if (rcu_dynticks_curr_cpu_in_eqs()) {
826 rcu_dynticks_task_exit();
828 rcu_dynticks_eqs_exit();
831 rcu_cleanup_after_idle();
835 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
836 rdp->dynticks_nmi_nesting,
837 rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
838 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
839 rdp->dynticks_nmi_nesting + incby);
844 * rcu_nmi_enter - inform RCU of entry to NMI context
846 void rcu_nmi_enter(void)
848 rcu_nmi_enter_common(false);
850 NOKPROBE_SYMBOL(rcu_nmi_enter);
853 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
855 * Enter an interrupt handler, which might possibly result in exiting
856 * idle mode, in other words, entering the mode in which read-side critical
857 * sections can occur. The caller must have disabled interrupts.
859 * Note that the Linux kernel is fully capable of entering an interrupt
860 * handler that it never exits, for example when doing upcalls to user mode!
861 * This code assumes that the idle loop never does upcalls to user mode.
862 * If your architecture's idle loop does do upcalls to user mode (or does
863 * anything else that results in unbalanced calls to the irq_enter() and
864 * irq_exit() functions), RCU will give you what you deserve, good and hard.
865 * But very infrequently and irreproducibly.
867 * Use things like work queues to work around this limitation.
869 * You have been warned.
871 * If you add or remove a call to rcu_irq_enter(), be sure to test with
872 * CONFIG_RCU_EQS_DEBUG=y.
874 void rcu_irq_enter(void)
876 lockdep_assert_irqs_disabled();
877 rcu_nmi_enter_common(true);
881 * Wrapper for rcu_irq_enter() where interrupts are enabled.
883 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
884 * with CONFIG_RCU_EQS_DEBUG=y.
886 void rcu_irq_enter_irqson(void)
890 local_irq_save(flags);
892 local_irq_restore(flags);
896 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
898 * Return true if RCU is watching the running CPU, which means that this
899 * CPU can safely enter RCU read-side critical sections. In other words,
900 * if the current CPU is not in its idle loop or is in an interrupt or
901 * NMI handler, return true.
903 bool notrace rcu_is_watching(void)
907 preempt_disable_notrace();
908 ret = !rcu_dynticks_curr_cpu_in_eqs();
909 preempt_enable_notrace();
912 EXPORT_SYMBOL_GPL(rcu_is_watching);
915 * If a holdout task is actually running, request an urgent quiescent
916 * state from its CPU. This is unsynchronized, so migrations can cause
917 * the request to go to the wrong CPU. Which is OK, all that will happen
918 * is that the CPU's next context switch will be a bit slower and next
919 * time around this task will generate another request.
921 void rcu_request_urgent_qs_task(struct task_struct *t)
928 return; /* This task is not running on that CPU. */
929 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
932 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
935 * Is the current CPU online as far as RCU is concerned?
937 * Disable preemption to avoid false positives that could otherwise
938 * happen due to the current CPU number being sampled, this task being
939 * preempted, its old CPU being taken offline, resuming on some other CPU,
940 * then determining that its old CPU is now offline.
942 * Disable checking if in an NMI handler because we cannot safely
943 * report errors from NMI handlers anyway. In addition, it is OK to use
944 * RCU on an offline processor during initial boot, hence the check for
945 * rcu_scheduler_fully_active.
947 bool rcu_lockdep_current_cpu_online(void)
949 struct rcu_data *rdp;
950 struct rcu_node *rnp;
953 if (in_nmi() || !rcu_scheduler_fully_active)
956 rdp = this_cpu_ptr(&rcu_data);
958 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
963 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
965 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
968 * We are reporting a quiescent state on behalf of some other CPU, so
969 * it is our responsibility to check for and handle potential overflow
970 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
971 * After all, the CPU might be in deep idle state, and thus executing no
974 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
976 raw_lockdep_assert_held_rcu_node(rnp);
977 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
979 WRITE_ONCE(rdp->gpwrap, true);
980 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
981 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
985 * Snapshot the specified CPU's dynticks counter so that we can later
986 * credit them with an implicit quiescent state. Return 1 if this CPU
987 * is in dynticks idle mode, which is an extended quiescent state.
989 static int dyntick_save_progress_counter(struct rcu_data *rdp)
991 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
992 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
993 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
994 rcu_gpnum_ovf(rdp->mynode, rdp);
1001 * Return true if the specified CPU has passed through a quiescent
1002 * state by virtue of being in or having passed through an dynticks
1003 * idle state since the last call to dyntick_save_progress_counter()
1004 * for this same CPU, or by virtue of having been offline.
1006 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1011 struct rcu_node *rnp = rdp->mynode;
1014 * If the CPU passed through or entered a dynticks idle phase with
1015 * no active irq/NMI handlers, then we can safely pretend that the CPU
1016 * already acknowledged the request to pass through a quiescent
1017 * state. Either way, that CPU cannot possibly be in an RCU
1018 * read-side critical section that started before the beginning
1019 * of the current RCU grace period.
1021 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1022 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1023 rcu_gpnum_ovf(rnp, rdp);
1027 /* If waiting too long on an offline CPU, complain. */
1028 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1029 time_after(jiffies, rcu_state.gp_start + HZ)) {
1031 struct rcu_node *rnp1;
1033 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1034 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1035 __func__, rnp->grplo, rnp->grphi, rnp->level,
1036 (long)rnp->gp_seq, (long)rnp->completedqs);
1037 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1038 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1039 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1040 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1041 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1042 __func__, rdp->cpu, ".o"[onl],
1043 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1044 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1045 return 1; /* Break things loose after complaining. */
1049 * A CPU running for an extended time within the kernel can
1050 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1051 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1052 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1053 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1054 * variable are safe because the assignments are repeated if this
1055 * CPU failed to pass through a quiescent state. This code
1056 * also checks .jiffies_resched in case jiffies_to_sched_qs
1059 jtsq = READ_ONCE(jiffies_to_sched_qs);
1060 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1061 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1062 if (!READ_ONCE(*rnhqp) &&
1063 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1064 time_after(jiffies, rcu_state.jiffies_resched))) {
1065 WRITE_ONCE(*rnhqp, true);
1066 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1067 smp_store_release(ruqp, true);
1068 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1069 WRITE_ONCE(*ruqp, true);
1073 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1074 * The above code handles this, but only for straight cond_resched().
1075 * And some in-kernel loops check need_resched() before calling
1076 * cond_resched(), which defeats the above code for CPUs that are
1077 * running in-kernel with scheduling-clock interrupts disabled.
1078 * So hit them over the head with the resched_cpu() hammer!
1080 if (tick_nohz_full_cpu(rdp->cpu) &&
1082 READ_ONCE(rdp->last_fqs_resched) + jtsq * 3)) {
1083 resched_cpu(rdp->cpu);
1084 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1088 * If more than halfway to RCU CPU stall-warning time, invoke
1089 * resched_cpu() more frequently to try to loosen things up a bit.
1090 * Also check to see if the CPU is getting hammered with interrupts,
1091 * but only once per grace period, just to keep the IPIs down to
1094 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1095 if (time_after(jiffies,
1096 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1097 resched_cpu(rdp->cpu);
1098 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1100 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1101 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1102 (rnp->ffmask & rdp->grpmask)) {
1103 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1104 rdp->rcu_iw_pending = true;
1105 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1106 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1113 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1114 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1115 unsigned long gp_seq_req, const char *s)
1117 trace_rcu_future_grace_period(rcu_state.name, rnp->gp_seq, gp_seq_req,
1118 rnp->level, rnp->grplo, rnp->grphi, s);
1122 * rcu_start_this_gp - Request the start of a particular grace period
1123 * @rnp_start: The leaf node of the CPU from which to start.
1124 * @rdp: The rcu_data corresponding to the CPU from which to start.
1125 * @gp_seq_req: The gp_seq of the grace period to start.
1127 * Start the specified grace period, as needed to handle newly arrived
1128 * callbacks. The required future grace periods are recorded in each
1129 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1130 * is reason to awaken the grace-period kthread.
1132 * The caller must hold the specified rcu_node structure's ->lock, which
1133 * is why the caller is responsible for waking the grace-period kthread.
1135 * Returns true if the GP thread needs to be awakened else false.
1137 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1138 unsigned long gp_seq_req)
1141 struct rcu_node *rnp;
1144 * Use funnel locking to either acquire the root rcu_node
1145 * structure's lock or bail out if the need for this grace period
1146 * has already been recorded -- or if that grace period has in
1147 * fact already started. If there is already a grace period in
1148 * progress in a non-leaf node, no recording is needed because the
1149 * end of the grace period will scan the leaf rcu_node structures.
1150 * Note that rnp_start->lock must not be released.
1152 raw_lockdep_assert_held_rcu_node(rnp_start);
1153 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1154 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1155 if (rnp != rnp_start)
1156 raw_spin_lock_rcu_node(rnp);
1157 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1158 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1159 (rnp != rnp_start &&
1160 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1161 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1165 rnp->gp_seq_needed = gp_seq_req;
1166 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1168 * We just marked the leaf or internal node, and a
1169 * grace period is in progress, which means that
1170 * rcu_gp_cleanup() will see the marking. Bail to
1171 * reduce contention.
1173 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1174 TPS("Startedleaf"));
1177 if (rnp != rnp_start && rnp->parent != NULL)
1178 raw_spin_unlock_rcu_node(rnp);
1180 break; /* At root, and perhaps also leaf. */
1183 /* If GP already in progress, just leave, otherwise start one. */
1184 if (rcu_gp_in_progress()) {
1185 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1188 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1189 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1190 rcu_state.gp_req_activity = jiffies;
1191 if (!rcu_state.gp_kthread) {
1192 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1195 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rcu_state.gp_seq), TPS("newreq"));
1196 ret = true; /* Caller must wake GP kthread. */
1198 /* Push furthest requested GP to leaf node and rcu_data structure. */
1199 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1200 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1201 rdp->gp_seq_needed = rnp->gp_seq_needed;
1203 if (rnp != rnp_start)
1204 raw_spin_unlock_rcu_node(rnp);
1209 * Clean up any old requests for the just-ended grace period. Also return
1210 * whether any additional grace periods have been requested.
1212 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1215 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1217 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1219 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1220 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1221 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1226 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1227 * an interrupt or softirq handler), and don't bother awakening when there
1228 * is nothing for the grace-period kthread to do (as in several CPUs raced
1229 * to awaken, and we lost), and finally don't try to awaken a kthread that
1230 * has not yet been created. If all those checks are passed, track some
1231 * debug information and awaken.
1233 * So why do the self-wakeup when in an interrupt or softirq handler
1234 * in the grace-period kthread's context? Because the kthread might have
1235 * been interrupted just as it was going to sleep, and just after the final
1236 * pre-sleep check of the awaken condition. In this case, a wakeup really
1237 * is required, and is therefore supplied.
1239 static void rcu_gp_kthread_wake(void)
1241 if ((current == rcu_state.gp_kthread &&
1242 !in_irq() && !in_serving_softirq()) ||
1243 !READ_ONCE(rcu_state.gp_flags) ||
1244 !rcu_state.gp_kthread)
1246 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1247 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1248 swake_up_one(&rcu_state.gp_wq);
1252 * If there is room, assign a ->gp_seq number to any callbacks on this
1253 * CPU that have not already been assigned. Also accelerate any callbacks
1254 * that were previously assigned a ->gp_seq number that has since proven
1255 * to be too conservative, which can happen if callbacks get assigned a
1256 * ->gp_seq number while RCU is idle, but with reference to a non-root
1257 * rcu_node structure. This function is idempotent, so it does not hurt
1258 * to call it repeatedly. Returns an flag saying that we should awaken
1259 * the RCU grace-period kthread.
1261 * The caller must hold rnp->lock with interrupts disabled.
1263 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1265 unsigned long gp_seq_req;
1268 rcu_lockdep_assert_cblist_protected(rdp);
1269 raw_lockdep_assert_held_rcu_node(rnp);
1271 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1272 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1276 * Callbacks are often registered with incomplete grace-period
1277 * information. Something about the fact that getting exact
1278 * information requires acquiring a global lock... RCU therefore
1279 * makes a conservative estimate of the grace period number at which
1280 * a given callback will become ready to invoke. The following
1281 * code checks this estimate and improves it when possible, thus
1282 * accelerating callback invocation to an earlier grace-period
1285 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1286 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1287 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1289 /* Trace depending on how much we were able to accelerate. */
1290 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1291 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1293 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1298 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1299 * rcu_node structure's ->lock be held. It consults the cached value
1300 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1301 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1302 * while holding the leaf rcu_node structure's ->lock.
1304 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1305 struct rcu_data *rdp)
1310 rcu_lockdep_assert_cblist_protected(rdp);
1311 c = rcu_seq_snap(&rcu_state.gp_seq);
1312 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1313 /* Old request still live, so mark recent callbacks. */
1314 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1317 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1318 needwake = rcu_accelerate_cbs(rnp, rdp);
1319 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1321 rcu_gp_kthread_wake();
1325 * Move any callbacks whose grace period has completed to the
1326 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1327 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1328 * sublist. This function is idempotent, so it does not hurt to
1329 * invoke it repeatedly. As long as it is not invoked -too- often...
1330 * Returns true if the RCU grace-period kthread needs to be awakened.
1332 * The caller must hold rnp->lock with interrupts disabled.
1334 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1336 rcu_lockdep_assert_cblist_protected(rdp);
1337 raw_lockdep_assert_held_rcu_node(rnp);
1339 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1340 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1344 * Find all callbacks whose ->gp_seq numbers indicate that they
1345 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1347 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1349 /* Classify any remaining callbacks. */
1350 return rcu_accelerate_cbs(rnp, rdp);
1354 * Move and classify callbacks, but only if doing so won't require
1355 * that the RCU grace-period kthread be awakened.
1357 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1358 struct rcu_data *rdp)
1360 rcu_lockdep_assert_cblist_protected(rdp);
1361 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1363 // The grace period cannot end while we hold the rcu_node lock.
1364 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1365 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1366 raw_spin_unlock_rcu_node(rnp);
1370 * Update CPU-local rcu_data state to record the beginnings and ends of
1371 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1372 * structure corresponding to the current CPU, and must have irqs disabled.
1373 * Returns true if the grace-period kthread needs to be awakened.
1375 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1379 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1380 rcu_segcblist_is_offloaded(&rdp->cblist);
1382 raw_lockdep_assert_held_rcu_node(rnp);
1384 if (rdp->gp_seq == rnp->gp_seq)
1385 return false; /* Nothing to do. */
1387 /* Handle the ends of any preceding grace periods first. */
1388 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1389 unlikely(READ_ONCE(rdp->gpwrap))) {
1391 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1392 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1395 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1398 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1399 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1400 unlikely(READ_ONCE(rdp->gpwrap))) {
1402 * If the current grace period is waiting for this CPU,
1403 * set up to detect a quiescent state, otherwise don't
1404 * go looking for one.
1406 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1407 need_gp = !!(rnp->qsmask & rdp->grpmask);
1408 rdp->cpu_no_qs.b.norm = need_gp;
1409 rdp->core_needs_qs = need_gp;
1410 zero_cpu_stall_ticks(rdp);
1412 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1413 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1414 rdp->gp_seq_needed = rnp->gp_seq_needed;
1415 WRITE_ONCE(rdp->gpwrap, false);
1416 rcu_gpnum_ovf(rnp, rdp);
1420 static void note_gp_changes(struct rcu_data *rdp)
1422 unsigned long flags;
1424 struct rcu_node *rnp;
1426 local_irq_save(flags);
1428 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1429 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1430 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1431 local_irq_restore(flags);
1434 needwake = __note_gp_changes(rnp, rdp);
1435 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1437 rcu_gp_kthread_wake();
1440 static void rcu_gp_slow(int delay)
1443 !(rcu_seq_ctr(rcu_state.gp_seq) %
1444 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1445 schedule_timeout_uninterruptible(delay);
1449 * Initialize a new grace period. Return false if no grace period required.
1451 static bool rcu_gp_init(void)
1453 unsigned long flags;
1454 unsigned long oldmask;
1456 struct rcu_data *rdp;
1457 struct rcu_node *rnp = rcu_get_root();
1459 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1460 raw_spin_lock_irq_rcu_node(rnp);
1461 if (!READ_ONCE(rcu_state.gp_flags)) {
1462 /* Spurious wakeup, tell caller to go back to sleep. */
1463 raw_spin_unlock_irq_rcu_node(rnp);
1466 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1468 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1470 * Grace period already in progress, don't start another.
1471 * Not supposed to be able to happen.
1473 raw_spin_unlock_irq_rcu_node(rnp);
1477 /* Advance to a new grace period and initialize state. */
1478 record_gp_stall_check_time();
1479 /* Record GP times before starting GP, hence rcu_seq_start(). */
1480 rcu_seq_start(&rcu_state.gp_seq);
1481 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1482 raw_spin_unlock_irq_rcu_node(rnp);
1485 * Apply per-leaf buffered online and offline operations to the
1486 * rcu_node tree. Note that this new grace period need not wait
1487 * for subsequent online CPUs, and that quiescent-state forcing
1488 * will handle subsequent offline CPUs.
1490 rcu_state.gp_state = RCU_GP_ONOFF;
1491 rcu_for_each_leaf_node(rnp) {
1492 raw_spin_lock(&rcu_state.ofl_lock);
1493 raw_spin_lock_irq_rcu_node(rnp);
1494 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1495 !rnp->wait_blkd_tasks) {
1496 /* Nothing to do on this leaf rcu_node structure. */
1497 raw_spin_unlock_irq_rcu_node(rnp);
1498 raw_spin_unlock(&rcu_state.ofl_lock);
1502 /* Record old state, apply changes to ->qsmaskinit field. */
1503 oldmask = rnp->qsmaskinit;
1504 rnp->qsmaskinit = rnp->qsmaskinitnext;
1506 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1507 if (!oldmask != !rnp->qsmaskinit) {
1508 if (!oldmask) { /* First online CPU for rcu_node. */
1509 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1510 rcu_init_new_rnp(rnp);
1511 } else if (rcu_preempt_has_tasks(rnp)) {
1512 rnp->wait_blkd_tasks = true; /* blocked tasks */
1513 } else { /* Last offline CPU and can propagate. */
1514 rcu_cleanup_dead_rnp(rnp);
1519 * If all waited-on tasks from prior grace period are
1520 * done, and if all this rcu_node structure's CPUs are
1521 * still offline, propagate up the rcu_node tree and
1522 * clear ->wait_blkd_tasks. Otherwise, if one of this
1523 * rcu_node structure's CPUs has since come back online,
1524 * simply clear ->wait_blkd_tasks.
1526 if (rnp->wait_blkd_tasks &&
1527 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1528 rnp->wait_blkd_tasks = false;
1529 if (!rnp->qsmaskinit)
1530 rcu_cleanup_dead_rnp(rnp);
1533 raw_spin_unlock_irq_rcu_node(rnp);
1534 raw_spin_unlock(&rcu_state.ofl_lock);
1536 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1539 * Set the quiescent-state-needed bits in all the rcu_node
1540 * structures for all currently online CPUs in breadth-first
1541 * order, starting from the root rcu_node structure, relying on the
1542 * layout of the tree within the rcu_state.node[] array. Note that
1543 * other CPUs will access only the leaves of the hierarchy, thus
1544 * seeing that no grace period is in progress, at least until the
1545 * corresponding leaf node has been initialized.
1547 * The grace period cannot complete until the initialization
1548 * process finishes, because this kthread handles both.
1550 rcu_state.gp_state = RCU_GP_INIT;
1551 rcu_for_each_node_breadth_first(rnp) {
1552 rcu_gp_slow(gp_init_delay);
1553 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1554 rdp = this_cpu_ptr(&rcu_data);
1555 rcu_preempt_check_blocked_tasks(rnp);
1556 rnp->qsmask = rnp->qsmaskinit;
1557 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1558 if (rnp == rdp->mynode)
1559 (void)__note_gp_changes(rnp, rdp);
1560 rcu_preempt_boost_start_gp(rnp);
1561 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1562 rnp->level, rnp->grplo,
1563 rnp->grphi, rnp->qsmask);
1564 /* Quiescent states for tasks on any now-offline CPUs. */
1565 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1566 rnp->rcu_gp_init_mask = mask;
1567 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1568 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1570 raw_spin_unlock_irq_rcu_node(rnp);
1571 cond_resched_tasks_rcu_qs();
1572 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1579 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1582 static bool rcu_gp_fqs_check_wake(int *gfp)
1584 struct rcu_node *rnp = rcu_get_root();
1586 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1587 *gfp = READ_ONCE(rcu_state.gp_flags);
1588 if (*gfp & RCU_GP_FLAG_FQS)
1591 /* The current grace period has completed. */
1592 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1599 * Do one round of quiescent-state forcing.
1601 static void rcu_gp_fqs(bool first_time)
1603 struct rcu_node *rnp = rcu_get_root();
1605 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1606 WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1608 /* Collect dyntick-idle snapshots. */
1609 force_qs_rnp(dyntick_save_progress_counter);
1611 /* Handle dyntick-idle and offline CPUs. */
1612 force_qs_rnp(rcu_implicit_dynticks_qs);
1614 /* Clear flag to prevent immediate re-entry. */
1615 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1616 raw_spin_lock_irq_rcu_node(rnp);
1617 WRITE_ONCE(rcu_state.gp_flags,
1618 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1619 raw_spin_unlock_irq_rcu_node(rnp);
1624 * Loop doing repeated quiescent-state forcing until the grace period ends.
1626 static void rcu_gp_fqs_loop(void)
1632 struct rcu_node *rnp = rcu_get_root();
1634 first_gp_fqs = true;
1635 j = READ_ONCE(jiffies_till_first_fqs);
1639 rcu_state.jiffies_force_qs = jiffies + j;
1640 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1641 jiffies + (j ? 3 * j : 2));
1643 trace_rcu_grace_period(rcu_state.name,
1644 READ_ONCE(rcu_state.gp_seq),
1646 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1647 ret = swait_event_idle_timeout_exclusive(
1648 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1649 rcu_state.gp_state = RCU_GP_DOING_FQS;
1650 /* Locking provides needed memory barriers. */
1651 /* If grace period done, leave loop. */
1652 if (!READ_ONCE(rnp->qsmask) &&
1653 !rcu_preempt_blocked_readers_cgp(rnp))
1655 /* If time for quiescent-state forcing, do it. */
1656 if (ULONG_CMP_GE(jiffies, rcu_state.jiffies_force_qs) ||
1657 (gf & RCU_GP_FLAG_FQS)) {
1658 trace_rcu_grace_period(rcu_state.name,
1659 READ_ONCE(rcu_state.gp_seq),
1661 rcu_gp_fqs(first_gp_fqs);
1662 first_gp_fqs = false;
1663 trace_rcu_grace_period(rcu_state.name,
1664 READ_ONCE(rcu_state.gp_seq),
1666 cond_resched_tasks_rcu_qs();
1667 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1668 ret = 0; /* Force full wait till next FQS. */
1669 j = READ_ONCE(jiffies_till_next_fqs);
1671 /* Deal with stray signal. */
1672 cond_resched_tasks_rcu_qs();
1673 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1674 WARN_ON(signal_pending(current));
1675 trace_rcu_grace_period(rcu_state.name,
1676 READ_ONCE(rcu_state.gp_seq),
1678 ret = 1; /* Keep old FQS timing. */
1680 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1683 j = rcu_state.jiffies_force_qs - j;
1689 * Clean up after the old grace period.
1691 static void rcu_gp_cleanup(void)
1693 unsigned long gp_duration;
1694 bool needgp = false;
1695 unsigned long new_gp_seq;
1697 struct rcu_data *rdp;
1698 struct rcu_node *rnp = rcu_get_root();
1699 struct swait_queue_head *sq;
1701 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1702 raw_spin_lock_irq_rcu_node(rnp);
1703 rcu_state.gp_end = jiffies;
1704 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1705 if (gp_duration > rcu_state.gp_max)
1706 rcu_state.gp_max = gp_duration;
1709 * We know the grace period is complete, but to everyone else
1710 * it appears to still be ongoing. But it is also the case
1711 * that to everyone else it looks like there is nothing that
1712 * they can do to advance the grace period. It is therefore
1713 * safe for us to drop the lock in order to mark the grace
1714 * period as completed in all of the rcu_node structures.
1716 raw_spin_unlock_irq_rcu_node(rnp);
1719 * Propagate new ->gp_seq value to rcu_node structures so that
1720 * other CPUs don't have to wait until the start of the next grace
1721 * period to process their callbacks. This also avoids some nasty
1722 * RCU grace-period initialization races by forcing the end of
1723 * the current grace period to be completely recorded in all of
1724 * the rcu_node structures before the beginning of the next grace
1725 * period is recorded in any of the rcu_node structures.
1727 new_gp_seq = rcu_state.gp_seq;
1728 rcu_seq_end(&new_gp_seq);
1729 rcu_for_each_node_breadth_first(rnp) {
1730 raw_spin_lock_irq_rcu_node(rnp);
1731 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1732 dump_blkd_tasks(rnp, 10);
1733 WARN_ON_ONCE(rnp->qsmask);
1734 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1735 rdp = this_cpu_ptr(&rcu_data);
1736 if (rnp == rdp->mynode)
1737 needgp = __note_gp_changes(rnp, rdp) || needgp;
1738 /* smp_mb() provided by prior unlock-lock pair. */
1739 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1740 sq = rcu_nocb_gp_get(rnp);
1741 raw_spin_unlock_irq_rcu_node(rnp);
1742 rcu_nocb_gp_cleanup(sq);
1743 cond_resched_tasks_rcu_qs();
1744 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1745 rcu_gp_slow(gp_cleanup_delay);
1747 rnp = rcu_get_root();
1748 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1750 /* Declare grace period done, trace first to use old GP number. */
1751 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1752 rcu_seq_end(&rcu_state.gp_seq);
1753 rcu_state.gp_state = RCU_GP_IDLE;
1754 /* Check for GP requests since above loop. */
1755 rdp = this_cpu_ptr(&rcu_data);
1756 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1757 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1758 TPS("CleanupMore"));
1761 /* Advance CBs to reduce false positives below. */
1762 offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1763 rcu_segcblist_is_offloaded(&rdp->cblist);
1764 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1765 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1766 rcu_state.gp_req_activity = jiffies;
1767 trace_rcu_grace_period(rcu_state.name,
1768 READ_ONCE(rcu_state.gp_seq),
1771 WRITE_ONCE(rcu_state.gp_flags,
1772 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1774 raw_spin_unlock_irq_rcu_node(rnp);
1778 * Body of kthread that handles grace periods.
1780 static int __noreturn rcu_gp_kthread(void *unused)
1782 rcu_bind_gp_kthread();
1785 /* Handle grace-period start. */
1787 trace_rcu_grace_period(rcu_state.name,
1788 READ_ONCE(rcu_state.gp_seq),
1790 rcu_state.gp_state = RCU_GP_WAIT_GPS;
1791 swait_event_idle_exclusive(rcu_state.gp_wq,
1792 READ_ONCE(rcu_state.gp_flags) &
1794 rcu_state.gp_state = RCU_GP_DONE_GPS;
1795 /* Locking provides needed memory barrier. */
1798 cond_resched_tasks_rcu_qs();
1799 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1800 WARN_ON(signal_pending(current));
1801 trace_rcu_grace_period(rcu_state.name,
1802 READ_ONCE(rcu_state.gp_seq),
1806 /* Handle quiescent-state forcing. */
1809 /* Handle grace-period end. */
1810 rcu_state.gp_state = RCU_GP_CLEANUP;
1812 rcu_state.gp_state = RCU_GP_CLEANED;
1817 * Report a full set of quiescent states to the rcu_state data structure.
1818 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1819 * another grace period is required. Whether we wake the grace-period
1820 * kthread or it awakens itself for the next round of quiescent-state
1821 * forcing, that kthread will clean up after the just-completed grace
1822 * period. Note that the caller must hold rnp->lock, which is released
1825 static void rcu_report_qs_rsp(unsigned long flags)
1826 __releases(rcu_get_root()->lock)
1828 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1829 WARN_ON_ONCE(!rcu_gp_in_progress());
1830 WRITE_ONCE(rcu_state.gp_flags,
1831 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1832 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1833 rcu_gp_kthread_wake();
1837 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1838 * Allows quiescent states for a group of CPUs to be reported at one go
1839 * to the specified rcu_node structure, though all the CPUs in the group
1840 * must be represented by the same rcu_node structure (which need not be a
1841 * leaf rcu_node structure, though it often will be). The gps parameter
1842 * is the grace-period snapshot, which means that the quiescent states
1843 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1844 * must be held upon entry, and it is released before return.
1846 * As a special case, if mask is zero, the bit-already-cleared check is
1847 * disabled. This allows propagating quiescent state due to resumed tasks
1848 * during grace-period initialization.
1850 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1851 unsigned long gps, unsigned long flags)
1852 __releases(rnp->lock)
1854 unsigned long oldmask = 0;
1855 struct rcu_node *rnp_c;
1857 raw_lockdep_assert_held_rcu_node(rnp);
1859 /* Walk up the rcu_node hierarchy. */
1861 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1864 * Our bit has already been cleared, or the
1865 * relevant grace period is already over, so done.
1867 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1870 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1871 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1872 rcu_preempt_blocked_readers_cgp(rnp));
1873 rnp->qsmask &= ~mask;
1874 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1875 mask, rnp->qsmask, rnp->level,
1876 rnp->grplo, rnp->grphi,
1878 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1880 /* Other bits still set at this level, so done. */
1881 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1884 rnp->completedqs = rnp->gp_seq;
1885 mask = rnp->grpmask;
1886 if (rnp->parent == NULL) {
1888 /* No more levels. Exit loop holding root lock. */
1892 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1895 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1896 oldmask = rnp_c->qsmask;
1900 * Get here if we are the last CPU to pass through a quiescent
1901 * state for this grace period. Invoke rcu_report_qs_rsp()
1902 * to clean up and start the next grace period if one is needed.
1904 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1908 * Record a quiescent state for all tasks that were previously queued
1909 * on the specified rcu_node structure and that were blocking the current
1910 * RCU grace period. The caller must hold the corresponding rnp->lock with
1911 * irqs disabled, and this lock is released upon return, but irqs remain
1914 static void __maybe_unused
1915 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1916 __releases(rnp->lock)
1920 struct rcu_node *rnp_p;
1922 raw_lockdep_assert_held_rcu_node(rnp);
1923 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPTION)) ||
1924 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1926 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1927 return; /* Still need more quiescent states! */
1930 rnp->completedqs = rnp->gp_seq;
1931 rnp_p = rnp->parent;
1932 if (rnp_p == NULL) {
1934 * Only one rcu_node structure in the tree, so don't
1935 * try to report up to its nonexistent parent!
1937 rcu_report_qs_rsp(flags);
1941 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1943 mask = rnp->grpmask;
1944 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1945 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1946 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1950 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1951 * structure. This must be called from the specified CPU.
1954 rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
1956 unsigned long flags;
1958 bool needwake = false;
1959 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1960 rcu_segcblist_is_offloaded(&rdp->cblist);
1961 struct rcu_node *rnp;
1964 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1965 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
1969 * The grace period in which this quiescent state was
1970 * recorded has ended, so don't report it upwards.
1971 * We will instead need a new quiescent state that lies
1972 * within the current grace period.
1974 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
1975 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1978 mask = rdp->grpmask;
1979 rdp->core_needs_qs = false;
1980 if ((rnp->qsmask & mask) == 0) {
1981 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1984 * This GP can't end until cpu checks in, so all of our
1985 * callbacks can be processed during the next GP.
1988 needwake = rcu_accelerate_cbs(rnp, rdp);
1990 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1991 /* ^^^ Released rnp->lock */
1993 rcu_gp_kthread_wake();
1998 * Check to see if there is a new grace period of which this CPU
1999 * is not yet aware, and if so, set up local rcu_data state for it.
2000 * Otherwise, see if this CPU has just passed through its first
2001 * quiescent state for this grace period, and record that fact if so.
2004 rcu_check_quiescent_state(struct rcu_data *rdp)
2006 /* Check for grace-period ends and beginnings. */
2007 note_gp_changes(rdp);
2010 * Does this CPU still need to do its part for current grace period?
2011 * If no, return and let the other CPUs do their part as well.
2013 if (!rdp->core_needs_qs)
2017 * Was there a quiescent state since the beginning of the grace
2018 * period? If no, then exit and wait for the next call.
2020 if (rdp->cpu_no_qs.b.norm)
2024 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2027 rcu_report_qs_rdp(rdp->cpu, rdp);
2031 * Near the end of the offline process. Trace the fact that this CPU
2034 int rcutree_dying_cpu(unsigned int cpu)
2037 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2038 struct rcu_node *rnp = rdp->mynode;
2040 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2043 blkd = !!(rnp->qsmask & rdp->grpmask);
2044 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq,
2045 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2050 * All CPUs for the specified rcu_node structure have gone offline,
2051 * and all tasks that were preempted within an RCU read-side critical
2052 * section while running on one of those CPUs have since exited their RCU
2053 * read-side critical section. Some other CPU is reporting this fact with
2054 * the specified rcu_node structure's ->lock held and interrupts disabled.
2055 * This function therefore goes up the tree of rcu_node structures,
2056 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2057 * the leaf rcu_node structure's ->qsmaskinit field has already been
2060 * This function does check that the specified rcu_node structure has
2061 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2062 * prematurely. That said, invoking it after the fact will cost you
2063 * a needless lock acquisition. So once it has done its work, don't
2066 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2069 struct rcu_node *rnp = rnp_leaf;
2071 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2072 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2073 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2074 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2077 mask = rnp->grpmask;
2081 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2082 rnp->qsmaskinit &= ~mask;
2083 /* Between grace periods, so better already be zero! */
2084 WARN_ON_ONCE(rnp->qsmask);
2085 if (rnp->qsmaskinit) {
2086 raw_spin_unlock_rcu_node(rnp);
2087 /* irqs remain disabled. */
2090 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2095 * The CPU has been completely removed, and some other CPU is reporting
2096 * this fact from process context. Do the remainder of the cleanup.
2097 * There can only be one CPU hotplug operation at a time, so no need for
2100 int rcutree_dead_cpu(unsigned int cpu)
2102 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2103 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2105 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2108 /* Adjust any no-longer-needed kthreads. */
2109 rcu_boost_kthread_setaffinity(rnp, -1);
2110 /* Do any needed no-CB deferred wakeups from this CPU. */
2111 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2116 * Invoke any RCU callbacks that have made it to the end of their grace
2117 * period. Thottle as specified by rdp->blimit.
2119 static void rcu_do_batch(struct rcu_data *rdp)
2121 unsigned long flags;
2122 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2123 rcu_segcblist_is_offloaded(&rdp->cblist);
2124 struct rcu_head *rhp;
2125 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2127 long pending, tlimit = 0;
2129 /* If no callbacks are ready, just return. */
2130 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2131 trace_rcu_batch_start(rcu_state.name,
2132 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2133 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2134 trace_rcu_batch_end(rcu_state.name, 0,
2135 !rcu_segcblist_empty(&rdp->cblist),
2136 need_resched(), is_idle_task(current),
2137 rcu_is_callbacks_kthread());
2142 * Extract the list of ready callbacks, disabling to prevent
2143 * races with call_rcu() from interrupt handlers. Leave the
2144 * callback counts, as rcu_barrier() needs to be conservative.
2146 local_irq_save(flags);
2148 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2149 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2150 bl = max(rdp->blimit, pending >> rcu_divisor);
2151 if (unlikely(bl > 100))
2152 tlimit = local_clock() + rcu_resched_ns;
2153 trace_rcu_batch_start(rcu_state.name,
2154 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2155 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2156 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2158 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2159 rcu_nocb_unlock_irqrestore(rdp, flags);
2161 /* Invoke callbacks. */
2162 rhp = rcu_cblist_dequeue(&rcl);
2163 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2164 debug_rcu_head_unqueue(rhp);
2165 if (__rcu_reclaim(rcu_state.name, rhp))
2166 rcu_cblist_dequeued_lazy(&rcl);
2168 * Stop only if limit reached and CPU has something to do.
2169 * Note: The rcl structure counts down from zero.
2171 if (-rcl.len >= bl && !offloaded &&
2173 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2175 if (unlikely(tlimit)) {
2176 /* only call local_clock() every 32 callbacks */
2177 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2179 /* Exceeded the time limit, so leave. */
2183 WARN_ON_ONCE(in_serving_softirq());
2185 lockdep_assert_irqs_enabled();
2186 cond_resched_tasks_rcu_qs();
2187 lockdep_assert_irqs_enabled();
2192 local_irq_save(flags);
2195 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2196 is_idle_task(current), rcu_is_callbacks_kthread());
2198 /* Update counts and requeue any remaining callbacks. */
2199 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2200 smp_mb(); /* List handling before counting for rcu_barrier(). */
2201 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2203 /* Reinstate batch limit if we have worked down the excess. */
2204 count = rcu_segcblist_n_cbs(&rdp->cblist);
2205 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2206 rdp->blimit = blimit;
2208 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2209 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2210 rdp->qlen_last_fqs_check = 0;
2211 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2212 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2213 rdp->qlen_last_fqs_check = count;
2216 * The following usually indicates a double call_rcu(). To track
2217 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2219 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2220 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2221 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2223 rcu_nocb_unlock_irqrestore(rdp, flags);
2225 /* Re-invoke RCU core processing if there are callbacks remaining. */
2226 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2231 * This function is invoked from each scheduling-clock interrupt,
2232 * and checks to see if this CPU is in a non-context-switch quiescent
2233 * state, for example, user mode or idle loop. It also schedules RCU
2234 * core processing. If the current grace period has gone on too long,
2235 * it will ask the scheduler to manufacture a context switch for the sole
2236 * purpose of providing a providing the needed quiescent state.
2238 void rcu_sched_clock_irq(int user)
2240 trace_rcu_utilization(TPS("Start scheduler-tick"));
2241 raw_cpu_inc(rcu_data.ticks_this_gp);
2242 /* The load-acquire pairs with the store-release setting to true. */
2243 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2244 /* Idle and userspace execution already are quiescent states. */
2245 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2246 set_tsk_need_resched(current);
2247 set_preempt_need_resched();
2249 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2251 rcu_flavor_sched_clock_irq(user);
2255 trace_rcu_utilization(TPS("End scheduler-tick"));
2259 * Scan the leaf rcu_node structures. For each structure on which all
2260 * CPUs have reported a quiescent state and on which there are tasks
2261 * blocking the current grace period, initiate RCU priority boosting.
2262 * Otherwise, invoke the specified function to check dyntick state for
2263 * each CPU that has not yet reported a quiescent state.
2265 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2268 unsigned long flags;
2270 struct rcu_node *rnp;
2272 rcu_for_each_leaf_node(rnp) {
2273 cond_resched_tasks_rcu_qs();
2275 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2276 if (rnp->qsmask == 0) {
2277 if (!IS_ENABLED(CONFIG_PREEMPTION) ||
2278 rcu_preempt_blocked_readers_cgp(rnp)) {
2280 * No point in scanning bits because they
2281 * are all zero. But we might need to
2282 * priority-boost blocked readers.
2284 rcu_initiate_boost(rnp, flags);
2285 /* rcu_initiate_boost() releases rnp->lock */
2288 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2291 for_each_leaf_node_possible_cpu(rnp, cpu) {
2292 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2293 if ((rnp->qsmask & bit) != 0) {
2294 if (f(per_cpu_ptr(&rcu_data, cpu)))
2299 /* Idle/offline CPUs, report (releases rnp->lock). */
2300 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2302 /* Nothing to do here, so just drop the lock. */
2303 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2309 * Force quiescent states on reluctant CPUs, and also detect which
2310 * CPUs are in dyntick-idle mode.
2312 void rcu_force_quiescent_state(void)
2314 unsigned long flags;
2316 struct rcu_node *rnp;
2317 struct rcu_node *rnp_old = NULL;
2319 /* Funnel through hierarchy to reduce memory contention. */
2320 rnp = __this_cpu_read(rcu_data.mynode);
2321 for (; rnp != NULL; rnp = rnp->parent) {
2322 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2323 !raw_spin_trylock(&rnp->fqslock);
2324 if (rnp_old != NULL)
2325 raw_spin_unlock(&rnp_old->fqslock);
2330 /* rnp_old == rcu_get_root(), rnp == NULL. */
2332 /* Reached the root of the rcu_node tree, acquire lock. */
2333 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2334 raw_spin_unlock(&rnp_old->fqslock);
2335 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2336 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2337 return; /* Someone beat us to it. */
2339 WRITE_ONCE(rcu_state.gp_flags,
2340 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2341 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2342 rcu_gp_kthread_wake();
2344 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2346 /* Perform RCU core processing work for the current CPU. */
2347 static __latent_entropy void rcu_core(void)
2349 unsigned long flags;
2350 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2351 struct rcu_node *rnp = rdp->mynode;
2352 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2353 rcu_segcblist_is_offloaded(&rdp->cblist);
2355 if (cpu_is_offline(smp_processor_id()))
2357 trace_rcu_utilization(TPS("Start RCU core"));
2358 WARN_ON_ONCE(!rdp->beenonline);
2360 /* Report any deferred quiescent states if preemption enabled. */
2361 if (!(preempt_count() & PREEMPT_MASK)) {
2362 rcu_preempt_deferred_qs(current);
2363 } else if (rcu_preempt_need_deferred_qs(current)) {
2364 set_tsk_need_resched(current);
2365 set_preempt_need_resched();
2368 /* Update RCU state based on any recent quiescent states. */
2369 rcu_check_quiescent_state(rdp);
2371 /* No grace period and unregistered callbacks? */
2372 if (!rcu_gp_in_progress() &&
2373 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2374 local_irq_save(flags);
2375 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2376 rcu_accelerate_cbs_unlocked(rnp, rdp);
2377 local_irq_restore(flags);
2380 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2382 /* If there are callbacks ready, invoke them. */
2383 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2384 likely(READ_ONCE(rcu_scheduler_fully_active)))
2387 /* Do any needed deferred wakeups of rcuo kthreads. */
2388 do_nocb_deferred_wakeup(rdp);
2389 trace_rcu_utilization(TPS("End RCU core"));
2392 static void rcu_core_si(struct softirq_action *h)
2397 static void rcu_wake_cond(struct task_struct *t, int status)
2400 * If the thread is yielding, only wake it when this
2401 * is invoked from idle
2403 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2407 static void invoke_rcu_core_kthread(void)
2409 struct task_struct *t;
2410 unsigned long flags;
2412 local_irq_save(flags);
2413 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2414 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2415 if (t != NULL && t != current)
2416 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2417 local_irq_restore(flags);
2421 * Wake up this CPU's rcuc kthread to do RCU core processing.
2423 static void invoke_rcu_core(void)
2425 if (!cpu_online(smp_processor_id()))
2428 raise_softirq(RCU_SOFTIRQ);
2430 invoke_rcu_core_kthread();
2433 static void rcu_cpu_kthread_park(unsigned int cpu)
2435 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2438 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2440 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2444 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2445 * the RCU softirq used in configurations of RCU that do not support RCU
2446 * priority boosting.
2448 static void rcu_cpu_kthread(unsigned int cpu)
2450 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2451 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2454 for (spincnt = 0; spincnt < 10; spincnt++) {
2455 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
2457 *statusp = RCU_KTHREAD_RUNNING;
2458 local_irq_disable();
2466 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2467 *statusp = RCU_KTHREAD_WAITING;
2471 *statusp = RCU_KTHREAD_YIELDING;
2472 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2473 schedule_timeout_interruptible(2);
2474 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2475 *statusp = RCU_KTHREAD_WAITING;
2478 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2479 .store = &rcu_data.rcu_cpu_kthread_task,
2480 .thread_should_run = rcu_cpu_kthread_should_run,
2481 .thread_fn = rcu_cpu_kthread,
2482 .thread_comm = "rcuc/%u",
2483 .setup = rcu_cpu_kthread_setup,
2484 .park = rcu_cpu_kthread_park,
2488 * Spawn per-CPU RCU core processing kthreads.
2490 static int __init rcu_spawn_core_kthreads(void)
2494 for_each_possible_cpu(cpu)
2495 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2496 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2498 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2499 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2504 * Handle any core-RCU processing required by a call_rcu() invocation.
2506 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2507 unsigned long flags)
2510 * If called from an extended quiescent state, invoke the RCU
2511 * core in order to force a re-evaluation of RCU's idleness.
2513 if (!rcu_is_watching())
2516 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2517 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2521 * Force the grace period if too many callbacks or too long waiting.
2522 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2523 * if some other CPU has recently done so. Also, don't bother
2524 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2525 * is the only one waiting for a grace period to complete.
2527 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2528 rdp->qlen_last_fqs_check + qhimark)) {
2530 /* Are we ignoring a completed grace period? */
2531 note_gp_changes(rdp);
2533 /* Start a new grace period if one not already started. */
2534 if (!rcu_gp_in_progress()) {
2535 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2537 /* Give the grace period a kick. */
2538 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2539 if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2540 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2541 rcu_force_quiescent_state();
2542 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2543 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2549 * RCU callback function to leak a callback.
2551 static void rcu_leak_callback(struct rcu_head *rhp)
2556 * Helper function for call_rcu() and friends. The cpu argument will
2557 * normally be -1, indicating "currently running CPU". It may specify
2558 * a CPU only if that CPU is a no-CBs CPU. Currently, only rcu_barrier()
2559 * is expected to specify a CPU.
2562 __call_rcu(struct rcu_head *head, rcu_callback_t func, bool lazy)
2564 unsigned long flags;
2565 struct rcu_data *rdp;
2568 /* Misaligned rcu_head! */
2569 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2571 if (debug_rcu_head_queue(head)) {
2573 * Probable double call_rcu(), so leak the callback.
2574 * Use rcu:rcu_callback trace event to find the previous
2575 * time callback was passed to __call_rcu().
2577 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2579 WRITE_ONCE(head->func, rcu_leak_callback);
2584 local_irq_save(flags);
2585 rdp = this_cpu_ptr(&rcu_data);
2587 /* Add the callback to our list. */
2588 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2589 // This can trigger due to call_rcu() from offline CPU:
2590 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2591 WARN_ON_ONCE(!rcu_is_watching());
2592 // Very early boot, before rcu_init(). Initialize if needed
2593 // and then drop through to queue the callback.
2594 if (rcu_segcblist_empty(&rdp->cblist))
2595 rcu_segcblist_init(&rdp->cblist);
2597 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2598 return; // Enqueued onto ->nocb_bypass, so just leave.
2599 /* If we get here, rcu_nocb_try_bypass() acquired ->nocb_lock. */
2600 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2601 if (__is_kfree_rcu_offset((unsigned long)func))
2602 trace_rcu_kfree_callback(rcu_state.name, head,
2603 (unsigned long)func,
2604 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2605 rcu_segcblist_n_cbs(&rdp->cblist));
2607 trace_rcu_callback(rcu_state.name, head,
2608 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2609 rcu_segcblist_n_cbs(&rdp->cblist));
2611 /* Go handle any RCU core processing required. */
2612 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2613 unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2614 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2616 __call_rcu_core(rdp, head, flags);
2617 local_irq_restore(flags);
2622 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2623 * @head: structure to be used for queueing the RCU updates.
2624 * @func: actual callback function to be invoked after the grace period
2626 * The callback function will be invoked some time after a full grace
2627 * period elapses, in other words after all pre-existing RCU read-side
2628 * critical sections have completed. However, the callback function
2629 * might well execute concurrently with RCU read-side critical sections
2630 * that started after call_rcu() was invoked. RCU read-side critical
2631 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2632 * may be nested. In addition, regions of code across which interrupts,
2633 * preemption, or softirqs have been disabled also serve as RCU read-side
2634 * critical sections. This includes hardware interrupt handlers, softirq
2635 * handlers, and NMI handlers.
2637 * Note that all CPUs must agree that the grace period extended beyond
2638 * all pre-existing RCU read-side critical section. On systems with more
2639 * than one CPU, this means that when "func()" is invoked, each CPU is
2640 * guaranteed to have executed a full memory barrier since the end of its
2641 * last RCU read-side critical section whose beginning preceded the call
2642 * to call_rcu(). It also means that each CPU executing an RCU read-side
2643 * critical section that continues beyond the start of "func()" must have
2644 * executed a memory barrier after the call_rcu() but before the beginning
2645 * of that RCU read-side critical section. Note that these guarantees
2646 * include CPUs that are offline, idle, or executing in user mode, as
2647 * well as CPUs that are executing in the kernel.
2649 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2650 * resulting RCU callback function "func()", then both CPU A and CPU B are
2651 * guaranteed to execute a full memory barrier during the time interval
2652 * between the call to call_rcu() and the invocation of "func()" -- even
2653 * if CPU A and CPU B are the same CPU (but again only if the system has
2654 * more than one CPU).
2656 void call_rcu(struct rcu_head *head, rcu_callback_t func)
2658 __call_rcu(head, func, 0);
2660 EXPORT_SYMBOL_GPL(call_rcu);
2663 * Queue an RCU callback for lazy invocation after a grace period.
2664 * This will likely be later named something like "call_rcu_lazy()",
2665 * but this change will require some way of tagging the lazy RCU
2666 * callbacks in the list of pending callbacks. Until then, this
2667 * function may only be called from __kfree_rcu().
2669 void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
2671 __call_rcu(head, func, 1);
2673 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2676 * During early boot, any blocking grace-period wait automatically
2677 * implies a grace period. Later on, this is never the case for PREEMPT.
2679 * Howevr, because a context switch is a grace period for !PREEMPT, any
2680 * blocking grace-period wait automatically implies a grace period if
2681 * there is only one CPU online at any point time during execution of
2682 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
2683 * occasionally incorrectly indicate that there are multiple CPUs online
2684 * when there was in fact only one the whole time, as this just adds some
2685 * overhead: RCU still operates correctly.
2687 static int rcu_blocking_is_gp(void)
2691 if (IS_ENABLED(CONFIG_PREEMPTION))
2692 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
2693 might_sleep(); /* Check for RCU read-side critical section. */
2695 ret = num_online_cpus() <= 1;
2701 * synchronize_rcu - wait until a grace period has elapsed.
2703 * Control will return to the caller some time after a full grace
2704 * period has elapsed, in other words after all currently executing RCU
2705 * read-side critical sections have completed. Note, however, that
2706 * upon return from synchronize_rcu(), the caller might well be executing
2707 * concurrently with new RCU read-side critical sections that began while
2708 * synchronize_rcu() was waiting. RCU read-side critical sections are
2709 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
2710 * In addition, regions of code across which interrupts, preemption, or
2711 * softirqs have been disabled also serve as RCU read-side critical
2712 * sections. This includes hardware interrupt handlers, softirq handlers,
2715 * Note that this guarantee implies further memory-ordering guarantees.
2716 * On systems with more than one CPU, when synchronize_rcu() returns,
2717 * each CPU is guaranteed to have executed a full memory barrier since
2718 * the end of its last RCU read-side critical section whose beginning
2719 * preceded the call to synchronize_rcu(). In addition, each CPU having
2720 * an RCU read-side critical section that extends beyond the return from
2721 * synchronize_rcu() is guaranteed to have executed a full memory barrier
2722 * after the beginning of synchronize_rcu() and before the beginning of
2723 * that RCU read-side critical section. Note that these guarantees include
2724 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2725 * that are executing in the kernel.
2727 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
2728 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2729 * to have executed a full memory barrier during the execution of
2730 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
2731 * again only if the system has more than one CPU).
2733 void synchronize_rcu(void)
2735 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
2736 lock_is_held(&rcu_lock_map) ||
2737 lock_is_held(&rcu_sched_lock_map),
2738 "Illegal synchronize_rcu() in RCU read-side critical section");
2739 if (rcu_blocking_is_gp())
2741 if (rcu_gp_is_expedited())
2742 synchronize_rcu_expedited();
2744 wait_rcu_gp(call_rcu);
2746 EXPORT_SYMBOL_GPL(synchronize_rcu);
2749 * get_state_synchronize_rcu - Snapshot current RCU state
2751 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2752 * to determine whether or not a full grace period has elapsed in the
2755 unsigned long get_state_synchronize_rcu(void)
2758 * Any prior manipulation of RCU-protected data must happen
2759 * before the load from ->gp_seq.
2762 return rcu_seq_snap(&rcu_state.gp_seq);
2764 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2767 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2769 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2771 * If a full RCU grace period has elapsed since the earlier call to
2772 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2773 * synchronize_rcu() to wait for a full grace period.
2775 * Yes, this function does not take counter wrap into account. But
2776 * counter wrap is harmless. If the counter wraps, we have waited for
2777 * more than 2 billion grace periods (and way more on a 64-bit system!),
2778 * so waiting for one additional grace period should be just fine.
2780 void cond_synchronize_rcu(unsigned long oldstate)
2782 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
2785 smp_mb(); /* Ensure GP ends before subsequent accesses. */
2787 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2790 * Check to see if there is any immediate RCU-related work to be done by
2791 * the current CPU, returning 1 if so and zero otherwise. The checks are
2792 * in order of increasing expense: checks that can be carried out against
2793 * CPU-local state are performed first. However, we must check for CPU
2794 * stalls first, else we might not get a chance.
2796 static int rcu_pending(void)
2798 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2799 struct rcu_node *rnp = rdp->mynode;
2801 /* Check for CPU stalls, if enabled. */
2802 check_cpu_stall(rdp);
2804 /* Does this CPU need a deferred NOCB wakeup? */
2805 if (rcu_nocb_need_deferred_wakeup(rdp))
2808 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2809 if (rcu_nohz_full_cpu())
2812 /* Is the RCU core waiting for a quiescent state from this CPU? */
2813 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
2816 /* Does this CPU have callbacks ready to invoke? */
2817 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2820 /* Has RCU gone idle with this CPU needing another grace period? */
2821 if (!rcu_gp_in_progress() &&
2822 rcu_segcblist_is_enabled(&rdp->cblist) &&
2823 (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
2824 !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
2825 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2828 /* Have RCU grace period completed or started? */
2829 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
2830 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
2838 * Helper function for rcu_barrier() tracing. If tracing is disabled,
2839 * the compiler is expected to optimize this away.
2841 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
2843 trace_rcu_barrier(rcu_state.name, s, cpu,
2844 atomic_read(&rcu_state.barrier_cpu_count), done);
2848 * RCU callback function for rcu_barrier(). If we are last, wake
2849 * up the task executing rcu_barrier().
2851 static void rcu_barrier_callback(struct rcu_head *rhp)
2853 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
2854 rcu_barrier_trace(TPS("LastCB"), -1,
2855 rcu_state.barrier_sequence);
2856 complete(&rcu_state.barrier_completion);
2858 rcu_barrier_trace(TPS("CB"), -1, rcu_state.barrier_sequence);
2863 * Called with preemption disabled, and from cross-cpu IRQ context.
2865 static void rcu_barrier_func(void *unused)
2867 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2869 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
2870 rdp->barrier_head.func = rcu_barrier_callback;
2871 debug_rcu_head_queue(&rdp->barrier_head);
2873 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
2874 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
2875 atomic_inc(&rcu_state.barrier_cpu_count);
2877 debug_rcu_head_unqueue(&rdp->barrier_head);
2878 rcu_barrier_trace(TPS("IRQNQ"), -1,
2879 rcu_state.barrier_sequence);
2881 rcu_nocb_unlock(rdp);
2885 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
2887 * Note that this primitive does not necessarily wait for an RCU grace period
2888 * to complete. For example, if there are no RCU callbacks queued anywhere
2889 * in the system, then rcu_barrier() is within its rights to return
2890 * immediately, without waiting for anything, much less an RCU grace period.
2892 void rcu_barrier(void)
2895 struct rcu_data *rdp;
2896 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
2898 rcu_barrier_trace(TPS("Begin"), -1, s);
2900 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2901 mutex_lock(&rcu_state.barrier_mutex);
2903 /* Did someone else do our work for us? */
2904 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
2905 rcu_barrier_trace(TPS("EarlyExit"), -1,
2906 rcu_state.barrier_sequence);
2907 smp_mb(); /* caller's subsequent code after above check. */
2908 mutex_unlock(&rcu_state.barrier_mutex);
2912 /* Mark the start of the barrier operation. */
2913 rcu_seq_start(&rcu_state.barrier_sequence);
2914 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
2917 * Initialize the count to one rather than to zero in order to
2918 * avoid a too-soon return to zero in case of a short grace period
2919 * (or preemption of this task). Exclude CPU-hotplug operations
2920 * to ensure that no offline CPU has callbacks queued.
2922 init_completion(&rcu_state.barrier_completion);
2923 atomic_set(&rcu_state.barrier_cpu_count, 1);
2927 * Force each CPU with callbacks to register a new callback.
2928 * When that callback is invoked, we will know that all of the
2929 * corresponding CPU's preceding callbacks have been invoked.
2931 for_each_possible_cpu(cpu) {
2932 rdp = per_cpu_ptr(&rcu_data, cpu);
2933 if (!cpu_online(cpu) &&
2934 !rcu_segcblist_is_offloaded(&rdp->cblist))
2936 if (rcu_segcblist_n_cbs(&rdp->cblist)) {
2937 rcu_barrier_trace(TPS("OnlineQ"), cpu,
2938 rcu_state.barrier_sequence);
2939 smp_call_function_single(cpu, rcu_barrier_func, NULL, 1);
2941 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
2942 rcu_state.barrier_sequence);
2948 * Now that we have an rcu_barrier_callback() callback on each
2949 * CPU, and thus each counted, remove the initial count.
2951 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count))
2952 complete(&rcu_state.barrier_completion);
2954 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2955 wait_for_completion(&rcu_state.barrier_completion);
2957 /* Mark the end of the barrier operation. */
2958 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
2959 rcu_seq_end(&rcu_state.barrier_sequence);
2961 /* Other rcu_barrier() invocations can now safely proceed. */
2962 mutex_unlock(&rcu_state.barrier_mutex);
2964 EXPORT_SYMBOL_GPL(rcu_barrier);
2967 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
2968 * first CPU in a given leaf rcu_node structure coming online. The caller
2969 * must hold the corresponding leaf rcu_node ->lock with interrrupts
2972 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
2976 struct rcu_node *rnp = rnp_leaf;
2978 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2979 WARN_ON_ONCE(rnp->wait_blkd_tasks);
2981 mask = rnp->grpmask;
2985 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
2986 oldmask = rnp->qsmaskinit;
2987 rnp->qsmaskinit |= mask;
2988 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
2995 * Do boot-time initialization of a CPU's per-CPU RCU data.
2998 rcu_boot_init_percpu_data(int cpu)
3000 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3002 /* Set up local state, ensuring consistent view of global state. */
3003 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3004 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
3005 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
3006 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3007 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3008 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3009 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3011 rcu_boot_init_nocb_percpu_data(rdp);
3015 * Invoked early in the CPU-online process, when pretty much all services
3016 * are available. The incoming CPU is not present.
3018 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3019 * offline event can be happening at a given time. Note also that we can
3020 * accept some slop in the rsp->gp_seq access due to the fact that this
3021 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3022 * And any offloaded callbacks are being numbered elsewhere.
3024 int rcutree_prepare_cpu(unsigned int cpu)
3026 unsigned long flags;
3027 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3028 struct rcu_node *rnp = rcu_get_root();
3030 /* Set up local state, ensuring consistent view of global state. */
3031 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3032 rdp->qlen_last_fqs_check = 0;
3033 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
3034 rdp->blimit = blimit;
3035 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3036 !rcu_segcblist_is_offloaded(&rdp->cblist))
3037 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3038 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
3039 rcu_dynticks_eqs_online();
3040 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3043 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3044 * propagation up the rcu_node tree will happen at the beginning
3045 * of the next grace period.
3048 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3049 rdp->beenonline = true; /* We have now been online. */
3050 rdp->gp_seq = rnp->gp_seq;
3051 rdp->gp_seq_needed = rnp->gp_seq;
3052 rdp->cpu_no_qs.b.norm = true;
3053 rdp->core_needs_qs = false;
3054 rdp->rcu_iw_pending = false;
3055 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3056 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3057 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3058 rcu_prepare_kthreads(cpu);
3059 rcu_spawn_cpu_nocb_kthread(cpu);
3065 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3067 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3069 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3071 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3075 * Near the end of the CPU-online process. Pretty much all services
3076 * enabled, and the CPU is now very much alive.
3078 int rcutree_online_cpu(unsigned int cpu)
3080 unsigned long flags;
3081 struct rcu_data *rdp;
3082 struct rcu_node *rnp;
3084 rdp = per_cpu_ptr(&rcu_data, cpu);
3086 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3087 rnp->ffmask |= rdp->grpmask;
3088 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3089 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3090 return 0; /* Too early in boot for scheduler work. */
3091 sync_sched_exp_online_cleanup(cpu);
3092 rcutree_affinity_setting(cpu, -1);
3097 * Near the beginning of the process. The CPU is still very much alive
3098 * with pretty much all services enabled.
3100 int rcutree_offline_cpu(unsigned int cpu)
3102 unsigned long flags;
3103 struct rcu_data *rdp;
3104 struct rcu_node *rnp;
3106 rdp = per_cpu_ptr(&rcu_data, cpu);
3108 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3109 rnp->ffmask &= ~rdp->grpmask;
3110 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3112 rcutree_affinity_setting(cpu, cpu);
3116 static DEFINE_PER_CPU(int, rcu_cpu_started);
3119 * Mark the specified CPU as being online so that subsequent grace periods
3120 * (both expedited and normal) will wait on it. Note that this means that
3121 * incoming CPUs are not allowed to use RCU read-side critical sections
3122 * until this function is called. Failing to observe this restriction
3123 * will result in lockdep splats.
3125 * Note that this function is special in that it is invoked directly
3126 * from the incoming CPU rather than from the cpuhp_step mechanism.
3127 * This is because this function must be invoked at a precise location.
3129 void rcu_cpu_starting(unsigned int cpu)
3131 unsigned long flags;
3134 unsigned long oldmask;
3135 struct rcu_data *rdp;
3136 struct rcu_node *rnp;
3138 if (per_cpu(rcu_cpu_started, cpu))
3141 per_cpu(rcu_cpu_started, cpu) = 1;
3143 rdp = per_cpu_ptr(&rcu_data, cpu);
3145 mask = rdp->grpmask;
3146 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3147 rnp->qsmaskinitnext |= mask;
3148 oldmask = rnp->expmaskinitnext;
3149 rnp->expmaskinitnext |= mask;
3150 oldmask ^= rnp->expmaskinitnext;
3151 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3152 /* Allow lockless access for expedited grace periods. */
3153 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */
3154 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3155 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3156 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3157 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3158 /* Report QS -after- changing ->qsmaskinitnext! */
3159 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3161 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3163 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3167 * The outgoing function has no further need of RCU, so remove it from
3168 * the rcu_node tree's ->qsmaskinitnext bit masks.
3170 * Note that this function is special in that it is invoked directly
3171 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3172 * This is because this function must be invoked at a precise location.
3174 void rcu_report_dead(unsigned int cpu)
3176 unsigned long flags;
3178 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3179 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3181 /* QS for any half-done expedited grace period. */
3183 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
3185 rcu_preempt_deferred_qs(current);
3187 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3188 mask = rdp->grpmask;
3189 raw_spin_lock(&rcu_state.ofl_lock);
3190 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3191 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3192 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3193 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3194 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3195 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3196 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3198 rnp->qsmaskinitnext &= ~mask;
3199 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3200 raw_spin_unlock(&rcu_state.ofl_lock);
3202 per_cpu(rcu_cpu_started, cpu) = 0;
3205 #ifdef CONFIG_HOTPLUG_CPU
3207 * The outgoing CPU has just passed through the dying-idle state, and we
3208 * are being invoked from the CPU that was IPIed to continue the offline
3209 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
3211 void rcutree_migrate_callbacks(int cpu)
3213 unsigned long flags;
3214 struct rcu_data *my_rdp;
3215 struct rcu_node *my_rnp;
3216 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3219 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
3220 rcu_segcblist_empty(&rdp->cblist))
3221 return; /* No callbacks to migrate. */
3223 local_irq_save(flags);
3224 my_rdp = this_cpu_ptr(&rcu_data);
3225 my_rnp = my_rdp->mynode;
3226 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
3227 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
3228 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
3229 /* Leverage recent GPs and set GP for new callbacks. */
3230 needwake = rcu_advance_cbs(my_rnp, rdp) ||
3231 rcu_advance_cbs(my_rnp, my_rdp);
3232 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3233 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
3234 rcu_segcblist_disable(&rdp->cblist);
3235 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3236 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3237 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
3238 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
3239 __call_rcu_nocb_wake(my_rdp, true, flags);
3241 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
3242 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
3245 rcu_gp_kthread_wake();
3246 lockdep_assert_irqs_enabled();
3247 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3248 !rcu_segcblist_empty(&rdp->cblist),
3249 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3250 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3251 rcu_segcblist_first_cb(&rdp->cblist));
3256 * On non-huge systems, use expedited RCU grace periods to make suspend
3257 * and hibernation run faster.
3259 static int rcu_pm_notify(struct notifier_block *self,
3260 unsigned long action, void *hcpu)
3263 case PM_HIBERNATION_PREPARE:
3264 case PM_SUSPEND_PREPARE:
3267 case PM_POST_HIBERNATION:
3268 case PM_POST_SUSPEND:
3269 rcu_unexpedite_gp();
3278 * Spawn the kthreads that handle RCU's grace periods.
3280 static int __init rcu_spawn_gp_kthread(void)
3282 unsigned long flags;
3283 int kthread_prio_in = kthread_prio;
3284 struct rcu_node *rnp;
3285 struct sched_param sp;
3286 struct task_struct *t;
3288 /* Force priority into range. */
3289 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
3290 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
3292 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3294 else if (kthread_prio < 0)
3296 else if (kthread_prio > 99)
3299 if (kthread_prio != kthread_prio_in)
3300 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3301 kthread_prio, kthread_prio_in);
3303 rcu_scheduler_fully_active = 1;
3304 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
3305 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
3308 sp.sched_priority = kthread_prio;
3309 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3311 rnp = rcu_get_root();
3312 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3313 rcu_state.gp_kthread = t;
3314 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3316 rcu_spawn_nocb_kthreads();
3317 rcu_spawn_boost_kthreads();
3318 rcu_spawn_core_kthreads();
3321 early_initcall(rcu_spawn_gp_kthread);
3324 * This function is invoked towards the end of the scheduler's
3325 * initialization process. Before this is called, the idle task might
3326 * contain synchronous grace-period primitives (during which time, this idle
3327 * task is booting the system, and such primitives are no-ops). After this
3328 * function is called, any synchronous grace-period primitives are run as
3329 * expedited, with the requesting task driving the grace period forward.
3330 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3331 * runtime RCU functionality.
3333 void rcu_scheduler_starting(void)
3335 WARN_ON(num_online_cpus() != 1);
3336 WARN_ON(nr_context_switches() > 0);
3337 rcu_test_sync_prims();
3338 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3339 rcu_test_sync_prims();
3343 * Helper function for rcu_init() that initializes the rcu_state structure.
3345 static void __init rcu_init_one(void)
3347 static const char * const buf[] = RCU_NODE_NAME_INIT;
3348 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3349 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3350 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3352 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3356 struct rcu_node *rnp;
3358 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3360 /* Silence gcc 4.8 false positive about array index out of range. */
3361 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3362 panic("rcu_init_one: rcu_num_lvls out of range");
3364 /* Initialize the level-tracking arrays. */
3366 for (i = 1; i < rcu_num_lvls; i++)
3367 rcu_state.level[i] =
3368 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
3369 rcu_init_levelspread(levelspread, num_rcu_lvl);
3371 /* Initialize the elements themselves, starting from the leaves. */
3373 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3374 cpustride *= levelspread[i];
3375 rnp = rcu_state.level[i];
3376 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3377 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3378 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3379 &rcu_node_class[i], buf[i]);
3380 raw_spin_lock_init(&rnp->fqslock);
3381 lockdep_set_class_and_name(&rnp->fqslock,
3382 &rcu_fqs_class[i], fqs[i]);
3383 rnp->gp_seq = rcu_state.gp_seq;
3384 rnp->gp_seq_needed = rcu_state.gp_seq;
3385 rnp->completedqs = rcu_state.gp_seq;
3387 rnp->qsmaskinit = 0;
3388 rnp->grplo = j * cpustride;
3389 rnp->grphi = (j + 1) * cpustride - 1;
3390 if (rnp->grphi >= nr_cpu_ids)
3391 rnp->grphi = nr_cpu_ids - 1;
3397 rnp->grpnum = j % levelspread[i - 1];
3398 rnp->grpmask = BIT(rnp->grpnum);
3399 rnp->parent = rcu_state.level[i - 1] +
3400 j / levelspread[i - 1];
3403 INIT_LIST_HEAD(&rnp->blkd_tasks);
3404 rcu_init_one_nocb(rnp);
3405 init_waitqueue_head(&rnp->exp_wq[0]);
3406 init_waitqueue_head(&rnp->exp_wq[1]);
3407 init_waitqueue_head(&rnp->exp_wq[2]);
3408 init_waitqueue_head(&rnp->exp_wq[3]);
3409 spin_lock_init(&rnp->exp_lock);
3413 init_swait_queue_head(&rcu_state.gp_wq);
3414 init_swait_queue_head(&rcu_state.expedited_wq);
3415 rnp = rcu_first_leaf_node();
3416 for_each_possible_cpu(i) {
3417 while (i > rnp->grphi)
3419 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
3420 rcu_boot_init_percpu_data(i);
3425 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3426 * replace the definitions in tree.h because those are needed to size
3427 * the ->node array in the rcu_state structure.
3429 void rcu_init_geometry(void)
3433 static unsigned long old_nr_cpu_ids;
3434 int rcu_capacity[RCU_NUM_LVLS];
3435 static bool initialized;
3439 * Warn if setup_nr_cpu_ids() had not yet been invoked,
3440 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
3442 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
3446 old_nr_cpu_ids = nr_cpu_ids;
3450 * Initialize any unspecified boot parameters.
3451 * The default values of jiffies_till_first_fqs and
3452 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3453 * value, which is a function of HZ, then adding one for each
3454 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3456 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3457 if (jiffies_till_first_fqs == ULONG_MAX)
3458 jiffies_till_first_fqs = d;
3459 if (jiffies_till_next_fqs == ULONG_MAX)
3460 jiffies_till_next_fqs = d;
3461 adjust_jiffies_till_sched_qs();
3463 /* If the compile-time values are accurate, just leave. */
3464 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3465 nr_cpu_ids == NR_CPUS)
3467 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3468 rcu_fanout_leaf, nr_cpu_ids);
3471 * The boot-time rcu_fanout_leaf parameter must be at least two
3472 * and cannot exceed the number of bits in the rcu_node masks.
3473 * Complain and fall back to the compile-time values if this
3474 * limit is exceeded.
3476 if (rcu_fanout_leaf < 2 ||
3477 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3478 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3484 * Compute number of nodes that can be handled an rcu_node tree
3485 * with the given number of levels.
3487 rcu_capacity[0] = rcu_fanout_leaf;
3488 for (i = 1; i < RCU_NUM_LVLS; i++)
3489 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
3492 * The tree must be able to accommodate the configured number of CPUs.
3493 * If this limit is exceeded, fall back to the compile-time values.
3495 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
3496 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3501 /* Calculate the number of levels in the tree. */
3502 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
3504 rcu_num_lvls = i + 1;
3506 /* Calculate the number of rcu_nodes at each level of the tree. */
3507 for (i = 0; i < rcu_num_lvls; i++) {
3508 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
3509 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
3512 /* Calculate the total number of rcu_node structures. */
3514 for (i = 0; i < rcu_num_lvls; i++)
3515 rcu_num_nodes += num_rcu_lvl[i];
3519 * Dump out the structure of the rcu_node combining tree associated
3520 * with the rcu_state structure.
3522 static void __init rcu_dump_rcu_node_tree(void)
3525 struct rcu_node *rnp;
3527 pr_info("rcu_node tree layout dump\n");
3529 rcu_for_each_node_breadth_first(rnp) {
3530 if (rnp->level != level) {
3535 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
3540 struct workqueue_struct *rcu_gp_wq;
3541 struct workqueue_struct *rcu_par_gp_wq;
3543 void __init rcu_init(void)
3547 rcu_early_boot_tests();
3549 rcu_bootup_announce();
3550 rcu_init_geometry();
3553 rcu_dump_rcu_node_tree();
3555 open_softirq(RCU_SOFTIRQ, rcu_core_si);
3558 * We don't need protection against CPU-hotplug here because
3559 * this is called early in boot, before either interrupts
3560 * or the scheduler are operational.
3562 pm_notifier(rcu_pm_notify, 0);
3563 for_each_online_cpu(cpu) {
3564 rcutree_prepare_cpu(cpu);
3565 rcu_cpu_starting(cpu);
3566 rcutree_online_cpu(cpu);
3569 /* Create workqueue for expedited GPs and for Tree SRCU. */
3570 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
3571 WARN_ON(!rcu_gp_wq);
3572 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
3573 WARN_ON(!rcu_par_gp_wq);
3577 #include "tree_stall.h"
3578 #include "tree_exp.h"
3579 #include "tree_plugin.h"