1 // SPDX-License-Identifier: GPL-2.0+
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
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>
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/panic.h>
36 #include <linux/panic_notifier.h>
37 #include <linux/percpu.h>
38 #include <linux/notifier.h>
39 #include <linux/cpu.h>
40 #include <linux/mutex.h>
41 #include <linux/time.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/wait.h>
44 #include <linux/kthread.h>
45 #include <uapi/linux/sched/types.h>
46 #include <linux/prefetch.h>
47 #include <linux/delay.h>
48 #include <linux/random.h>
49 #include <linux/trace_events.h>
50 #include <linux/suspend.h>
51 #include <linux/ftrace.h>
52 #include <linux/tick.h>
53 #include <linux/sysrq.h>
54 #include <linux/kprobes.h>
55 #include <linux/gfp.h>
56 #include <linux/oom.h>
57 #include <linux/smpboot.h>
58 #include <linux/jiffies.h>
59 #include <linux/slab.h>
60 #include <linux/sched/isolation.h>
61 #include <linux/sched/clock.h>
62 #include <linux/vmalloc.h>
64 #include <linux/kasan.h>
65 #include "../time/tick-internal.h"
70 #ifdef MODULE_PARAM_PREFIX
71 #undef MODULE_PARAM_PREFIX
73 #define MODULE_PARAM_PREFIX "rcutree."
75 /* Data structures. */
77 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
78 .dynticks_nesting = 1,
79 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
80 .dynticks = ATOMIC_INIT(1),
81 #ifdef CONFIG_RCU_NOCB_CPU
82 .cblist.flags = SEGCBLIST_SOFTIRQ_ONLY,
85 static struct rcu_state rcu_state = {
86 .level = { &rcu_state.node[0] },
87 .gp_state = RCU_GP_IDLE,
88 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
89 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
92 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
93 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
94 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
97 /* Dump rcu_node combining tree at boot to verify correct setup. */
98 static bool dump_tree;
99 module_param(dump_tree, bool, 0444);
100 /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
101 static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
102 #ifndef CONFIG_PREEMPT_RT
103 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);
153 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
154 static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
156 /* rcuc/rcub kthread realtime priority */
157 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
158 module_param(kthread_prio, int, 0444);
160 /* Delay in jiffies for grace-period initialization delays, debug only. */
162 static int gp_preinit_delay;
163 module_param(gp_preinit_delay, int, 0444);
164 static int gp_init_delay;
165 module_param(gp_init_delay, int, 0444);
166 static int gp_cleanup_delay;
167 module_param(gp_cleanup_delay, int, 0444);
169 // Add delay to rcu_read_unlock() for strict grace periods.
170 static int rcu_unlock_delay;
171 #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
172 module_param(rcu_unlock_delay, int, 0444);
176 * This rcu parameter is runtime-read-only. It reflects
177 * a minimum allowed number of objects which can be cached
178 * per-CPU. Object size is equal to one page. This value
179 * can be changed at boot time.
181 static int rcu_min_cached_objs = 5;
182 module_param(rcu_min_cached_objs, int, 0444);
184 // A page shrinker can ask for pages to be freed to make them
185 // available for other parts of the system. This usually happens
186 // under low memory conditions, and in that case we should also
187 // defer page-cache filling for a short time period.
189 // The default value is 5 seconds, which is long enough to reduce
190 // interference with the shrinker while it asks other systems to
191 // drain their caches.
192 static int rcu_delay_page_cache_fill_msec = 5000;
193 module_param(rcu_delay_page_cache_fill_msec, int, 0444);
195 /* Retrieve RCU kthreads priority for rcutorture */
196 int rcu_get_gp_kthreads_prio(void)
200 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
203 * Number of grace periods between delays, normalized by the duration of
204 * the delay. The longer the delay, the more the grace periods between
205 * each delay. The reason for this normalization is that it means that,
206 * for non-zero delays, the overall slowdown of grace periods is constant
207 * regardless of the duration of the delay. This arrangement balances
208 * the need for long delays to increase some race probabilities with the
209 * need for fast grace periods to increase other race probabilities.
211 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
214 * Compute the mask of online CPUs for the specified rcu_node structure.
215 * This will not be stable unless the rcu_node structure's ->lock is
216 * held, but the bit corresponding to the current CPU will be stable
219 static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
221 return READ_ONCE(rnp->qsmaskinitnext);
225 * Return true if an RCU grace period is in progress. The READ_ONCE()s
226 * permit this function to be invoked without holding the root rcu_node
227 * structure's ->lock, but of course results can be subject to change.
229 static int rcu_gp_in_progress(void)
231 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
235 * Return the number of callbacks queued on the specified CPU.
236 * Handles both the nocbs and normal cases.
238 static long rcu_get_n_cbs_cpu(int cpu)
240 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
242 if (rcu_segcblist_is_enabled(&rdp->cblist))
243 return rcu_segcblist_n_cbs(&rdp->cblist);
247 void rcu_softirq_qs(void)
250 rcu_preempt_deferred_qs(current);
251 rcu_tasks_qs(current, false);
255 * Increment the current CPU's rcu_data structure's ->dynticks field
256 * with ordering. Return the new value.
258 static noinline noinstr unsigned long rcu_dynticks_inc(int incby)
260 return arch_atomic_add_return(incby, this_cpu_ptr(&rcu_data.dynticks));
264 * Record entry into an extended quiescent state. This is only to be
265 * called when not already in an extended quiescent state, that is,
266 * RCU is watching prior to the call to this function and is no longer
267 * watching upon return.
269 static noinstr void rcu_dynticks_eqs_enter(void)
274 * CPUs seeing atomic_add_return() must see prior RCU read-side
275 * critical sections, and we also must force ordering with the
278 rcu_dynticks_task_trace_enter(); // Before ->dynticks update!
279 seq = rcu_dynticks_inc(1);
280 // RCU is no longer watching. Better be in extended quiescent state!
281 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && (seq & 0x1));
285 * Record exit from an extended quiescent state. This is only to be
286 * called from an extended quiescent state, that is, RCU is not watching
287 * prior to the call to this function and is watching upon return.
289 static noinstr void rcu_dynticks_eqs_exit(void)
294 * CPUs seeing atomic_add_return() must see prior idle sojourns,
295 * and we also must force ordering with the next RCU read-side
298 seq = rcu_dynticks_inc(1);
299 // RCU is now watching. Better not be in an extended quiescent state!
300 rcu_dynticks_task_trace_exit(); // After ->dynticks update!
301 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !(seq & 0x1));
305 * Reset the current CPU's ->dynticks counter to indicate that the
306 * newly onlined CPU is no longer in an extended quiescent state.
307 * This will either leave the counter unchanged, or increment it
308 * to the next non-quiescent value.
310 * The non-atomic test/increment sequence works because the upper bits
311 * of the ->dynticks counter are manipulated only by the corresponding CPU,
312 * or when the corresponding CPU is offline.
314 static void rcu_dynticks_eqs_online(void)
316 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
318 if (atomic_read(&rdp->dynticks) & 0x1)
324 * Is the current CPU in an extended quiescent state?
326 * No ordering, as we are sampling CPU-local information.
328 static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void)
330 return !(arch_atomic_read(this_cpu_ptr(&rcu_data.dynticks)) & 0x1);
334 * Snapshot the ->dynticks counter with full ordering so as to allow
335 * stable comparison of this counter with past and future snapshots.
337 static int rcu_dynticks_snap(struct rcu_data *rdp)
339 smp_mb(); // Fundamental RCU ordering guarantee.
340 return atomic_read_acquire(&rdp->dynticks);
344 * Return true if the snapshot returned from rcu_dynticks_snap()
345 * indicates that RCU is in an extended quiescent state.
347 static bool rcu_dynticks_in_eqs(int snap)
349 return !(snap & 0x1);
352 /* Return true if the specified CPU is currently idle from an RCU viewpoint. */
353 bool rcu_is_idle_cpu(int cpu)
355 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
357 return rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
361 * Return true if the CPU corresponding to the specified rcu_data
362 * structure has spent some time in an extended quiescent state since
363 * rcu_dynticks_snap() returned the specified snapshot.
365 static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
367 return snap != rcu_dynticks_snap(rdp);
371 * Return true if the referenced integer is zero while the specified
372 * CPU remains within a single extended quiescent state.
374 bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
376 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
379 // If not quiescent, force back to earlier extended quiescent state.
380 snap = atomic_read(&rdp->dynticks) & ~0x1;
382 smp_rmb(); // Order ->dynticks and *vp reads.
384 return false; // Non-zero, so report failure;
385 smp_rmb(); // Order *vp read and ->dynticks re-read.
387 // If still in the same extended quiescent state, we are good!
388 return snap == atomic_read(&rdp->dynticks);
392 * Let the RCU core know that this CPU has gone through the scheduler,
393 * which is a quiescent state. This is called when the need for a
394 * quiescent state is urgent, so we burn an atomic operation and full
395 * memory barriers to let the RCU core know about it, regardless of what
396 * this CPU might (or might not) do in the near future.
398 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
400 * The caller must have disabled interrupts and must not be idle.
402 notrace void rcu_momentary_dyntick_idle(void)
406 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
407 seq = rcu_dynticks_inc(2);
408 /* It is illegal to call this from idle state. */
409 WARN_ON_ONCE(!(seq & 0x1));
410 rcu_preempt_deferred_qs(current);
412 EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
415 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
417 * If the current CPU is idle and running at a first-level (not nested)
418 * interrupt, or directly, from idle, return true.
420 * The caller must have at least disabled IRQs.
422 static int rcu_is_cpu_rrupt_from_idle(void)
427 * Usually called from the tick; but also used from smp_function_call()
428 * for expedited grace periods. This latter can result in running from
429 * the idle task, instead of an actual IPI.
431 lockdep_assert_irqs_disabled();
433 /* Check for counter underflows */
434 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
435 "RCU dynticks_nesting counter underflow!");
436 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
437 "RCU dynticks_nmi_nesting counter underflow/zero!");
439 /* Are we at first interrupt nesting level? */
440 nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting);
445 * If we're not in an interrupt, we must be in the idle task!
447 WARN_ON_ONCE(!nesting && !is_idle_task(current));
449 /* Does CPU appear to be idle from an RCU standpoint? */
450 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
453 #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
454 // Maximum callbacks per rcu_do_batch ...
455 #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
456 static long blimit = DEFAULT_RCU_BLIMIT;
457 #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
458 static long qhimark = DEFAULT_RCU_QHIMARK;
459 #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
460 static long qlowmark = DEFAULT_RCU_QLOMARK;
461 #define DEFAULT_RCU_QOVLD_MULT 2
462 #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
463 static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
464 static long qovld_calc = -1; // No pre-initialization lock acquisitions!
466 module_param(blimit, long, 0444);
467 module_param(qhimark, long, 0444);
468 module_param(qlowmark, long, 0444);
469 module_param(qovld, long, 0444);
471 static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
472 static ulong jiffies_till_next_fqs = ULONG_MAX;
473 static bool rcu_kick_kthreads;
474 static int rcu_divisor = 7;
475 module_param(rcu_divisor, int, 0644);
477 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
478 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
479 module_param(rcu_resched_ns, long, 0644);
482 * How long the grace period must be before we start recruiting
483 * quiescent-state help from rcu_note_context_switch().
485 static ulong jiffies_till_sched_qs = ULONG_MAX;
486 module_param(jiffies_till_sched_qs, ulong, 0444);
487 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
488 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
491 * Make sure that we give the grace-period kthread time to detect any
492 * idle CPUs before taking active measures to force quiescent states.
493 * However, don't go below 100 milliseconds, adjusted upwards for really
496 static void adjust_jiffies_till_sched_qs(void)
500 /* If jiffies_till_sched_qs was specified, respect the request. */
501 if (jiffies_till_sched_qs != ULONG_MAX) {
502 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
505 /* Otherwise, set to third fqs scan, but bound below on large system. */
506 j = READ_ONCE(jiffies_till_first_fqs) +
507 2 * READ_ONCE(jiffies_till_next_fqs);
508 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
509 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
510 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
511 WRITE_ONCE(jiffies_to_sched_qs, j);
514 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
517 int ret = kstrtoul(val, 0, &j);
520 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
521 adjust_jiffies_till_sched_qs();
526 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
529 int ret = kstrtoul(val, 0, &j);
532 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
533 adjust_jiffies_till_sched_qs();
538 static const struct kernel_param_ops first_fqs_jiffies_ops = {
539 .set = param_set_first_fqs_jiffies,
540 .get = param_get_ulong,
543 static const struct kernel_param_ops next_fqs_jiffies_ops = {
544 .set = param_set_next_fqs_jiffies,
545 .get = param_get_ulong,
548 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
549 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
550 module_param(rcu_kick_kthreads, bool, 0644);
552 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
553 static int rcu_pending(int user);
556 * Return the number of RCU GPs completed thus far for debug & stats.
558 unsigned long rcu_get_gp_seq(void)
560 return READ_ONCE(rcu_state.gp_seq);
562 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
565 * Return the number of RCU expedited batches completed thus far for
566 * debug & stats. Odd numbers mean that a batch is in progress, even
567 * numbers mean idle. The value returned will thus be roughly double
568 * the cumulative batches since boot.
570 unsigned long rcu_exp_batches_completed(void)
572 return rcu_state.expedited_sequence;
574 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
577 * Return the root node of the rcu_state structure.
579 static struct rcu_node *rcu_get_root(void)
581 return &rcu_state.node[0];
585 * Send along grace-period-related data for rcutorture diagnostics.
587 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
588 unsigned long *gp_seq)
592 *flags = READ_ONCE(rcu_state.gp_flags);
593 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
599 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
602 * Enter an RCU extended quiescent state, which can be either the
603 * idle loop or adaptive-tickless usermode execution.
605 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
606 * the possibility of usermode upcalls having messed up our count
607 * of interrupt nesting level during the prior busy period.
609 static noinstr void rcu_eqs_enter(bool user)
611 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
613 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
614 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
615 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
616 rdp->dynticks_nesting == 0);
617 if (rdp->dynticks_nesting != 1) {
618 // RCU will still be watching, so just do accounting and leave.
619 rdp->dynticks_nesting--;
623 lockdep_assert_irqs_disabled();
624 instrumentation_begin();
625 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
626 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
627 rcu_prepare_for_idle();
628 rcu_preempt_deferred_qs(current);
630 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
631 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
633 instrumentation_end();
634 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
635 // RCU is watching here ...
636 rcu_dynticks_eqs_enter();
637 // ... but is no longer watching here.
638 rcu_dynticks_task_enter();
642 * rcu_idle_enter - inform RCU that current CPU is entering idle
644 * Enter idle mode, in other words, -leave- the mode in which RCU
645 * read-side critical sections can occur. (Though RCU read-side
646 * critical sections can occur in irq handlers in idle, a possibility
647 * handled by irq_enter() and irq_exit().)
649 * If you add or remove a call to rcu_idle_enter(), be sure to test with
650 * CONFIG_RCU_EQS_DEBUG=y.
652 void rcu_idle_enter(void)
654 lockdep_assert_irqs_disabled();
655 rcu_eqs_enter(false);
657 EXPORT_SYMBOL_GPL(rcu_idle_enter);
659 #ifdef CONFIG_NO_HZ_FULL
661 #if !defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)
663 * An empty function that will trigger a reschedule on
664 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
666 static void late_wakeup_func(struct irq_work *work)
670 static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
671 IRQ_WORK_INIT(late_wakeup_func);
676 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
677 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
679 * In these cases the late RCU wake ups aren't supported in the resched loops and our
680 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
681 * get re-enabled again.
683 noinstr static void rcu_irq_work_resched(void)
685 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
687 if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
690 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
693 instrumentation_begin();
694 if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
695 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
697 instrumentation_end();
701 static inline void rcu_irq_work_resched(void) { }
705 * rcu_user_enter - inform RCU that we are resuming userspace.
707 * Enter RCU idle mode right before resuming userspace. No use of RCU
708 * is permitted between this call and rcu_user_exit(). This way the
709 * CPU doesn't need to maintain the tick for RCU maintenance purposes
710 * when the CPU runs in userspace.
712 * If you add or remove a call to rcu_user_enter(), be sure to test with
713 * CONFIG_RCU_EQS_DEBUG=y.
715 noinstr void rcu_user_enter(void)
717 lockdep_assert_irqs_disabled();
720 * Other than generic entry implementation, we may be past the last
721 * rescheduling opportunity in the entry code. Trigger a self IPI
722 * that will fire and reschedule once we resume in user/guest mode.
724 rcu_irq_work_resched();
728 #endif /* CONFIG_NO_HZ_FULL */
731 * rcu_nmi_exit - inform RCU of exit from NMI context
733 * If we are returning from the outermost NMI handler that interrupted an
734 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
735 * to let the RCU grace-period handling know that the CPU is back to
738 * If you add or remove a call to rcu_nmi_exit(), be sure to test
739 * with CONFIG_RCU_EQS_DEBUG=y.
741 noinstr void rcu_nmi_exit(void)
743 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
745 instrumentation_begin();
747 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
748 * (We are exiting an NMI handler, so RCU better be paying attention
751 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
752 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
755 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
756 * leave it in non-RCU-idle state.
758 if (rdp->dynticks_nmi_nesting != 1) {
759 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
760 atomic_read(&rdp->dynticks));
761 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
762 rdp->dynticks_nmi_nesting - 2);
763 instrumentation_end();
767 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
768 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
769 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
772 rcu_prepare_for_idle();
774 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
775 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
776 instrumentation_end();
778 // RCU is watching here ...
779 rcu_dynticks_eqs_enter();
780 // ... but is no longer watching here.
783 rcu_dynticks_task_enter();
787 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
789 * Exit from an interrupt handler, which might possibly result in entering
790 * idle mode, in other words, leaving the mode in which read-side critical
791 * sections can occur. The caller must have disabled interrupts.
793 * This code assumes that the idle loop never does anything that might
794 * result in unbalanced calls to irq_enter() and irq_exit(). If your
795 * architecture's idle loop violates this assumption, RCU will give you what
796 * you deserve, good and hard. But very infrequently and irreproducibly.
798 * Use things like work queues to work around this limitation.
800 * You have been warned.
802 * If you add or remove a call to rcu_irq_exit(), be sure to test with
803 * CONFIG_RCU_EQS_DEBUG=y.
805 void noinstr rcu_irq_exit(void)
807 lockdep_assert_irqs_disabled();
811 #ifdef CONFIG_PROVE_RCU
813 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
815 void rcu_irq_exit_check_preempt(void)
817 lockdep_assert_irqs_disabled();
819 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
820 "RCU dynticks_nesting counter underflow/zero!");
821 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
823 "Bad RCU dynticks_nmi_nesting counter\n");
824 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
825 "RCU in extended quiescent state!");
827 #endif /* #ifdef CONFIG_PROVE_RCU */
830 * Wrapper for rcu_irq_exit() where interrupts are enabled.
832 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
833 * with CONFIG_RCU_EQS_DEBUG=y.
835 void rcu_irq_exit_irqson(void)
839 local_irq_save(flags);
841 local_irq_restore(flags);
845 * Exit an RCU extended quiescent state, which can be either the
846 * idle loop or adaptive-tickless usermode execution.
848 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
849 * allow for the possibility of usermode upcalls messing up our count of
850 * interrupt nesting level during the busy period that is just now starting.
852 static void noinstr rcu_eqs_exit(bool user)
854 struct rcu_data *rdp;
857 lockdep_assert_irqs_disabled();
858 rdp = this_cpu_ptr(&rcu_data);
859 oldval = rdp->dynticks_nesting;
860 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
862 // RCU was already watching, so just do accounting and leave.
863 rdp->dynticks_nesting++;
866 rcu_dynticks_task_exit();
867 // RCU is not watching here ...
868 rcu_dynticks_eqs_exit();
869 // ... but is watching here.
870 instrumentation_begin();
872 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
873 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
875 rcu_cleanup_after_idle();
876 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
877 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
878 WRITE_ONCE(rdp->dynticks_nesting, 1);
879 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
880 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
881 instrumentation_end();
885 * rcu_idle_exit - inform RCU that current CPU is leaving idle
887 * Exit idle mode, in other words, -enter- the mode in which RCU
888 * read-side critical sections can occur.
890 * If you add or remove a call to rcu_idle_exit(), be sure to test with
891 * CONFIG_RCU_EQS_DEBUG=y.
893 void rcu_idle_exit(void)
897 local_irq_save(flags);
899 local_irq_restore(flags);
901 EXPORT_SYMBOL_GPL(rcu_idle_exit);
903 #ifdef CONFIG_NO_HZ_FULL
905 * rcu_user_exit - inform RCU that we are exiting userspace.
907 * Exit RCU idle mode while entering the kernel because it can
908 * run a RCU read side critical section anytime.
910 * If you add or remove a call to rcu_user_exit(), be sure to test with
911 * CONFIG_RCU_EQS_DEBUG=y.
913 void noinstr rcu_user_exit(void)
919 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
921 * The scheduler tick is not normally enabled when CPUs enter the kernel
922 * from nohz_full userspace execution. After all, nohz_full userspace
923 * execution is an RCU quiescent state and the time executing in the kernel
924 * is quite short. Except of course when it isn't. And it is not hard to
925 * cause a large system to spend tens of seconds or even minutes looping
926 * in the kernel, which can cause a number of problems, include RCU CPU
929 * Therefore, if a nohz_full CPU fails to report a quiescent state
930 * in a timely manner, the RCU grace-period kthread sets that CPU's
931 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
932 * exception will invoke this function, which will turn on the scheduler
933 * tick, which will enable RCU to detect that CPU's quiescent states,
934 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
935 * The tick will be disabled once a quiescent state is reported for
938 * Of course, in carefully tuned systems, there might never be an
939 * interrupt or exception. In that case, the RCU grace-period kthread
940 * will eventually cause one to happen. However, in less carefully
941 * controlled environments, this function allows RCU to get what it
942 * needs without creating otherwise useless interruptions.
944 void __rcu_irq_enter_check_tick(void)
946 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
948 // If we're here from NMI there's nothing to do.
952 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
953 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
955 if (!tick_nohz_full_cpu(rdp->cpu) ||
956 !READ_ONCE(rdp->rcu_urgent_qs) ||
957 READ_ONCE(rdp->rcu_forced_tick)) {
958 // RCU doesn't need nohz_full help from this CPU, or it is
959 // already getting that help.
963 // We get here only when not in an extended quiescent state and
964 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
965 // already watching and (2) The fact that we are in an interrupt
966 // handler and that the rcu_node lock is an irq-disabled lock
967 // prevents self-deadlock. So we can safely recheck under the lock.
968 // Note that the nohz_full state currently cannot change.
969 raw_spin_lock_rcu_node(rdp->mynode);
970 if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
971 // A nohz_full CPU is in the kernel and RCU needs a
972 // quiescent state. Turn on the tick!
973 WRITE_ONCE(rdp->rcu_forced_tick, true);
974 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
976 raw_spin_unlock_rcu_node(rdp->mynode);
978 NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
979 #endif /* CONFIG_NO_HZ_FULL */
982 * rcu_nmi_enter - inform RCU of entry to NMI context
984 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
985 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
986 * that the CPU is active. This implementation permits nested NMIs, as
987 * long as the nesting level does not overflow an int. (You will probably
988 * run out of stack space first.)
990 * If you add or remove a call to rcu_nmi_enter(), be sure to test
991 * with CONFIG_RCU_EQS_DEBUG=y.
993 noinstr void rcu_nmi_enter(void)
996 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
998 /* Complain about underflow. */
999 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
1002 * If idle from RCU viewpoint, atomically increment ->dynticks
1003 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1004 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1005 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1006 * to be in the outermost NMI handler that interrupted an RCU-idle
1007 * period (observation due to Andy Lutomirski).
1009 if (rcu_dynticks_curr_cpu_in_eqs()) {
1012 rcu_dynticks_task_exit();
1014 // RCU is not watching here ...
1015 rcu_dynticks_eqs_exit();
1016 // ... but is watching here.
1019 instrumentation_begin();
1020 rcu_cleanup_after_idle();
1021 instrumentation_end();
1024 instrumentation_begin();
1025 // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
1026 instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
1027 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
1028 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
1031 } else if (!in_nmi()) {
1032 instrumentation_begin();
1033 rcu_irq_enter_check_tick();
1035 instrumentation_begin();
1038 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1039 rdp->dynticks_nmi_nesting,
1040 rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
1041 instrumentation_end();
1042 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
1043 rdp->dynticks_nmi_nesting + incby);
1048 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1050 * Enter an interrupt handler, which might possibly result in exiting
1051 * idle mode, in other words, entering the mode in which read-side critical
1052 * sections can occur. The caller must have disabled interrupts.
1054 * Note that the Linux kernel is fully capable of entering an interrupt
1055 * handler that it never exits, for example when doing upcalls to user mode!
1056 * This code assumes that the idle loop never does upcalls to user mode.
1057 * If your architecture's idle loop does do upcalls to user mode (or does
1058 * anything else that results in unbalanced calls to the irq_enter() and
1059 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1060 * But very infrequently and irreproducibly.
1062 * Use things like work queues to work around this limitation.
1064 * You have been warned.
1066 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1067 * CONFIG_RCU_EQS_DEBUG=y.
1069 noinstr void rcu_irq_enter(void)
1071 lockdep_assert_irqs_disabled();
1076 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1078 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1079 * with CONFIG_RCU_EQS_DEBUG=y.
1081 void rcu_irq_enter_irqson(void)
1083 unsigned long flags;
1085 local_irq_save(flags);
1087 local_irq_restore(flags);
1091 * If any sort of urgency was applied to the current CPU (for example,
1092 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
1093 * to get to a quiescent state, disable it.
1095 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
1097 raw_lockdep_assert_held_rcu_node(rdp->mynode);
1098 WRITE_ONCE(rdp->rcu_urgent_qs, false);
1099 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
1100 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
1101 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1102 WRITE_ONCE(rdp->rcu_forced_tick, false);
1107 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
1109 * Return true if RCU is watching the running CPU, which means that this
1110 * CPU can safely enter RCU read-side critical sections. In other words,
1111 * if the current CPU is not in its idle loop or is in an interrupt or
1112 * NMI handler, return true.
1114 * Make notrace because it can be called by the internal functions of
1115 * ftrace, and making this notrace removes unnecessary recursion calls.
1117 notrace bool rcu_is_watching(void)
1121 preempt_disable_notrace();
1122 ret = !rcu_dynticks_curr_cpu_in_eqs();
1123 preempt_enable_notrace();
1126 EXPORT_SYMBOL_GPL(rcu_is_watching);
1129 * If a holdout task is actually running, request an urgent quiescent
1130 * state from its CPU. This is unsynchronized, so migrations can cause
1131 * the request to go to the wrong CPU. Which is OK, all that will happen
1132 * is that the CPU's next context switch will be a bit slower and next
1133 * time around this task will generate another request.
1135 void rcu_request_urgent_qs_task(struct task_struct *t)
1142 return; /* This task is not running on that CPU. */
1143 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
1146 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1149 * Is the current CPU online as far as RCU is concerned?
1151 * Disable preemption to avoid false positives that could otherwise
1152 * happen due to the current CPU number being sampled, this task being
1153 * preempted, its old CPU being taken offline, resuming on some other CPU,
1154 * then determining that its old CPU is now offline.
1156 * Disable checking if in an NMI handler because we cannot safely
1157 * report errors from NMI handlers anyway. In addition, it is OK to use
1158 * RCU on an offline processor during initial boot, hence the check for
1159 * rcu_scheduler_fully_active.
1161 bool rcu_lockdep_current_cpu_online(void)
1163 struct rcu_data *rdp;
1164 struct rcu_node *rnp;
1167 if (in_nmi() || !rcu_scheduler_fully_active)
1169 preempt_disable_notrace();
1170 rdp = this_cpu_ptr(&rcu_data);
1172 if (rdp->grpmask & rcu_rnp_online_cpus(rnp) || READ_ONCE(rnp->ofl_seq) & 0x1)
1174 preempt_enable_notrace();
1177 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1179 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1182 * When trying to report a quiescent state on behalf of some other CPU,
1183 * it is our responsibility to check for and handle potential overflow
1184 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1185 * After all, the CPU might be in deep idle state, and thus executing no
1188 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1190 raw_lockdep_assert_held_rcu_node(rnp);
1191 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1193 WRITE_ONCE(rdp->gpwrap, true);
1194 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1195 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1199 * Snapshot the specified CPU's dynticks counter so that we can later
1200 * credit them with an implicit quiescent state. Return 1 if this CPU
1201 * is in dynticks idle mode, which is an extended quiescent state.
1203 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1205 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1206 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1207 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1208 rcu_gpnum_ovf(rdp->mynode, rdp);
1215 * Return true if the specified CPU has passed through a quiescent
1216 * state by virtue of being in or having passed through an dynticks
1217 * idle state since the last call to dyntick_save_progress_counter()
1218 * for this same CPU, or by virtue of having been offline.
1220 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1225 struct rcu_node *rnp = rdp->mynode;
1228 * If the CPU passed through or entered a dynticks idle phase with
1229 * no active irq/NMI handlers, then we can safely pretend that the CPU
1230 * already acknowledged the request to pass through a quiescent
1231 * state. Either way, that CPU cannot possibly be in an RCU
1232 * read-side critical section that started before the beginning
1233 * of the current RCU grace period.
1235 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1236 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1237 rcu_gpnum_ovf(rnp, rdp);
1242 * Complain if a CPU that is considered to be offline from RCU's
1243 * perspective has not yet reported a quiescent state. After all,
1244 * the offline CPU should have reported a quiescent state during
1245 * the CPU-offline process, or, failing that, by rcu_gp_init()
1246 * if it ran concurrently with either the CPU going offline or the
1247 * last task on a leaf rcu_node structure exiting its RCU read-side
1248 * critical section while all CPUs corresponding to that structure
1249 * are offline. This added warning detects bugs in any of these
1252 * The rcu_node structure's ->lock is held here, which excludes
1253 * the relevant portions the CPU-hotplug code, the grace-period
1254 * initialization code, and the rcu_read_unlock() code paths.
1256 * For more detail, please refer to the "Hotplug CPU" section
1257 * of RCU's Requirements documentation.
1259 if (WARN_ON_ONCE(!(rdp->grpmask & rcu_rnp_online_cpus(rnp)))) {
1261 struct rcu_node *rnp1;
1263 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1264 __func__, rnp->grplo, rnp->grphi, rnp->level,
1265 (long)rnp->gp_seq, (long)rnp->completedqs);
1266 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1267 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1268 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1269 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1270 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1271 __func__, rdp->cpu, ".o"[onl],
1272 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1273 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1274 return 1; /* Break things loose after complaining. */
1278 * A CPU running for an extended time within the kernel can
1279 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1280 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1281 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1282 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1283 * variable are safe because the assignments are repeated if this
1284 * CPU failed to pass through a quiescent state. This code
1285 * also checks .jiffies_resched in case jiffies_to_sched_qs
1288 jtsq = READ_ONCE(jiffies_to_sched_qs);
1289 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1290 rnhqp = per_cpu_ptr(&rcu_data.rcu_need_heavy_qs, rdp->cpu);
1291 if (!READ_ONCE(*rnhqp) &&
1292 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1293 time_after(jiffies, rcu_state.jiffies_resched) ||
1294 rcu_state.cbovld)) {
1295 WRITE_ONCE(*rnhqp, true);
1296 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1297 smp_store_release(ruqp, true);
1298 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1299 WRITE_ONCE(*ruqp, true);
1303 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1304 * The above code handles this, but only for straight cond_resched().
1305 * And some in-kernel loops check need_resched() before calling
1306 * cond_resched(), which defeats the above code for CPUs that are
1307 * running in-kernel with scheduling-clock interrupts disabled.
1308 * So hit them over the head with the resched_cpu() hammer!
1310 if (tick_nohz_full_cpu(rdp->cpu) &&
1311 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
1312 rcu_state.cbovld)) {
1313 WRITE_ONCE(*ruqp, true);
1314 resched_cpu(rdp->cpu);
1315 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1319 * If more than halfway to RCU CPU stall-warning time, invoke
1320 * resched_cpu() more frequently to try to loosen things up a bit.
1321 * Also check to see if the CPU is getting hammered with interrupts,
1322 * but only once per grace period, just to keep the IPIs down to
1325 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1326 if (time_after(jiffies,
1327 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1328 resched_cpu(rdp->cpu);
1329 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1331 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1332 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1333 (rnp->ffmask & rdp->grpmask)) {
1334 rdp->rcu_iw_pending = true;
1335 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1336 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1343 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1344 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1345 unsigned long gp_seq_req, const char *s)
1347 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
1348 gp_seq_req, rnp->level,
1349 rnp->grplo, rnp->grphi, s);
1353 * rcu_start_this_gp - Request the start of a particular grace period
1354 * @rnp_start: The leaf node of the CPU from which to start.
1355 * @rdp: The rcu_data corresponding to the CPU from which to start.
1356 * @gp_seq_req: The gp_seq of the grace period to start.
1358 * Start the specified grace period, as needed to handle newly arrived
1359 * callbacks. The required future grace periods are recorded in each
1360 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1361 * is reason to awaken the grace-period kthread.
1363 * The caller must hold the specified rcu_node structure's ->lock, which
1364 * is why the caller is responsible for waking the grace-period kthread.
1366 * Returns true if the GP thread needs to be awakened else false.
1368 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1369 unsigned long gp_seq_req)
1372 struct rcu_node *rnp;
1375 * Use funnel locking to either acquire the root rcu_node
1376 * structure's lock or bail out if the need for this grace period
1377 * has already been recorded -- or if that grace period has in
1378 * fact already started. If there is already a grace period in
1379 * progress in a non-leaf node, no recording is needed because the
1380 * end of the grace period will scan the leaf rcu_node structures.
1381 * Note that rnp_start->lock must not be released.
1383 raw_lockdep_assert_held_rcu_node(rnp_start);
1384 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1385 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1386 if (rnp != rnp_start)
1387 raw_spin_lock_rcu_node(rnp);
1388 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1389 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1390 (rnp != rnp_start &&
1391 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1392 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1396 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1397 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1399 * We just marked the leaf or internal node, and a
1400 * grace period is in progress, which means that
1401 * rcu_gp_cleanup() will see the marking. Bail to
1402 * reduce contention.
1404 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1405 TPS("Startedleaf"));
1408 if (rnp != rnp_start && rnp->parent != NULL)
1409 raw_spin_unlock_rcu_node(rnp);
1411 break; /* At root, and perhaps also leaf. */
1414 /* If GP already in progress, just leave, otherwise start one. */
1415 if (rcu_gp_in_progress()) {
1416 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1419 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1420 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1421 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1422 if (!READ_ONCE(rcu_state.gp_kthread)) {
1423 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1426 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1427 ret = true; /* Caller must wake GP kthread. */
1429 /* Push furthest requested GP to leaf node and rcu_data structure. */
1430 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1431 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1432 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1434 if (rnp != rnp_start)
1435 raw_spin_unlock_rcu_node(rnp);
1440 * Clean up any old requests for the just-ended grace period. Also return
1441 * whether any additional grace periods have been requested.
1443 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1446 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1448 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1450 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1451 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1452 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1457 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1458 * interrupt or softirq handler, in which case we just might immediately
1459 * sleep upon return, resulting in a grace-period hang), and don't bother
1460 * awakening when there is nothing for the grace-period kthread to do
1461 * (as in several CPUs raced to awaken, we lost), and finally don't try
1462 * to awaken a kthread that has not yet been created. If all those checks
1463 * are passed, track some debug information and awaken.
1465 * So why do the self-wakeup when in an interrupt or softirq handler
1466 * in the grace-period kthread's context? Because the kthread might have
1467 * been interrupted just as it was going to sleep, and just after the final
1468 * pre-sleep check of the awaken condition. In this case, a wakeup really
1469 * is required, and is therefore supplied.
1471 static void rcu_gp_kthread_wake(void)
1473 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1475 if ((current == t && !in_irq() && !in_serving_softirq()) ||
1476 !READ_ONCE(rcu_state.gp_flags) || !t)
1478 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1479 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1480 swake_up_one(&rcu_state.gp_wq);
1484 * If there is room, assign a ->gp_seq number to any callbacks on this
1485 * CPU that have not already been assigned. Also accelerate any callbacks
1486 * that were previously assigned a ->gp_seq number that has since proven
1487 * to be too conservative, which can happen if callbacks get assigned a
1488 * ->gp_seq number while RCU is idle, but with reference to a non-root
1489 * rcu_node structure. This function is idempotent, so it does not hurt
1490 * to call it repeatedly. Returns an flag saying that we should awaken
1491 * the RCU grace-period kthread.
1493 * The caller must hold rnp->lock with interrupts disabled.
1495 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1497 unsigned long gp_seq_req;
1500 rcu_lockdep_assert_cblist_protected(rdp);
1501 raw_lockdep_assert_held_rcu_node(rnp);
1503 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1504 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1507 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1510 * Callbacks are often registered with incomplete grace-period
1511 * information. Something about the fact that getting exact
1512 * information requires acquiring a global lock... RCU therefore
1513 * makes a conservative estimate of the grace period number at which
1514 * a given callback will become ready to invoke. The following
1515 * code checks this estimate and improves it when possible, thus
1516 * accelerating callback invocation to an earlier grace-period
1519 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1520 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1521 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1523 /* Trace depending on how much we were able to accelerate. */
1524 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1525 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1527 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1529 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1535 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1536 * rcu_node structure's ->lock be held. It consults the cached value
1537 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1538 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1539 * while holding the leaf rcu_node structure's ->lock.
1541 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1542 struct rcu_data *rdp)
1547 rcu_lockdep_assert_cblist_protected(rdp);
1548 c = rcu_seq_snap(&rcu_state.gp_seq);
1549 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1550 /* Old request still live, so mark recent callbacks. */
1551 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1554 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1555 needwake = rcu_accelerate_cbs(rnp, rdp);
1556 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1558 rcu_gp_kthread_wake();
1562 * Move any callbacks whose grace period has completed to the
1563 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1564 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1565 * sublist. This function is idempotent, so it does not hurt to
1566 * invoke it repeatedly. As long as it is not invoked -too- often...
1567 * Returns true if the RCU grace-period kthread needs to be awakened.
1569 * The caller must hold rnp->lock with interrupts disabled.
1571 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1573 rcu_lockdep_assert_cblist_protected(rdp);
1574 raw_lockdep_assert_held_rcu_node(rnp);
1576 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1577 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1581 * Find all callbacks whose ->gp_seq numbers indicate that they
1582 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1584 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1586 /* Classify any remaining callbacks. */
1587 return rcu_accelerate_cbs(rnp, rdp);
1591 * Move and classify callbacks, but only if doing so won't require
1592 * that the RCU grace-period kthread be awakened.
1594 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1595 struct rcu_data *rdp)
1597 rcu_lockdep_assert_cblist_protected(rdp);
1598 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1600 // The grace period cannot end while we hold the rcu_node lock.
1601 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1602 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1603 raw_spin_unlock_rcu_node(rnp);
1607 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1608 * quiescent state. This is intended to be invoked when the CPU notices
1609 * a new grace period.
1611 static void rcu_strict_gp_check_qs(void)
1613 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1620 * Update CPU-local rcu_data state to record the beginnings and ends of
1621 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1622 * structure corresponding to the current CPU, and must have irqs disabled.
1623 * Returns true if the grace-period kthread needs to be awakened.
1625 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1629 const bool offloaded = rcu_rdp_is_offloaded(rdp);
1631 raw_lockdep_assert_held_rcu_node(rnp);
1633 if (rdp->gp_seq == rnp->gp_seq)
1634 return false; /* Nothing to do. */
1636 /* Handle the ends of any preceding grace periods first. */
1637 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1638 unlikely(READ_ONCE(rdp->gpwrap))) {
1640 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1641 rdp->core_needs_qs = false;
1642 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1645 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1646 if (rdp->core_needs_qs)
1647 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1650 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1651 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1652 unlikely(READ_ONCE(rdp->gpwrap))) {
1654 * If the current grace period is waiting for this CPU,
1655 * set up to detect a quiescent state, otherwise don't
1656 * go looking for one.
1658 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1659 need_qs = !!(rnp->qsmask & rdp->grpmask);
1660 rdp->cpu_no_qs.b.norm = need_qs;
1661 rdp->core_needs_qs = need_qs;
1662 zero_cpu_stall_ticks(rdp);
1664 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1665 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1666 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1667 WRITE_ONCE(rdp->gpwrap, false);
1668 rcu_gpnum_ovf(rnp, rdp);
1672 static void note_gp_changes(struct rcu_data *rdp)
1674 unsigned long flags;
1676 struct rcu_node *rnp;
1678 local_irq_save(flags);
1680 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1681 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1682 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1683 local_irq_restore(flags);
1686 needwake = __note_gp_changes(rnp, rdp);
1687 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1688 rcu_strict_gp_check_qs();
1690 rcu_gp_kthread_wake();
1693 static void rcu_gp_slow(int delay)
1696 !(rcu_seq_ctr(rcu_state.gp_seq) %
1697 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1698 schedule_timeout_idle(delay);
1701 static unsigned long sleep_duration;
1703 /* Allow rcutorture to stall the grace-period kthread. */
1704 void rcu_gp_set_torture_wait(int duration)
1706 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1707 WRITE_ONCE(sleep_duration, duration);
1709 EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1711 /* Actually implement the aforementioned wait. */
1712 static void rcu_gp_torture_wait(void)
1714 unsigned long duration;
1716 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1718 duration = xchg(&sleep_duration, 0UL);
1720 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1721 schedule_timeout_idle(duration);
1722 pr_alert("%s: Wait complete\n", __func__);
1727 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1730 static void rcu_strict_gp_boundary(void *unused)
1736 * Initialize a new grace period. Return false if no grace period required.
1738 static noinline_for_stack bool rcu_gp_init(void)
1740 unsigned long firstseq;
1741 unsigned long flags;
1742 unsigned long oldmask;
1744 struct rcu_data *rdp;
1745 struct rcu_node *rnp = rcu_get_root();
1747 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1748 raw_spin_lock_irq_rcu_node(rnp);
1749 if (!READ_ONCE(rcu_state.gp_flags)) {
1750 /* Spurious wakeup, tell caller to go back to sleep. */
1751 raw_spin_unlock_irq_rcu_node(rnp);
1754 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1756 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1758 * Grace period already in progress, don't start another.
1759 * Not supposed to be able to happen.
1761 raw_spin_unlock_irq_rcu_node(rnp);
1765 /* Advance to a new grace period and initialize state. */
1766 record_gp_stall_check_time();
1767 /* Record GP times before starting GP, hence rcu_seq_start(). */
1768 rcu_seq_start(&rcu_state.gp_seq);
1769 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1770 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1771 raw_spin_unlock_irq_rcu_node(rnp);
1774 * Apply per-leaf buffered online and offline operations to
1775 * the rcu_node tree. Note that this new grace period need not
1776 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1777 * offlining path, when combined with checks in this function,
1778 * will handle CPUs that are currently going offline or that will
1779 * go offline later. Please also refer to "Hotplug CPU" section
1780 * of RCU's Requirements documentation.
1782 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1783 rcu_for_each_leaf_node(rnp) {
1784 smp_mb(); // Pair with barriers used when updating ->ofl_seq to odd values.
1785 firstseq = READ_ONCE(rnp->ofl_seq);
1787 while (firstseq == READ_ONCE(rnp->ofl_seq))
1788 schedule_timeout_idle(1); // Can't wake unless RCU is watching.
1789 smp_mb(); // Pair with barriers used when updating ->ofl_seq to even values.
1790 raw_spin_lock(&rcu_state.ofl_lock);
1791 raw_spin_lock_irq_rcu_node(rnp);
1792 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1793 !rnp->wait_blkd_tasks) {
1794 /* Nothing to do on this leaf rcu_node structure. */
1795 raw_spin_unlock_irq_rcu_node(rnp);
1796 raw_spin_unlock(&rcu_state.ofl_lock);
1800 /* Record old state, apply changes to ->qsmaskinit field. */
1801 oldmask = rnp->qsmaskinit;
1802 rnp->qsmaskinit = rnp->qsmaskinitnext;
1804 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1805 if (!oldmask != !rnp->qsmaskinit) {
1806 if (!oldmask) { /* First online CPU for rcu_node. */
1807 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1808 rcu_init_new_rnp(rnp);
1809 } else if (rcu_preempt_has_tasks(rnp)) {
1810 rnp->wait_blkd_tasks = true; /* blocked tasks */
1811 } else { /* Last offline CPU and can propagate. */
1812 rcu_cleanup_dead_rnp(rnp);
1817 * If all waited-on tasks from prior grace period are
1818 * done, and if all this rcu_node structure's CPUs are
1819 * still offline, propagate up the rcu_node tree and
1820 * clear ->wait_blkd_tasks. Otherwise, if one of this
1821 * rcu_node structure's CPUs has since come back online,
1822 * simply clear ->wait_blkd_tasks.
1824 if (rnp->wait_blkd_tasks &&
1825 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1826 rnp->wait_blkd_tasks = false;
1827 if (!rnp->qsmaskinit)
1828 rcu_cleanup_dead_rnp(rnp);
1831 raw_spin_unlock_irq_rcu_node(rnp);
1832 raw_spin_unlock(&rcu_state.ofl_lock);
1834 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1837 * Set the quiescent-state-needed bits in all the rcu_node
1838 * structures for all currently online CPUs in breadth-first
1839 * order, starting from the root rcu_node structure, relying on the
1840 * layout of the tree within the rcu_state.node[] array. Note that
1841 * other CPUs will access only the leaves of the hierarchy, thus
1842 * seeing that no grace period is in progress, at least until the
1843 * corresponding leaf node has been initialized.
1845 * The grace period cannot complete until the initialization
1846 * process finishes, because this kthread handles both.
1848 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1849 rcu_for_each_node_breadth_first(rnp) {
1850 rcu_gp_slow(gp_init_delay);
1851 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1852 rdp = this_cpu_ptr(&rcu_data);
1853 rcu_preempt_check_blocked_tasks(rnp);
1854 rnp->qsmask = rnp->qsmaskinit;
1855 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1856 if (rnp == rdp->mynode)
1857 (void)__note_gp_changes(rnp, rdp);
1858 rcu_preempt_boost_start_gp(rnp);
1859 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1860 rnp->level, rnp->grplo,
1861 rnp->grphi, rnp->qsmask);
1862 /* Quiescent states for tasks on any now-offline CPUs. */
1863 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1864 rnp->rcu_gp_init_mask = mask;
1865 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1866 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1868 raw_spin_unlock_irq_rcu_node(rnp);
1869 cond_resched_tasks_rcu_qs();
1870 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1873 // If strict, make all CPUs aware of new grace period.
1874 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1875 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1881 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1884 static bool rcu_gp_fqs_check_wake(int *gfp)
1886 struct rcu_node *rnp = rcu_get_root();
1888 // If under overload conditions, force an immediate FQS scan.
1889 if (*gfp & RCU_GP_FLAG_OVLD)
1892 // Someone like call_rcu() requested a force-quiescent-state scan.
1893 *gfp = READ_ONCE(rcu_state.gp_flags);
1894 if (*gfp & RCU_GP_FLAG_FQS)
1897 // The current grace period has completed.
1898 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1905 * Do one round of quiescent-state forcing.
1907 static void rcu_gp_fqs(bool first_time)
1909 struct rcu_node *rnp = rcu_get_root();
1911 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1912 WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1914 /* Collect dyntick-idle snapshots. */
1915 force_qs_rnp(dyntick_save_progress_counter);
1917 /* Handle dyntick-idle and offline CPUs. */
1918 force_qs_rnp(rcu_implicit_dynticks_qs);
1920 /* Clear flag to prevent immediate re-entry. */
1921 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1922 raw_spin_lock_irq_rcu_node(rnp);
1923 WRITE_ONCE(rcu_state.gp_flags,
1924 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1925 raw_spin_unlock_irq_rcu_node(rnp);
1930 * Loop doing repeated quiescent-state forcing until the grace period ends.
1932 static noinline_for_stack void rcu_gp_fqs_loop(void)
1938 struct rcu_node *rnp = rcu_get_root();
1940 first_gp_fqs = true;
1941 j = READ_ONCE(jiffies_till_first_fqs);
1942 if (rcu_state.cbovld)
1943 gf = RCU_GP_FLAG_OVLD;
1947 WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1949 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1950 * update; required for stall checks.
1953 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1954 jiffies + (j ? 3 * j : 2));
1956 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1958 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1959 (void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
1960 rcu_gp_fqs_check_wake(&gf), j);
1961 rcu_gp_torture_wait();
1962 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
1963 /* Locking provides needed memory barriers. */
1964 /* If grace period done, leave loop. */
1965 if (!READ_ONCE(rnp->qsmask) &&
1966 !rcu_preempt_blocked_readers_cgp(rnp))
1968 /* If time for quiescent-state forcing, do it. */
1969 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1970 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1971 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1973 rcu_gp_fqs(first_gp_fqs);
1976 first_gp_fqs = false;
1977 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1979 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1981 cond_resched_tasks_rcu_qs();
1982 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1983 ret = 0; /* Force full wait till next FQS. */
1984 j = READ_ONCE(jiffies_till_next_fqs);
1986 /* Deal with stray signal. */
1987 cond_resched_tasks_rcu_qs();
1988 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1989 WARN_ON(signal_pending(current));
1990 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1992 ret = 1; /* Keep old FQS timing. */
1994 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1997 j = rcu_state.jiffies_force_qs - j;
2004 * Clean up after the old grace period.
2006 static noinline void rcu_gp_cleanup(void)
2009 bool needgp = false;
2010 unsigned long gp_duration;
2011 unsigned long new_gp_seq;
2013 struct rcu_data *rdp;
2014 struct rcu_node *rnp = rcu_get_root();
2015 struct swait_queue_head *sq;
2017 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2018 raw_spin_lock_irq_rcu_node(rnp);
2019 rcu_state.gp_end = jiffies;
2020 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
2021 if (gp_duration > rcu_state.gp_max)
2022 rcu_state.gp_max = gp_duration;
2025 * We know the grace period is complete, but to everyone else
2026 * it appears to still be ongoing. But it is also the case
2027 * that to everyone else it looks like there is nothing that
2028 * they can do to advance the grace period. It is therefore
2029 * safe for us to drop the lock in order to mark the grace
2030 * period as completed in all of the rcu_node structures.
2032 raw_spin_unlock_irq_rcu_node(rnp);
2035 * Propagate new ->gp_seq value to rcu_node structures so that
2036 * other CPUs don't have to wait until the start of the next grace
2037 * period to process their callbacks. This also avoids some nasty
2038 * RCU grace-period initialization races by forcing the end of
2039 * the current grace period to be completely recorded in all of
2040 * the rcu_node structures before the beginning of the next grace
2041 * period is recorded in any of the rcu_node structures.
2043 new_gp_seq = rcu_state.gp_seq;
2044 rcu_seq_end(&new_gp_seq);
2045 rcu_for_each_node_breadth_first(rnp) {
2046 raw_spin_lock_irq_rcu_node(rnp);
2047 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2048 dump_blkd_tasks(rnp, 10);
2049 WARN_ON_ONCE(rnp->qsmask);
2050 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2051 rdp = this_cpu_ptr(&rcu_data);
2052 if (rnp == rdp->mynode)
2053 needgp = __note_gp_changes(rnp, rdp) || needgp;
2054 /* smp_mb() provided by prior unlock-lock pair. */
2055 needgp = rcu_future_gp_cleanup(rnp) || needgp;
2056 // Reset overload indication for CPUs no longer overloaded
2057 if (rcu_is_leaf_node(rnp))
2058 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
2059 rdp = per_cpu_ptr(&rcu_data, cpu);
2060 check_cb_ovld_locked(rdp, rnp);
2062 sq = rcu_nocb_gp_get(rnp);
2063 raw_spin_unlock_irq_rcu_node(rnp);
2064 rcu_nocb_gp_cleanup(sq);
2065 cond_resched_tasks_rcu_qs();
2066 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2067 rcu_gp_slow(gp_cleanup_delay);
2069 rnp = rcu_get_root();
2070 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2072 /* Declare grace period done, trace first to use old GP number. */
2073 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2074 rcu_seq_end(&rcu_state.gp_seq);
2075 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2076 WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
2077 /* Check for GP requests since above loop. */
2078 rdp = this_cpu_ptr(&rcu_data);
2079 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2080 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2081 TPS("CleanupMore"));
2084 /* Advance CBs to reduce false positives below. */
2085 offloaded = rcu_rdp_is_offloaded(rdp);
2086 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
2087 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2088 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2089 trace_rcu_grace_period(rcu_state.name,
2093 WRITE_ONCE(rcu_state.gp_flags,
2094 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2096 raw_spin_unlock_irq_rcu_node(rnp);
2098 // If strict, make all CPUs aware of the end of the old grace period.
2099 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2100 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
2104 * Body of kthread that handles grace periods.
2106 static int __noreturn rcu_gp_kthread(void *unused)
2108 rcu_bind_gp_kthread();
2111 /* Handle grace-period start. */
2113 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2115 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
2116 swait_event_idle_exclusive(rcu_state.gp_wq,
2117 READ_ONCE(rcu_state.gp_flags) &
2119 rcu_gp_torture_wait();
2120 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
2121 /* Locking provides needed memory barrier. */
2124 cond_resched_tasks_rcu_qs();
2125 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2126 WARN_ON(signal_pending(current));
2127 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2131 /* Handle quiescent-state forcing. */
2134 /* Handle grace-period end. */
2135 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
2137 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
2142 * Report a full set of quiescent states to the rcu_state data structure.
2143 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2144 * another grace period is required. Whether we wake the grace-period
2145 * kthread or it awakens itself for the next round of quiescent-state
2146 * forcing, that kthread will clean up after the just-completed grace
2147 * period. Note that the caller must hold rnp->lock, which is released
2150 static void rcu_report_qs_rsp(unsigned long flags)
2151 __releases(rcu_get_root()->lock)
2153 raw_lockdep_assert_held_rcu_node(rcu_get_root());
2154 WARN_ON_ONCE(!rcu_gp_in_progress());
2155 WRITE_ONCE(rcu_state.gp_flags,
2156 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2157 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2158 rcu_gp_kthread_wake();
2162 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2163 * Allows quiescent states for a group of CPUs to be reported at one go
2164 * to the specified rcu_node structure, though all the CPUs in the group
2165 * must be represented by the same rcu_node structure (which need not be a
2166 * leaf rcu_node structure, though it often will be). The gps parameter
2167 * is the grace-period snapshot, which means that the quiescent states
2168 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2169 * must be held upon entry, and it is released before return.
2171 * As a special case, if mask is zero, the bit-already-cleared check is
2172 * disabled. This allows propagating quiescent state due to resumed tasks
2173 * during grace-period initialization.
2175 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2176 unsigned long gps, unsigned long flags)
2177 __releases(rnp->lock)
2179 unsigned long oldmask = 0;
2180 struct rcu_node *rnp_c;
2182 raw_lockdep_assert_held_rcu_node(rnp);
2184 /* Walk up the rcu_node hierarchy. */
2186 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2189 * Our bit has already been cleared, or the
2190 * relevant grace period is already over, so done.
2192 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2195 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2196 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2197 rcu_preempt_blocked_readers_cgp(rnp));
2198 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2199 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2200 mask, rnp->qsmask, rnp->level,
2201 rnp->grplo, rnp->grphi,
2203 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2205 /* Other bits still set at this level, so done. */
2206 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2209 rnp->completedqs = rnp->gp_seq;
2210 mask = rnp->grpmask;
2211 if (rnp->parent == NULL) {
2213 /* No more levels. Exit loop holding root lock. */
2217 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2220 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2221 oldmask = READ_ONCE(rnp_c->qsmask);
2225 * Get here if we are the last CPU to pass through a quiescent
2226 * state for this grace period. Invoke rcu_report_qs_rsp()
2227 * to clean up and start the next grace period if one is needed.
2229 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2233 * Record a quiescent state for all tasks that were previously queued
2234 * on the specified rcu_node structure and that were blocking the current
2235 * RCU grace period. The caller must hold the corresponding rnp->lock with
2236 * irqs disabled, and this lock is released upon return, but irqs remain
2239 static void __maybe_unused
2240 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2241 __releases(rnp->lock)
2245 struct rcu_node *rnp_p;
2247 raw_lockdep_assert_held_rcu_node(rnp);
2248 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2249 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2251 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2252 return; /* Still need more quiescent states! */
2255 rnp->completedqs = rnp->gp_seq;
2256 rnp_p = rnp->parent;
2257 if (rnp_p == NULL) {
2259 * Only one rcu_node structure in the tree, so don't
2260 * try to report up to its nonexistent parent!
2262 rcu_report_qs_rsp(flags);
2266 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2268 mask = rnp->grpmask;
2269 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2270 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2271 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2275 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2276 * structure. This must be called from the specified CPU.
2279 rcu_report_qs_rdp(struct rcu_data *rdp)
2281 unsigned long flags;
2283 bool needwake = false;
2284 const bool offloaded = rcu_rdp_is_offloaded(rdp);
2285 struct rcu_node *rnp;
2287 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2289 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2290 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2294 * The grace period in which this quiescent state was
2295 * recorded has ended, so don't report it upwards.
2296 * We will instead need a new quiescent state that lies
2297 * within the current grace period.
2299 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2300 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2303 mask = rdp->grpmask;
2304 rdp->core_needs_qs = false;
2305 if ((rnp->qsmask & mask) == 0) {
2306 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2309 * This GP can't end until cpu checks in, so all of our
2310 * callbacks can be processed during the next GP.
2313 needwake = rcu_accelerate_cbs(rnp, rdp);
2315 rcu_disable_urgency_upon_qs(rdp);
2316 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2317 /* ^^^ Released rnp->lock */
2319 rcu_gp_kthread_wake();
2324 * Check to see if there is a new grace period of which this CPU
2325 * is not yet aware, and if so, set up local rcu_data state for it.
2326 * Otherwise, see if this CPU has just passed through its first
2327 * quiescent state for this grace period, and record that fact if so.
2330 rcu_check_quiescent_state(struct rcu_data *rdp)
2332 /* Check for grace-period ends and beginnings. */
2333 note_gp_changes(rdp);
2336 * Does this CPU still need to do its part for current grace period?
2337 * If no, return and let the other CPUs do their part as well.
2339 if (!rdp->core_needs_qs)
2343 * Was there a quiescent state since the beginning of the grace
2344 * period? If no, then exit and wait for the next call.
2346 if (rdp->cpu_no_qs.b.norm)
2350 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2353 rcu_report_qs_rdp(rdp);
2357 * Near the end of the offline process. Trace the fact that this CPU
2360 int rcutree_dying_cpu(unsigned int cpu)
2363 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2364 struct rcu_node *rnp = rdp->mynode;
2366 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2369 blkd = !!(rnp->qsmask & rdp->grpmask);
2370 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2371 blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
2376 * All CPUs for the specified rcu_node structure have gone offline,
2377 * and all tasks that were preempted within an RCU read-side critical
2378 * section while running on one of those CPUs have since exited their RCU
2379 * read-side critical section. Some other CPU is reporting this fact with
2380 * the specified rcu_node structure's ->lock held and interrupts disabled.
2381 * This function therefore goes up the tree of rcu_node structures,
2382 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2383 * the leaf rcu_node structure's ->qsmaskinit field has already been
2386 * This function does check that the specified rcu_node structure has
2387 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2388 * prematurely. That said, invoking it after the fact will cost you
2389 * a needless lock acquisition. So once it has done its work, don't
2392 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2395 struct rcu_node *rnp = rnp_leaf;
2397 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2398 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2399 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2400 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2403 mask = rnp->grpmask;
2407 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2408 rnp->qsmaskinit &= ~mask;
2409 /* Between grace periods, so better already be zero! */
2410 WARN_ON_ONCE(rnp->qsmask);
2411 if (rnp->qsmaskinit) {
2412 raw_spin_unlock_rcu_node(rnp);
2413 /* irqs remain disabled. */
2416 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2421 * The CPU has been completely removed, and some other CPU is reporting
2422 * this fact from process context. Do the remainder of the cleanup.
2423 * There can only be one CPU hotplug operation at a time, so no need for
2426 int rcutree_dead_cpu(unsigned int cpu)
2428 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2429 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2431 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2434 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
2435 /* Adjust any no-longer-needed kthreads. */
2436 rcu_boost_kthread_setaffinity(rnp, -1);
2437 // Stop-machine done, so allow nohz_full to disable tick.
2438 tick_dep_clear(TICK_DEP_BIT_RCU);
2443 * Invoke any RCU callbacks that have made it to the end of their grace
2444 * period. Throttle as specified by rdp->blimit.
2446 static void rcu_do_batch(struct rcu_data *rdp)
2449 bool __maybe_unused empty;
2450 unsigned long flags;
2451 const bool offloaded = rcu_rdp_is_offloaded(rdp);
2452 struct rcu_head *rhp;
2453 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2455 long pending, tlimit = 0;
2457 /* If no callbacks are ready, just return. */
2458 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2459 trace_rcu_batch_start(rcu_state.name,
2460 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2461 trace_rcu_batch_end(rcu_state.name, 0,
2462 !rcu_segcblist_empty(&rdp->cblist),
2463 need_resched(), is_idle_task(current),
2464 rcu_is_callbacks_kthread());
2469 * Extract the list of ready callbacks, disabling to prevent
2470 * races with call_rcu() from interrupt handlers. Leave the
2471 * callback counts, as rcu_barrier() needs to be conservative.
2473 local_irq_save(flags);
2475 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2476 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2477 div = READ_ONCE(rcu_divisor);
2478 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2479 bl = max(rdp->blimit, pending >> div);
2480 if (in_serving_softirq() && unlikely(bl > 100)) {
2481 long rrn = READ_ONCE(rcu_resched_ns);
2483 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2484 tlimit = local_clock() + rrn;
2486 trace_rcu_batch_start(rcu_state.name,
2487 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2488 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2490 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2492 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2493 rcu_nocb_unlock_irqrestore(rdp, flags);
2495 /* Invoke callbacks. */
2496 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2497 rhp = rcu_cblist_dequeue(&rcl);
2499 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2503 debug_rcu_head_unqueue(rhp);
2505 rcu_lock_acquire(&rcu_callback_map);
2506 trace_rcu_invoke_callback(rcu_state.name, rhp);
2509 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2512 rcu_lock_release(&rcu_callback_map);
2515 * Stop only if limit reached and CPU has something to do.
2517 if (in_serving_softirq()) {
2518 if (count >= bl && (need_resched() ||
2519 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2523 * Make sure we don't spend too much time here and deprive other
2524 * softirq vectors of CPU cycles.
2526 if (unlikely(tlimit)) {
2527 /* only call local_clock() every 32 callbacks */
2528 if (likely((count & 31) || local_clock() < tlimit))
2530 /* Exceeded the time limit, so leave. */
2535 lockdep_assert_irqs_enabled();
2536 cond_resched_tasks_rcu_qs();
2537 lockdep_assert_irqs_enabled();
2542 local_irq_save(flags);
2544 rdp->n_cbs_invoked += count;
2545 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2546 is_idle_task(current), rcu_is_callbacks_kthread());
2548 /* Update counts and requeue any remaining callbacks. */
2549 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2550 rcu_segcblist_add_len(&rdp->cblist, -count);
2552 /* Reinstate batch limit if we have worked down the excess. */
2553 count = rcu_segcblist_n_cbs(&rdp->cblist);
2554 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2555 rdp->blimit = blimit;
2557 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2558 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2559 rdp->qlen_last_fqs_check = 0;
2560 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2561 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2562 rdp->qlen_last_fqs_check = count;
2565 * The following usually indicates a double call_rcu(). To track
2566 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2568 empty = rcu_segcblist_empty(&rdp->cblist);
2569 WARN_ON_ONCE(count == 0 && !empty);
2570 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2571 count != 0 && empty);
2572 WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2573 WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2575 rcu_nocb_unlock_irqrestore(rdp, flags);
2577 /* Re-invoke RCU core processing if there are callbacks remaining. */
2578 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2580 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2584 * This function is invoked from each scheduling-clock interrupt,
2585 * and checks to see if this CPU is in a non-context-switch quiescent
2586 * state, for example, user mode or idle loop. It also schedules RCU
2587 * core processing. If the current grace period has gone on too long,
2588 * it will ask the scheduler to manufacture a context switch for the sole
2589 * purpose of providing the needed quiescent state.
2591 void rcu_sched_clock_irq(int user)
2593 trace_rcu_utilization(TPS("Start scheduler-tick"));
2594 lockdep_assert_irqs_disabled();
2595 raw_cpu_inc(rcu_data.ticks_this_gp);
2596 /* The load-acquire pairs with the store-release setting to true. */
2597 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2598 /* Idle and userspace execution already are quiescent states. */
2599 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2600 set_tsk_need_resched(current);
2601 set_preempt_need_resched();
2603 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2605 rcu_flavor_sched_clock_irq(user);
2606 if (rcu_pending(user))
2608 lockdep_assert_irqs_disabled();
2610 trace_rcu_utilization(TPS("End scheduler-tick"));
2614 * Scan the leaf rcu_node structures. For each structure on which all
2615 * CPUs have reported a quiescent state and on which there are tasks
2616 * blocking the current grace period, initiate RCU priority boosting.
2617 * Otherwise, invoke the specified function to check dyntick state for
2618 * each CPU that has not yet reported a quiescent state.
2620 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2623 unsigned long flags;
2625 struct rcu_data *rdp;
2626 struct rcu_node *rnp;
2628 rcu_state.cbovld = rcu_state.cbovldnext;
2629 rcu_state.cbovldnext = false;
2630 rcu_for_each_leaf_node(rnp) {
2631 cond_resched_tasks_rcu_qs();
2633 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2634 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2635 if (rnp->qsmask == 0) {
2636 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2638 * No point in scanning bits because they
2639 * are all zero. But we might need to
2640 * priority-boost blocked readers.
2642 rcu_initiate_boost(rnp, flags);
2643 /* rcu_initiate_boost() releases rnp->lock */
2646 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2649 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2650 rdp = per_cpu_ptr(&rcu_data, cpu);
2652 mask |= rdp->grpmask;
2653 rcu_disable_urgency_upon_qs(rdp);
2657 /* Idle/offline CPUs, report (releases rnp->lock). */
2658 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2660 /* Nothing to do here, so just drop the lock. */
2661 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2667 * Force quiescent states on reluctant CPUs, and also detect which
2668 * CPUs are in dyntick-idle mode.
2670 void rcu_force_quiescent_state(void)
2672 unsigned long flags;
2674 struct rcu_node *rnp;
2675 struct rcu_node *rnp_old = NULL;
2677 /* Funnel through hierarchy to reduce memory contention. */
2678 rnp = raw_cpu_read(rcu_data.mynode);
2679 for (; rnp != NULL; rnp = rnp->parent) {
2680 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2681 !raw_spin_trylock(&rnp->fqslock);
2682 if (rnp_old != NULL)
2683 raw_spin_unlock(&rnp_old->fqslock);
2688 /* rnp_old == rcu_get_root(), rnp == NULL. */
2690 /* Reached the root of the rcu_node tree, acquire lock. */
2691 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2692 raw_spin_unlock(&rnp_old->fqslock);
2693 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2694 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2695 return; /* Someone beat us to it. */
2697 WRITE_ONCE(rcu_state.gp_flags,
2698 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2699 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2700 rcu_gp_kthread_wake();
2702 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2704 // Workqueue handler for an RCU reader for kernels enforcing struct RCU
2706 static void strict_work_handler(struct work_struct *work)
2712 /* Perform RCU core processing work for the current CPU. */
2713 static __latent_entropy void rcu_core(void)
2715 unsigned long flags;
2716 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2717 struct rcu_node *rnp = rdp->mynode;
2718 const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2720 if (cpu_is_offline(smp_processor_id()))
2722 trace_rcu_utilization(TPS("Start RCU core"));
2723 WARN_ON_ONCE(!rdp->beenonline);
2725 /* Report any deferred quiescent states if preemption enabled. */
2726 if (!(preempt_count() & PREEMPT_MASK)) {
2727 rcu_preempt_deferred_qs(current);
2728 } else if (rcu_preempt_need_deferred_qs(current)) {
2729 set_tsk_need_resched(current);
2730 set_preempt_need_resched();
2733 /* Update RCU state based on any recent quiescent states. */
2734 rcu_check_quiescent_state(rdp);
2736 /* No grace period and unregistered callbacks? */
2737 if (!rcu_gp_in_progress() &&
2738 rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2739 rcu_nocb_lock_irqsave(rdp, flags);
2740 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2741 rcu_accelerate_cbs_unlocked(rnp, rdp);
2742 rcu_nocb_unlock_irqrestore(rdp, flags);
2745 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2747 /* If there are callbacks ready, invoke them. */
2748 if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2749 likely(READ_ONCE(rcu_scheduler_fully_active)))
2752 /* Do any needed deferred wakeups of rcuo kthreads. */
2753 do_nocb_deferred_wakeup(rdp);
2754 trace_rcu_utilization(TPS("End RCU core"));
2756 // If strict GPs, schedule an RCU reader in a clean environment.
2757 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2758 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2761 static void rcu_core_si(struct softirq_action *h)
2766 static void rcu_wake_cond(struct task_struct *t, int status)
2769 * If the thread is yielding, only wake it when this
2770 * is invoked from idle
2772 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2776 static void invoke_rcu_core_kthread(void)
2778 struct task_struct *t;
2779 unsigned long flags;
2781 local_irq_save(flags);
2782 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2783 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2784 if (t != NULL && t != current)
2785 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2786 local_irq_restore(flags);
2790 * Wake up this CPU's rcuc kthread to do RCU core processing.
2792 static void invoke_rcu_core(void)
2794 if (!cpu_online(smp_processor_id()))
2797 raise_softirq(RCU_SOFTIRQ);
2799 invoke_rcu_core_kthread();
2802 static void rcu_cpu_kthread_park(unsigned int cpu)
2804 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2807 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2809 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2813 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2814 * the RCU softirq used in configurations of RCU that do not support RCU
2815 * priority boosting.
2817 static void rcu_cpu_kthread(unsigned int cpu)
2819 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2820 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2823 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2824 for (spincnt = 0; spincnt < 10; spincnt++) {
2826 *statusp = RCU_KTHREAD_RUNNING;
2827 local_irq_disable();
2835 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2836 *statusp = RCU_KTHREAD_WAITING;
2840 *statusp = RCU_KTHREAD_YIELDING;
2841 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2842 schedule_timeout_idle(2);
2843 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2844 *statusp = RCU_KTHREAD_WAITING;
2847 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2848 .store = &rcu_data.rcu_cpu_kthread_task,
2849 .thread_should_run = rcu_cpu_kthread_should_run,
2850 .thread_fn = rcu_cpu_kthread,
2851 .thread_comm = "rcuc/%u",
2852 .setup = rcu_cpu_kthread_setup,
2853 .park = rcu_cpu_kthread_park,
2857 * Spawn per-CPU RCU core processing kthreads.
2859 static int __init rcu_spawn_core_kthreads(void)
2863 for_each_possible_cpu(cpu)
2864 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2865 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2867 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2868 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2873 * Handle any core-RCU processing required by a call_rcu() invocation.
2875 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2876 unsigned long flags)
2879 * If called from an extended quiescent state, invoke the RCU
2880 * core in order to force a re-evaluation of RCU's idleness.
2882 if (!rcu_is_watching())
2885 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2886 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2890 * Force the grace period if too many callbacks or too long waiting.
2891 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2892 * if some other CPU has recently done so. Also, don't bother
2893 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2894 * is the only one waiting for a grace period to complete.
2896 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2897 rdp->qlen_last_fqs_check + qhimark)) {
2899 /* Are we ignoring a completed grace period? */
2900 note_gp_changes(rdp);
2902 /* Start a new grace period if one not already started. */
2903 if (!rcu_gp_in_progress()) {
2904 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2906 /* Give the grace period a kick. */
2907 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2908 if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2909 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2910 rcu_force_quiescent_state();
2911 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2912 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2918 * RCU callback function to leak a callback.
2920 static void rcu_leak_callback(struct rcu_head *rhp)
2925 * Check and if necessary update the leaf rcu_node structure's
2926 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2927 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2928 * structure's ->lock.
2930 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2932 raw_lockdep_assert_held_rcu_node(rnp);
2933 if (qovld_calc <= 0)
2934 return; // Early boot and wildcard value set.
2935 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2936 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2938 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2942 * Check and if necessary update the leaf rcu_node structure's
2943 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2944 * number of queued RCU callbacks. No locks need be held, but the
2945 * caller must have disabled interrupts.
2947 * Note that this function ignores the possibility that there are a lot
2948 * of callbacks all of which have already seen the end of their respective
2949 * grace periods. This omission is due to the need for no-CBs CPUs to
2950 * be holding ->nocb_lock to do this check, which is too heavy for a
2951 * common-case operation.
2953 static void check_cb_ovld(struct rcu_data *rdp)
2955 struct rcu_node *const rnp = rdp->mynode;
2957 if (qovld_calc <= 0 ||
2958 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2959 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2960 return; // Early boot wildcard value or already set correctly.
2961 raw_spin_lock_rcu_node(rnp);
2962 check_cb_ovld_locked(rdp, rnp);
2963 raw_spin_unlock_rcu_node(rnp);
2966 /* Helper function for call_rcu() and friends. */
2968 __call_rcu(struct rcu_head *head, rcu_callback_t func)
2970 static atomic_t doublefrees;
2971 unsigned long flags;
2972 struct rcu_data *rdp;
2975 /* Misaligned rcu_head! */
2976 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2978 if (debug_rcu_head_queue(head)) {
2980 * Probable double call_rcu(), so leak the callback.
2981 * Use rcu:rcu_callback trace event to find the previous
2982 * time callback was passed to __call_rcu().
2984 if (atomic_inc_return(&doublefrees) < 4) {
2985 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
2988 WRITE_ONCE(head->func, rcu_leak_callback);
2993 local_irq_save(flags);
2994 kasan_record_aux_stack(head);
2995 rdp = this_cpu_ptr(&rcu_data);
2997 /* Add the callback to our list. */
2998 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2999 // This can trigger due to call_rcu() from offline CPU:
3000 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
3001 WARN_ON_ONCE(!rcu_is_watching());
3002 // Very early boot, before rcu_init(). Initialize if needed
3003 // and then drop through to queue the callback.
3004 if (rcu_segcblist_empty(&rdp->cblist))
3005 rcu_segcblist_init(&rdp->cblist);
3009 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
3010 return; // Enqueued onto ->nocb_bypass, so just leave.
3011 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
3012 rcu_segcblist_enqueue(&rdp->cblist, head);
3013 if (__is_kvfree_rcu_offset((unsigned long)func))
3014 trace_rcu_kvfree_callback(rcu_state.name, head,
3015 (unsigned long)func,
3016 rcu_segcblist_n_cbs(&rdp->cblist));
3018 trace_rcu_callback(rcu_state.name, head,
3019 rcu_segcblist_n_cbs(&rdp->cblist));
3021 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
3023 /* Go handle any RCU core processing required. */
3024 if (unlikely(rcu_rdp_is_offloaded(rdp))) {
3025 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
3027 __call_rcu_core(rdp, head, flags);
3028 local_irq_restore(flags);
3033 * call_rcu() - Queue an RCU callback for invocation after a grace period.
3034 * @head: structure to be used for queueing the RCU updates.
3035 * @func: actual callback function to be invoked after the grace period
3037 * The callback function will be invoked some time after a full grace
3038 * period elapses, in other words after all pre-existing RCU read-side
3039 * critical sections have completed. However, the callback function
3040 * might well execute concurrently with RCU read-side critical sections
3041 * that started after call_rcu() was invoked.
3043 * RCU read-side critical sections are delimited by rcu_read_lock()
3044 * and rcu_read_unlock(), and may be nested. In addition, but only in
3045 * v5.0 and later, regions of code across which interrupts, preemption,
3046 * or softirqs have been disabled also serve as RCU read-side critical
3047 * sections. This includes hardware interrupt handlers, softirq handlers,
3050 * Note that all CPUs must agree that the grace period extended beyond
3051 * all pre-existing RCU read-side critical section. On systems with more
3052 * than one CPU, this means that when "func()" is invoked, each CPU is
3053 * guaranteed to have executed a full memory barrier since the end of its
3054 * last RCU read-side critical section whose beginning preceded the call
3055 * to call_rcu(). It also means that each CPU executing an RCU read-side
3056 * critical section that continues beyond the start of "func()" must have
3057 * executed a memory barrier after the call_rcu() but before the beginning
3058 * of that RCU read-side critical section. Note that these guarantees
3059 * include CPUs that are offline, idle, or executing in user mode, as
3060 * well as CPUs that are executing in the kernel.
3062 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
3063 * resulting RCU callback function "func()", then both CPU A and CPU B are
3064 * guaranteed to execute a full memory barrier during the time interval
3065 * between the call to call_rcu() and the invocation of "func()" -- even
3066 * if CPU A and CPU B are the same CPU (but again only if the system has
3067 * more than one CPU).
3069 * Implementation of these memory-ordering guarantees is described here:
3070 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3072 void call_rcu(struct rcu_head *head, rcu_callback_t func)
3074 __call_rcu(head, func);
3076 EXPORT_SYMBOL_GPL(call_rcu);
3079 /* Maximum number of jiffies to wait before draining a batch. */
3080 #define KFREE_DRAIN_JIFFIES (HZ / 50)
3081 #define KFREE_N_BATCHES 2
3082 #define FREE_N_CHANNELS 2
3085 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
3086 * @nr_records: Number of active pointers in the array
3087 * @next: Next bulk object in the block chain
3088 * @records: Array of the kvfree_rcu() pointers
3090 struct kvfree_rcu_bulk_data {
3091 unsigned long nr_records;
3092 struct kvfree_rcu_bulk_data *next;
3097 * This macro defines how many entries the "records" array
3098 * will contain. It is based on the fact that the size of
3099 * kvfree_rcu_bulk_data structure becomes exactly one page.
3101 #define KVFREE_BULK_MAX_ENTR \
3102 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
3105 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3106 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3107 * @head_free: List of kfree_rcu() objects waiting for a grace period
3108 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
3109 * @krcp: Pointer to @kfree_rcu_cpu structure
3112 struct kfree_rcu_cpu_work {
3113 struct rcu_work rcu_work;
3114 struct rcu_head *head_free;
3115 struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
3116 struct kfree_rcu_cpu *krcp;
3120 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3121 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3122 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
3123 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3124 * @lock: Synchronize access to this structure
3125 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3126 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
3127 * @initialized: The @rcu_work fields have been initialized
3128 * @count: Number of objects for which GP not started
3130 * A simple cache list that contains objects for reuse purpose.
3131 * In order to save some per-cpu space the list is singular.
3132 * Even though it is lockless an access has to be protected by the
3134 * @page_cache_work: A work to refill the cache when it is empty
3135 * @backoff_page_cache_fill: Delay cache refills
3136 * @work_in_progress: Indicates that page_cache_work is running
3137 * @hrtimer: A hrtimer for scheduling a page_cache_work
3138 * @nr_bkv_objs: number of allocated objects at @bkvcache.
3140 * This is a per-CPU structure. The reason that it is not included in
3141 * the rcu_data structure is to permit this code to be extracted from
3142 * the RCU files. Such extraction could allow further optimization of
3143 * the interactions with the slab allocators.
3145 struct kfree_rcu_cpu {
3146 struct rcu_head *head;
3147 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
3148 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3149 raw_spinlock_t lock;
3150 struct delayed_work monitor_work;
3155 struct delayed_work page_cache_work;
3156 atomic_t backoff_page_cache_fill;
3157 atomic_t work_in_progress;
3158 struct hrtimer hrtimer;
3160 struct llist_head bkvcache;
3164 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3165 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3168 static __always_inline void
3169 debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
3171 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3174 for (i = 0; i < bhead->nr_records; i++)
3175 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3179 static inline struct kfree_rcu_cpu *
3180 krc_this_cpu_lock(unsigned long *flags)
3182 struct kfree_rcu_cpu *krcp;
3184 local_irq_save(*flags); // For safely calling this_cpu_ptr().
3185 krcp = this_cpu_ptr(&krc);
3186 raw_spin_lock(&krcp->lock);
3192 krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3194 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3197 static inline struct kvfree_rcu_bulk_data *
3198 get_cached_bnode(struct kfree_rcu_cpu *krcp)
3200 if (!krcp->nr_bkv_objs)
3203 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
3204 return (struct kvfree_rcu_bulk_data *)
3205 llist_del_first(&krcp->bkvcache);
3209 put_cached_bnode(struct kfree_rcu_cpu *krcp,
3210 struct kvfree_rcu_bulk_data *bnode)
3213 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3216 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3217 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
3222 drain_page_cache(struct kfree_rcu_cpu *krcp)
3224 unsigned long flags;
3225 struct llist_node *page_list, *pos, *n;
3228 raw_spin_lock_irqsave(&krcp->lock, flags);
3229 page_list = llist_del_all(&krcp->bkvcache);
3230 WRITE_ONCE(krcp->nr_bkv_objs, 0);
3231 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3233 llist_for_each_safe(pos, n, page_list) {
3234 free_page((unsigned long)pos);
3242 * This function is invoked in workqueue context after a grace period.
3243 * It frees all the objects queued on ->bkvhead_free or ->head_free.
3245 static void kfree_rcu_work(struct work_struct *work)
3247 unsigned long flags;
3248 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3249 struct rcu_head *head, *next;
3250 struct kfree_rcu_cpu *krcp;
3251 struct kfree_rcu_cpu_work *krwp;
3254 krwp = container_of(to_rcu_work(work),
3255 struct kfree_rcu_cpu_work, rcu_work);
3258 raw_spin_lock_irqsave(&krcp->lock, flags);
3259 // Channels 1 and 2.
3260 for (i = 0; i < FREE_N_CHANNELS; i++) {
3261 bkvhead[i] = krwp->bkvhead_free[i];
3262 krwp->bkvhead_free[i] = NULL;
3266 head = krwp->head_free;
3267 krwp->head_free = NULL;
3268 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3270 // Handle the first two channels.
3271 for (i = 0; i < FREE_N_CHANNELS; i++) {
3272 for (; bkvhead[i]; bkvhead[i] = bnext) {
3273 bnext = bkvhead[i]->next;
3274 debug_rcu_bhead_unqueue(bkvhead[i]);
3276 rcu_lock_acquire(&rcu_callback_map);
3277 if (i == 0) { // kmalloc() / kfree().
3278 trace_rcu_invoke_kfree_bulk_callback(
3279 rcu_state.name, bkvhead[i]->nr_records,
3280 bkvhead[i]->records);
3282 kfree_bulk(bkvhead[i]->nr_records,
3283 bkvhead[i]->records);
3284 } else { // vmalloc() / vfree().
3285 for (j = 0; j < bkvhead[i]->nr_records; j++) {
3286 trace_rcu_invoke_kvfree_callback(
3288 bkvhead[i]->records[j], 0);
3290 vfree(bkvhead[i]->records[j]);
3293 rcu_lock_release(&rcu_callback_map);
3295 raw_spin_lock_irqsave(&krcp->lock, flags);
3296 if (put_cached_bnode(krcp, bkvhead[i]))
3298 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3301 free_page((unsigned long) bkvhead[i]);
3303 cond_resched_tasks_rcu_qs();
3308 * This is used when the "bulk" path can not be used for the
3309 * double-argument of kvfree_rcu(). This happens when the
3310 * page-cache is empty, which means that objects are instead
3311 * queued on a linked list through their rcu_head structures.
3312 * This list is named "Channel 3".
3314 for (; head; head = next) {
3315 unsigned long offset = (unsigned long)head->func;
3316 void *ptr = (void *)head - offset;
3319 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3320 rcu_lock_acquire(&rcu_callback_map);
3321 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3323 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3326 rcu_lock_release(&rcu_callback_map);
3327 cond_resched_tasks_rcu_qs();
3332 need_offload_krc(struct kfree_rcu_cpu *krcp)
3336 for (i = 0; i < FREE_N_CHANNELS; i++)
3337 if (krcp->bkvhead[i])
3340 return !!krcp->head;
3344 need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp)
3348 for (i = 0; i < FREE_N_CHANNELS; i++)
3349 if (krwp->bkvhead_free[i])
3352 return !!krwp->head_free;
3356 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3358 static void kfree_rcu_monitor(struct work_struct *work)
3360 struct kfree_rcu_cpu *krcp = container_of(work,
3361 struct kfree_rcu_cpu, monitor_work.work);
3362 unsigned long flags;
3365 raw_spin_lock_irqsave(&krcp->lock, flags);
3367 // Attempt to start a new batch.
3368 for (i = 0; i < KFREE_N_BATCHES; i++) {
3369 struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3371 // Try to detach bulk_head or head and attach it, only when
3372 // all channels are free. Any channel is not free means at krwp
3373 // there is on-going rcu work to handle krwp's free business.
3374 if (need_wait_for_krwp_work(krwp))
3377 if (need_offload_krc(krcp)) {
3378 // Channel 1 corresponds to the SLAB-pointer bulk path.
3379 // Channel 2 corresponds to vmalloc-pointer bulk path.
3380 for (j = 0; j < FREE_N_CHANNELS; j++) {
3381 if (!krwp->bkvhead_free[j]) {
3382 krwp->bkvhead_free[j] = krcp->bkvhead[j];
3383 krcp->bkvhead[j] = NULL;
3387 // Channel 3 corresponds to both SLAB and vmalloc
3388 // objects queued on the linked list.
3389 if (!krwp->head_free) {
3390 krwp->head_free = krcp->head;
3394 WRITE_ONCE(krcp->count, 0);
3396 // One work is per one batch, so there are three
3397 // "free channels", the batch can handle. It can
3398 // be that the work is in the pending state when
3399 // channels have been detached following by each
3401 queue_rcu_work(system_wq, &krwp->rcu_work);
3405 // If there is nothing to detach, it means that our job is
3406 // successfully done here. In case of having at least one
3407 // of the channels that is still busy we should rearm the
3408 // work to repeat an attempt. Because previous batches are
3409 // still in progress.
3410 if (!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head)
3411 krcp->monitor_todo = false;
3413 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3415 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3418 static enum hrtimer_restart
3419 schedule_page_work_fn(struct hrtimer *t)
3421 struct kfree_rcu_cpu *krcp =
3422 container_of(t, struct kfree_rcu_cpu, hrtimer);
3424 queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3425 return HRTIMER_NORESTART;
3428 static void fill_page_cache_func(struct work_struct *work)
3430 struct kvfree_rcu_bulk_data *bnode;
3431 struct kfree_rcu_cpu *krcp =
3432 container_of(work, struct kfree_rcu_cpu,
3433 page_cache_work.work);
3434 unsigned long flags;
3439 nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3440 1 : rcu_min_cached_objs;
3442 for (i = 0; i < nr_pages; i++) {
3443 bnode = (struct kvfree_rcu_bulk_data *)
3444 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3449 raw_spin_lock_irqsave(&krcp->lock, flags);
3450 pushed = put_cached_bnode(krcp, bnode);
3451 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3454 free_page((unsigned long) bnode);
3459 atomic_set(&krcp->work_in_progress, 0);
3460 atomic_set(&krcp->backoff_page_cache_fill, 0);
3464 run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3466 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3467 !atomic_xchg(&krcp->work_in_progress, 1)) {
3468 if (atomic_read(&krcp->backoff_page_cache_fill)) {
3469 queue_delayed_work(system_wq,
3470 &krcp->page_cache_work,
3471 msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3473 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3474 krcp->hrtimer.function = schedule_page_work_fn;
3475 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3480 // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3481 // state specified by flags. If can_alloc is true, the caller must
3482 // be schedulable and not be holding any locks or mutexes that might be
3483 // acquired by the memory allocator or anything that it might invoke.
3484 // Returns true if ptr was successfully recorded, else the caller must
3487 add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3488 unsigned long *flags, void *ptr, bool can_alloc)
3490 struct kvfree_rcu_bulk_data *bnode;
3493 *krcp = krc_this_cpu_lock(flags);
3494 if (unlikely(!(*krcp)->initialized))
3497 idx = !!is_vmalloc_addr(ptr);
3499 /* Check if a new block is required. */
3500 if (!(*krcp)->bkvhead[idx] ||
3501 (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3502 bnode = get_cached_bnode(*krcp);
3503 if (!bnode && can_alloc) {
3504 krc_this_cpu_unlock(*krcp, *flags);
3506 // __GFP_NORETRY - allows a light-weight direct reclaim
3507 // what is OK from minimizing of fallback hitting point of
3508 // view. Apart of that it forbids any OOM invoking what is
3509 // also beneficial since we are about to release memory soon.
3511 // __GFP_NOMEMALLOC - prevents from consuming of all the
3512 // memory reserves. Please note we have a fallback path.
3514 // __GFP_NOWARN - it is supposed that an allocation can
3515 // be failed under low memory or high memory pressure
3517 bnode = (struct kvfree_rcu_bulk_data *)
3518 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3519 *krcp = krc_this_cpu_lock(flags);
3525 /* Initialize the new block. */
3526 bnode->nr_records = 0;
3527 bnode->next = (*krcp)->bkvhead[idx];
3529 /* Attach it to the head. */
3530 (*krcp)->bkvhead[idx] = bnode;
3533 /* Finally insert. */
3534 (*krcp)->bkvhead[idx]->records
3535 [(*krcp)->bkvhead[idx]->nr_records++] = ptr;
3541 * Queue a request for lazy invocation of the appropriate free routine
3542 * after a grace period. Please note that three paths are maintained,
3543 * two for the common case using arrays of pointers and a third one that
3544 * is used only when the main paths cannot be used, for example, due to
3547 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3548 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3549 * be free'd in workqueue context. This allows us to: batch requests together to
3550 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3552 void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3554 unsigned long flags;
3555 struct kfree_rcu_cpu *krcp;
3560 ptr = (void *) head - (unsigned long) func;
3563 * Please note there is a limitation for the head-less
3564 * variant, that is why there is a clear rule for such
3565 * objects: it can be used from might_sleep() context
3566 * only. For other places please embed an rcu_head to
3570 ptr = (unsigned long *) func;
3573 // Queue the object but don't yet schedule the batch.
3574 if (debug_rcu_head_queue(ptr)) {
3575 // Probable double kfree_rcu(), just leak.
3576 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3579 // Mark as success and leave.
3583 kasan_record_aux_stack(ptr);
3584 success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3586 run_page_cache_worker(krcp);
3589 // Inline if kvfree_rcu(one_arg) call.
3593 head->next = krcp->head;
3598 WRITE_ONCE(krcp->count, krcp->count + 1);
3600 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3601 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3602 !krcp->monitor_todo) {
3603 krcp->monitor_todo = true;
3604 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3608 krc_this_cpu_unlock(krcp, flags);
3611 * Inline kvfree() after synchronize_rcu(). We can do
3612 * it from might_sleep() context only, so the current
3613 * CPU can pass the QS state.
3616 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3621 EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3623 static unsigned long
3624 kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3627 unsigned long count = 0;
3629 /* Snapshot count of all CPUs */
3630 for_each_possible_cpu(cpu) {
3631 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3633 count += READ_ONCE(krcp->count);
3634 count += READ_ONCE(krcp->nr_bkv_objs);
3635 atomic_set(&krcp->backoff_page_cache_fill, 1);
3641 static unsigned long
3642 kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3646 for_each_possible_cpu(cpu) {
3648 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3650 count = krcp->count;
3651 count += drain_page_cache(krcp);
3652 kfree_rcu_monitor(&krcp->monitor_work.work);
3654 sc->nr_to_scan -= count;
3657 if (sc->nr_to_scan <= 0)
3661 return freed == 0 ? SHRINK_STOP : freed;
3664 static struct shrinker kfree_rcu_shrinker = {
3665 .count_objects = kfree_rcu_shrink_count,
3666 .scan_objects = kfree_rcu_shrink_scan,
3668 .seeks = DEFAULT_SEEKS,
3671 void __init kfree_rcu_scheduler_running(void)
3674 unsigned long flags;
3676 for_each_possible_cpu(cpu) {
3677 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3679 raw_spin_lock_irqsave(&krcp->lock, flags);
3680 if ((!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head) ||
3681 krcp->monitor_todo) {
3682 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3685 krcp->monitor_todo = true;
3686 schedule_delayed_work_on(cpu, &krcp->monitor_work,
3687 KFREE_DRAIN_JIFFIES);
3688 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3693 * During early boot, any blocking grace-period wait automatically
3694 * implies a grace period. Later on, this is never the case for PREEMPTION.
3696 * However, because a context switch is a grace period for !PREEMPTION, any
3697 * blocking grace-period wait automatically implies a grace period if
3698 * there is only one CPU online at any point time during execution of
3699 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
3700 * occasionally incorrectly indicate that there are multiple CPUs online
3701 * when there was in fact only one the whole time, as this just adds some
3702 * overhead: RCU still operates correctly.
3704 static int rcu_blocking_is_gp(void)
3708 if (IS_ENABLED(CONFIG_PREEMPTION))
3709 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
3710 might_sleep(); /* Check for RCU read-side critical section. */
3713 * If the rcu_state.n_online_cpus counter is equal to one,
3714 * there is only one CPU, and that CPU sees all prior accesses
3715 * made by any CPU that was online at the time of its access.
3716 * Furthermore, if this counter is equal to one, its value cannot
3717 * change until after the preempt_enable() below.
3719 * Furthermore, if rcu_state.n_online_cpus is equal to one here,
3720 * all later CPUs (both this one and any that come online later
3721 * on) are guaranteed to see all accesses prior to this point
3722 * in the code, without the need for additional memory barriers.
3723 * Those memory barriers are provided by CPU-hotplug code.
3725 ret = READ_ONCE(rcu_state.n_online_cpus) <= 1;
3731 * synchronize_rcu - wait until a grace period has elapsed.
3733 * Control will return to the caller some time after a full grace
3734 * period has elapsed, in other words after all currently executing RCU
3735 * read-side critical sections have completed. Note, however, that
3736 * upon return from synchronize_rcu(), the caller might well be executing
3737 * concurrently with new RCU read-side critical sections that began while
3738 * synchronize_rcu() was waiting.
3740 * RCU read-side critical sections are delimited by rcu_read_lock()
3741 * and rcu_read_unlock(), and may be nested. In addition, but only in
3742 * v5.0 and later, regions of code across which interrupts, preemption,
3743 * or softirqs have been disabled also serve as RCU read-side critical
3744 * sections. This includes hardware interrupt handlers, softirq handlers,
3747 * Note that this guarantee implies further memory-ordering guarantees.
3748 * On systems with more than one CPU, when synchronize_rcu() returns,
3749 * each CPU is guaranteed to have executed a full memory barrier since
3750 * the end of its last RCU read-side critical section whose beginning
3751 * preceded the call to synchronize_rcu(). In addition, each CPU having
3752 * an RCU read-side critical section that extends beyond the return from
3753 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3754 * after the beginning of synchronize_rcu() and before the beginning of
3755 * that RCU read-side critical section. Note that these guarantees include
3756 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3757 * that are executing in the kernel.
3759 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3760 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3761 * to have executed a full memory barrier during the execution of
3762 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3763 * again only if the system has more than one CPU).
3765 * Implementation of these memory-ordering guarantees is described here:
3766 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3768 void synchronize_rcu(void)
3770 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3771 lock_is_held(&rcu_lock_map) ||
3772 lock_is_held(&rcu_sched_lock_map),
3773 "Illegal synchronize_rcu() in RCU read-side critical section");
3774 if (rcu_blocking_is_gp())
3775 return; // Context allows vacuous grace periods.
3776 if (rcu_gp_is_expedited())
3777 synchronize_rcu_expedited();
3779 wait_rcu_gp(call_rcu);
3781 EXPORT_SYMBOL_GPL(synchronize_rcu);
3784 * get_state_synchronize_rcu - Snapshot current RCU state
3786 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3787 * or poll_state_synchronize_rcu() to determine whether or not a full
3788 * grace period has elapsed in the meantime.
3790 unsigned long get_state_synchronize_rcu(void)
3793 * Any prior manipulation of RCU-protected data must happen
3794 * before the load from ->gp_seq.
3797 return rcu_seq_snap(&rcu_state.gp_seq);
3799 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3802 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3804 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3805 * or poll_state_synchronize_rcu() to determine whether or not a full
3806 * grace period has elapsed in the meantime. If the needed grace period
3807 * is not already slated to start, notifies RCU core of the need for that
3810 * Interrupts must be enabled for the case where it is necessary to awaken
3811 * the grace-period kthread.
3813 unsigned long start_poll_synchronize_rcu(void)
3815 unsigned long flags;
3816 unsigned long gp_seq = get_state_synchronize_rcu();
3818 struct rcu_data *rdp;
3819 struct rcu_node *rnp;
3821 lockdep_assert_irqs_enabled();
3822 local_irq_save(flags);
3823 rdp = this_cpu_ptr(&rcu_data);
3825 raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3826 needwake = rcu_start_this_gp(rnp, rdp, gp_seq);
3827 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3829 rcu_gp_kthread_wake();
3832 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3835 * poll_state_synchronize_rcu - Conditionally wait for an RCU grace period
3837 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3839 * If a full RCU grace period has elapsed since the earlier call from
3840 * which oldstate was obtained, return @true, otherwise return @false.
3841 * If @false is returned, it is the caller's responsibility to invoke this
3842 * function later on until it does return @true. Alternatively, the caller
3843 * can explicitly wait for a grace period, for example, by passing @oldstate
3844 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3846 * Yes, this function does not take counter wrap into account.
3847 * But counter wrap is harmless. If the counter wraps, we have waited for
3848 * more than 2 billion grace periods (and way more on a 64-bit system!).
3849 * Those needing to keep oldstate values for very long time periods
3850 * (many hours even on 32-bit systems) should check them occasionally
3851 * and either refresh them or set a flag indicating that the grace period
3854 * This function provides the same memory-ordering guarantees that
3855 * would be provided by a synchronize_rcu() that was invoked at the call
3856 * to the function that provided @oldstate, and that returned at the end
3859 bool poll_state_synchronize_rcu(unsigned long oldstate)
3861 if (rcu_seq_done(&rcu_state.gp_seq, oldstate)) {
3862 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3867 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3870 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3872 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3874 * If a full RCU grace period has elapsed since the earlier call to
3875 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3876 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3878 * Yes, this function does not take counter wrap into account. But
3879 * counter wrap is harmless. If the counter wraps, we have waited for
3880 * more than 2 billion grace periods (and way more on a 64-bit system!),
3881 * so waiting for one additional grace period should be just fine.
3883 * This function provides the same memory-ordering guarantees that
3884 * would be provided by a synchronize_rcu() that was invoked at the call
3885 * to the function that provided @oldstate, and that returned at the end
3888 void cond_synchronize_rcu(unsigned long oldstate)
3890 if (!poll_state_synchronize_rcu(oldstate))
3893 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3896 * Check to see if there is any immediate RCU-related work to be done by
3897 * the current CPU, returning 1 if so and zero otherwise. The checks are
3898 * in order of increasing expense: checks that can be carried out against
3899 * CPU-local state are performed first. However, we must check for CPU
3900 * stalls first, else we might not get a chance.
3902 static int rcu_pending(int user)
3904 bool gp_in_progress;
3905 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3906 struct rcu_node *rnp = rdp->mynode;
3908 lockdep_assert_irqs_disabled();
3910 /* Check for CPU stalls, if enabled. */
3911 check_cpu_stall(rdp);
3913 /* Does this CPU need a deferred NOCB wakeup? */
3914 if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3917 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3918 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3921 /* Is the RCU core waiting for a quiescent state from this CPU? */
3922 gp_in_progress = rcu_gp_in_progress();
3923 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3926 /* Does this CPU have callbacks ready to invoke? */
3927 if (!rcu_rdp_is_offloaded(rdp) &&
3928 rcu_segcblist_ready_cbs(&rdp->cblist))
3931 /* Has RCU gone idle with this CPU needing another grace period? */
3932 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3933 !rcu_rdp_is_offloaded(rdp) &&
3934 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3937 /* Have RCU grace period completed or started? */
3938 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3939 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3947 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3948 * the compiler is expected to optimize this away.
3950 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3952 trace_rcu_barrier(rcu_state.name, s, cpu,
3953 atomic_read(&rcu_state.barrier_cpu_count), done);
3957 * RCU callback function for rcu_barrier(). If we are last, wake
3958 * up the task executing rcu_barrier().
3960 * Note that the value of rcu_state.barrier_sequence must be captured
3961 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3962 * other CPUs might count the value down to zero before this CPU gets
3963 * around to invoking rcu_barrier_trace(), which might result in bogus
3964 * data from the next instance of rcu_barrier().
3966 static void rcu_barrier_callback(struct rcu_head *rhp)
3968 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3970 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3971 rcu_barrier_trace(TPS("LastCB"), -1, s);
3972 complete(&rcu_state.barrier_completion);
3974 rcu_barrier_trace(TPS("CB"), -1, s);
3979 * Called with preemption disabled, and from cross-cpu IRQ context.
3981 static void rcu_barrier_func(void *cpu_in)
3983 uintptr_t cpu = (uintptr_t)cpu_in;
3984 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3986 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3987 rdp->barrier_head.func = rcu_barrier_callback;
3988 debug_rcu_head_queue(&rdp->barrier_head);
3990 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
3991 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
3992 atomic_inc(&rcu_state.barrier_cpu_count);
3994 debug_rcu_head_unqueue(&rdp->barrier_head);
3995 rcu_barrier_trace(TPS("IRQNQ"), -1,
3996 rcu_state.barrier_sequence);
3998 rcu_nocb_unlock(rdp);
4002 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
4004 * Note that this primitive does not necessarily wait for an RCU grace period
4005 * to complete. For example, if there are no RCU callbacks queued anywhere
4006 * in the system, then rcu_barrier() is within its rights to return
4007 * immediately, without waiting for anything, much less an RCU grace period.
4009 void rcu_barrier(void)
4012 struct rcu_data *rdp;
4013 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4015 rcu_barrier_trace(TPS("Begin"), -1, s);
4017 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4018 mutex_lock(&rcu_state.barrier_mutex);
4020 /* Did someone else do our work for us? */
4021 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4022 rcu_barrier_trace(TPS("EarlyExit"), -1,
4023 rcu_state.barrier_sequence);
4024 smp_mb(); /* caller's subsequent code after above check. */
4025 mutex_unlock(&rcu_state.barrier_mutex);
4029 /* Mark the start of the barrier operation. */
4030 rcu_seq_start(&rcu_state.barrier_sequence);
4031 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4034 * Initialize the count to two rather than to zero in order
4035 * to avoid a too-soon return to zero in case of an immediate
4036 * invocation of the just-enqueued callback (or preemption of
4037 * this task). Exclude CPU-hotplug operations to ensure that no
4038 * offline non-offloaded CPU has callbacks queued.
4040 init_completion(&rcu_state.barrier_completion);
4041 atomic_set(&rcu_state.barrier_cpu_count, 2);
4045 * Force each CPU with callbacks to register a new callback.
4046 * When that callback is invoked, we will know that all of the
4047 * corresponding CPU's preceding callbacks have been invoked.
4049 for_each_possible_cpu(cpu) {
4050 rdp = per_cpu_ptr(&rcu_data, cpu);
4051 if (cpu_is_offline(cpu) &&
4052 !rcu_rdp_is_offloaded(rdp))
4054 if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
4055 rcu_barrier_trace(TPS("OnlineQ"), cpu,
4056 rcu_state.barrier_sequence);
4057 smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
4058 } else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
4059 cpu_is_offline(cpu)) {
4060 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
4061 rcu_state.barrier_sequence);
4062 local_irq_disable();
4063 rcu_barrier_func((void *)cpu);
4065 } else if (cpu_is_offline(cpu)) {
4066 rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
4067 rcu_state.barrier_sequence);
4069 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
4070 rcu_state.barrier_sequence);
4076 * Now that we have an rcu_barrier_callback() callback on each
4077 * CPU, and thus each counted, remove the initial count.
4079 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4080 complete(&rcu_state.barrier_completion);
4082 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4083 wait_for_completion(&rcu_state.barrier_completion);
4085 /* Mark the end of the barrier operation. */
4086 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4087 rcu_seq_end(&rcu_state.barrier_sequence);
4089 /* Other rcu_barrier() invocations can now safely proceed. */
4090 mutex_unlock(&rcu_state.barrier_mutex);
4092 EXPORT_SYMBOL_GPL(rcu_barrier);
4095 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4096 * first CPU in a given leaf rcu_node structure coming online. The caller
4097 * must hold the corresponding leaf rcu_node ->lock with interrupts
4100 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4104 struct rcu_node *rnp = rnp_leaf;
4106 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4107 WARN_ON_ONCE(rnp->wait_blkd_tasks);
4109 mask = rnp->grpmask;
4113 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4114 oldmask = rnp->qsmaskinit;
4115 rnp->qsmaskinit |= mask;
4116 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4123 * Do boot-time initialization of a CPU's per-CPU RCU data.
4126 rcu_boot_init_percpu_data(int cpu)
4128 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4130 /* Set up local state, ensuring consistent view of global state. */
4131 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4132 INIT_WORK(&rdp->strict_work, strict_work_handler);
4133 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
4134 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
4135 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4136 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4137 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4138 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4140 rcu_boot_init_nocb_percpu_data(rdp);
4144 * Invoked early in the CPU-online process, when pretty much all services
4145 * are available. The incoming CPU is not present.
4147 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4148 * offline event can be happening at a given time. Note also that we can
4149 * accept some slop in the rsp->gp_seq access due to the fact that this
4150 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4151 * And any offloaded callbacks are being numbered elsewhere.
4153 int rcutree_prepare_cpu(unsigned int cpu)
4155 unsigned long flags;
4156 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4157 struct rcu_node *rnp = rcu_get_root();
4159 /* Set up local state, ensuring consistent view of global state. */
4160 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4161 rdp->qlen_last_fqs_check = 0;
4162 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4163 rdp->blimit = blimit;
4164 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
4165 rcu_dynticks_eqs_online();
4166 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4169 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4172 if (!rcu_segcblist_is_enabled(&rdp->cblist))
4173 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4176 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4177 * propagation up the rcu_node tree will happen at the beginning
4178 * of the next grace period.
4181 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4182 rdp->beenonline = true; /* We have now been online. */
4183 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4184 rdp->gp_seq_needed = rdp->gp_seq;
4185 rdp->cpu_no_qs.b.norm = true;
4186 rdp->core_needs_qs = false;
4187 rdp->rcu_iw_pending = false;
4188 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4189 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4190 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4191 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4192 rcu_spawn_one_boost_kthread(rnp);
4193 rcu_spawn_cpu_nocb_kthread(cpu);
4194 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4200 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4202 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4204 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4206 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4210 * Near the end of the CPU-online process. Pretty much all services
4211 * enabled, and the CPU is now very much alive.
4213 int rcutree_online_cpu(unsigned int cpu)
4215 unsigned long flags;
4216 struct rcu_data *rdp;
4217 struct rcu_node *rnp;
4219 rdp = per_cpu_ptr(&rcu_data, cpu);
4221 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4222 rnp->ffmask |= rdp->grpmask;
4223 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4224 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4225 return 0; /* Too early in boot for scheduler work. */
4226 sync_sched_exp_online_cleanup(cpu);
4227 rcutree_affinity_setting(cpu, -1);
4229 // Stop-machine done, so allow nohz_full to disable tick.
4230 tick_dep_clear(TICK_DEP_BIT_RCU);
4235 * Near the beginning of the process. The CPU is still very much alive
4236 * with pretty much all services enabled.
4238 int rcutree_offline_cpu(unsigned int cpu)
4240 unsigned long flags;
4241 struct rcu_data *rdp;
4242 struct rcu_node *rnp;
4244 rdp = per_cpu_ptr(&rcu_data, cpu);
4246 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4247 rnp->ffmask &= ~rdp->grpmask;
4248 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4250 rcutree_affinity_setting(cpu, cpu);
4252 // nohz_full CPUs need the tick for stop-machine to work quickly
4253 tick_dep_set(TICK_DEP_BIT_RCU);
4258 * Mark the specified CPU as being online so that subsequent grace periods
4259 * (both expedited and normal) will wait on it. Note that this means that
4260 * incoming CPUs are not allowed to use RCU read-side critical sections
4261 * until this function is called. Failing to observe this restriction
4262 * will result in lockdep splats.
4264 * Note that this function is special in that it is invoked directly
4265 * from the incoming CPU rather than from the cpuhp_step mechanism.
4266 * This is because this function must be invoked at a precise location.
4268 void rcu_cpu_starting(unsigned int cpu)
4270 unsigned long flags;
4272 struct rcu_data *rdp;
4273 struct rcu_node *rnp;
4276 rdp = per_cpu_ptr(&rcu_data, cpu);
4277 if (rdp->cpu_started)
4279 rdp->cpu_started = true;
4282 mask = rdp->grpmask;
4283 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4284 WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4285 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4286 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4287 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4288 newcpu = !(rnp->expmaskinitnext & mask);
4289 rnp->expmaskinitnext |= mask;
4290 /* Allow lockless access for expedited grace periods. */
4291 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4292 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4293 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4294 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4295 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4297 /* An incoming CPU should never be blocking a grace period. */
4298 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4299 rcu_disable_urgency_upon_qs(rdp);
4300 /* Report QS -after- changing ->qsmaskinitnext! */
4301 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4303 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4305 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4306 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4307 WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4308 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4312 * The outgoing function has no further need of RCU, so remove it from
4313 * the rcu_node tree's ->qsmaskinitnext bit masks.
4315 * Note that this function is special in that it is invoked directly
4316 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4317 * This is because this function must be invoked at a precise location.
4319 void rcu_report_dead(unsigned int cpu)
4321 unsigned long flags;
4323 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4324 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4326 // Do any dangling deferred wakeups.
4327 do_nocb_deferred_wakeup(rdp);
4329 /* QS for any half-done expedited grace period. */
4331 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
4333 rcu_preempt_deferred_qs(current);
4335 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4336 mask = rdp->grpmask;
4337 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4338 WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4339 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4340 raw_spin_lock(&rcu_state.ofl_lock);
4341 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4342 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4343 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4344 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4345 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4346 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4347 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4349 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4350 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4351 raw_spin_unlock(&rcu_state.ofl_lock);
4352 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4353 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4354 WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4356 rdp->cpu_started = false;
4359 #ifdef CONFIG_HOTPLUG_CPU
4361 * The outgoing CPU has just passed through the dying-idle state, and we
4362 * are being invoked from the CPU that was IPIed to continue the offline
4363 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4365 void rcutree_migrate_callbacks(int cpu)
4367 unsigned long flags;
4368 struct rcu_data *my_rdp;
4369 struct rcu_node *my_rnp;
4370 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4373 if (rcu_rdp_is_offloaded(rdp) ||
4374 rcu_segcblist_empty(&rdp->cblist))
4375 return; /* No callbacks to migrate. */
4377 local_irq_save(flags);
4378 my_rdp = this_cpu_ptr(&rcu_data);
4379 my_rnp = my_rdp->mynode;
4380 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4381 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4382 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4383 /* Leverage recent GPs and set GP for new callbacks. */
4384 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4385 rcu_advance_cbs(my_rnp, my_rdp);
4386 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4387 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4388 rcu_segcblist_disable(&rdp->cblist);
4389 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
4390 !rcu_segcblist_n_cbs(&my_rdp->cblist));
4391 if (rcu_rdp_is_offloaded(my_rdp)) {
4392 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4393 __call_rcu_nocb_wake(my_rdp, true, flags);
4395 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4396 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4399 rcu_gp_kthread_wake();
4400 lockdep_assert_irqs_enabled();
4401 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4402 !rcu_segcblist_empty(&rdp->cblist),
4403 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4404 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4405 rcu_segcblist_first_cb(&rdp->cblist));
4410 * On non-huge systems, use expedited RCU grace periods to make suspend
4411 * and hibernation run faster.
4413 static int rcu_pm_notify(struct notifier_block *self,
4414 unsigned long action, void *hcpu)
4417 case PM_HIBERNATION_PREPARE:
4418 case PM_SUSPEND_PREPARE:
4421 case PM_POST_HIBERNATION:
4422 case PM_POST_SUSPEND:
4423 rcu_unexpedite_gp();
4432 * Spawn the kthreads that handle RCU's grace periods.
4434 static int __init rcu_spawn_gp_kthread(void)
4436 unsigned long flags;
4437 int kthread_prio_in = kthread_prio;
4438 struct rcu_node *rnp;
4439 struct sched_param sp;
4440 struct task_struct *t;
4442 /* Force priority into range. */
4443 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4444 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4446 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4448 else if (kthread_prio < 0)
4450 else if (kthread_prio > 99)
4453 if (kthread_prio != kthread_prio_in)
4454 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4455 kthread_prio, kthread_prio_in);
4457 rcu_scheduler_fully_active = 1;
4458 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4459 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4462 sp.sched_priority = kthread_prio;
4463 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4465 rnp = rcu_get_root();
4466 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4467 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4468 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4469 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4470 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4471 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4473 rcu_spawn_nocb_kthreads();
4474 rcu_spawn_boost_kthreads();
4475 rcu_spawn_core_kthreads();
4478 early_initcall(rcu_spawn_gp_kthread);
4481 * This function is invoked towards the end of the scheduler's
4482 * initialization process. Before this is called, the idle task might
4483 * contain synchronous grace-period primitives (during which time, this idle
4484 * task is booting the system, and such primitives are no-ops). After this
4485 * function is called, any synchronous grace-period primitives are run as
4486 * expedited, with the requesting task driving the grace period forward.
4487 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4488 * runtime RCU functionality.
4490 void rcu_scheduler_starting(void)
4492 WARN_ON(num_online_cpus() != 1);
4493 WARN_ON(nr_context_switches() > 0);
4494 rcu_test_sync_prims();
4495 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4496 rcu_test_sync_prims();
4500 * Helper function for rcu_init() that initializes the rcu_state structure.
4502 static void __init rcu_init_one(void)
4504 static const char * const buf[] = RCU_NODE_NAME_INIT;
4505 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4506 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4507 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4509 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4513 struct rcu_node *rnp;
4515 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4517 /* Silence gcc 4.8 false positive about array index out of range. */
4518 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4519 panic("rcu_init_one: rcu_num_lvls out of range");
4521 /* Initialize the level-tracking arrays. */
4523 for (i = 1; i < rcu_num_lvls; i++)
4524 rcu_state.level[i] =
4525 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4526 rcu_init_levelspread(levelspread, num_rcu_lvl);
4528 /* Initialize the elements themselves, starting from the leaves. */
4530 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4531 cpustride *= levelspread[i];
4532 rnp = rcu_state.level[i];
4533 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4534 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4535 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4536 &rcu_node_class[i], buf[i]);
4537 raw_spin_lock_init(&rnp->fqslock);
4538 lockdep_set_class_and_name(&rnp->fqslock,
4539 &rcu_fqs_class[i], fqs[i]);
4540 rnp->gp_seq = rcu_state.gp_seq;
4541 rnp->gp_seq_needed = rcu_state.gp_seq;
4542 rnp->completedqs = rcu_state.gp_seq;
4544 rnp->qsmaskinit = 0;
4545 rnp->grplo = j * cpustride;
4546 rnp->grphi = (j + 1) * cpustride - 1;
4547 if (rnp->grphi >= nr_cpu_ids)
4548 rnp->grphi = nr_cpu_ids - 1;
4554 rnp->grpnum = j % levelspread[i - 1];
4555 rnp->grpmask = BIT(rnp->grpnum);
4556 rnp->parent = rcu_state.level[i - 1] +
4557 j / levelspread[i - 1];
4560 INIT_LIST_HEAD(&rnp->blkd_tasks);
4561 rcu_init_one_nocb(rnp);
4562 init_waitqueue_head(&rnp->exp_wq[0]);
4563 init_waitqueue_head(&rnp->exp_wq[1]);
4564 init_waitqueue_head(&rnp->exp_wq[2]);
4565 init_waitqueue_head(&rnp->exp_wq[3]);
4566 spin_lock_init(&rnp->exp_lock);
4570 init_swait_queue_head(&rcu_state.gp_wq);
4571 init_swait_queue_head(&rcu_state.expedited_wq);
4572 rnp = rcu_first_leaf_node();
4573 for_each_possible_cpu(i) {
4574 while (i > rnp->grphi)
4576 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4577 rcu_boot_init_percpu_data(i);
4582 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4583 * replace the definitions in tree.h because those are needed to size
4584 * the ->node array in the rcu_state structure.
4586 void rcu_init_geometry(void)
4590 static unsigned long old_nr_cpu_ids;
4591 int rcu_capacity[RCU_NUM_LVLS];
4592 static bool initialized;
4596 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4597 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4599 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4603 old_nr_cpu_ids = nr_cpu_ids;
4607 * Initialize any unspecified boot parameters.
4608 * The default values of jiffies_till_first_fqs and
4609 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4610 * value, which is a function of HZ, then adding one for each
4611 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4613 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4614 if (jiffies_till_first_fqs == ULONG_MAX)
4615 jiffies_till_first_fqs = d;
4616 if (jiffies_till_next_fqs == ULONG_MAX)
4617 jiffies_till_next_fqs = d;
4618 adjust_jiffies_till_sched_qs();
4620 /* If the compile-time values are accurate, just leave. */
4621 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4622 nr_cpu_ids == NR_CPUS)
4624 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4625 rcu_fanout_leaf, nr_cpu_ids);
4628 * The boot-time rcu_fanout_leaf parameter must be at least two
4629 * and cannot exceed the number of bits in the rcu_node masks.
4630 * Complain and fall back to the compile-time values if this
4631 * limit is exceeded.
4633 if (rcu_fanout_leaf < 2 ||
4634 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4635 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4641 * Compute number of nodes that can be handled an rcu_node tree
4642 * with the given number of levels.
4644 rcu_capacity[0] = rcu_fanout_leaf;
4645 for (i = 1; i < RCU_NUM_LVLS; i++)
4646 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4649 * The tree must be able to accommodate the configured number of CPUs.
4650 * If this limit is exceeded, fall back to the compile-time values.
4652 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4653 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4658 /* Calculate the number of levels in the tree. */
4659 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4661 rcu_num_lvls = i + 1;
4663 /* Calculate the number of rcu_nodes at each level of the tree. */
4664 for (i = 0; i < rcu_num_lvls; i++) {
4665 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4666 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4669 /* Calculate the total number of rcu_node structures. */
4671 for (i = 0; i < rcu_num_lvls; i++)
4672 rcu_num_nodes += num_rcu_lvl[i];
4676 * Dump out the structure of the rcu_node combining tree associated
4677 * with the rcu_state structure.
4679 static void __init rcu_dump_rcu_node_tree(void)
4682 struct rcu_node *rnp;
4684 pr_info("rcu_node tree layout dump\n");
4686 rcu_for_each_node_breadth_first(rnp) {
4687 if (rnp->level != level) {
4692 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4697 struct workqueue_struct *rcu_gp_wq;
4698 struct workqueue_struct *rcu_par_gp_wq;
4700 static void __init kfree_rcu_batch_init(void)
4705 /* Clamp it to [0:100] seconds interval. */
4706 if (rcu_delay_page_cache_fill_msec < 0 ||
4707 rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
4709 rcu_delay_page_cache_fill_msec =
4710 clamp(rcu_delay_page_cache_fill_msec, 0,
4711 (int) (100 * MSEC_PER_SEC));
4713 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4714 rcu_delay_page_cache_fill_msec);
4717 for_each_possible_cpu(cpu) {
4718 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4720 for (i = 0; i < KFREE_N_BATCHES; i++) {
4721 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4722 krcp->krw_arr[i].krcp = krcp;
4725 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4726 INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
4727 krcp->initialized = true;
4729 if (register_shrinker(&kfree_rcu_shrinker))
4730 pr_err("Failed to register kfree_rcu() shrinker!\n");
4733 void __init rcu_init(void)
4737 rcu_early_boot_tests();
4739 kfree_rcu_batch_init();
4740 rcu_bootup_announce();
4741 rcu_init_geometry();
4744 rcu_dump_rcu_node_tree();
4746 open_softirq(RCU_SOFTIRQ, rcu_core_si);
4749 * We don't need protection against CPU-hotplug here because
4750 * this is called early in boot, before either interrupts
4751 * or the scheduler are operational.
4753 pm_notifier(rcu_pm_notify, 0);
4754 for_each_online_cpu(cpu) {
4755 rcutree_prepare_cpu(cpu);
4756 rcu_cpu_starting(cpu);
4757 rcutree_online_cpu(cpu);
4760 /* Create workqueue for Tree SRCU and for expedited GPs. */
4761 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4762 WARN_ON(!rcu_gp_wq);
4763 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4764 WARN_ON(!rcu_par_gp_wq);
4766 /* Fill in default value for rcutree.qovld boot parameter. */
4767 /* -After- the rcu_node ->lock fields are initialized! */
4769 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4774 #include "tree_stall.h"
4775 #include "tree_exp.h"
4776 #include "tree_nocb.h"
4777 #include "tree_plugin.h"