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 void 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);
378 EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
381 * rcu_is_cpu_rrupt_from_idle - see if interrupted from idle
383 * If the current CPU is idle and running at a first-level (not nested)
384 * interrupt from idle, return true. The caller must have at least
385 * disabled preemption.
387 static int rcu_is_cpu_rrupt_from_idle(void)
389 /* Called only from within the scheduling-clock interrupt */
390 lockdep_assert_in_irq();
392 /* Check for counter underflows */
393 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
394 "RCU dynticks_nesting counter underflow!");
395 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
396 "RCU dynticks_nmi_nesting counter underflow/zero!");
398 /* Are we at first interrupt nesting level? */
399 if (__this_cpu_read(rcu_data.dynticks_nmi_nesting) != 1)
402 /* Does CPU appear to be idle from an RCU standpoint? */
403 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
406 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch ... */
407 #define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */
408 static long blimit = DEFAULT_RCU_BLIMIT;
409 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
410 static long qhimark = DEFAULT_RCU_QHIMARK;
411 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
412 static long qlowmark = DEFAULT_RCU_QLOMARK;
414 module_param(blimit, long, 0444);
415 module_param(qhimark, long, 0444);
416 module_param(qlowmark, long, 0444);
418 static ulong jiffies_till_first_fqs = ULONG_MAX;
419 static ulong jiffies_till_next_fqs = ULONG_MAX;
420 static bool rcu_kick_kthreads;
421 static int rcu_divisor = 7;
422 module_param(rcu_divisor, int, 0644);
424 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
425 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
426 module_param(rcu_resched_ns, long, 0644);
429 * How long the grace period must be before we start recruiting
430 * quiescent-state help from rcu_note_context_switch().
432 static ulong jiffies_till_sched_qs = ULONG_MAX;
433 module_param(jiffies_till_sched_qs, ulong, 0444);
434 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
435 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
438 * Make sure that we give the grace-period kthread time to detect any
439 * idle CPUs before taking active measures to force quiescent states.
440 * However, don't go below 100 milliseconds, adjusted upwards for really
443 static void adjust_jiffies_till_sched_qs(void)
447 /* If jiffies_till_sched_qs was specified, respect the request. */
448 if (jiffies_till_sched_qs != ULONG_MAX) {
449 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
452 /* Otherwise, set to third fqs scan, but bound below on large system. */
453 j = READ_ONCE(jiffies_till_first_fqs) +
454 2 * READ_ONCE(jiffies_till_next_fqs);
455 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
456 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
457 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
458 WRITE_ONCE(jiffies_to_sched_qs, j);
461 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
464 int ret = kstrtoul(val, 0, &j);
467 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
468 adjust_jiffies_till_sched_qs();
473 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
476 int ret = kstrtoul(val, 0, &j);
479 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
480 adjust_jiffies_till_sched_qs();
485 static struct kernel_param_ops first_fqs_jiffies_ops = {
486 .set = param_set_first_fqs_jiffies,
487 .get = param_get_ulong,
490 static struct kernel_param_ops next_fqs_jiffies_ops = {
491 .set = param_set_next_fqs_jiffies,
492 .get = param_get_ulong,
495 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
496 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
497 module_param(rcu_kick_kthreads, bool, 0644);
499 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
500 static int rcu_pending(int user);
503 * Return the number of RCU GPs completed thus far for debug & stats.
505 unsigned long rcu_get_gp_seq(void)
507 return READ_ONCE(rcu_state.gp_seq);
509 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
512 * Return the number of RCU expedited batches completed thus far for
513 * debug & stats. Odd numbers mean that a batch is in progress, even
514 * numbers mean idle. The value returned will thus be roughly double
515 * the cumulative batches since boot.
517 unsigned long rcu_exp_batches_completed(void)
519 return rcu_state.expedited_sequence;
521 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
524 * Return the root node of the rcu_state structure.
526 static struct rcu_node *rcu_get_root(void)
528 return &rcu_state.node[0];
532 * Convert a ->gp_state value to a character string.
534 static const char *gp_state_getname(short gs)
536 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
538 return gp_state_names[gs];
542 * Send along grace-period-related data for rcutorture diagnostics.
544 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
545 unsigned long *gp_seq)
549 *flags = READ_ONCE(rcu_state.gp_flags);
550 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
556 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
559 * Enter an RCU extended quiescent state, which can be either the
560 * idle loop or adaptive-tickless usermode execution.
562 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
563 * the possibility of usermode upcalls having messed up our count
564 * of interrupt nesting level during the prior busy period.
566 static void rcu_eqs_enter(bool user)
568 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
570 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
571 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
572 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
573 rdp->dynticks_nesting == 0);
574 if (rdp->dynticks_nesting != 1) {
575 rdp->dynticks_nesting--;
579 lockdep_assert_irqs_disabled();
580 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, rdp->dynticks);
581 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
582 rdp = this_cpu_ptr(&rcu_data);
583 do_nocb_deferred_wakeup(rdp);
584 rcu_prepare_for_idle();
585 rcu_preempt_deferred_qs(current);
586 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
587 rcu_dynticks_eqs_enter();
588 rcu_dynticks_task_enter();
592 * rcu_idle_enter - inform RCU that current CPU is entering idle
594 * Enter idle mode, in other words, -leave- the mode in which RCU
595 * read-side critical sections can occur. (Though RCU read-side
596 * critical sections can occur in irq handlers in idle, a possibility
597 * handled by irq_enter() and irq_exit().)
599 * If you add or remove a call to rcu_idle_enter(), be sure to test with
600 * CONFIG_RCU_EQS_DEBUG=y.
602 void rcu_idle_enter(void)
604 lockdep_assert_irqs_disabled();
605 rcu_eqs_enter(false);
608 #ifdef CONFIG_NO_HZ_FULL
610 * rcu_user_enter - inform RCU that we are resuming userspace.
612 * Enter RCU idle mode right before resuming userspace. No use of RCU
613 * is permitted between this call and rcu_user_exit(). This way the
614 * CPU doesn't need to maintain the tick for RCU maintenance purposes
615 * when the CPU runs in userspace.
617 * If you add or remove a call to rcu_user_enter(), be sure to test with
618 * CONFIG_RCU_EQS_DEBUG=y.
620 void rcu_user_enter(void)
622 lockdep_assert_irqs_disabled();
625 #endif /* CONFIG_NO_HZ_FULL */
628 * If we are returning from the outermost NMI handler that interrupted an
629 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
630 * to let the RCU grace-period handling know that the CPU is back to
633 * If you add or remove a call to rcu_nmi_exit_common(), be sure to test
634 * with CONFIG_RCU_EQS_DEBUG=y.
636 static __always_inline void rcu_nmi_exit_common(bool irq)
638 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
641 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
642 * (We are exiting an NMI handler, so RCU better be paying attention
645 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
646 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
649 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
650 * leave it in non-RCU-idle state.
652 if (rdp->dynticks_nmi_nesting != 1) {
653 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2, rdp->dynticks);
654 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
655 rdp->dynticks_nmi_nesting - 2);
659 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
660 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, rdp->dynticks);
661 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
664 rcu_prepare_for_idle();
666 rcu_dynticks_eqs_enter();
669 rcu_dynticks_task_enter();
673 * rcu_nmi_exit - inform RCU of exit from NMI context
675 * If you add or remove a call to rcu_nmi_exit(), be sure to test
676 * with CONFIG_RCU_EQS_DEBUG=y.
678 void rcu_nmi_exit(void)
680 rcu_nmi_exit_common(false);
684 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
686 * Exit from an interrupt handler, which might possibly result in entering
687 * idle mode, in other words, leaving the mode in which read-side critical
688 * sections can occur. The caller must have disabled interrupts.
690 * This code assumes that the idle loop never does anything that might
691 * result in unbalanced calls to irq_enter() and irq_exit(). If your
692 * architecture's idle loop violates this assumption, RCU will give you what
693 * you deserve, good and hard. But very infrequently and irreproducibly.
695 * Use things like work queues to work around this limitation.
697 * You have been warned.
699 * If you add or remove a call to rcu_irq_exit(), be sure to test with
700 * CONFIG_RCU_EQS_DEBUG=y.
702 void rcu_irq_exit(void)
704 lockdep_assert_irqs_disabled();
705 rcu_nmi_exit_common(true);
709 * Wrapper for rcu_irq_exit() where interrupts are enabled.
711 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
712 * with CONFIG_RCU_EQS_DEBUG=y.
714 void rcu_irq_exit_irqson(void)
718 local_irq_save(flags);
720 local_irq_restore(flags);
724 * Exit an RCU extended quiescent state, which can be either the
725 * idle loop or adaptive-tickless usermode execution.
727 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
728 * allow for the possibility of usermode upcalls messing up our count of
729 * interrupt nesting level during the busy period that is just now starting.
731 static void rcu_eqs_exit(bool user)
733 struct rcu_data *rdp;
736 lockdep_assert_irqs_disabled();
737 rdp = this_cpu_ptr(&rcu_data);
738 oldval = rdp->dynticks_nesting;
739 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
741 rdp->dynticks_nesting++;
744 rcu_dynticks_task_exit();
745 rcu_dynticks_eqs_exit();
746 rcu_cleanup_after_idle();
747 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, rdp->dynticks);
748 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
749 WRITE_ONCE(rdp->dynticks_nesting, 1);
750 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
751 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
755 * rcu_idle_exit - inform RCU that current CPU is leaving idle
757 * Exit idle mode, in other words, -enter- the mode in which RCU
758 * read-side critical sections can occur.
760 * If you add or remove a call to rcu_idle_exit(), be sure to test with
761 * CONFIG_RCU_EQS_DEBUG=y.
763 void rcu_idle_exit(void)
767 local_irq_save(flags);
769 local_irq_restore(flags);
772 #ifdef CONFIG_NO_HZ_FULL
774 * rcu_user_exit - inform RCU that we are exiting userspace.
776 * Exit RCU idle mode while entering the kernel because it can
777 * run a RCU read side critical section anytime.
779 * If you add or remove a call to rcu_user_exit(), be sure to test with
780 * CONFIG_RCU_EQS_DEBUG=y.
782 void rcu_user_exit(void)
786 #endif /* CONFIG_NO_HZ_FULL */
789 * rcu_nmi_enter_common - inform RCU of entry to NMI context
790 * @irq: Is this call from rcu_irq_enter?
792 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
793 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
794 * that the CPU is active. This implementation permits nested NMIs, as
795 * long as the nesting level does not overflow an int. (You will probably
796 * run out of stack space first.)
798 * If you add or remove a call to rcu_nmi_enter_common(), be sure to test
799 * with CONFIG_RCU_EQS_DEBUG=y.
801 static __always_inline void rcu_nmi_enter_common(bool irq)
803 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
806 /* Complain about underflow. */
807 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
810 * If idle from RCU viewpoint, atomically increment ->dynticks
811 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
812 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
813 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
814 * to be in the outermost NMI handler that interrupted an RCU-idle
815 * period (observation due to Andy Lutomirski).
817 if (rcu_dynticks_curr_cpu_in_eqs()) {
820 rcu_dynticks_task_exit();
822 rcu_dynticks_eqs_exit();
825 rcu_cleanup_after_idle();
828 } else if (tick_nohz_full_cpu(rdp->cpu) &&
829 rdp->dynticks_nmi_nesting == DYNTICK_IRQ_NONIDLE &&
830 READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) {
831 rdp->rcu_forced_tick = true;
832 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
834 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
835 rdp->dynticks_nmi_nesting,
836 rdp->dynticks_nmi_nesting + incby, rdp->dynticks);
837 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
838 rdp->dynticks_nmi_nesting + incby);
843 * rcu_nmi_enter - inform RCU of entry to NMI context
845 void rcu_nmi_enter(void)
847 rcu_nmi_enter_common(false);
849 NOKPROBE_SYMBOL(rcu_nmi_enter);
852 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
854 * Enter an interrupt handler, which might possibly result in exiting
855 * idle mode, in other words, entering the mode in which read-side critical
856 * sections can occur. The caller must have disabled interrupts.
858 * Note that the Linux kernel is fully capable of entering an interrupt
859 * handler that it never exits, for example when doing upcalls to user mode!
860 * This code assumes that the idle loop never does upcalls to user mode.
861 * If your architecture's idle loop does do upcalls to user mode (or does
862 * anything else that results in unbalanced calls to the irq_enter() and
863 * irq_exit() functions), RCU will give you what you deserve, good and hard.
864 * But very infrequently and irreproducibly.
866 * Use things like work queues to work around this limitation.
868 * You have been warned.
870 * If you add or remove a call to rcu_irq_enter(), be sure to test with
871 * CONFIG_RCU_EQS_DEBUG=y.
873 void rcu_irq_enter(void)
875 lockdep_assert_irqs_disabled();
876 rcu_nmi_enter_common(true);
880 * Wrapper for rcu_irq_enter() where interrupts are enabled.
882 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
883 * with CONFIG_RCU_EQS_DEBUG=y.
885 void rcu_irq_enter_irqson(void)
889 local_irq_save(flags);
891 local_irq_restore(flags);
895 * If any sort of urgency was applied to the current CPU (for example,
896 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
897 * to get to a quiescent state, disable it.
899 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
901 WRITE_ONCE(rdp->rcu_urgent_qs, false);
902 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
903 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
904 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
905 rdp->rcu_forced_tick = false;
910 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
912 * Return true if RCU is watching the running CPU, which means that this
913 * CPU can safely enter RCU read-side critical sections. In other words,
914 * if the current CPU is not in its idle loop or is in an interrupt or
915 * NMI handler, return true.
917 bool notrace rcu_is_watching(void)
921 preempt_disable_notrace();
922 ret = !rcu_dynticks_curr_cpu_in_eqs();
923 preempt_enable_notrace();
926 EXPORT_SYMBOL_GPL(rcu_is_watching);
929 * If a holdout task is actually running, request an urgent quiescent
930 * state from its CPU. This is unsynchronized, so migrations can cause
931 * the request to go to the wrong CPU. Which is OK, all that will happen
932 * is that the CPU's next context switch will be a bit slower and next
933 * time around this task will generate another request.
935 void rcu_request_urgent_qs_task(struct task_struct *t)
942 return; /* This task is not running on that CPU. */
943 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
946 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
949 * Is the current CPU online as far as RCU is concerned?
951 * Disable preemption to avoid false positives that could otherwise
952 * happen due to the current CPU number being sampled, this task being
953 * preempted, its old CPU being taken offline, resuming on some other CPU,
954 * then determining that its old CPU is now offline.
956 * Disable checking if in an NMI handler because we cannot safely
957 * report errors from NMI handlers anyway. In addition, it is OK to use
958 * RCU on an offline processor during initial boot, hence the check for
959 * rcu_scheduler_fully_active.
961 bool rcu_lockdep_current_cpu_online(void)
963 struct rcu_data *rdp;
964 struct rcu_node *rnp;
967 if (in_nmi() || !rcu_scheduler_fully_active)
970 rdp = this_cpu_ptr(&rcu_data);
972 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
977 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
979 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
982 * We are reporting a quiescent state on behalf of some other CPU, so
983 * it is our responsibility to check for and handle potential overflow
984 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
985 * After all, the CPU might be in deep idle state, and thus executing no
988 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
990 raw_lockdep_assert_held_rcu_node(rnp);
991 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
993 WRITE_ONCE(rdp->gpwrap, true);
994 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
995 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
999 * Snapshot the specified CPU's dynticks counter so that we can later
1000 * credit them with an implicit quiescent state. Return 1 if this CPU
1001 * is in dynticks idle mode, which is an extended quiescent state.
1003 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1005 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1006 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1007 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1008 rcu_gpnum_ovf(rdp->mynode, rdp);
1015 * Return true if the specified CPU has passed through a quiescent
1016 * state by virtue of being in or having passed through an dynticks
1017 * idle state since the last call to dyntick_save_progress_counter()
1018 * for this same CPU, or by virtue of having been offline.
1020 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1025 struct rcu_node *rnp = rdp->mynode;
1028 * If the CPU passed through or entered a dynticks idle phase with
1029 * no active irq/NMI handlers, then we can safely pretend that the CPU
1030 * already acknowledged the request to pass through a quiescent
1031 * state. Either way, that CPU cannot possibly be in an RCU
1032 * read-side critical section that started before the beginning
1033 * of the current RCU grace period.
1035 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1036 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1037 rcu_gpnum_ovf(rnp, rdp);
1041 /* If waiting too long on an offline CPU, complain. */
1042 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1043 time_after(jiffies, rcu_state.gp_start + HZ)) {
1045 struct rcu_node *rnp1;
1047 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1048 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1049 __func__, rnp->grplo, rnp->grphi, rnp->level,
1050 (long)rnp->gp_seq, (long)rnp->completedqs);
1051 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1052 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1053 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1054 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1055 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1056 __func__, rdp->cpu, ".o"[onl],
1057 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1058 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1059 return 1; /* Break things loose after complaining. */
1063 * A CPU running for an extended time within the kernel can
1064 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1065 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1066 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1067 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1068 * variable are safe because the assignments are repeated if this
1069 * CPU failed to pass through a quiescent state. This code
1070 * also checks .jiffies_resched in case jiffies_to_sched_qs
1073 jtsq = READ_ONCE(jiffies_to_sched_qs);
1074 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1075 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1076 if (!READ_ONCE(*rnhqp) &&
1077 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1078 time_after(jiffies, rcu_state.jiffies_resched))) {
1079 WRITE_ONCE(*rnhqp, true);
1080 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1081 smp_store_release(ruqp, true);
1082 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1083 WRITE_ONCE(*ruqp, true);
1087 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1088 * The above code handles this, but only for straight cond_resched().
1089 * And some in-kernel loops check need_resched() before calling
1090 * cond_resched(), which defeats the above code for CPUs that are
1091 * running in-kernel with scheduling-clock interrupts disabled.
1092 * So hit them over the head with the resched_cpu() hammer!
1094 if (tick_nohz_full_cpu(rdp->cpu) &&
1096 READ_ONCE(rdp->last_fqs_resched) + jtsq * 3)) {
1097 WRITE_ONCE(*ruqp, true);
1098 resched_cpu(rdp->cpu);
1099 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1103 * If more than halfway to RCU CPU stall-warning time, invoke
1104 * resched_cpu() more frequently to try to loosen things up a bit.
1105 * Also check to see if the CPU is getting hammered with interrupts,
1106 * but only once per grace period, just to keep the IPIs down to
1109 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1110 if (time_after(jiffies,
1111 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1112 resched_cpu(rdp->cpu);
1113 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1115 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1116 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1117 (rnp->ffmask & rdp->grpmask)) {
1118 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1119 rdp->rcu_iw_pending = true;
1120 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1121 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1128 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1129 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1130 unsigned long gp_seq_req, const char *s)
1132 trace_rcu_future_grace_period(rcu_state.name, rnp->gp_seq, gp_seq_req,
1133 rnp->level, rnp->grplo, rnp->grphi, s);
1137 * rcu_start_this_gp - Request the start of a particular grace period
1138 * @rnp_start: The leaf node of the CPU from which to start.
1139 * @rdp: The rcu_data corresponding to the CPU from which to start.
1140 * @gp_seq_req: The gp_seq of the grace period to start.
1142 * Start the specified grace period, as needed to handle newly arrived
1143 * callbacks. The required future grace periods are recorded in each
1144 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1145 * is reason to awaken the grace-period kthread.
1147 * The caller must hold the specified rcu_node structure's ->lock, which
1148 * is why the caller is responsible for waking the grace-period kthread.
1150 * Returns true if the GP thread needs to be awakened else false.
1152 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1153 unsigned long gp_seq_req)
1156 struct rcu_node *rnp;
1159 * Use funnel locking to either acquire the root rcu_node
1160 * structure's lock or bail out if the need for this grace period
1161 * has already been recorded -- or if that grace period has in
1162 * fact already started. If there is already a grace period in
1163 * progress in a non-leaf node, no recording is needed because the
1164 * end of the grace period will scan the leaf rcu_node structures.
1165 * Note that rnp_start->lock must not be released.
1167 raw_lockdep_assert_held_rcu_node(rnp_start);
1168 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1169 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1170 if (rnp != rnp_start)
1171 raw_spin_lock_rcu_node(rnp);
1172 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1173 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1174 (rnp != rnp_start &&
1175 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1176 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1180 rnp->gp_seq_needed = gp_seq_req;
1181 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1183 * We just marked the leaf or internal node, and a
1184 * grace period is in progress, which means that
1185 * rcu_gp_cleanup() will see the marking. Bail to
1186 * reduce contention.
1188 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1189 TPS("Startedleaf"));
1192 if (rnp != rnp_start && rnp->parent != NULL)
1193 raw_spin_unlock_rcu_node(rnp);
1195 break; /* At root, and perhaps also leaf. */
1198 /* If GP already in progress, just leave, otherwise start one. */
1199 if (rcu_gp_in_progress()) {
1200 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1203 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1204 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1205 rcu_state.gp_req_activity = jiffies;
1206 if (!rcu_state.gp_kthread) {
1207 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1210 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rcu_state.gp_seq), TPS("newreq"));
1211 ret = true; /* Caller must wake GP kthread. */
1213 /* Push furthest requested GP to leaf node and rcu_data structure. */
1214 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1215 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1216 rdp->gp_seq_needed = rnp->gp_seq_needed;
1218 if (rnp != rnp_start)
1219 raw_spin_unlock_rcu_node(rnp);
1224 * Clean up any old requests for the just-ended grace period. Also return
1225 * whether any additional grace periods have been requested.
1227 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1230 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1232 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1234 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1235 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1236 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1241 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1242 * an interrupt or softirq handler), and don't bother awakening when there
1243 * is nothing for the grace-period kthread to do (as in several CPUs raced
1244 * to awaken, and we lost), and finally don't try to awaken a kthread that
1245 * has not yet been created. If all those checks are passed, track some
1246 * debug information and awaken.
1248 * So why do the self-wakeup when in an interrupt or softirq handler
1249 * in the grace-period kthread's context? Because the kthread might have
1250 * been interrupted just as it was going to sleep, and just after the final
1251 * pre-sleep check of the awaken condition. In this case, a wakeup really
1252 * is required, and is therefore supplied.
1254 static void rcu_gp_kthread_wake(void)
1256 if ((current == rcu_state.gp_kthread &&
1257 !in_irq() && !in_serving_softirq()) ||
1258 !READ_ONCE(rcu_state.gp_flags) ||
1259 !rcu_state.gp_kthread)
1261 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1262 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1263 swake_up_one(&rcu_state.gp_wq);
1267 * If there is room, assign a ->gp_seq number to any callbacks on this
1268 * CPU that have not already been assigned. Also accelerate any callbacks
1269 * that were previously assigned a ->gp_seq number that has since proven
1270 * to be too conservative, which can happen if callbacks get assigned a
1271 * ->gp_seq number while RCU is idle, but with reference to a non-root
1272 * rcu_node structure. This function is idempotent, so it does not hurt
1273 * to call it repeatedly. Returns an flag saying that we should awaken
1274 * the RCU grace-period kthread.
1276 * The caller must hold rnp->lock with interrupts disabled.
1278 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1280 unsigned long gp_seq_req;
1283 rcu_lockdep_assert_cblist_protected(rdp);
1284 raw_lockdep_assert_held_rcu_node(rnp);
1286 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1287 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1291 * Callbacks are often registered with incomplete grace-period
1292 * information. Something about the fact that getting exact
1293 * information requires acquiring a global lock... RCU therefore
1294 * makes a conservative estimate of the grace period number at which
1295 * a given callback will become ready to invoke. The following
1296 * code checks this estimate and improves it when possible, thus
1297 * accelerating callback invocation to an earlier grace-period
1300 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1301 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1302 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1304 /* Trace depending on how much we were able to accelerate. */
1305 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1306 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1308 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1313 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1314 * rcu_node structure's ->lock be held. It consults the cached value
1315 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1316 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1317 * while holding the leaf rcu_node structure's ->lock.
1319 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1320 struct rcu_data *rdp)
1325 rcu_lockdep_assert_cblist_protected(rdp);
1326 c = rcu_seq_snap(&rcu_state.gp_seq);
1327 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1328 /* Old request still live, so mark recent callbacks. */
1329 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1332 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1333 needwake = rcu_accelerate_cbs(rnp, rdp);
1334 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1336 rcu_gp_kthread_wake();
1340 * Move any callbacks whose grace period has completed to the
1341 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1342 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1343 * sublist. This function is idempotent, so it does not hurt to
1344 * invoke it repeatedly. As long as it is not invoked -too- often...
1345 * Returns true if the RCU grace-period kthread needs to be awakened.
1347 * The caller must hold rnp->lock with interrupts disabled.
1349 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1351 rcu_lockdep_assert_cblist_protected(rdp);
1352 raw_lockdep_assert_held_rcu_node(rnp);
1354 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1355 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1359 * Find all callbacks whose ->gp_seq numbers indicate that they
1360 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1362 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1364 /* Classify any remaining callbacks. */
1365 return rcu_accelerate_cbs(rnp, rdp);
1369 * Move and classify callbacks, but only if doing so won't require
1370 * that the RCU grace-period kthread be awakened.
1372 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1373 struct rcu_data *rdp)
1375 rcu_lockdep_assert_cblist_protected(rdp);
1376 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1377 !raw_spin_trylock_rcu_node(rnp))
1379 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1380 raw_spin_unlock_rcu_node(rnp);
1384 * Update CPU-local rcu_data state to record the beginnings and ends of
1385 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1386 * structure corresponding to the current CPU, and must have irqs disabled.
1387 * Returns true if the grace-period kthread needs to be awakened.
1389 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1393 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1394 rcu_segcblist_is_offloaded(&rdp->cblist);
1396 raw_lockdep_assert_held_rcu_node(rnp);
1398 if (rdp->gp_seq == rnp->gp_seq)
1399 return false; /* Nothing to do. */
1401 /* Handle the ends of any preceding grace periods first. */
1402 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1403 unlikely(READ_ONCE(rdp->gpwrap))) {
1405 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1406 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1409 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1412 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1413 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1414 unlikely(READ_ONCE(rdp->gpwrap))) {
1416 * If the current grace period is waiting for this CPU,
1417 * set up to detect a quiescent state, otherwise don't
1418 * go looking for one.
1420 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1421 need_gp = !!(rnp->qsmask & rdp->grpmask);
1422 rdp->cpu_no_qs.b.norm = need_gp;
1423 rdp->core_needs_qs = need_gp;
1424 zero_cpu_stall_ticks(rdp);
1426 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1427 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1428 rdp->gp_seq_needed = rnp->gp_seq_needed;
1429 WRITE_ONCE(rdp->gpwrap, false);
1430 rcu_gpnum_ovf(rnp, rdp);
1434 static void note_gp_changes(struct rcu_data *rdp)
1436 unsigned long flags;
1438 struct rcu_node *rnp;
1440 local_irq_save(flags);
1442 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1443 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1444 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1445 local_irq_restore(flags);
1448 needwake = __note_gp_changes(rnp, rdp);
1449 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1451 rcu_gp_kthread_wake();
1454 static void rcu_gp_slow(int delay)
1457 !(rcu_seq_ctr(rcu_state.gp_seq) %
1458 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1459 schedule_timeout_uninterruptible(delay);
1463 * Initialize a new grace period. Return false if no grace period required.
1465 static bool rcu_gp_init(void)
1467 unsigned long flags;
1468 unsigned long oldmask;
1470 struct rcu_data *rdp;
1471 struct rcu_node *rnp = rcu_get_root();
1473 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1474 raw_spin_lock_irq_rcu_node(rnp);
1475 if (!READ_ONCE(rcu_state.gp_flags)) {
1476 /* Spurious wakeup, tell caller to go back to sleep. */
1477 raw_spin_unlock_irq_rcu_node(rnp);
1480 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1482 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1484 * Grace period already in progress, don't start another.
1485 * Not supposed to be able to happen.
1487 raw_spin_unlock_irq_rcu_node(rnp);
1491 /* Advance to a new grace period and initialize state. */
1492 record_gp_stall_check_time();
1493 /* Record GP times before starting GP, hence rcu_seq_start(). */
1494 rcu_seq_start(&rcu_state.gp_seq);
1495 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1496 raw_spin_unlock_irq_rcu_node(rnp);
1499 * Apply per-leaf buffered online and offline operations to the
1500 * rcu_node tree. Note that this new grace period need not wait
1501 * for subsequent online CPUs, and that quiescent-state forcing
1502 * will handle subsequent offline CPUs.
1504 rcu_state.gp_state = RCU_GP_ONOFF;
1505 rcu_for_each_leaf_node(rnp) {
1506 raw_spin_lock(&rcu_state.ofl_lock);
1507 raw_spin_lock_irq_rcu_node(rnp);
1508 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1509 !rnp->wait_blkd_tasks) {
1510 /* Nothing to do on this leaf rcu_node structure. */
1511 raw_spin_unlock_irq_rcu_node(rnp);
1512 raw_spin_unlock(&rcu_state.ofl_lock);
1516 /* Record old state, apply changes to ->qsmaskinit field. */
1517 oldmask = rnp->qsmaskinit;
1518 rnp->qsmaskinit = rnp->qsmaskinitnext;
1520 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1521 if (!oldmask != !rnp->qsmaskinit) {
1522 if (!oldmask) { /* First online CPU for rcu_node. */
1523 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1524 rcu_init_new_rnp(rnp);
1525 } else if (rcu_preempt_has_tasks(rnp)) {
1526 rnp->wait_blkd_tasks = true; /* blocked tasks */
1527 } else { /* Last offline CPU and can propagate. */
1528 rcu_cleanup_dead_rnp(rnp);
1533 * If all waited-on tasks from prior grace period are
1534 * done, and if all this rcu_node structure's CPUs are
1535 * still offline, propagate up the rcu_node tree and
1536 * clear ->wait_blkd_tasks. Otherwise, if one of this
1537 * rcu_node structure's CPUs has since come back online,
1538 * simply clear ->wait_blkd_tasks.
1540 if (rnp->wait_blkd_tasks &&
1541 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1542 rnp->wait_blkd_tasks = false;
1543 if (!rnp->qsmaskinit)
1544 rcu_cleanup_dead_rnp(rnp);
1547 raw_spin_unlock_irq_rcu_node(rnp);
1548 raw_spin_unlock(&rcu_state.ofl_lock);
1550 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1553 * Set the quiescent-state-needed bits in all the rcu_node
1554 * structures for all currently online CPUs in breadth-first
1555 * order, starting from the root rcu_node structure, relying on the
1556 * layout of the tree within the rcu_state.node[] array. Note that
1557 * other CPUs will access only the leaves of the hierarchy, thus
1558 * seeing that no grace period is in progress, at least until the
1559 * corresponding leaf node has been initialized.
1561 * The grace period cannot complete until the initialization
1562 * process finishes, because this kthread handles both.
1564 rcu_state.gp_state = RCU_GP_INIT;
1565 rcu_for_each_node_breadth_first(rnp) {
1566 rcu_gp_slow(gp_init_delay);
1567 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1568 rdp = this_cpu_ptr(&rcu_data);
1569 rcu_preempt_check_blocked_tasks(rnp);
1570 rnp->qsmask = rnp->qsmaskinit;
1571 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1572 if (rnp == rdp->mynode)
1573 (void)__note_gp_changes(rnp, rdp);
1574 rcu_preempt_boost_start_gp(rnp);
1575 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1576 rnp->level, rnp->grplo,
1577 rnp->grphi, rnp->qsmask);
1578 /* Quiescent states for tasks on any now-offline CPUs. */
1579 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1580 rnp->rcu_gp_init_mask = mask;
1581 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1582 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1584 raw_spin_unlock_irq_rcu_node(rnp);
1585 cond_resched_tasks_rcu_qs();
1586 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1593 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1596 static bool rcu_gp_fqs_check_wake(int *gfp)
1598 struct rcu_node *rnp = rcu_get_root();
1600 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1601 *gfp = READ_ONCE(rcu_state.gp_flags);
1602 if (*gfp & RCU_GP_FLAG_FQS)
1605 /* The current grace period has completed. */
1606 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1613 * Do one round of quiescent-state forcing.
1615 static void rcu_gp_fqs(bool first_time)
1617 struct rcu_node *rnp = rcu_get_root();
1619 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1620 rcu_state.n_force_qs++;
1622 /* Collect dyntick-idle snapshots. */
1623 force_qs_rnp(dyntick_save_progress_counter);
1625 /* Handle dyntick-idle and offline CPUs. */
1626 force_qs_rnp(rcu_implicit_dynticks_qs);
1628 /* Clear flag to prevent immediate re-entry. */
1629 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1630 raw_spin_lock_irq_rcu_node(rnp);
1631 WRITE_ONCE(rcu_state.gp_flags,
1632 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1633 raw_spin_unlock_irq_rcu_node(rnp);
1638 * Loop doing repeated quiescent-state forcing until the grace period ends.
1640 static void rcu_gp_fqs_loop(void)
1646 struct rcu_node *rnp = rcu_get_root();
1648 first_gp_fqs = true;
1649 j = READ_ONCE(jiffies_till_first_fqs);
1653 rcu_state.jiffies_force_qs = jiffies + j;
1654 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1655 jiffies + (j ? 3 * j : 2));
1657 trace_rcu_grace_period(rcu_state.name,
1658 READ_ONCE(rcu_state.gp_seq),
1660 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1661 ret = swait_event_idle_timeout_exclusive(
1662 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1663 rcu_state.gp_state = RCU_GP_DOING_FQS;
1664 /* Locking provides needed memory barriers. */
1665 /* If grace period done, leave loop. */
1666 if (!READ_ONCE(rnp->qsmask) &&
1667 !rcu_preempt_blocked_readers_cgp(rnp))
1669 /* If time for quiescent-state forcing, do it. */
1670 if (ULONG_CMP_GE(jiffies, rcu_state.jiffies_force_qs) ||
1671 (gf & RCU_GP_FLAG_FQS)) {
1672 trace_rcu_grace_period(rcu_state.name,
1673 READ_ONCE(rcu_state.gp_seq),
1675 rcu_gp_fqs(first_gp_fqs);
1676 first_gp_fqs = false;
1677 trace_rcu_grace_period(rcu_state.name,
1678 READ_ONCE(rcu_state.gp_seq),
1680 cond_resched_tasks_rcu_qs();
1681 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1682 ret = 0; /* Force full wait till next FQS. */
1683 j = READ_ONCE(jiffies_till_next_fqs);
1685 /* Deal with stray signal. */
1686 cond_resched_tasks_rcu_qs();
1687 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1688 WARN_ON(signal_pending(current));
1689 trace_rcu_grace_period(rcu_state.name,
1690 READ_ONCE(rcu_state.gp_seq),
1692 ret = 1; /* Keep old FQS timing. */
1694 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1697 j = rcu_state.jiffies_force_qs - j;
1703 * Clean up after the old grace period.
1705 static void rcu_gp_cleanup(void)
1707 unsigned long gp_duration;
1708 bool needgp = false;
1709 unsigned long new_gp_seq;
1711 struct rcu_data *rdp;
1712 struct rcu_node *rnp = rcu_get_root();
1713 struct swait_queue_head *sq;
1715 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1716 raw_spin_lock_irq_rcu_node(rnp);
1717 rcu_state.gp_end = jiffies;
1718 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1719 if (gp_duration > rcu_state.gp_max)
1720 rcu_state.gp_max = gp_duration;
1723 * We know the grace period is complete, but to everyone else
1724 * it appears to still be ongoing. But it is also the case
1725 * that to everyone else it looks like there is nothing that
1726 * they can do to advance the grace period. It is therefore
1727 * safe for us to drop the lock in order to mark the grace
1728 * period as completed in all of the rcu_node structures.
1730 raw_spin_unlock_irq_rcu_node(rnp);
1733 * Propagate new ->gp_seq value to rcu_node structures so that
1734 * other CPUs don't have to wait until the start of the next grace
1735 * period to process their callbacks. This also avoids some nasty
1736 * RCU grace-period initialization races by forcing the end of
1737 * the current grace period to be completely recorded in all of
1738 * the rcu_node structures before the beginning of the next grace
1739 * period is recorded in any of the rcu_node structures.
1741 new_gp_seq = rcu_state.gp_seq;
1742 rcu_seq_end(&new_gp_seq);
1743 rcu_for_each_node_breadth_first(rnp) {
1744 raw_spin_lock_irq_rcu_node(rnp);
1745 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1746 dump_blkd_tasks(rnp, 10);
1747 WARN_ON_ONCE(rnp->qsmask);
1748 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1749 rdp = this_cpu_ptr(&rcu_data);
1750 if (rnp == rdp->mynode)
1751 needgp = __note_gp_changes(rnp, rdp) || needgp;
1752 /* smp_mb() provided by prior unlock-lock pair. */
1753 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1754 sq = rcu_nocb_gp_get(rnp);
1755 raw_spin_unlock_irq_rcu_node(rnp);
1756 rcu_nocb_gp_cleanup(sq);
1757 cond_resched_tasks_rcu_qs();
1758 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1759 rcu_gp_slow(gp_cleanup_delay);
1761 rnp = rcu_get_root();
1762 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1764 /* Declare grace period done, trace first to use old GP number. */
1765 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1766 rcu_seq_end(&rcu_state.gp_seq);
1767 rcu_state.gp_state = RCU_GP_IDLE;
1768 /* Check for GP requests since above loop. */
1769 rdp = this_cpu_ptr(&rcu_data);
1770 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1771 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1772 TPS("CleanupMore"));
1775 /* Advance CBs to reduce false positives below. */
1776 offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1777 rcu_segcblist_is_offloaded(&rdp->cblist);
1778 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1779 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1780 rcu_state.gp_req_activity = jiffies;
1781 trace_rcu_grace_period(rcu_state.name,
1782 READ_ONCE(rcu_state.gp_seq),
1785 WRITE_ONCE(rcu_state.gp_flags,
1786 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1788 raw_spin_unlock_irq_rcu_node(rnp);
1792 * Body of kthread that handles grace periods.
1794 static int __noreturn rcu_gp_kthread(void *unused)
1796 rcu_bind_gp_kthread();
1799 /* Handle grace-period start. */
1801 trace_rcu_grace_period(rcu_state.name,
1802 READ_ONCE(rcu_state.gp_seq),
1804 rcu_state.gp_state = RCU_GP_WAIT_GPS;
1805 swait_event_idle_exclusive(rcu_state.gp_wq,
1806 READ_ONCE(rcu_state.gp_flags) &
1808 rcu_state.gp_state = RCU_GP_DONE_GPS;
1809 /* Locking provides needed memory barrier. */
1812 cond_resched_tasks_rcu_qs();
1813 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1814 WARN_ON(signal_pending(current));
1815 trace_rcu_grace_period(rcu_state.name,
1816 READ_ONCE(rcu_state.gp_seq),
1820 /* Handle quiescent-state forcing. */
1823 /* Handle grace-period end. */
1824 rcu_state.gp_state = RCU_GP_CLEANUP;
1826 rcu_state.gp_state = RCU_GP_CLEANED;
1831 * Report a full set of quiescent states to the rcu_state data structure.
1832 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1833 * another grace period is required. Whether we wake the grace-period
1834 * kthread or it awakens itself for the next round of quiescent-state
1835 * forcing, that kthread will clean up after the just-completed grace
1836 * period. Note that the caller must hold rnp->lock, which is released
1839 static void rcu_report_qs_rsp(unsigned long flags)
1840 __releases(rcu_get_root()->lock)
1842 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1843 WARN_ON_ONCE(!rcu_gp_in_progress());
1844 WRITE_ONCE(rcu_state.gp_flags,
1845 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1846 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1847 rcu_gp_kthread_wake();
1851 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1852 * Allows quiescent states for a group of CPUs to be reported at one go
1853 * to the specified rcu_node structure, though all the CPUs in the group
1854 * must be represented by the same rcu_node structure (which need not be a
1855 * leaf rcu_node structure, though it often will be). The gps parameter
1856 * is the grace-period snapshot, which means that the quiescent states
1857 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1858 * must be held upon entry, and it is released before return.
1860 * As a special case, if mask is zero, the bit-already-cleared check is
1861 * disabled. This allows propagating quiescent state due to resumed tasks
1862 * during grace-period initialization.
1864 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1865 unsigned long gps, unsigned long flags)
1866 __releases(rnp->lock)
1868 unsigned long oldmask = 0;
1869 struct rcu_node *rnp_c;
1871 raw_lockdep_assert_held_rcu_node(rnp);
1873 /* Walk up the rcu_node hierarchy. */
1875 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1878 * Our bit has already been cleared, or the
1879 * relevant grace period is already over, so done.
1881 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1884 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1885 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1886 rcu_preempt_blocked_readers_cgp(rnp));
1887 rnp->qsmask &= ~mask;
1888 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1889 mask, rnp->qsmask, rnp->level,
1890 rnp->grplo, rnp->grphi,
1892 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1894 /* Other bits still set at this level, so done. */
1895 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1898 rnp->completedqs = rnp->gp_seq;
1899 mask = rnp->grpmask;
1900 if (rnp->parent == NULL) {
1902 /* No more levels. Exit loop holding root lock. */
1906 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1909 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1910 oldmask = rnp_c->qsmask;
1914 * Get here if we are the last CPU to pass through a quiescent
1915 * state for this grace period. Invoke rcu_report_qs_rsp()
1916 * to clean up and start the next grace period if one is needed.
1918 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1922 * Record a quiescent state for all tasks that were previously queued
1923 * on the specified rcu_node structure and that were blocking the current
1924 * RCU grace period. The caller must hold the corresponding rnp->lock with
1925 * irqs disabled, and this lock is released upon return, but irqs remain
1928 static void __maybe_unused
1929 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1930 __releases(rnp->lock)
1934 struct rcu_node *rnp_p;
1936 raw_lockdep_assert_held_rcu_node(rnp);
1937 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPTION)) ||
1938 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1940 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1941 return; /* Still need more quiescent states! */
1944 rnp->completedqs = rnp->gp_seq;
1945 rnp_p = rnp->parent;
1946 if (rnp_p == NULL) {
1948 * Only one rcu_node structure in the tree, so don't
1949 * try to report up to its nonexistent parent!
1951 rcu_report_qs_rsp(flags);
1955 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1957 mask = rnp->grpmask;
1958 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1959 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1960 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1964 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1965 * structure. This must be called from the specified CPU.
1968 rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
1970 unsigned long flags;
1972 bool needwake = false;
1973 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1974 rcu_segcblist_is_offloaded(&rdp->cblist);
1975 struct rcu_node *rnp;
1978 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1979 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
1983 * The grace period in which this quiescent state was
1984 * recorded has ended, so don't report it upwards.
1985 * We will instead need a new quiescent state that lies
1986 * within the current grace period.
1988 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
1989 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1992 mask = rdp->grpmask;
1993 if ((rnp->qsmask & mask) == 0) {
1994 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1997 * This GP can't end until cpu checks in, so all of our
1998 * callbacks can be processed during the next GP.
2001 needwake = rcu_accelerate_cbs(rnp, rdp);
2003 rcu_disable_urgency_upon_qs(rdp);
2004 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2005 /* ^^^ Released rnp->lock */
2007 rcu_gp_kthread_wake();
2012 * Check to see if there is a new grace period of which this CPU
2013 * is not yet aware, and if so, set up local rcu_data state for it.
2014 * Otherwise, see if this CPU has just passed through its first
2015 * quiescent state for this grace period, and record that fact if so.
2018 rcu_check_quiescent_state(struct rcu_data *rdp)
2020 /* Check for grace-period ends and beginnings. */
2021 note_gp_changes(rdp);
2024 * Does this CPU still need to do its part for current grace period?
2025 * If no, return and let the other CPUs do their part as well.
2027 if (!rdp->core_needs_qs)
2031 * Was there a quiescent state since the beginning of the grace
2032 * period? If no, then exit and wait for the next call.
2034 if (rdp->cpu_no_qs.b.norm)
2038 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2041 rcu_report_qs_rdp(rdp->cpu, rdp);
2045 * Near the end of the offline process. Trace the fact that this CPU
2048 int rcutree_dying_cpu(unsigned int cpu)
2051 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2052 struct rcu_node *rnp = rdp->mynode;
2054 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2057 blkd = !!(rnp->qsmask & rdp->grpmask);
2058 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq,
2059 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2064 * All CPUs for the specified rcu_node structure have gone offline,
2065 * and all tasks that were preempted within an RCU read-side critical
2066 * section while running on one of those CPUs have since exited their RCU
2067 * read-side critical section. Some other CPU is reporting this fact with
2068 * the specified rcu_node structure's ->lock held and interrupts disabled.
2069 * This function therefore goes up the tree of rcu_node structures,
2070 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2071 * the leaf rcu_node structure's ->qsmaskinit field has already been
2074 * This function does check that the specified rcu_node structure has
2075 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2076 * prematurely. That said, invoking it after the fact will cost you
2077 * a needless lock acquisition. So once it has done its work, don't
2080 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2083 struct rcu_node *rnp = rnp_leaf;
2085 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2086 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2087 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2088 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2091 mask = rnp->grpmask;
2095 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2096 rnp->qsmaskinit &= ~mask;
2097 /* Between grace periods, so better already be zero! */
2098 WARN_ON_ONCE(rnp->qsmask);
2099 if (rnp->qsmaskinit) {
2100 raw_spin_unlock_rcu_node(rnp);
2101 /* irqs remain disabled. */
2104 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2109 * The CPU has been completely removed, and some other CPU is reporting
2110 * this fact from process context. Do the remainder of the cleanup.
2111 * There can only be one CPU hotplug operation at a time, so no need for
2114 int rcutree_dead_cpu(unsigned int cpu)
2116 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2117 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2119 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2122 /* Adjust any no-longer-needed kthreads. */
2123 rcu_boost_kthread_setaffinity(rnp, -1);
2124 /* Do any needed no-CB deferred wakeups from this CPU. */
2125 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2127 // Stop-machine done, so allow nohz_full to disable tick.
2128 tick_dep_clear(TICK_DEP_BIT_RCU);
2133 * Invoke any RCU callbacks that have made it to the end of their grace
2134 * period. Thottle as specified by rdp->blimit.
2136 static void rcu_do_batch(struct rcu_data *rdp)
2138 unsigned long flags;
2139 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2140 rcu_segcblist_is_offloaded(&rdp->cblist);
2141 struct rcu_head *rhp;
2142 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2144 long pending, tlimit = 0;
2146 /* If no callbacks are ready, just return. */
2147 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2148 trace_rcu_batch_start(rcu_state.name,
2149 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2150 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2151 trace_rcu_batch_end(rcu_state.name, 0,
2152 !rcu_segcblist_empty(&rdp->cblist),
2153 need_resched(), is_idle_task(current),
2154 rcu_is_callbacks_kthread());
2159 * Extract the list of ready callbacks, disabling to prevent
2160 * races with call_rcu() from interrupt handlers. Leave the
2161 * callback counts, as rcu_barrier() needs to be conservative.
2163 local_irq_save(flags);
2165 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2166 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2167 bl = max(rdp->blimit, pending >> rcu_divisor);
2168 if (unlikely(bl > 100))
2169 tlimit = local_clock() + rcu_resched_ns;
2170 trace_rcu_batch_start(rcu_state.name,
2171 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2172 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2173 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2175 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2176 rcu_nocb_unlock_irqrestore(rdp, flags);
2178 /* Invoke callbacks. */
2179 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2180 rhp = rcu_cblist_dequeue(&rcl);
2181 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2182 debug_rcu_head_unqueue(rhp);
2183 if (__rcu_reclaim(rcu_state.name, rhp))
2184 rcu_cblist_dequeued_lazy(&rcl);
2186 * Stop only if limit reached and CPU has something to do.
2187 * Note: The rcl structure counts down from zero.
2189 if (-rcl.len >= bl && !offloaded &&
2191 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2193 if (unlikely(tlimit)) {
2194 /* only call local_clock() every 32 callbacks */
2195 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2197 /* Exceeded the time limit, so leave. */
2201 WARN_ON_ONCE(in_serving_softirq());
2203 lockdep_assert_irqs_enabled();
2204 cond_resched_tasks_rcu_qs();
2205 lockdep_assert_irqs_enabled();
2210 local_irq_save(flags);
2213 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2214 is_idle_task(current), rcu_is_callbacks_kthread());
2216 /* Update counts and requeue any remaining callbacks. */
2217 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2218 smp_mb(); /* List handling before counting for rcu_barrier(). */
2219 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2221 /* Reinstate batch limit if we have worked down the excess. */
2222 count = rcu_segcblist_n_cbs(&rdp->cblist);
2223 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2224 rdp->blimit = blimit;
2226 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2227 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2228 rdp->qlen_last_fqs_check = 0;
2229 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2230 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2231 rdp->qlen_last_fqs_check = count;
2234 * The following usually indicates a double call_rcu(). To track
2235 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2237 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2238 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2239 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2241 rcu_nocb_unlock_irqrestore(rdp, flags);
2243 /* Re-invoke RCU core processing if there are callbacks remaining. */
2244 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2246 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2250 * This function is invoked from each scheduling-clock interrupt,
2251 * and checks to see if this CPU is in a non-context-switch quiescent
2252 * state, for example, user mode or idle loop. It also schedules RCU
2253 * core processing. If the current grace period has gone on too long,
2254 * it will ask the scheduler to manufacture a context switch for the sole
2255 * purpose of providing a providing the needed quiescent state.
2257 void rcu_sched_clock_irq(int user)
2259 trace_rcu_utilization(TPS("Start scheduler-tick"));
2260 raw_cpu_inc(rcu_data.ticks_this_gp);
2261 /* The load-acquire pairs with the store-release setting to true. */
2262 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2263 /* Idle and userspace execution already are quiescent states. */
2264 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2265 set_tsk_need_resched(current);
2266 set_preempt_need_resched();
2268 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2270 rcu_flavor_sched_clock_irq(user);
2271 if (rcu_pending(user))
2274 trace_rcu_utilization(TPS("End scheduler-tick"));
2278 * Scan the leaf rcu_node structures. For each structure on which all
2279 * CPUs have reported a quiescent state and on which there are tasks
2280 * blocking the current grace period, initiate RCU priority boosting.
2281 * Otherwise, invoke the specified function to check dyntick state for
2282 * each CPU that has not yet reported a quiescent state.
2284 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2287 unsigned long flags;
2289 struct rcu_data *rdp;
2290 struct rcu_node *rnp;
2292 rcu_for_each_leaf_node(rnp) {
2293 cond_resched_tasks_rcu_qs();
2295 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2296 if (rnp->qsmask == 0) {
2297 if (!IS_ENABLED(CONFIG_PREEMPTION) ||
2298 rcu_preempt_blocked_readers_cgp(rnp)) {
2300 * No point in scanning bits because they
2301 * are all zero. But we might need to
2302 * priority-boost blocked readers.
2304 rcu_initiate_boost(rnp, flags);
2305 /* rcu_initiate_boost() releases rnp->lock */
2308 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2311 for_each_leaf_node_possible_cpu(rnp, cpu) {
2312 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2313 if ((rnp->qsmask & bit) != 0) {
2314 rdp = per_cpu_ptr(&rcu_data, cpu);
2317 rcu_disable_urgency_upon_qs(rdp);
2322 /* Idle/offline CPUs, report (releases rnp->lock). */
2323 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2325 /* Nothing to do here, so just drop the lock. */
2326 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2332 * Force quiescent states on reluctant CPUs, and also detect which
2333 * CPUs are in dyntick-idle mode.
2335 void rcu_force_quiescent_state(void)
2337 unsigned long flags;
2339 struct rcu_node *rnp;
2340 struct rcu_node *rnp_old = NULL;
2342 /* Funnel through hierarchy to reduce memory contention. */
2343 rnp = __this_cpu_read(rcu_data.mynode);
2344 for (; rnp != NULL; rnp = rnp->parent) {
2345 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2346 !raw_spin_trylock(&rnp->fqslock);
2347 if (rnp_old != NULL)
2348 raw_spin_unlock(&rnp_old->fqslock);
2353 /* rnp_old == rcu_get_root(), rnp == NULL. */
2355 /* Reached the root of the rcu_node tree, acquire lock. */
2356 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2357 raw_spin_unlock(&rnp_old->fqslock);
2358 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2359 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2360 return; /* Someone beat us to it. */
2362 WRITE_ONCE(rcu_state.gp_flags,
2363 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2364 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2365 rcu_gp_kthread_wake();
2367 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2369 /* Perform RCU core processing work for the current CPU. */
2370 static __latent_entropy void rcu_core(void)
2372 unsigned long flags;
2373 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2374 struct rcu_node *rnp = rdp->mynode;
2375 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2376 rcu_segcblist_is_offloaded(&rdp->cblist);
2378 if (cpu_is_offline(smp_processor_id()))
2380 trace_rcu_utilization(TPS("Start RCU core"));
2381 WARN_ON_ONCE(!rdp->beenonline);
2383 /* Report any deferred quiescent states if preemption enabled. */
2384 if (!(preempt_count() & PREEMPT_MASK)) {
2385 rcu_preempt_deferred_qs(current);
2386 } else if (rcu_preempt_need_deferred_qs(current)) {
2387 set_tsk_need_resched(current);
2388 set_preempt_need_resched();
2391 /* Update RCU state based on any recent quiescent states. */
2392 rcu_check_quiescent_state(rdp);
2394 /* No grace period and unregistered callbacks? */
2395 if (!rcu_gp_in_progress() &&
2396 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2397 local_irq_save(flags);
2398 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2399 rcu_accelerate_cbs_unlocked(rnp, rdp);
2400 local_irq_restore(flags);
2403 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2405 /* If there are callbacks ready, invoke them. */
2406 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2407 likely(READ_ONCE(rcu_scheduler_fully_active)))
2410 /* Do any needed deferred wakeups of rcuo kthreads. */
2411 do_nocb_deferred_wakeup(rdp);
2412 trace_rcu_utilization(TPS("End RCU core"));
2415 static void rcu_core_si(struct softirq_action *h)
2420 static void rcu_wake_cond(struct task_struct *t, int status)
2423 * If the thread is yielding, only wake it when this
2424 * is invoked from idle
2426 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2430 static void invoke_rcu_core_kthread(void)
2432 struct task_struct *t;
2433 unsigned long flags;
2435 local_irq_save(flags);
2436 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2437 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2438 if (t != NULL && t != current)
2439 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2440 local_irq_restore(flags);
2444 * Wake up this CPU's rcuc kthread to do RCU core processing.
2446 static void invoke_rcu_core(void)
2448 if (!cpu_online(smp_processor_id()))
2451 raise_softirq(RCU_SOFTIRQ);
2453 invoke_rcu_core_kthread();
2456 static void rcu_cpu_kthread_park(unsigned int cpu)
2458 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2461 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2463 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2467 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2468 * the RCU softirq used in configurations of RCU that do not support RCU
2469 * priority boosting.
2471 static void rcu_cpu_kthread(unsigned int cpu)
2473 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2474 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2477 for (spincnt = 0; spincnt < 10; spincnt++) {
2478 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
2480 *statusp = RCU_KTHREAD_RUNNING;
2481 local_irq_disable();
2489 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2490 *statusp = RCU_KTHREAD_WAITING;
2494 *statusp = RCU_KTHREAD_YIELDING;
2495 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2496 schedule_timeout_interruptible(2);
2497 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2498 *statusp = RCU_KTHREAD_WAITING;
2501 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2502 .store = &rcu_data.rcu_cpu_kthread_task,
2503 .thread_should_run = rcu_cpu_kthread_should_run,
2504 .thread_fn = rcu_cpu_kthread,
2505 .thread_comm = "rcuc/%u",
2506 .setup = rcu_cpu_kthread_setup,
2507 .park = rcu_cpu_kthread_park,
2511 * Spawn per-CPU RCU core processing kthreads.
2513 static int __init rcu_spawn_core_kthreads(void)
2517 for_each_possible_cpu(cpu)
2518 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2519 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2521 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2522 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2525 early_initcall(rcu_spawn_core_kthreads);
2528 * Handle any core-RCU processing required by a call_rcu() invocation.
2530 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2531 unsigned long flags)
2534 * If called from an extended quiescent state, invoke the RCU
2535 * core in order to force a re-evaluation of RCU's idleness.
2537 if (!rcu_is_watching())
2540 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2541 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2545 * Force the grace period if too many callbacks or too long waiting.
2546 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2547 * if some other CPU has recently done so. Also, don't bother
2548 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2549 * is the only one waiting for a grace period to complete.
2551 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2552 rdp->qlen_last_fqs_check + qhimark)) {
2554 /* Are we ignoring a completed grace period? */
2555 note_gp_changes(rdp);
2557 /* Start a new grace period if one not already started. */
2558 if (!rcu_gp_in_progress()) {
2559 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2561 /* Give the grace period a kick. */
2562 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2563 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2564 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2565 rcu_force_quiescent_state();
2566 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2567 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2573 * RCU callback function to leak a callback.
2575 static void rcu_leak_callback(struct rcu_head *rhp)
2580 * Helper function for call_rcu() and friends. The cpu argument will
2581 * normally be -1, indicating "currently running CPU". It may specify
2582 * a CPU only if that CPU is a no-CBs CPU. Currently, only rcu_barrier()
2583 * is expected to specify a CPU.
2586 __call_rcu(struct rcu_head *head, rcu_callback_t func, bool lazy)
2588 unsigned long flags;
2589 struct rcu_data *rdp;
2592 /* Misaligned rcu_head! */
2593 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2595 if (debug_rcu_head_queue(head)) {
2597 * Probable double call_rcu(), so leak the callback.
2598 * Use rcu:rcu_callback trace event to find the previous
2599 * time callback was passed to __call_rcu().
2601 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2603 WRITE_ONCE(head->func, rcu_leak_callback);
2608 local_irq_save(flags);
2609 rdp = this_cpu_ptr(&rcu_data);
2611 /* Add the callback to our list. */
2612 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2613 // This can trigger due to call_rcu() from offline CPU:
2614 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2615 WARN_ON_ONCE(!rcu_is_watching());
2616 // Very early boot, before rcu_init(). Initialize if needed
2617 // and then drop through to queue the callback.
2618 if (rcu_segcblist_empty(&rdp->cblist))
2619 rcu_segcblist_init(&rdp->cblist);
2621 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2622 return; // Enqueued onto ->nocb_bypass, so just leave.
2623 /* If we get here, rcu_nocb_try_bypass() acquired ->nocb_lock. */
2624 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2625 if (__is_kfree_rcu_offset((unsigned long)func))
2626 trace_rcu_kfree_callback(rcu_state.name, head,
2627 (unsigned long)func,
2628 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2629 rcu_segcblist_n_cbs(&rdp->cblist));
2631 trace_rcu_callback(rcu_state.name, head,
2632 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2633 rcu_segcblist_n_cbs(&rdp->cblist));
2635 /* Go handle any RCU core processing required. */
2636 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2637 unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2638 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2640 __call_rcu_core(rdp, head, flags);
2641 local_irq_restore(flags);
2646 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2647 * @head: structure to be used for queueing the RCU updates.
2648 * @func: actual callback function to be invoked after the grace period
2650 * The callback function will be invoked some time after a full grace
2651 * period elapses, in other words after all pre-existing RCU read-side
2652 * critical sections have completed. However, the callback function
2653 * might well execute concurrently with RCU read-side critical sections
2654 * that started after call_rcu() was invoked. RCU read-side critical
2655 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2656 * may be nested. In addition, regions of code across which interrupts,
2657 * preemption, or softirqs have been disabled also serve as RCU read-side
2658 * critical sections. This includes hardware interrupt handlers, softirq
2659 * handlers, and NMI handlers.
2661 * Note that all CPUs must agree that the grace period extended beyond
2662 * all pre-existing RCU read-side critical section. On systems with more
2663 * than one CPU, this means that when "func()" is invoked, each CPU is
2664 * guaranteed to have executed a full memory barrier since the end of its
2665 * last RCU read-side critical section whose beginning preceded the call
2666 * to call_rcu(). It also means that each CPU executing an RCU read-side
2667 * critical section that continues beyond the start of "func()" must have
2668 * executed a memory barrier after the call_rcu() but before the beginning
2669 * of that RCU read-side critical section. Note that these guarantees
2670 * include CPUs that are offline, idle, or executing in user mode, as
2671 * well as CPUs that are executing in the kernel.
2673 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2674 * resulting RCU callback function "func()", then both CPU A and CPU B are
2675 * guaranteed to execute a full memory barrier during the time interval
2676 * between the call to call_rcu() and the invocation of "func()" -- even
2677 * if CPU A and CPU B are the same CPU (but again only if the system has
2678 * more than one CPU).
2680 void call_rcu(struct rcu_head *head, rcu_callback_t func)
2682 __call_rcu(head, func, 0);
2684 EXPORT_SYMBOL_GPL(call_rcu);
2687 /* Maximum number of jiffies to wait before draining a batch. */
2688 #define KFREE_DRAIN_JIFFIES (HZ / 50)
2689 #define KFREE_N_BATCHES 2
2692 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2693 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2694 * @head_free: List of kfree_rcu() objects waiting for a grace period
2695 * @krcp: Pointer to @kfree_rcu_cpu structure
2698 struct kfree_rcu_cpu_work {
2699 struct rcu_work rcu_work;
2700 struct rcu_head *head_free;
2701 struct kfree_rcu_cpu *krcp;
2705 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2706 * @head: List of kfree_rcu() objects not yet waiting for a grace period
2707 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2708 * @lock: Synchronize access to this structure
2709 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
2710 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
2711 * @initialized: The @lock and @rcu_work fields have been initialized
2713 * This is a per-CPU structure. The reason that it is not included in
2714 * the rcu_data structure is to permit this code to be extracted from
2715 * the RCU files. Such extraction could allow further optimization of
2716 * the interactions with the slab allocators.
2718 struct kfree_rcu_cpu {
2719 struct rcu_head *head;
2720 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
2722 struct delayed_work monitor_work;
2727 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc);
2730 * This function is invoked in workqueue context after a grace period.
2731 * It frees all the objects queued on ->head_free.
2733 static void kfree_rcu_work(struct work_struct *work)
2735 unsigned long flags;
2736 struct rcu_head *head, *next;
2737 struct kfree_rcu_cpu *krcp;
2738 struct kfree_rcu_cpu_work *krwp;
2740 krwp = container_of(to_rcu_work(work),
2741 struct kfree_rcu_cpu_work, rcu_work);
2743 spin_lock_irqsave(&krcp->lock, flags);
2744 head = krwp->head_free;
2745 krwp->head_free = NULL;
2746 spin_unlock_irqrestore(&krcp->lock, flags);
2748 // List "head" is now private, so traverse locklessly.
2749 for (; head; head = next) {
2751 // Potentially optimize with kfree_bulk in future.
2752 __rcu_reclaim(rcu_state.name, head);
2753 cond_resched_tasks_rcu_qs();
2758 * Schedule the kfree batch RCU work to run in workqueue context after a GP.
2760 * This function is invoked by kfree_rcu_monitor() when the KFREE_DRAIN_JIFFIES
2761 * timeout has been reached.
2763 static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
2766 struct kfree_rcu_cpu_work *krwp = NULL;
2768 lockdep_assert_held(&krcp->lock);
2769 for (i = 0; i < KFREE_N_BATCHES; i++)
2770 if (!krcp->krw_arr[i].head_free) {
2771 krwp = &(krcp->krw_arr[i]);
2775 // If a previous RCU batch is in progress, we cannot immediately
2776 // queue another one, so return false to tell caller to retry.
2780 krwp->head_free = krcp->head;
2782 INIT_RCU_WORK(&krwp->rcu_work, kfree_rcu_work);
2783 queue_rcu_work(system_wq, &krwp->rcu_work);
2787 static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp,
2788 unsigned long flags)
2790 // Attempt to start a new batch.
2791 krcp->monitor_todo = false;
2792 if (queue_kfree_rcu_work(krcp)) {
2793 // Success! Our job is done here.
2794 spin_unlock_irqrestore(&krcp->lock, flags);
2798 // Previous RCU batch still in progress, try again later.
2799 krcp->monitor_todo = true;
2800 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
2801 spin_unlock_irqrestore(&krcp->lock, flags);
2805 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
2806 * It invokes kfree_rcu_drain_unlock() to attempt to start another batch.
2808 static void kfree_rcu_monitor(struct work_struct *work)
2810 unsigned long flags;
2811 struct kfree_rcu_cpu *krcp = container_of(work, struct kfree_rcu_cpu,
2814 spin_lock_irqsave(&krcp->lock, flags);
2815 if (krcp->monitor_todo)
2816 kfree_rcu_drain_unlock(krcp, flags);
2818 spin_unlock_irqrestore(&krcp->lock, flags);
2822 * This version of kfree_call_rcu does not do batching of kfree_rcu() requests.
2823 * Used only by rcuperf torture test for comparison with kfree_rcu_batch().
2825 void kfree_call_rcu_nobatch(struct rcu_head *head, rcu_callback_t func)
2827 __call_rcu(head, func, 1);
2829 EXPORT_SYMBOL_GPL(kfree_call_rcu_nobatch);
2832 * Queue a request for lazy invocation of kfree() after a grace period.
2834 * Each kfree_call_rcu() request is added to a batch. The batch will be drained
2835 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch
2836 * will be kfree'd in workqueue context. This allows us to:
2838 * 1. Batch requests together to reduce the number of grace periods during
2839 * heavy kfree_rcu() load.
2841 * 2. It makes it possible to use kfree_bulk() on a large number of
2842 * kfree_rcu() requests thus reducing cache misses and the per-object
2843 * overhead of kfree().
2845 void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
2847 unsigned long flags;
2848 struct kfree_rcu_cpu *krcp;
2852 local_irq_save(flags); // For safely calling this_cpu_ptr().
2853 krcp = this_cpu_ptr(&krc);
2854 if (krcp->initialized)
2855 spin_lock(&krcp->lock);
2857 // Queue the object but don't yet schedule the batch.
2859 head->next = krcp->head;
2862 // Set timer to drain after KFREE_DRAIN_JIFFIES.
2863 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
2864 !krcp->monitor_todo) {
2865 krcp->monitor_todo = true;
2866 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
2869 if (krcp->initialized)
2870 spin_unlock(&krcp->lock);
2871 local_irq_restore(flags);
2873 EXPORT_SYMBOL_GPL(kfree_call_rcu);
2875 void __init kfree_rcu_scheduler_running(void)
2878 unsigned long flags;
2880 for_each_online_cpu(cpu) {
2881 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
2883 spin_lock_irqsave(&krcp->lock, flags);
2884 if (!krcp->head || krcp->monitor_todo) {
2885 spin_unlock_irqrestore(&krcp->lock, flags);
2888 krcp->monitor_todo = true;
2889 schedule_delayed_work_on(cpu, &krcp->monitor_work,
2890 KFREE_DRAIN_JIFFIES);
2891 spin_unlock_irqrestore(&krcp->lock, flags);
2896 * During early boot, any blocking grace-period wait automatically
2897 * implies a grace period. Later on, this is never the case for PREEMPT.
2899 * Howevr, because a context switch is a grace period for !PREEMPT, any
2900 * blocking grace-period wait automatically implies a grace period if
2901 * there is only one CPU online at any point time during execution of
2902 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
2903 * occasionally incorrectly indicate that there are multiple CPUs online
2904 * when there was in fact only one the whole time, as this just adds some
2905 * overhead: RCU still operates correctly.
2907 static int rcu_blocking_is_gp(void)
2911 if (IS_ENABLED(CONFIG_PREEMPTION))
2912 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
2913 might_sleep(); /* Check for RCU read-side critical section. */
2915 ret = num_online_cpus() <= 1;
2921 * synchronize_rcu - wait until a grace period has elapsed.
2923 * Control will return to the caller some time after a full grace
2924 * period has elapsed, in other words after all currently executing RCU
2925 * read-side critical sections have completed. Note, however, that
2926 * upon return from synchronize_rcu(), the caller might well be executing
2927 * concurrently with new RCU read-side critical sections that began while
2928 * synchronize_rcu() was waiting. RCU read-side critical sections are
2929 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
2930 * In addition, regions of code across which interrupts, preemption, or
2931 * softirqs have been disabled also serve as RCU read-side critical
2932 * sections. This includes hardware interrupt handlers, softirq handlers,
2935 * Note that this guarantee implies further memory-ordering guarantees.
2936 * On systems with more than one CPU, when synchronize_rcu() returns,
2937 * each CPU is guaranteed to have executed a full memory barrier since
2938 * the end of its last RCU read-side critical section whose beginning
2939 * preceded the call to synchronize_rcu(). In addition, each CPU having
2940 * an RCU read-side critical section that extends beyond the return from
2941 * synchronize_rcu() is guaranteed to have executed a full memory barrier
2942 * after the beginning of synchronize_rcu() and before the beginning of
2943 * that RCU read-side critical section. Note that these guarantees include
2944 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2945 * that are executing in the kernel.
2947 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
2948 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2949 * to have executed a full memory barrier during the execution of
2950 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
2951 * again only if the system has more than one CPU).
2953 void synchronize_rcu(void)
2955 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
2956 lock_is_held(&rcu_lock_map) ||
2957 lock_is_held(&rcu_sched_lock_map),
2958 "Illegal synchronize_rcu() in RCU read-side critical section");
2959 if (rcu_blocking_is_gp())
2961 if (rcu_gp_is_expedited())
2962 synchronize_rcu_expedited();
2964 wait_rcu_gp(call_rcu);
2966 EXPORT_SYMBOL_GPL(synchronize_rcu);
2969 * get_state_synchronize_rcu - Snapshot current RCU state
2971 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2972 * to determine whether or not a full grace period has elapsed in the
2975 unsigned long get_state_synchronize_rcu(void)
2978 * Any prior manipulation of RCU-protected data must happen
2979 * before the load from ->gp_seq.
2982 return rcu_seq_snap(&rcu_state.gp_seq);
2984 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2987 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2989 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2991 * If a full RCU grace period has elapsed since the earlier call to
2992 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2993 * synchronize_rcu() to wait for a full grace period.
2995 * Yes, this function does not take counter wrap into account. But
2996 * counter wrap is harmless. If the counter wraps, we have waited for
2997 * more than 2 billion grace periods (and way more on a 64-bit system!),
2998 * so waiting for one additional grace period should be just fine.
3000 void cond_synchronize_rcu(unsigned long oldstate)
3002 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
3005 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3007 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3010 * Check to see if there is any immediate RCU-related work to be done by
3011 * the current CPU, returning 1 if so and zero otherwise. The checks are
3012 * in order of increasing expense: checks that can be carried out against
3013 * CPU-local state are performed first. However, we must check for CPU
3014 * stalls first, else we might not get a chance.
3016 static int rcu_pending(int user)
3018 bool gp_in_progress;
3019 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3020 struct rcu_node *rnp = rdp->mynode;
3022 /* Check for CPU stalls, if enabled. */
3023 check_cpu_stall(rdp);
3025 /* Does this CPU need a deferred NOCB wakeup? */
3026 if (rcu_nocb_need_deferred_wakeup(rdp))
3029 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3030 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3033 /* Is the RCU core waiting for a quiescent state from this CPU? */
3034 gp_in_progress = rcu_gp_in_progress();
3035 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3038 /* Does this CPU have callbacks ready to invoke? */
3039 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3042 /* Has RCU gone idle with this CPU needing another grace period? */
3043 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3044 (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
3045 !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
3046 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3049 /* Have RCU grace period completed or started? */
3050 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3051 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3059 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3060 * the compiler is expected to optimize this away.
3062 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3064 trace_rcu_barrier(rcu_state.name, s, cpu,
3065 atomic_read(&rcu_state.barrier_cpu_count), done);
3069 * RCU callback function for rcu_barrier(). If we are last, wake
3070 * up the task executing rcu_barrier().
3072 static void rcu_barrier_callback(struct rcu_head *rhp)
3074 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3075 rcu_barrier_trace(TPS("LastCB"), -1,
3076 rcu_state.barrier_sequence);
3077 complete(&rcu_state.barrier_completion);
3079 rcu_barrier_trace(TPS("CB"), -1, rcu_state.barrier_sequence);
3084 * Called with preemption disabled, and from cross-cpu IRQ context.
3086 static void rcu_barrier_func(void *unused)
3088 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
3090 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3091 rdp->barrier_head.func = rcu_barrier_callback;
3092 debug_rcu_head_queue(&rdp->barrier_head);
3094 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
3095 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3096 atomic_inc(&rcu_state.barrier_cpu_count);
3098 debug_rcu_head_unqueue(&rdp->barrier_head);
3099 rcu_barrier_trace(TPS("IRQNQ"), -1,
3100 rcu_state.barrier_sequence);
3102 rcu_nocb_unlock(rdp);
3106 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3108 * Note that this primitive does not necessarily wait for an RCU grace period
3109 * to complete. For example, if there are no RCU callbacks queued anywhere
3110 * in the system, then rcu_barrier() is within its rights to return
3111 * immediately, without waiting for anything, much less an RCU grace period.
3113 void rcu_barrier(void)
3116 struct rcu_data *rdp;
3117 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3119 rcu_barrier_trace(TPS("Begin"), -1, s);
3121 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3122 mutex_lock(&rcu_state.barrier_mutex);
3124 /* Did someone else do our work for us? */
3125 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
3126 rcu_barrier_trace(TPS("EarlyExit"), -1,
3127 rcu_state.barrier_sequence);
3128 smp_mb(); /* caller's subsequent code after above check. */
3129 mutex_unlock(&rcu_state.barrier_mutex);
3133 /* Mark the start of the barrier operation. */
3134 rcu_seq_start(&rcu_state.barrier_sequence);
3135 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
3138 * Initialize the count to one rather than to zero in order to
3139 * avoid a too-soon return to zero in case of a short grace period
3140 * (or preemption of this task). Exclude CPU-hotplug operations
3141 * to ensure that no offline CPU has callbacks queued.
3143 init_completion(&rcu_state.barrier_completion);
3144 atomic_set(&rcu_state.barrier_cpu_count, 1);
3148 * Force each CPU with callbacks to register a new callback.
3149 * When that callback is invoked, we will know that all of the
3150 * corresponding CPU's preceding callbacks have been invoked.
3152 for_each_possible_cpu(cpu) {
3153 rdp = per_cpu_ptr(&rcu_data, cpu);
3154 if (!cpu_online(cpu) &&
3155 !rcu_segcblist_is_offloaded(&rdp->cblist))
3157 if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3158 rcu_barrier_trace(TPS("OnlineQ"), cpu,
3159 rcu_state.barrier_sequence);
3160 smp_call_function_single(cpu, rcu_barrier_func, NULL, 1);
3162 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
3163 rcu_state.barrier_sequence);
3169 * Now that we have an rcu_barrier_callback() callback on each
3170 * CPU, and thus each counted, remove the initial count.
3172 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count))
3173 complete(&rcu_state.barrier_completion);
3175 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3176 wait_for_completion(&rcu_state.barrier_completion);
3178 /* Mark the end of the barrier operation. */
3179 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
3180 rcu_seq_end(&rcu_state.barrier_sequence);
3182 /* Other rcu_barrier() invocations can now safely proceed. */
3183 mutex_unlock(&rcu_state.barrier_mutex);
3185 EXPORT_SYMBOL_GPL(rcu_barrier);
3188 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3189 * first CPU in a given leaf rcu_node structure coming online. The caller
3190 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3193 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3197 struct rcu_node *rnp = rnp_leaf;
3199 raw_lockdep_assert_held_rcu_node(rnp_leaf);
3200 WARN_ON_ONCE(rnp->wait_blkd_tasks);
3202 mask = rnp->grpmask;
3206 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3207 oldmask = rnp->qsmaskinit;
3208 rnp->qsmaskinit |= mask;
3209 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3216 * Do boot-time initialization of a CPU's per-CPU RCU data.
3219 rcu_boot_init_percpu_data(int cpu)
3221 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3223 /* Set up local state, ensuring consistent view of global state. */
3224 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3225 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
3226 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
3227 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3228 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3229 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3230 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3232 rcu_boot_init_nocb_percpu_data(rdp);
3236 * Invoked early in the CPU-online process, when pretty much all services
3237 * are available. The incoming CPU is not present.
3239 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3240 * offline event can be happening at a given time. Note also that we can
3241 * accept some slop in the rsp->gp_seq access due to the fact that this
3242 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3243 * And any offloaded callbacks are being numbered elsewhere.
3245 int rcutree_prepare_cpu(unsigned int cpu)
3247 unsigned long flags;
3248 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3249 struct rcu_node *rnp = rcu_get_root();
3251 /* Set up local state, ensuring consistent view of global state. */
3252 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3253 rdp->qlen_last_fqs_check = 0;
3254 rdp->n_force_qs_snap = rcu_state.n_force_qs;
3255 rdp->blimit = blimit;
3256 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3257 !rcu_segcblist_is_offloaded(&rdp->cblist))
3258 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3259 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
3260 rcu_dynticks_eqs_online();
3261 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3264 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3265 * propagation up the rcu_node tree will happen at the beginning
3266 * of the next grace period.
3269 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3270 rdp->beenonline = true; /* We have now been online. */
3271 rdp->gp_seq = rnp->gp_seq;
3272 rdp->gp_seq_needed = rnp->gp_seq;
3273 rdp->cpu_no_qs.b.norm = true;
3274 rdp->core_needs_qs = false;
3275 rdp->rcu_iw_pending = false;
3276 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3277 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3278 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3279 rcu_prepare_kthreads(cpu);
3280 rcu_spawn_cpu_nocb_kthread(cpu);
3286 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3288 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3290 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3292 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3296 * Near the end of the CPU-online process. Pretty much all services
3297 * enabled, and the CPU is now very much alive.
3299 int rcutree_online_cpu(unsigned int cpu)
3301 unsigned long flags;
3302 struct rcu_data *rdp;
3303 struct rcu_node *rnp;
3305 rdp = per_cpu_ptr(&rcu_data, cpu);
3307 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3308 rnp->ffmask |= rdp->grpmask;
3309 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3310 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3311 return 0; /* Too early in boot for scheduler work. */
3312 sync_sched_exp_online_cleanup(cpu);
3313 rcutree_affinity_setting(cpu, -1);
3315 // Stop-machine done, so allow nohz_full to disable tick.
3316 tick_dep_clear(TICK_DEP_BIT_RCU);
3321 * Near the beginning of the process. The CPU is still very much alive
3322 * with pretty much all services enabled.
3324 int rcutree_offline_cpu(unsigned int cpu)
3326 unsigned long flags;
3327 struct rcu_data *rdp;
3328 struct rcu_node *rnp;
3330 rdp = per_cpu_ptr(&rcu_data, cpu);
3332 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3333 rnp->ffmask &= ~rdp->grpmask;
3334 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3336 rcutree_affinity_setting(cpu, cpu);
3338 // nohz_full CPUs need the tick for stop-machine to work quickly
3339 tick_dep_set(TICK_DEP_BIT_RCU);
3343 static DEFINE_PER_CPU(int, rcu_cpu_started);
3346 * Mark the specified CPU as being online so that subsequent grace periods
3347 * (both expedited and normal) will wait on it. Note that this means that
3348 * incoming CPUs are not allowed to use RCU read-side critical sections
3349 * until this function is called. Failing to observe this restriction
3350 * will result in lockdep splats.
3352 * Note that this function is special in that it is invoked directly
3353 * from the incoming CPU rather than from the cpuhp_step mechanism.
3354 * This is because this function must be invoked at a precise location.
3356 void rcu_cpu_starting(unsigned int cpu)
3358 unsigned long flags;
3361 unsigned long oldmask;
3362 struct rcu_data *rdp;
3363 struct rcu_node *rnp;
3365 if (per_cpu(rcu_cpu_started, cpu))
3368 per_cpu(rcu_cpu_started, cpu) = 1;
3370 rdp = per_cpu_ptr(&rcu_data, cpu);
3372 mask = rdp->grpmask;
3373 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3374 rnp->qsmaskinitnext |= mask;
3375 oldmask = rnp->expmaskinitnext;
3376 rnp->expmaskinitnext |= mask;
3377 oldmask ^= rnp->expmaskinitnext;
3378 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3379 /* Allow lockless access for expedited grace periods. */
3380 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */
3381 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3382 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3383 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3384 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3385 rcu_disable_urgency_upon_qs(rdp);
3386 /* Report QS -after- changing ->qsmaskinitnext! */
3387 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3389 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3391 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3394 #ifdef CONFIG_HOTPLUG_CPU
3396 * The outgoing function has no further need of RCU, so remove it from
3397 * the rcu_node tree's ->qsmaskinitnext bit masks.
3399 * Note that this function is special in that it is invoked directly
3400 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3401 * This is because this function must be invoked at a precise location.
3403 void rcu_report_dead(unsigned int cpu)
3405 unsigned long flags;
3407 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3408 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3410 /* QS for any half-done expedited grace period. */
3412 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
3414 rcu_preempt_deferred_qs(current);
3416 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3417 mask = rdp->grpmask;
3418 raw_spin_lock(&rcu_state.ofl_lock);
3419 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3420 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3421 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3422 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3423 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3424 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3425 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3427 rnp->qsmaskinitnext &= ~mask;
3428 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3429 raw_spin_unlock(&rcu_state.ofl_lock);
3431 per_cpu(rcu_cpu_started, cpu) = 0;
3435 * The outgoing CPU has just passed through the dying-idle state, and we
3436 * are being invoked from the CPU that was IPIed to continue the offline
3437 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
3439 void rcutree_migrate_callbacks(int cpu)
3441 unsigned long flags;
3442 struct rcu_data *my_rdp;
3443 struct rcu_node *my_rnp;
3444 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3447 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
3448 rcu_segcblist_empty(&rdp->cblist))
3449 return; /* No callbacks to migrate. */
3451 local_irq_save(flags);
3452 my_rdp = this_cpu_ptr(&rcu_data);
3453 my_rnp = my_rdp->mynode;
3454 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
3455 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
3456 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
3457 /* Leverage recent GPs and set GP for new callbacks. */
3458 needwake = rcu_advance_cbs(my_rnp, rdp) ||
3459 rcu_advance_cbs(my_rnp, my_rdp);
3460 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3461 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
3462 rcu_segcblist_disable(&rdp->cblist);
3463 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3464 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3465 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
3466 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
3467 __call_rcu_nocb_wake(my_rdp, true, flags);
3469 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
3470 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
3473 rcu_gp_kthread_wake();
3474 lockdep_assert_irqs_enabled();
3475 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3476 !rcu_segcblist_empty(&rdp->cblist),
3477 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3478 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3479 rcu_segcblist_first_cb(&rdp->cblist));
3484 * On non-huge systems, use expedited RCU grace periods to make suspend
3485 * and hibernation run faster.
3487 static int rcu_pm_notify(struct notifier_block *self,
3488 unsigned long action, void *hcpu)
3491 case PM_HIBERNATION_PREPARE:
3492 case PM_SUSPEND_PREPARE:
3495 case PM_POST_HIBERNATION:
3496 case PM_POST_SUSPEND:
3497 rcu_unexpedite_gp();
3506 * Spawn the kthreads that handle RCU's grace periods.
3508 static int __init rcu_spawn_gp_kthread(void)
3510 unsigned long flags;
3511 int kthread_prio_in = kthread_prio;
3512 struct rcu_node *rnp;
3513 struct sched_param sp;
3514 struct task_struct *t;
3516 /* Force priority into range. */
3517 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
3518 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
3520 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3522 else if (kthread_prio < 0)
3524 else if (kthread_prio > 99)
3527 if (kthread_prio != kthread_prio_in)
3528 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3529 kthread_prio, kthread_prio_in);
3531 rcu_scheduler_fully_active = 1;
3532 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
3533 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
3536 sp.sched_priority = kthread_prio;
3537 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3539 rnp = rcu_get_root();
3540 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3541 rcu_state.gp_kthread = t;
3542 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3544 rcu_spawn_nocb_kthreads();
3545 rcu_spawn_boost_kthreads();
3548 early_initcall(rcu_spawn_gp_kthread);
3551 * This function is invoked towards the end of the scheduler's
3552 * initialization process. Before this is called, the idle task might
3553 * contain synchronous grace-period primitives (during which time, this idle
3554 * task is booting the system, and such primitives are no-ops). After this
3555 * function is called, any synchronous grace-period primitives are run as
3556 * expedited, with the requesting task driving the grace period forward.
3557 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3558 * runtime RCU functionality.
3560 void rcu_scheduler_starting(void)
3562 WARN_ON(num_online_cpus() != 1);
3563 WARN_ON(nr_context_switches() > 0);
3564 rcu_test_sync_prims();
3565 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3566 rcu_test_sync_prims();
3570 * Helper function for rcu_init() that initializes the rcu_state structure.
3572 static void __init rcu_init_one(void)
3574 static const char * const buf[] = RCU_NODE_NAME_INIT;
3575 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3576 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3577 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3579 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3583 struct rcu_node *rnp;
3585 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3587 /* Silence gcc 4.8 false positive about array index out of range. */
3588 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3589 panic("rcu_init_one: rcu_num_lvls out of range");
3591 /* Initialize the level-tracking arrays. */
3593 for (i = 1; i < rcu_num_lvls; i++)
3594 rcu_state.level[i] =
3595 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
3596 rcu_init_levelspread(levelspread, num_rcu_lvl);
3598 /* Initialize the elements themselves, starting from the leaves. */
3600 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3601 cpustride *= levelspread[i];
3602 rnp = rcu_state.level[i];
3603 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3604 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3605 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3606 &rcu_node_class[i], buf[i]);
3607 raw_spin_lock_init(&rnp->fqslock);
3608 lockdep_set_class_and_name(&rnp->fqslock,
3609 &rcu_fqs_class[i], fqs[i]);
3610 rnp->gp_seq = rcu_state.gp_seq;
3611 rnp->gp_seq_needed = rcu_state.gp_seq;
3612 rnp->completedqs = rcu_state.gp_seq;
3614 rnp->qsmaskinit = 0;
3615 rnp->grplo = j * cpustride;
3616 rnp->grphi = (j + 1) * cpustride - 1;
3617 if (rnp->grphi >= nr_cpu_ids)
3618 rnp->grphi = nr_cpu_ids - 1;
3624 rnp->grpnum = j % levelspread[i - 1];
3625 rnp->grpmask = BIT(rnp->grpnum);
3626 rnp->parent = rcu_state.level[i - 1] +
3627 j / levelspread[i - 1];
3630 INIT_LIST_HEAD(&rnp->blkd_tasks);
3631 rcu_init_one_nocb(rnp);
3632 init_waitqueue_head(&rnp->exp_wq[0]);
3633 init_waitqueue_head(&rnp->exp_wq[1]);
3634 init_waitqueue_head(&rnp->exp_wq[2]);
3635 init_waitqueue_head(&rnp->exp_wq[3]);
3636 spin_lock_init(&rnp->exp_lock);
3640 init_swait_queue_head(&rcu_state.gp_wq);
3641 init_swait_queue_head(&rcu_state.expedited_wq);
3642 rnp = rcu_first_leaf_node();
3643 for_each_possible_cpu(i) {
3644 while (i > rnp->grphi)
3646 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
3647 rcu_boot_init_percpu_data(i);
3652 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3653 * replace the definitions in tree.h because those are needed to size
3654 * the ->node array in the rcu_state structure.
3656 static void __init rcu_init_geometry(void)
3660 int rcu_capacity[RCU_NUM_LVLS];
3663 * Initialize any unspecified boot parameters.
3664 * The default values of jiffies_till_first_fqs and
3665 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3666 * value, which is a function of HZ, then adding one for each
3667 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3669 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3670 if (jiffies_till_first_fqs == ULONG_MAX)
3671 jiffies_till_first_fqs = d;
3672 if (jiffies_till_next_fqs == ULONG_MAX)
3673 jiffies_till_next_fqs = d;
3674 adjust_jiffies_till_sched_qs();
3676 /* If the compile-time values are accurate, just leave. */
3677 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3678 nr_cpu_ids == NR_CPUS)
3680 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3681 rcu_fanout_leaf, nr_cpu_ids);
3684 * The boot-time rcu_fanout_leaf parameter must be at least two
3685 * and cannot exceed the number of bits in the rcu_node masks.
3686 * Complain and fall back to the compile-time values if this
3687 * limit is exceeded.
3689 if (rcu_fanout_leaf < 2 ||
3690 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3691 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3697 * Compute number of nodes that can be handled an rcu_node tree
3698 * with the given number of levels.
3700 rcu_capacity[0] = rcu_fanout_leaf;
3701 for (i = 1; i < RCU_NUM_LVLS; i++)
3702 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
3705 * The tree must be able to accommodate the configured number of CPUs.
3706 * If this limit is exceeded, fall back to the compile-time values.
3708 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
3709 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3714 /* Calculate the number of levels in the tree. */
3715 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
3717 rcu_num_lvls = i + 1;
3719 /* Calculate the number of rcu_nodes at each level of the tree. */
3720 for (i = 0; i < rcu_num_lvls; i++) {
3721 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
3722 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
3725 /* Calculate the total number of rcu_node structures. */
3727 for (i = 0; i < rcu_num_lvls; i++)
3728 rcu_num_nodes += num_rcu_lvl[i];
3732 * Dump out the structure of the rcu_node combining tree associated
3733 * with the rcu_state structure.
3735 static void __init rcu_dump_rcu_node_tree(void)
3738 struct rcu_node *rnp;
3740 pr_info("rcu_node tree layout dump\n");
3742 rcu_for_each_node_breadth_first(rnp) {
3743 if (rnp->level != level) {
3748 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
3753 struct workqueue_struct *rcu_gp_wq;
3754 struct workqueue_struct *rcu_par_gp_wq;
3756 static void __init kfree_rcu_batch_init(void)
3761 for_each_possible_cpu(cpu) {
3762 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3764 spin_lock_init(&krcp->lock);
3765 for (i = 0; i < KFREE_N_BATCHES; i++)
3766 krcp->krw_arr[i].krcp = krcp;
3767 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
3768 krcp->initialized = true;
3772 void __init rcu_init(void)
3776 rcu_early_boot_tests();
3778 kfree_rcu_batch_init();
3779 rcu_bootup_announce();
3780 rcu_init_geometry();
3783 rcu_dump_rcu_node_tree();
3785 open_softirq(RCU_SOFTIRQ, rcu_core_si);
3788 * We don't need protection against CPU-hotplug here because
3789 * this is called early in boot, before either interrupts
3790 * or the scheduler are operational.
3792 pm_notifier(rcu_pm_notify, 0);
3793 for_each_online_cpu(cpu) {
3794 rcutree_prepare_cpu(cpu);
3795 rcu_cpu_starting(cpu);
3796 rcutree_online_cpu(cpu);
3799 /* Create workqueue for expedited GPs and for Tree SRCU. */
3800 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
3801 WARN_ON(!rcu_gp_wq);
3802 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
3803 WARN_ON(!rcu_par_gp_wq);
3807 #include "tree_stall.h"
3808 #include "tree_exp.h"
3809 #include "tree_plugin.h"