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 <linux/context_tracking.h>
66 #include "../time/tick-internal.h"
71 #ifdef MODULE_PARAM_PREFIX
72 #undef MODULE_PARAM_PREFIX
74 #define MODULE_PARAM_PREFIX "rcutree."
76 /* Data structures. */
78 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
80 #ifdef CONFIG_RCU_NOCB_CPU
81 .cblist.flags = SEGCBLIST_RCU_CORE,
84 static struct rcu_state rcu_state = {
85 .level = { &rcu_state.node[0] },
86 .gp_state = RCU_GP_IDLE,
87 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
88 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
89 .barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock),
92 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
93 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
94 .ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
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);
157 * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
158 * real-time priority(enabling/disabling) is controlled by
159 * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
161 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
162 module_param(kthread_prio, int, 0444);
164 /* Delay in jiffies for grace-period initialization delays, debug only. */
166 static int gp_preinit_delay;
167 module_param(gp_preinit_delay, int, 0444);
168 static int gp_init_delay;
169 module_param(gp_init_delay, int, 0444);
170 static int gp_cleanup_delay;
171 module_param(gp_cleanup_delay, int, 0444);
173 // Add delay to rcu_read_unlock() for strict grace periods.
174 static int rcu_unlock_delay;
175 #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
176 module_param(rcu_unlock_delay, int, 0444);
180 * This rcu parameter is runtime-read-only. It reflects
181 * a minimum allowed number of objects which can be cached
182 * per-CPU. Object size is equal to one page. This value
183 * can be changed at boot time.
185 static int rcu_min_cached_objs = 5;
186 module_param(rcu_min_cached_objs, int, 0444);
188 // A page shrinker can ask for pages to be freed to make them
189 // available for other parts of the system. This usually happens
190 // under low memory conditions, and in that case we should also
191 // defer page-cache filling for a short time period.
193 // The default value is 5 seconds, which is long enough to reduce
194 // interference with the shrinker while it asks other systems to
195 // drain their caches.
196 static int rcu_delay_page_cache_fill_msec = 5000;
197 module_param(rcu_delay_page_cache_fill_msec, int, 0444);
199 /* Retrieve RCU kthreads priority for rcutorture */
200 int rcu_get_gp_kthreads_prio(void)
204 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
207 * Number of grace periods between delays, normalized by the duration of
208 * the delay. The longer the delay, the more the grace periods between
209 * each delay. The reason for this normalization is that it means that,
210 * for non-zero delays, the overall slowdown of grace periods is constant
211 * regardless of the duration of the delay. This arrangement balances
212 * the need for long delays to increase some race probabilities with the
213 * need for fast grace periods to increase other race probabilities.
215 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
218 * Compute the mask of online CPUs for the specified rcu_node structure.
219 * This will not be stable unless the rcu_node structure's ->lock is
220 * held, but the bit corresponding to the current CPU will be stable
223 static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
225 return READ_ONCE(rnp->qsmaskinitnext);
229 * Is the CPU corresponding to the specified rcu_data structure online
230 * from RCU's perspective? This perspective is given by that structure's
231 * ->qsmaskinitnext field rather than by the global cpu_online_mask.
233 static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
235 return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
239 * Return true if an RCU grace period is in progress. The READ_ONCE()s
240 * permit this function to be invoked without holding the root rcu_node
241 * structure's ->lock, but of course results can be subject to change.
243 static int rcu_gp_in_progress(void)
245 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
249 * Return the number of callbacks queued on the specified CPU.
250 * Handles both the nocbs and normal cases.
252 static long rcu_get_n_cbs_cpu(int cpu)
254 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
256 if (rcu_segcblist_is_enabled(&rdp->cblist))
257 return rcu_segcblist_n_cbs(&rdp->cblist);
261 void rcu_softirq_qs(void)
264 rcu_preempt_deferred_qs(current);
265 rcu_tasks_qs(current, false);
269 * Reset the current CPU's ->dynticks counter to indicate that the
270 * newly onlined CPU is no longer in an extended quiescent state.
271 * This will either leave the counter unchanged, or increment it
272 * to the next non-quiescent value.
274 * The non-atomic test/increment sequence works because the upper bits
275 * of the ->dynticks counter are manipulated only by the corresponding CPU,
276 * or when the corresponding CPU is offline.
278 static void rcu_dynticks_eqs_online(void)
280 if (ct_dynticks() & RCU_DYNTICKS_IDX)
282 ct_state_inc(RCU_DYNTICKS_IDX);
286 * Snapshot the ->dynticks counter with full ordering so as to allow
287 * stable comparison of this counter with past and future snapshots.
289 static int rcu_dynticks_snap(int cpu)
291 smp_mb(); // Fundamental RCU ordering guarantee.
292 return ct_dynticks_cpu_acquire(cpu);
296 * Return true if the snapshot returned from rcu_dynticks_snap()
297 * indicates that RCU is in an extended quiescent state.
299 static bool rcu_dynticks_in_eqs(int snap)
301 return !(snap & RCU_DYNTICKS_IDX);
304 /* Return true if the specified CPU is currently idle from an RCU viewpoint. */
305 bool rcu_is_idle_cpu(int cpu)
307 return rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu));
311 * Return true if the CPU corresponding to the specified rcu_data
312 * structure has spent some time in an extended quiescent state since
313 * rcu_dynticks_snap() returned the specified snapshot.
315 static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
317 return snap != rcu_dynticks_snap(rdp->cpu);
321 * Return true if the referenced integer is zero while the specified
322 * CPU remains within a single extended quiescent state.
324 bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
328 // If not quiescent, force back to earlier extended quiescent state.
329 snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX;
330 smp_rmb(); // Order ->dynticks and *vp reads.
332 return false; // Non-zero, so report failure;
333 smp_rmb(); // Order *vp read and ->dynticks re-read.
335 // If still in the same extended quiescent state, we are good!
336 return snap == ct_dynticks_cpu(cpu);
340 * Let the RCU core know that this CPU has gone through the scheduler,
341 * which is a quiescent state. This is called when the need for a
342 * quiescent state is urgent, so we burn an atomic operation and full
343 * memory barriers to let the RCU core know about it, regardless of what
344 * this CPU might (or might not) do in the near future.
346 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
348 * The caller must have disabled interrupts and must not be idle.
350 notrace void rcu_momentary_dyntick_idle(void)
354 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
355 seq = ct_state_inc(2 * RCU_DYNTICKS_IDX);
356 /* It is illegal to call this from idle state. */
357 WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX));
358 rcu_preempt_deferred_qs(current);
360 EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
363 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
365 * If the current CPU is idle and running at a first-level (not nested)
366 * interrupt, or directly, from idle, return true.
368 * The caller must have at least disabled IRQs.
370 static int rcu_is_cpu_rrupt_from_idle(void)
375 * Usually called from the tick; but also used from smp_function_call()
376 * for expedited grace periods. This latter can result in running from
377 * the idle task, instead of an actual IPI.
379 lockdep_assert_irqs_disabled();
381 /* Check for counter underflows */
382 RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
383 "RCU dynticks_nesting counter underflow!");
384 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
385 "RCU dynticks_nmi_nesting counter underflow/zero!");
387 /* Are we at first interrupt nesting level? */
388 nesting = ct_dynticks_nmi_nesting();
393 * If we're not in an interrupt, we must be in the idle task!
395 WARN_ON_ONCE(!nesting && !is_idle_task(current));
397 /* Does CPU appear to be idle from an RCU standpoint? */
398 return ct_dynticks_nesting() == 0;
401 #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
402 // Maximum callbacks per rcu_do_batch ...
403 #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
404 static long blimit = DEFAULT_RCU_BLIMIT;
405 #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
406 static long qhimark = DEFAULT_RCU_QHIMARK;
407 #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
408 static long qlowmark = DEFAULT_RCU_QLOMARK;
409 #define DEFAULT_RCU_QOVLD_MULT 2
410 #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
411 static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
412 static long qovld_calc = -1; // No pre-initialization lock acquisitions!
414 module_param(blimit, long, 0444);
415 module_param(qhimark, long, 0444);
416 module_param(qlowmark, long, 0444);
417 module_param(qovld, long, 0444);
419 static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
420 static ulong jiffies_till_next_fqs = ULONG_MAX;
421 static bool rcu_kick_kthreads;
422 static int rcu_divisor = 7;
423 module_param(rcu_divisor, int, 0644);
425 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
426 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
427 module_param(rcu_resched_ns, long, 0644);
430 * How long the grace period must be before we start recruiting
431 * quiescent-state help from rcu_note_context_switch().
433 static ulong jiffies_till_sched_qs = ULONG_MAX;
434 module_param(jiffies_till_sched_qs, ulong, 0444);
435 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
436 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
439 * Make sure that we give the grace-period kthread time to detect any
440 * idle CPUs before taking active measures to force quiescent states.
441 * However, don't go below 100 milliseconds, adjusted upwards for really
444 static void adjust_jiffies_till_sched_qs(void)
448 /* If jiffies_till_sched_qs was specified, respect the request. */
449 if (jiffies_till_sched_qs != ULONG_MAX) {
450 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
453 /* Otherwise, set to third fqs scan, but bound below on large system. */
454 j = READ_ONCE(jiffies_till_first_fqs) +
455 2 * READ_ONCE(jiffies_till_next_fqs);
456 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
457 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
458 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
459 WRITE_ONCE(jiffies_to_sched_qs, j);
462 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
465 int ret = kstrtoul(val, 0, &j);
468 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
469 adjust_jiffies_till_sched_qs();
474 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
477 int ret = kstrtoul(val, 0, &j);
480 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
481 adjust_jiffies_till_sched_qs();
486 static const struct kernel_param_ops first_fqs_jiffies_ops = {
487 .set = param_set_first_fqs_jiffies,
488 .get = param_get_ulong,
491 static const struct kernel_param_ops next_fqs_jiffies_ops = {
492 .set = param_set_next_fqs_jiffies,
493 .get = param_get_ulong,
496 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
497 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
498 module_param(rcu_kick_kthreads, bool, 0644);
500 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
501 static int rcu_pending(int user);
504 * Return the number of RCU GPs completed thus far for debug & stats.
506 unsigned long rcu_get_gp_seq(void)
508 return READ_ONCE(rcu_state.gp_seq);
510 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
513 * Return the number of RCU expedited batches completed thus far for
514 * debug & stats. Odd numbers mean that a batch is in progress, even
515 * numbers mean idle. The value returned will thus be roughly double
516 * the cumulative batches since boot.
518 unsigned long rcu_exp_batches_completed(void)
520 return rcu_state.expedited_sequence;
522 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
525 * Return the root node of the rcu_state structure.
527 static struct rcu_node *rcu_get_root(void)
529 return &rcu_state.node[0];
533 * Send along grace-period-related data for rcutorture diagnostics.
535 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
536 unsigned long *gp_seq)
540 *flags = READ_ONCE(rcu_state.gp_flags);
541 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
547 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
549 #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
551 * An empty function that will trigger a reschedule on
552 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
554 static void late_wakeup_func(struct irq_work *work)
558 static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
559 IRQ_WORK_INIT(late_wakeup_func);
564 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
565 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
567 * In these cases the late RCU wake ups aren't supported in the resched loops and our
568 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
569 * get re-enabled again.
571 noinstr void rcu_irq_work_resched(void)
573 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
575 if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
578 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
581 instrumentation_begin();
582 if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
583 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
585 instrumentation_end();
587 #endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
589 #ifdef CONFIG_PROVE_RCU
591 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
593 void rcu_irq_exit_check_preempt(void)
595 lockdep_assert_irqs_disabled();
597 RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
598 "RCU dynticks_nesting counter underflow/zero!");
599 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
601 "Bad RCU dynticks_nmi_nesting counter\n");
602 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
603 "RCU in extended quiescent state!");
605 #endif /* #ifdef CONFIG_PROVE_RCU */
607 #ifdef CONFIG_NO_HZ_FULL
609 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
611 * The scheduler tick is not normally enabled when CPUs enter the kernel
612 * from nohz_full userspace execution. After all, nohz_full userspace
613 * execution is an RCU quiescent state and the time executing in the kernel
614 * is quite short. Except of course when it isn't. And it is not hard to
615 * cause a large system to spend tens of seconds or even minutes looping
616 * in the kernel, which can cause a number of problems, include RCU CPU
619 * Therefore, if a nohz_full CPU fails to report a quiescent state
620 * in a timely manner, the RCU grace-period kthread sets that CPU's
621 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
622 * exception will invoke this function, which will turn on the scheduler
623 * tick, which will enable RCU to detect that CPU's quiescent states,
624 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
625 * The tick will be disabled once a quiescent state is reported for
628 * Of course, in carefully tuned systems, there might never be an
629 * interrupt or exception. In that case, the RCU grace-period kthread
630 * will eventually cause one to happen. However, in less carefully
631 * controlled environments, this function allows RCU to get what it
632 * needs without creating otherwise useless interruptions.
634 void __rcu_irq_enter_check_tick(void)
636 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
638 // If we're here from NMI there's nothing to do.
642 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
643 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
645 if (!tick_nohz_full_cpu(rdp->cpu) ||
646 !READ_ONCE(rdp->rcu_urgent_qs) ||
647 READ_ONCE(rdp->rcu_forced_tick)) {
648 // RCU doesn't need nohz_full help from this CPU, or it is
649 // already getting that help.
653 // We get here only when not in an extended quiescent state and
654 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
655 // already watching and (2) The fact that we are in an interrupt
656 // handler and that the rcu_node lock is an irq-disabled lock
657 // prevents self-deadlock. So we can safely recheck under the lock.
658 // Note that the nohz_full state currently cannot change.
659 raw_spin_lock_rcu_node(rdp->mynode);
660 if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
661 // A nohz_full CPU is in the kernel and RCU needs a
662 // quiescent state. Turn on the tick!
663 WRITE_ONCE(rdp->rcu_forced_tick, true);
664 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
666 raw_spin_unlock_rcu_node(rdp->mynode);
668 #endif /* CONFIG_NO_HZ_FULL */
671 * Check to see if any future non-offloaded RCU-related work will need
672 * to be done by the current CPU, even if none need be done immediately,
673 * returning 1 if so. This function is part of the RCU implementation;
674 * it is -not- an exported member of the RCU API. This is used by
675 * the idle-entry code to figure out whether it is safe to disable the
676 * scheduler-clock interrupt.
678 * Just check whether or not this CPU has non-offloaded RCU callbacks
681 int rcu_needs_cpu(void)
683 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
684 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
688 * If any sort of urgency was applied to the current CPU (for example,
689 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
690 * to get to a quiescent state, disable it.
692 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
694 raw_lockdep_assert_held_rcu_node(rdp->mynode);
695 WRITE_ONCE(rdp->rcu_urgent_qs, false);
696 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
697 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
698 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
699 WRITE_ONCE(rdp->rcu_forced_tick, false);
704 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
706 * Return true if RCU is watching the running CPU, which means that this
707 * CPU can safely enter RCU read-side critical sections. In other words,
708 * if the current CPU is not in its idle loop or is in an interrupt or
709 * NMI handler, return true.
711 * Make notrace because it can be called by the internal functions of
712 * ftrace, and making this notrace removes unnecessary recursion calls.
714 notrace bool rcu_is_watching(void)
718 preempt_disable_notrace();
719 ret = !rcu_dynticks_curr_cpu_in_eqs();
720 preempt_enable_notrace();
723 EXPORT_SYMBOL_GPL(rcu_is_watching);
726 * If a holdout task is actually running, request an urgent quiescent
727 * state from its CPU. This is unsynchronized, so migrations can cause
728 * the request to go to the wrong CPU. Which is OK, all that will happen
729 * is that the CPU's next context switch will be a bit slower and next
730 * time around this task will generate another request.
732 void rcu_request_urgent_qs_task(struct task_struct *t)
739 return; /* This task is not running on that CPU. */
740 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
743 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
746 * Is the current CPU online as far as RCU is concerned?
748 * Disable preemption to avoid false positives that could otherwise
749 * happen due to the current CPU number being sampled, this task being
750 * preempted, its old CPU being taken offline, resuming on some other CPU,
751 * then determining that its old CPU is now offline.
753 * Disable checking if in an NMI handler because we cannot safely
754 * report errors from NMI handlers anyway. In addition, it is OK to use
755 * RCU on an offline processor during initial boot, hence the check for
756 * rcu_scheduler_fully_active.
758 bool rcu_lockdep_current_cpu_online(void)
760 struct rcu_data *rdp;
763 if (in_nmi() || !rcu_scheduler_fully_active)
765 preempt_disable_notrace();
766 rdp = this_cpu_ptr(&rcu_data);
768 * Strictly, we care here about the case where the current CPU is
769 * in rcu_cpu_starting() and thus has an excuse for rdp->grpmask
770 * not being up to date. So arch_spin_is_locked() might have a
771 * false positive if it's held by some *other* CPU, but that's
772 * OK because that just means a false *negative* on the warning.
774 if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
776 preempt_enable_notrace();
779 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
781 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
784 * When trying to report a quiescent state on behalf of some other CPU,
785 * it is our responsibility to check for and handle potential overflow
786 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
787 * After all, the CPU might be in deep idle state, and thus executing no
790 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
792 raw_lockdep_assert_held_rcu_node(rnp);
793 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
795 WRITE_ONCE(rdp->gpwrap, true);
796 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
797 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
801 * Snapshot the specified CPU's dynticks counter so that we can later
802 * credit them with an implicit quiescent state. Return 1 if this CPU
803 * is in dynticks idle mode, which is an extended quiescent state.
805 static int dyntick_save_progress_counter(struct rcu_data *rdp)
807 rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
808 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
809 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
810 rcu_gpnum_ovf(rdp->mynode, rdp);
817 * Return true if the specified CPU has passed through a quiescent
818 * state by virtue of being in or having passed through an dynticks
819 * idle state since the last call to dyntick_save_progress_counter()
820 * for this same CPU, or by virtue of having been offline.
822 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
825 struct rcu_node *rnp = rdp->mynode;
828 * If the CPU passed through or entered a dynticks idle phase with
829 * no active irq/NMI handlers, then we can safely pretend that the CPU
830 * already acknowledged the request to pass through a quiescent
831 * state. Either way, that CPU cannot possibly be in an RCU
832 * read-side critical section that started before the beginning
833 * of the current RCU grace period.
835 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
836 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
837 rcu_gpnum_ovf(rnp, rdp);
842 * Complain if a CPU that is considered to be offline from RCU's
843 * perspective has not yet reported a quiescent state. After all,
844 * the offline CPU should have reported a quiescent state during
845 * the CPU-offline process, or, failing that, by rcu_gp_init()
846 * if it ran concurrently with either the CPU going offline or the
847 * last task on a leaf rcu_node structure exiting its RCU read-side
848 * critical section while all CPUs corresponding to that structure
849 * are offline. This added warning detects bugs in any of these
852 * The rcu_node structure's ->lock is held here, which excludes
853 * the relevant portions the CPU-hotplug code, the grace-period
854 * initialization code, and the rcu_read_unlock() code paths.
856 * For more detail, please refer to the "Hotplug CPU" section
857 * of RCU's Requirements documentation.
859 if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
860 struct rcu_node *rnp1;
862 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
863 __func__, rnp->grplo, rnp->grphi, rnp->level,
864 (long)rnp->gp_seq, (long)rnp->completedqs);
865 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
866 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
867 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
868 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
869 __func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
870 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
871 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
872 return 1; /* Break things loose after complaining. */
876 * A CPU running for an extended time within the kernel can
877 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
878 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
879 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
880 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
881 * variable are safe because the assignments are repeated if this
882 * CPU failed to pass through a quiescent state. This code
883 * also checks .jiffies_resched in case jiffies_to_sched_qs
886 jtsq = READ_ONCE(jiffies_to_sched_qs);
887 if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
888 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
889 time_after(jiffies, rcu_state.jiffies_resched) ||
891 WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
892 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
893 smp_store_release(&rdp->rcu_urgent_qs, true);
894 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
895 WRITE_ONCE(rdp->rcu_urgent_qs, true);
899 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
900 * The above code handles this, but only for straight cond_resched().
901 * And some in-kernel loops check need_resched() before calling
902 * cond_resched(), which defeats the above code for CPUs that are
903 * running in-kernel with scheduling-clock interrupts disabled.
904 * So hit them over the head with the resched_cpu() hammer!
906 if (tick_nohz_full_cpu(rdp->cpu) &&
907 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
909 WRITE_ONCE(rdp->rcu_urgent_qs, true);
910 resched_cpu(rdp->cpu);
911 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
915 * If more than halfway to RCU CPU stall-warning time, invoke
916 * resched_cpu() more frequently to try to loosen things up a bit.
917 * Also check to see if the CPU is getting hammered with interrupts,
918 * but only once per grace period, just to keep the IPIs down to
921 if (time_after(jiffies, rcu_state.jiffies_resched)) {
922 if (time_after(jiffies,
923 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
924 resched_cpu(rdp->cpu);
925 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
927 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
928 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
929 (rnp->ffmask & rdp->grpmask)) {
930 rdp->rcu_iw_pending = true;
931 rdp->rcu_iw_gp_seq = rnp->gp_seq;
932 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
939 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
940 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
941 unsigned long gp_seq_req, const char *s)
943 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
944 gp_seq_req, rnp->level,
945 rnp->grplo, rnp->grphi, s);
949 * rcu_start_this_gp - Request the start of a particular grace period
950 * @rnp_start: The leaf node of the CPU from which to start.
951 * @rdp: The rcu_data corresponding to the CPU from which to start.
952 * @gp_seq_req: The gp_seq of the grace period to start.
954 * Start the specified grace period, as needed to handle newly arrived
955 * callbacks. The required future grace periods are recorded in each
956 * rcu_node structure's ->gp_seq_needed field. Returns true if there
957 * is reason to awaken the grace-period kthread.
959 * The caller must hold the specified rcu_node structure's ->lock, which
960 * is why the caller is responsible for waking the grace-period kthread.
962 * Returns true if the GP thread needs to be awakened else false.
964 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
965 unsigned long gp_seq_req)
968 struct rcu_node *rnp;
971 * Use funnel locking to either acquire the root rcu_node
972 * structure's lock or bail out if the need for this grace period
973 * has already been recorded -- or if that grace period has in
974 * fact already started. If there is already a grace period in
975 * progress in a non-leaf node, no recording is needed because the
976 * end of the grace period will scan the leaf rcu_node structures.
977 * Note that rnp_start->lock must not be released.
979 raw_lockdep_assert_held_rcu_node(rnp_start);
980 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
981 for (rnp = rnp_start; 1; rnp = rnp->parent) {
982 if (rnp != rnp_start)
983 raw_spin_lock_rcu_node(rnp);
984 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
985 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
987 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
988 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
992 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
993 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
995 * We just marked the leaf or internal node, and a
996 * grace period is in progress, which means that
997 * rcu_gp_cleanup() will see the marking. Bail to
1000 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1001 TPS("Startedleaf"));
1004 if (rnp != rnp_start && rnp->parent != NULL)
1005 raw_spin_unlock_rcu_node(rnp);
1007 break; /* At root, and perhaps also leaf. */
1010 /* If GP already in progress, just leave, otherwise start one. */
1011 if (rcu_gp_in_progress()) {
1012 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1015 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1016 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1017 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1018 if (!READ_ONCE(rcu_state.gp_kthread)) {
1019 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1022 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1023 ret = true; /* Caller must wake GP kthread. */
1025 /* Push furthest requested GP to leaf node and rcu_data structure. */
1026 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1027 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1028 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1030 if (rnp != rnp_start)
1031 raw_spin_unlock_rcu_node(rnp);
1036 * Clean up any old requests for the just-ended grace period. Also return
1037 * whether any additional grace periods have been requested.
1039 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1042 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1044 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1046 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1047 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1048 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1053 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1054 * interrupt or softirq handler, in which case we just might immediately
1055 * sleep upon return, resulting in a grace-period hang), and don't bother
1056 * awakening when there is nothing for the grace-period kthread to do
1057 * (as in several CPUs raced to awaken, we lost), and finally don't try
1058 * to awaken a kthread that has not yet been created. If all those checks
1059 * are passed, track some debug information and awaken.
1061 * So why do the self-wakeup when in an interrupt or softirq handler
1062 * in the grace-period kthread's context? Because the kthread might have
1063 * been interrupted just as it was going to sleep, and just after the final
1064 * pre-sleep check of the awaken condition. In this case, a wakeup really
1065 * is required, and is therefore supplied.
1067 static void rcu_gp_kthread_wake(void)
1069 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1071 if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
1072 !READ_ONCE(rcu_state.gp_flags) || !t)
1074 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1075 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1076 swake_up_one(&rcu_state.gp_wq);
1080 * If there is room, assign a ->gp_seq number to any callbacks on this
1081 * CPU that have not already been assigned. Also accelerate any callbacks
1082 * that were previously assigned a ->gp_seq number that has since proven
1083 * to be too conservative, which can happen if callbacks get assigned a
1084 * ->gp_seq number while RCU is idle, but with reference to a non-root
1085 * rcu_node structure. This function is idempotent, so it does not hurt
1086 * to call it repeatedly. Returns an flag saying that we should awaken
1087 * the RCU grace-period kthread.
1089 * The caller must hold rnp->lock with interrupts disabled.
1091 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1093 unsigned long gp_seq_req;
1096 rcu_lockdep_assert_cblist_protected(rdp);
1097 raw_lockdep_assert_held_rcu_node(rnp);
1099 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1100 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1103 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1106 * Callbacks are often registered with incomplete grace-period
1107 * information. Something about the fact that getting exact
1108 * information requires acquiring a global lock... RCU therefore
1109 * makes a conservative estimate of the grace period number at which
1110 * a given callback will become ready to invoke. The following
1111 * code checks this estimate and improves it when possible, thus
1112 * accelerating callback invocation to an earlier grace-period
1115 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1116 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1117 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1119 /* Trace depending on how much we were able to accelerate. */
1120 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1121 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1123 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1125 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1131 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1132 * rcu_node structure's ->lock be held. It consults the cached value
1133 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1134 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1135 * while holding the leaf rcu_node structure's ->lock.
1137 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1138 struct rcu_data *rdp)
1143 rcu_lockdep_assert_cblist_protected(rdp);
1144 c = rcu_seq_snap(&rcu_state.gp_seq);
1145 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1146 /* Old request still live, so mark recent callbacks. */
1147 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1150 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1151 needwake = rcu_accelerate_cbs(rnp, rdp);
1152 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1154 rcu_gp_kthread_wake();
1158 * Move any callbacks whose grace period has completed to the
1159 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1160 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1161 * sublist. This function is idempotent, so it does not hurt to
1162 * invoke it repeatedly. As long as it is not invoked -too- often...
1163 * Returns true if the RCU grace-period kthread needs to be awakened.
1165 * The caller must hold rnp->lock with interrupts disabled.
1167 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1169 rcu_lockdep_assert_cblist_protected(rdp);
1170 raw_lockdep_assert_held_rcu_node(rnp);
1172 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1173 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1177 * Find all callbacks whose ->gp_seq numbers indicate that they
1178 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1180 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1182 /* Classify any remaining callbacks. */
1183 return rcu_accelerate_cbs(rnp, rdp);
1187 * Move and classify callbacks, but only if doing so won't require
1188 * that the RCU grace-period kthread be awakened.
1190 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1191 struct rcu_data *rdp)
1193 rcu_lockdep_assert_cblist_protected(rdp);
1194 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1196 // The grace period cannot end while we hold the rcu_node lock.
1197 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1198 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1199 raw_spin_unlock_rcu_node(rnp);
1203 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1204 * quiescent state. This is intended to be invoked when the CPU notices
1205 * a new grace period.
1207 static void rcu_strict_gp_check_qs(void)
1209 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1216 * Update CPU-local rcu_data state to record the beginnings and ends of
1217 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1218 * structure corresponding to the current CPU, and must have irqs disabled.
1219 * Returns true if the grace-period kthread needs to be awakened.
1221 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1225 const bool offloaded = rcu_rdp_is_offloaded(rdp);
1227 raw_lockdep_assert_held_rcu_node(rnp);
1229 if (rdp->gp_seq == rnp->gp_seq)
1230 return false; /* Nothing to do. */
1232 /* Handle the ends of any preceding grace periods first. */
1233 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1234 unlikely(READ_ONCE(rdp->gpwrap))) {
1236 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1237 rdp->core_needs_qs = false;
1238 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1241 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1242 if (rdp->core_needs_qs)
1243 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1246 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1247 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1248 unlikely(READ_ONCE(rdp->gpwrap))) {
1250 * If the current grace period is waiting for this CPU,
1251 * set up to detect a quiescent state, otherwise don't
1252 * go looking for one.
1254 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1255 need_qs = !!(rnp->qsmask & rdp->grpmask);
1256 rdp->cpu_no_qs.b.norm = need_qs;
1257 rdp->core_needs_qs = need_qs;
1258 zero_cpu_stall_ticks(rdp);
1260 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1261 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1262 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1263 if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
1264 WRITE_ONCE(rdp->last_sched_clock, jiffies);
1265 WRITE_ONCE(rdp->gpwrap, false);
1266 rcu_gpnum_ovf(rnp, rdp);
1270 static void note_gp_changes(struct rcu_data *rdp)
1272 unsigned long flags;
1274 struct rcu_node *rnp;
1276 local_irq_save(flags);
1278 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1279 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1280 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1281 local_irq_restore(flags);
1284 needwake = __note_gp_changes(rnp, rdp);
1285 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1286 rcu_strict_gp_check_qs();
1288 rcu_gp_kthread_wake();
1291 static atomic_t *rcu_gp_slow_suppress;
1293 /* Register a counter to suppress debugging grace-period delays. */
1294 void rcu_gp_slow_register(atomic_t *rgssp)
1296 WARN_ON_ONCE(rcu_gp_slow_suppress);
1298 WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
1300 EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
1302 /* Unregister a counter, with NULL for not caring which. */
1303 void rcu_gp_slow_unregister(atomic_t *rgssp)
1305 WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress);
1307 WRITE_ONCE(rcu_gp_slow_suppress, NULL);
1309 EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
1311 static bool rcu_gp_slow_is_suppressed(void)
1313 atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
1315 return rgssp && atomic_read(rgssp);
1318 static void rcu_gp_slow(int delay)
1320 if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
1321 !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1322 schedule_timeout_idle(delay);
1325 static unsigned long sleep_duration;
1327 /* Allow rcutorture to stall the grace-period kthread. */
1328 void rcu_gp_set_torture_wait(int duration)
1330 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1331 WRITE_ONCE(sleep_duration, duration);
1333 EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1335 /* Actually implement the aforementioned wait. */
1336 static void rcu_gp_torture_wait(void)
1338 unsigned long duration;
1340 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1342 duration = xchg(&sleep_duration, 0UL);
1344 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1345 schedule_timeout_idle(duration);
1346 pr_alert("%s: Wait complete\n", __func__);
1351 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1354 static void rcu_strict_gp_boundary(void *unused)
1359 // Has rcu_init() been invoked? This is used (for example) to determine
1360 // whether spinlocks may be acquired safely.
1361 static bool rcu_init_invoked(void)
1363 return !!rcu_state.n_online_cpus;
1366 // Make the polled API aware of the beginning of a grace period.
1367 static void rcu_poll_gp_seq_start(unsigned long *snap)
1369 struct rcu_node *rnp = rcu_get_root();
1371 if (rcu_init_invoked())
1372 raw_lockdep_assert_held_rcu_node(rnp);
1374 // If RCU was idle, note beginning of GP.
1375 if (!rcu_seq_state(rcu_state.gp_seq_polled))
1376 rcu_seq_start(&rcu_state.gp_seq_polled);
1378 // Either way, record current state.
1379 *snap = rcu_state.gp_seq_polled;
1382 // Make the polled API aware of the end of a grace period.
1383 static void rcu_poll_gp_seq_end(unsigned long *snap)
1385 struct rcu_node *rnp = rcu_get_root();
1387 if (rcu_init_invoked())
1388 raw_lockdep_assert_held_rcu_node(rnp);
1390 // If the previously noted GP is still in effect, record the
1391 // end of that GP. Either way, zero counter to avoid counter-wrap
1393 if (*snap && *snap == rcu_state.gp_seq_polled) {
1394 rcu_seq_end(&rcu_state.gp_seq_polled);
1395 rcu_state.gp_seq_polled_snap = 0;
1396 rcu_state.gp_seq_polled_exp_snap = 0;
1402 // Make the polled API aware of the beginning of a grace period, but
1403 // where caller does not hold the root rcu_node structure's lock.
1404 static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
1406 unsigned long flags;
1407 struct rcu_node *rnp = rcu_get_root();
1409 if (rcu_init_invoked()) {
1410 lockdep_assert_irqs_enabled();
1411 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1413 rcu_poll_gp_seq_start(snap);
1414 if (rcu_init_invoked())
1415 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1418 // Make the polled API aware of the end of a grace period, but where
1419 // caller does not hold the root rcu_node structure's lock.
1420 static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
1422 unsigned long flags;
1423 struct rcu_node *rnp = rcu_get_root();
1425 if (rcu_init_invoked()) {
1426 lockdep_assert_irqs_enabled();
1427 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1429 rcu_poll_gp_seq_end(snap);
1430 if (rcu_init_invoked())
1431 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1435 * Initialize a new grace period. Return false if no grace period required.
1437 static noinline_for_stack bool rcu_gp_init(void)
1439 unsigned long flags;
1440 unsigned long oldmask;
1442 struct rcu_data *rdp;
1443 struct rcu_node *rnp = rcu_get_root();
1445 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1446 raw_spin_lock_irq_rcu_node(rnp);
1447 if (!READ_ONCE(rcu_state.gp_flags)) {
1448 /* Spurious wakeup, tell caller to go back to sleep. */
1449 raw_spin_unlock_irq_rcu_node(rnp);
1452 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1454 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1456 * Grace period already in progress, don't start another.
1457 * Not supposed to be able to happen.
1459 raw_spin_unlock_irq_rcu_node(rnp);
1463 /* Advance to a new grace period and initialize state. */
1464 record_gp_stall_check_time();
1465 /* Record GP times before starting GP, hence rcu_seq_start(). */
1466 rcu_seq_start(&rcu_state.gp_seq);
1467 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1468 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1469 rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
1470 raw_spin_unlock_irq_rcu_node(rnp);
1473 * Apply per-leaf buffered online and offline operations to
1474 * the rcu_node tree. Note that this new grace period need not
1475 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1476 * offlining path, when combined with checks in this function,
1477 * will handle CPUs that are currently going offline or that will
1478 * go offline later. Please also refer to "Hotplug CPU" section
1479 * of RCU's Requirements documentation.
1481 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1482 /* Exclude CPU hotplug operations. */
1483 rcu_for_each_leaf_node(rnp) {
1484 local_irq_save(flags);
1485 arch_spin_lock(&rcu_state.ofl_lock);
1486 raw_spin_lock_rcu_node(rnp);
1487 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1488 !rnp->wait_blkd_tasks) {
1489 /* Nothing to do on this leaf rcu_node structure. */
1490 raw_spin_unlock_rcu_node(rnp);
1491 arch_spin_unlock(&rcu_state.ofl_lock);
1492 local_irq_restore(flags);
1496 /* Record old state, apply changes to ->qsmaskinit field. */
1497 oldmask = rnp->qsmaskinit;
1498 rnp->qsmaskinit = rnp->qsmaskinitnext;
1500 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1501 if (!oldmask != !rnp->qsmaskinit) {
1502 if (!oldmask) { /* First online CPU for rcu_node. */
1503 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1504 rcu_init_new_rnp(rnp);
1505 } else if (rcu_preempt_has_tasks(rnp)) {
1506 rnp->wait_blkd_tasks = true; /* blocked tasks */
1507 } else { /* Last offline CPU and can propagate. */
1508 rcu_cleanup_dead_rnp(rnp);
1513 * If all waited-on tasks from prior grace period are
1514 * done, and if all this rcu_node structure's CPUs are
1515 * still offline, propagate up the rcu_node tree and
1516 * clear ->wait_blkd_tasks. Otherwise, if one of this
1517 * rcu_node structure's CPUs has since come back online,
1518 * simply clear ->wait_blkd_tasks.
1520 if (rnp->wait_blkd_tasks &&
1521 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1522 rnp->wait_blkd_tasks = false;
1523 if (!rnp->qsmaskinit)
1524 rcu_cleanup_dead_rnp(rnp);
1527 raw_spin_unlock_rcu_node(rnp);
1528 arch_spin_unlock(&rcu_state.ofl_lock);
1529 local_irq_restore(flags);
1531 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1534 * Set the quiescent-state-needed bits in all the rcu_node
1535 * structures for all currently online CPUs in breadth-first
1536 * order, starting from the root rcu_node structure, relying on the
1537 * layout of the tree within the rcu_state.node[] array. Note that
1538 * other CPUs will access only the leaves of the hierarchy, thus
1539 * seeing that no grace period is in progress, at least until the
1540 * corresponding leaf node has been initialized.
1542 * The grace period cannot complete until the initialization
1543 * process finishes, because this kthread handles both.
1545 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1546 rcu_for_each_node_breadth_first(rnp) {
1547 rcu_gp_slow(gp_init_delay);
1548 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1549 rdp = this_cpu_ptr(&rcu_data);
1550 rcu_preempt_check_blocked_tasks(rnp);
1551 rnp->qsmask = rnp->qsmaskinit;
1552 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1553 if (rnp == rdp->mynode)
1554 (void)__note_gp_changes(rnp, rdp);
1555 rcu_preempt_boost_start_gp(rnp);
1556 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1557 rnp->level, rnp->grplo,
1558 rnp->grphi, rnp->qsmask);
1559 /* Quiescent states for tasks on any now-offline CPUs. */
1560 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1561 rnp->rcu_gp_init_mask = mask;
1562 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1563 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1565 raw_spin_unlock_irq_rcu_node(rnp);
1566 cond_resched_tasks_rcu_qs();
1567 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1570 // If strict, make all CPUs aware of new grace period.
1571 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1572 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1578 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1581 static bool rcu_gp_fqs_check_wake(int *gfp)
1583 struct rcu_node *rnp = rcu_get_root();
1585 // If under overload conditions, force an immediate FQS scan.
1586 if (*gfp & RCU_GP_FLAG_OVLD)
1589 // Someone like call_rcu() requested a force-quiescent-state scan.
1590 *gfp = READ_ONCE(rcu_state.gp_flags);
1591 if (*gfp & RCU_GP_FLAG_FQS)
1594 // The current grace period has completed.
1595 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1602 * Do one round of quiescent-state forcing.
1604 static void rcu_gp_fqs(bool first_time)
1606 struct rcu_node *rnp = rcu_get_root();
1608 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1609 WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1611 /* Collect dyntick-idle snapshots. */
1612 force_qs_rnp(dyntick_save_progress_counter);
1614 /* Handle dyntick-idle and offline CPUs. */
1615 force_qs_rnp(rcu_implicit_dynticks_qs);
1617 /* Clear flag to prevent immediate re-entry. */
1618 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1619 raw_spin_lock_irq_rcu_node(rnp);
1620 WRITE_ONCE(rcu_state.gp_flags,
1621 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1622 raw_spin_unlock_irq_rcu_node(rnp);
1627 * Loop doing repeated quiescent-state forcing until the grace period ends.
1629 static noinline_for_stack void rcu_gp_fqs_loop(void)
1631 bool first_gp_fqs = true;
1635 struct rcu_node *rnp = rcu_get_root();
1637 j = READ_ONCE(jiffies_till_first_fqs);
1638 if (rcu_state.cbovld)
1639 gf = RCU_GP_FLAG_OVLD;
1642 if (rcu_state.cbovld) {
1647 if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
1648 WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1650 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1651 * update; required for stall checks.
1654 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1655 jiffies + (j ? 3 * j : 2));
1657 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1659 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1660 (void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
1661 rcu_gp_fqs_check_wake(&gf), j);
1662 rcu_gp_torture_wait();
1663 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
1664 /* Locking provides needed memory barriers. */
1666 * Exit the loop if the root rcu_node structure indicates that the grace period
1667 * has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check
1668 * is required only for single-node rcu_node trees because readers blocking
1669 * the current grace period are queued only on leaf rcu_node structures.
1670 * For multi-node trees, checking the root node's ->qsmask suffices, because a
1671 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
1672 * the corresponding leaf nodes have passed through their quiescent state.
1674 if (!READ_ONCE(rnp->qsmask) &&
1675 !rcu_preempt_blocked_readers_cgp(rnp))
1677 /* If time for quiescent-state forcing, do it. */
1678 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1679 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1680 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1682 rcu_gp_fqs(first_gp_fqs);
1685 first_gp_fqs = false;
1686 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1688 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1690 cond_resched_tasks_rcu_qs();
1691 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1692 ret = 0; /* Force full wait till next FQS. */
1693 j = READ_ONCE(jiffies_till_next_fqs);
1695 /* Deal with stray signal. */
1696 cond_resched_tasks_rcu_qs();
1697 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1698 WARN_ON(signal_pending(current));
1699 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1701 ret = 1; /* Keep old FQS timing. */
1703 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1706 j = rcu_state.jiffies_force_qs - j;
1713 * Clean up after the old grace period.
1715 static noinline void rcu_gp_cleanup(void)
1718 bool needgp = false;
1719 unsigned long gp_duration;
1720 unsigned long new_gp_seq;
1722 struct rcu_data *rdp;
1723 struct rcu_node *rnp = rcu_get_root();
1724 struct swait_queue_head *sq;
1726 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1727 raw_spin_lock_irq_rcu_node(rnp);
1728 rcu_state.gp_end = jiffies;
1729 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1730 if (gp_duration > rcu_state.gp_max)
1731 rcu_state.gp_max = gp_duration;
1734 * We know the grace period is complete, but to everyone else
1735 * it appears to still be ongoing. But it is also the case
1736 * that to everyone else it looks like there is nothing that
1737 * they can do to advance the grace period. It is therefore
1738 * safe for us to drop the lock in order to mark the grace
1739 * period as completed in all of the rcu_node structures.
1741 rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
1742 raw_spin_unlock_irq_rcu_node(rnp);
1745 * Propagate new ->gp_seq value to rcu_node structures so that
1746 * other CPUs don't have to wait until the start of the next grace
1747 * period to process their callbacks. This also avoids some nasty
1748 * RCU grace-period initialization races by forcing the end of
1749 * the current grace period to be completely recorded in all of
1750 * the rcu_node structures before the beginning of the next grace
1751 * period is recorded in any of the rcu_node structures.
1753 new_gp_seq = rcu_state.gp_seq;
1754 rcu_seq_end(&new_gp_seq);
1755 rcu_for_each_node_breadth_first(rnp) {
1756 raw_spin_lock_irq_rcu_node(rnp);
1757 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1758 dump_blkd_tasks(rnp, 10);
1759 WARN_ON_ONCE(rnp->qsmask);
1760 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1762 smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
1763 rdp = this_cpu_ptr(&rcu_data);
1764 if (rnp == rdp->mynode)
1765 needgp = __note_gp_changes(rnp, rdp) || needgp;
1766 /* smp_mb() provided by prior unlock-lock pair. */
1767 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1768 // Reset overload indication for CPUs no longer overloaded
1769 if (rcu_is_leaf_node(rnp))
1770 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1771 rdp = per_cpu_ptr(&rcu_data, cpu);
1772 check_cb_ovld_locked(rdp, rnp);
1774 sq = rcu_nocb_gp_get(rnp);
1775 raw_spin_unlock_irq_rcu_node(rnp);
1776 rcu_nocb_gp_cleanup(sq);
1777 cond_resched_tasks_rcu_qs();
1778 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1779 rcu_gp_slow(gp_cleanup_delay);
1781 rnp = rcu_get_root();
1782 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1784 /* Declare grace period done, trace first to use old GP number. */
1785 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1786 rcu_seq_end(&rcu_state.gp_seq);
1787 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1788 WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
1789 /* Check for GP requests since above loop. */
1790 rdp = this_cpu_ptr(&rcu_data);
1791 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1792 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1793 TPS("CleanupMore"));
1796 /* Advance CBs to reduce false positives below. */
1797 offloaded = rcu_rdp_is_offloaded(rdp);
1798 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1800 // We get here if a grace period was needed (“needgp”)
1801 // and the above call to rcu_accelerate_cbs() did not set
1802 // the RCU_GP_FLAG_INIT bit in ->gp_state (which records
1803 // the need for another grace period). The purpose
1804 // of the “offloaded” check is to avoid invoking
1805 // rcu_accelerate_cbs() on an offloaded CPU because we do not
1806 // hold the ->nocb_lock needed to safely access an offloaded
1807 // ->cblist. We do not want to acquire that lock because
1808 // it can be heavily contended during callback floods.
1810 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1811 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1812 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
1815 // We get here either if there is no need for an
1816 // additional grace period or if rcu_accelerate_cbs() has
1817 // already set the RCU_GP_FLAG_INIT bit in ->gp_flags.
1818 // So all we need to do is to clear all of the other
1821 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1823 raw_spin_unlock_irq_rcu_node(rnp);
1825 // If strict, make all CPUs aware of the end of the old grace period.
1826 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1827 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1831 * Body of kthread that handles grace periods.
1833 static int __noreturn rcu_gp_kthread(void *unused)
1835 rcu_bind_gp_kthread();
1838 /* Handle grace-period start. */
1840 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1842 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
1843 swait_event_idle_exclusive(rcu_state.gp_wq,
1844 READ_ONCE(rcu_state.gp_flags) &
1846 rcu_gp_torture_wait();
1847 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
1848 /* Locking provides needed memory barrier. */
1851 cond_resched_tasks_rcu_qs();
1852 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1853 WARN_ON(signal_pending(current));
1854 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1858 /* Handle quiescent-state forcing. */
1861 /* Handle grace-period end. */
1862 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
1864 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
1869 * Report a full set of quiescent states to the rcu_state data structure.
1870 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1871 * another grace period is required. Whether we wake the grace-period
1872 * kthread or it awakens itself for the next round of quiescent-state
1873 * forcing, that kthread will clean up after the just-completed grace
1874 * period. Note that the caller must hold rnp->lock, which is released
1877 static void rcu_report_qs_rsp(unsigned long flags)
1878 __releases(rcu_get_root()->lock)
1880 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1881 WARN_ON_ONCE(!rcu_gp_in_progress());
1882 WRITE_ONCE(rcu_state.gp_flags,
1883 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1884 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1885 rcu_gp_kthread_wake();
1889 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1890 * Allows quiescent states for a group of CPUs to be reported at one go
1891 * to the specified rcu_node structure, though all the CPUs in the group
1892 * must be represented by the same rcu_node structure (which need not be a
1893 * leaf rcu_node structure, though it often will be). The gps parameter
1894 * is the grace-period snapshot, which means that the quiescent states
1895 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1896 * must be held upon entry, and it is released before return.
1898 * As a special case, if mask is zero, the bit-already-cleared check is
1899 * disabled. This allows propagating quiescent state due to resumed tasks
1900 * during grace-period initialization.
1902 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1903 unsigned long gps, unsigned long flags)
1904 __releases(rnp->lock)
1906 unsigned long oldmask = 0;
1907 struct rcu_node *rnp_c;
1909 raw_lockdep_assert_held_rcu_node(rnp);
1911 /* Walk up the rcu_node hierarchy. */
1913 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1916 * Our bit has already been cleared, or the
1917 * relevant grace period is already over, so done.
1919 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1922 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1923 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1924 rcu_preempt_blocked_readers_cgp(rnp));
1925 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
1926 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1927 mask, rnp->qsmask, rnp->level,
1928 rnp->grplo, rnp->grphi,
1930 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1932 /* Other bits still set at this level, so done. */
1933 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1936 rnp->completedqs = rnp->gp_seq;
1937 mask = rnp->grpmask;
1938 if (rnp->parent == NULL) {
1940 /* No more levels. Exit loop holding root lock. */
1944 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1947 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1948 oldmask = READ_ONCE(rnp_c->qsmask);
1952 * Get here if we are the last CPU to pass through a quiescent
1953 * state for this grace period. Invoke rcu_report_qs_rsp()
1954 * to clean up and start the next grace period if one is needed.
1956 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1960 * Record a quiescent state for all tasks that were previously queued
1961 * on the specified rcu_node structure and that were blocking the current
1962 * RCU grace period. The caller must hold the corresponding rnp->lock with
1963 * irqs disabled, and this lock is released upon return, but irqs remain
1966 static void __maybe_unused
1967 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1968 __releases(rnp->lock)
1972 struct rcu_node *rnp_p;
1974 raw_lockdep_assert_held_rcu_node(rnp);
1975 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
1976 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1978 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1979 return; /* Still need more quiescent states! */
1982 rnp->completedqs = rnp->gp_seq;
1983 rnp_p = rnp->parent;
1984 if (rnp_p == NULL) {
1986 * Only one rcu_node structure in the tree, so don't
1987 * try to report up to its nonexistent parent!
1989 rcu_report_qs_rsp(flags);
1993 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1995 mask = rnp->grpmask;
1996 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1997 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1998 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2002 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2003 * structure. This must be called from the specified CPU.
2006 rcu_report_qs_rdp(struct rcu_data *rdp)
2008 unsigned long flags;
2010 bool needwake = false;
2011 bool needacc = false;
2012 struct rcu_node *rnp;
2014 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2016 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2017 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2021 * The grace period in which this quiescent state was
2022 * recorded has ended, so don't report it upwards.
2023 * We will instead need a new quiescent state that lies
2024 * within the current grace period.
2026 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2027 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2030 mask = rdp->grpmask;
2031 rdp->core_needs_qs = false;
2032 if ((rnp->qsmask & mask) == 0) {
2033 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2036 * This GP can't end until cpu checks in, so all of our
2037 * callbacks can be processed during the next GP.
2039 * NOCB kthreads have their own way to deal with that...
2041 if (!rcu_rdp_is_offloaded(rdp)) {
2042 needwake = rcu_accelerate_cbs(rnp, rdp);
2043 } else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
2045 * ...but NOCB kthreads may miss or delay callbacks acceleration
2046 * if in the middle of a (de-)offloading process.
2051 rcu_disable_urgency_upon_qs(rdp);
2052 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2053 /* ^^^ Released rnp->lock */
2055 rcu_gp_kthread_wake();
2058 rcu_nocb_lock_irqsave(rdp, flags);
2059 rcu_accelerate_cbs_unlocked(rnp, rdp);
2060 rcu_nocb_unlock_irqrestore(rdp, flags);
2066 * Check to see if there is a new grace period of which this CPU
2067 * is not yet aware, and if so, set up local rcu_data state for it.
2068 * Otherwise, see if this CPU has just passed through its first
2069 * quiescent state for this grace period, and record that fact if so.
2072 rcu_check_quiescent_state(struct rcu_data *rdp)
2074 /* Check for grace-period ends and beginnings. */
2075 note_gp_changes(rdp);
2078 * Does this CPU still need to do its part for current grace period?
2079 * If no, return and let the other CPUs do their part as well.
2081 if (!rdp->core_needs_qs)
2085 * Was there a quiescent state since the beginning of the grace
2086 * period? If no, then exit and wait for the next call.
2088 if (rdp->cpu_no_qs.b.norm)
2092 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2095 rcu_report_qs_rdp(rdp);
2099 * Near the end of the offline process. Trace the fact that this CPU
2102 int rcutree_dying_cpu(unsigned int cpu)
2105 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2106 struct rcu_node *rnp = rdp->mynode;
2108 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2111 blkd = !!(rnp->qsmask & rdp->grpmask);
2112 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2113 blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
2118 * All CPUs for the specified rcu_node structure have gone offline,
2119 * and all tasks that were preempted within an RCU read-side critical
2120 * section while running on one of those CPUs have since exited their RCU
2121 * read-side critical section. Some other CPU is reporting this fact with
2122 * the specified rcu_node structure's ->lock held and interrupts disabled.
2123 * This function therefore goes up the tree of rcu_node structures,
2124 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2125 * the leaf rcu_node structure's ->qsmaskinit field has already been
2128 * This function does check that the specified rcu_node structure has
2129 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2130 * prematurely. That said, invoking it after the fact will cost you
2131 * a needless lock acquisition. So once it has done its work, don't
2134 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2137 struct rcu_node *rnp = rnp_leaf;
2139 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2140 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2141 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2142 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2145 mask = rnp->grpmask;
2149 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2150 rnp->qsmaskinit &= ~mask;
2151 /* Between grace periods, so better already be zero! */
2152 WARN_ON_ONCE(rnp->qsmask);
2153 if (rnp->qsmaskinit) {
2154 raw_spin_unlock_rcu_node(rnp);
2155 /* irqs remain disabled. */
2158 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2163 * The CPU has been completely removed, and some other CPU is reporting
2164 * this fact from process context. Do the remainder of the cleanup.
2165 * There can only be one CPU hotplug operation at a time, so no need for
2168 int rcutree_dead_cpu(unsigned int cpu)
2170 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2171 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2173 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2176 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
2177 /* Adjust any no-longer-needed kthreads. */
2178 rcu_boost_kthread_setaffinity(rnp, -1);
2179 // Stop-machine done, so allow nohz_full to disable tick.
2180 tick_dep_clear(TICK_DEP_BIT_RCU);
2185 * Invoke any RCU callbacks that have made it to the end of their grace
2186 * period. Throttle as specified by rdp->blimit.
2188 static void rcu_do_batch(struct rcu_data *rdp)
2191 bool __maybe_unused empty;
2192 unsigned long flags;
2193 struct rcu_head *rhp;
2194 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2196 long pending, tlimit = 0;
2198 /* If no callbacks are ready, just return. */
2199 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2200 trace_rcu_batch_start(rcu_state.name,
2201 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2202 trace_rcu_batch_end(rcu_state.name, 0,
2203 !rcu_segcblist_empty(&rdp->cblist),
2204 need_resched(), is_idle_task(current),
2205 rcu_is_callbacks_kthread(rdp));
2210 * Extract the list of ready callbacks, disabling IRQs to prevent
2211 * races with call_rcu() from interrupt handlers. Leave the
2212 * callback counts, as rcu_barrier() needs to be conservative.
2214 rcu_nocb_lock_irqsave(rdp, flags);
2215 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2216 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2217 div = READ_ONCE(rcu_divisor);
2218 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2219 bl = max(rdp->blimit, pending >> div);
2220 if (in_serving_softirq() && unlikely(bl > 100)) {
2221 long rrn = READ_ONCE(rcu_resched_ns);
2223 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2224 tlimit = local_clock() + rrn;
2226 trace_rcu_batch_start(rcu_state.name,
2227 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2228 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2229 if (rcu_rdp_is_offloaded(rdp))
2230 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2232 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2233 rcu_nocb_unlock_irqrestore(rdp, flags);
2235 /* Invoke callbacks. */
2236 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2237 rhp = rcu_cblist_dequeue(&rcl);
2239 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2243 debug_rcu_head_unqueue(rhp);
2245 rcu_lock_acquire(&rcu_callback_map);
2246 trace_rcu_invoke_callback(rcu_state.name, rhp);
2249 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2252 rcu_lock_release(&rcu_callback_map);
2255 * Stop only if limit reached and CPU has something to do.
2257 if (in_serving_softirq()) {
2258 if (count >= bl && (need_resched() || !is_idle_task(current)))
2261 * Make sure we don't spend too much time here and deprive other
2262 * softirq vectors of CPU cycles.
2264 if (unlikely(tlimit)) {
2265 /* only call local_clock() every 32 callbacks */
2266 if (likely((count & 31) || local_clock() < tlimit))
2268 /* Exceeded the time limit, so leave. */
2273 lockdep_assert_irqs_enabled();
2274 cond_resched_tasks_rcu_qs();
2275 lockdep_assert_irqs_enabled();
2280 rcu_nocb_lock_irqsave(rdp, flags);
2281 rdp->n_cbs_invoked += count;
2282 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2283 is_idle_task(current), rcu_is_callbacks_kthread(rdp));
2285 /* Update counts and requeue any remaining callbacks. */
2286 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2287 rcu_segcblist_add_len(&rdp->cblist, -count);
2289 /* Reinstate batch limit if we have worked down the excess. */
2290 count = rcu_segcblist_n_cbs(&rdp->cblist);
2291 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2292 rdp->blimit = blimit;
2294 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2295 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2296 rdp->qlen_last_fqs_check = 0;
2297 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2298 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2299 rdp->qlen_last_fqs_check = count;
2302 * The following usually indicates a double call_rcu(). To track
2303 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2305 empty = rcu_segcblist_empty(&rdp->cblist);
2306 WARN_ON_ONCE(count == 0 && !empty);
2307 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2308 count != 0 && empty);
2309 WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2310 WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2312 rcu_nocb_unlock_irqrestore(rdp, flags);
2314 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2318 * This function is invoked from each scheduling-clock interrupt,
2319 * and checks to see if this CPU is in a non-context-switch quiescent
2320 * state, for example, user mode or idle loop. It also schedules RCU
2321 * core processing. If the current grace period has gone on too long,
2322 * it will ask the scheduler to manufacture a context switch for the sole
2323 * purpose of providing the needed quiescent state.
2325 void rcu_sched_clock_irq(int user)
2329 if (IS_ENABLED(CONFIG_PROVE_RCU)) {
2331 WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
2332 __this_cpu_write(rcu_data.last_sched_clock, j);
2334 trace_rcu_utilization(TPS("Start scheduler-tick"));
2335 lockdep_assert_irqs_disabled();
2336 raw_cpu_inc(rcu_data.ticks_this_gp);
2337 /* The load-acquire pairs with the store-release setting to true. */
2338 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2339 /* Idle and userspace execution already are quiescent states. */
2340 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2341 set_tsk_need_resched(current);
2342 set_preempt_need_resched();
2344 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2346 rcu_flavor_sched_clock_irq(user);
2347 if (rcu_pending(user))
2349 if (user || rcu_is_cpu_rrupt_from_idle())
2350 rcu_note_voluntary_context_switch(current);
2351 lockdep_assert_irqs_disabled();
2353 trace_rcu_utilization(TPS("End scheduler-tick"));
2357 * Scan the leaf rcu_node structures. For each structure on which all
2358 * CPUs have reported a quiescent state and on which there are tasks
2359 * blocking the current grace period, initiate RCU priority boosting.
2360 * Otherwise, invoke the specified function to check dyntick state for
2361 * each CPU that has not yet reported a quiescent state.
2363 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2366 unsigned long flags;
2368 struct rcu_data *rdp;
2369 struct rcu_node *rnp;
2371 rcu_state.cbovld = rcu_state.cbovldnext;
2372 rcu_state.cbovldnext = false;
2373 rcu_for_each_leaf_node(rnp) {
2374 cond_resched_tasks_rcu_qs();
2376 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2377 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2378 if (rnp->qsmask == 0) {
2379 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2381 * No point in scanning bits because they
2382 * are all zero. But we might need to
2383 * priority-boost blocked readers.
2385 rcu_initiate_boost(rnp, flags);
2386 /* rcu_initiate_boost() releases rnp->lock */
2389 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2392 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2393 rdp = per_cpu_ptr(&rcu_data, cpu);
2395 mask |= rdp->grpmask;
2396 rcu_disable_urgency_upon_qs(rdp);
2400 /* Idle/offline CPUs, report (releases rnp->lock). */
2401 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2403 /* Nothing to do here, so just drop the lock. */
2404 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2410 * Force quiescent states on reluctant CPUs, and also detect which
2411 * CPUs are in dyntick-idle mode.
2413 void rcu_force_quiescent_state(void)
2415 unsigned long flags;
2417 struct rcu_node *rnp;
2418 struct rcu_node *rnp_old = NULL;
2420 /* Funnel through hierarchy to reduce memory contention. */
2421 rnp = __this_cpu_read(rcu_data.mynode);
2422 for (; rnp != NULL; rnp = rnp->parent) {
2423 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2424 !raw_spin_trylock(&rnp->fqslock);
2425 if (rnp_old != NULL)
2426 raw_spin_unlock(&rnp_old->fqslock);
2431 /* rnp_old == rcu_get_root(), rnp == NULL. */
2433 /* Reached the root of the rcu_node tree, acquire lock. */
2434 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2435 raw_spin_unlock(&rnp_old->fqslock);
2436 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2437 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2438 return; /* Someone beat us to it. */
2440 WRITE_ONCE(rcu_state.gp_flags,
2441 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2442 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2443 rcu_gp_kthread_wake();
2445 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2447 // Workqueue handler for an RCU reader for kernels enforcing struct RCU
2449 static void strict_work_handler(struct work_struct *work)
2455 /* Perform RCU core processing work for the current CPU. */
2456 static __latent_entropy void rcu_core(void)
2458 unsigned long flags;
2459 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2460 struct rcu_node *rnp = rdp->mynode;
2462 * On RT rcu_core() can be preempted when IRQs aren't disabled.
2463 * Therefore this function can race with concurrent NOCB (de-)offloading
2464 * on this CPU and the below condition must be considered volatile.
2465 * However if we race with:
2467 * _ Offloading: In the worst case we accelerate or process callbacks
2468 * concurrently with NOCB kthreads. We are guaranteed to
2469 * call rcu_nocb_lock() if that happens.
2471 * _ Deoffloading: In the worst case we miss callbacks acceleration or
2472 * processing. This is fine because the early stage
2473 * of deoffloading invokes rcu_core() after setting
2474 * SEGCBLIST_RCU_CORE. So we guarantee that we'll process
2475 * what could have been dismissed without the need to wait
2476 * for the next rcu_pending() check in the next jiffy.
2478 const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2480 if (cpu_is_offline(smp_processor_id()))
2482 trace_rcu_utilization(TPS("Start RCU core"));
2483 WARN_ON_ONCE(!rdp->beenonline);
2485 /* Report any deferred quiescent states if preemption enabled. */
2486 if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
2487 rcu_preempt_deferred_qs(current);
2488 } else if (rcu_preempt_need_deferred_qs(current)) {
2489 set_tsk_need_resched(current);
2490 set_preempt_need_resched();
2493 /* Update RCU state based on any recent quiescent states. */
2494 rcu_check_quiescent_state(rdp);
2496 /* No grace period and unregistered callbacks? */
2497 if (!rcu_gp_in_progress() &&
2498 rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2499 rcu_nocb_lock_irqsave(rdp, flags);
2500 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2501 rcu_accelerate_cbs_unlocked(rnp, rdp);
2502 rcu_nocb_unlock_irqrestore(rdp, flags);
2505 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2507 /* If there are callbacks ready, invoke them. */
2508 if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2509 likely(READ_ONCE(rcu_scheduler_fully_active))) {
2511 /* Re-invoke RCU core processing if there are callbacks remaining. */
2512 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2516 /* Do any needed deferred wakeups of rcuo kthreads. */
2517 do_nocb_deferred_wakeup(rdp);
2518 trace_rcu_utilization(TPS("End RCU core"));
2520 // If strict GPs, schedule an RCU reader in a clean environment.
2521 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2522 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2525 static void rcu_core_si(struct softirq_action *h)
2530 static void rcu_wake_cond(struct task_struct *t, int status)
2533 * If the thread is yielding, only wake it when this
2534 * is invoked from idle
2536 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2540 static void invoke_rcu_core_kthread(void)
2542 struct task_struct *t;
2543 unsigned long flags;
2545 local_irq_save(flags);
2546 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2547 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2548 if (t != NULL && t != current)
2549 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2550 local_irq_restore(flags);
2554 * Wake up this CPU's rcuc kthread to do RCU core processing.
2556 static void invoke_rcu_core(void)
2558 if (!cpu_online(smp_processor_id()))
2561 raise_softirq(RCU_SOFTIRQ);
2563 invoke_rcu_core_kthread();
2566 static void rcu_cpu_kthread_park(unsigned int cpu)
2568 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2571 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2573 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2577 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2578 * the RCU softirq used in configurations of RCU that do not support RCU
2579 * priority boosting.
2581 static void rcu_cpu_kthread(unsigned int cpu)
2583 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2584 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2585 unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
2588 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2589 for (spincnt = 0; spincnt < 10; spincnt++) {
2590 WRITE_ONCE(*j, jiffies);
2592 *statusp = RCU_KTHREAD_RUNNING;
2593 local_irq_disable();
2601 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2602 *statusp = RCU_KTHREAD_WAITING;
2606 *statusp = RCU_KTHREAD_YIELDING;
2607 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2608 schedule_timeout_idle(2);
2609 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2610 *statusp = RCU_KTHREAD_WAITING;
2611 WRITE_ONCE(*j, jiffies);
2614 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2615 .store = &rcu_data.rcu_cpu_kthread_task,
2616 .thread_should_run = rcu_cpu_kthread_should_run,
2617 .thread_fn = rcu_cpu_kthread,
2618 .thread_comm = "rcuc/%u",
2619 .setup = rcu_cpu_kthread_setup,
2620 .park = rcu_cpu_kthread_park,
2624 * Spawn per-CPU RCU core processing kthreads.
2626 static int __init rcu_spawn_core_kthreads(void)
2630 for_each_possible_cpu(cpu)
2631 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2634 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2635 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2640 * Handle any core-RCU processing required by a call_rcu() invocation.
2642 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2643 unsigned long flags)
2646 * If called from an extended quiescent state, invoke the RCU
2647 * core in order to force a re-evaluation of RCU's idleness.
2649 if (!rcu_is_watching())
2652 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2653 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2657 * Force the grace period if too many callbacks or too long waiting.
2658 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2659 * if some other CPU has recently done so. Also, don't bother
2660 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2661 * is the only one waiting for a grace period to complete.
2663 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2664 rdp->qlen_last_fqs_check + qhimark)) {
2666 /* Are we ignoring a completed grace period? */
2667 note_gp_changes(rdp);
2669 /* Start a new grace period if one not already started. */
2670 if (!rcu_gp_in_progress()) {
2671 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2673 /* Give the grace period a kick. */
2674 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2675 if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2676 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2677 rcu_force_quiescent_state();
2678 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2679 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2685 * RCU callback function to leak a callback.
2687 static void rcu_leak_callback(struct rcu_head *rhp)
2692 * Check and if necessary update the leaf rcu_node structure's
2693 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2694 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2695 * structure's ->lock.
2697 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2699 raw_lockdep_assert_held_rcu_node(rnp);
2700 if (qovld_calc <= 0)
2701 return; // Early boot and wildcard value set.
2702 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2703 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2705 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2709 * Check and if necessary update the leaf rcu_node structure's
2710 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2711 * number of queued RCU callbacks. No locks need be held, but the
2712 * caller must have disabled interrupts.
2714 * Note that this function ignores the possibility that there are a lot
2715 * of callbacks all of which have already seen the end of their respective
2716 * grace periods. This omission is due to the need for no-CBs CPUs to
2717 * be holding ->nocb_lock to do this check, which is too heavy for a
2718 * common-case operation.
2720 static void check_cb_ovld(struct rcu_data *rdp)
2722 struct rcu_node *const rnp = rdp->mynode;
2724 if (qovld_calc <= 0 ||
2725 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2726 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2727 return; // Early boot wildcard value or already set correctly.
2728 raw_spin_lock_rcu_node(rnp);
2729 check_cb_ovld_locked(rdp, rnp);
2730 raw_spin_unlock_rcu_node(rnp);
2734 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2735 * @head: structure to be used for queueing the RCU updates.
2736 * @func: actual callback function to be invoked after the grace period
2738 * The callback function will be invoked some time after a full grace
2739 * period elapses, in other words after all pre-existing RCU read-side
2740 * critical sections have completed. However, the callback function
2741 * might well execute concurrently with RCU read-side critical sections
2742 * that started after call_rcu() was invoked.
2744 * RCU read-side critical sections are delimited by rcu_read_lock()
2745 * and rcu_read_unlock(), and may be nested. In addition, but only in
2746 * v5.0 and later, regions of code across which interrupts, preemption,
2747 * or softirqs have been disabled also serve as RCU read-side critical
2748 * sections. This includes hardware interrupt handlers, softirq handlers,
2751 * Note that all CPUs must agree that the grace period extended beyond
2752 * all pre-existing RCU read-side critical section. On systems with more
2753 * than one CPU, this means that when "func()" is invoked, each CPU is
2754 * guaranteed to have executed a full memory barrier since the end of its
2755 * last RCU read-side critical section whose beginning preceded the call
2756 * to call_rcu(). It also means that each CPU executing an RCU read-side
2757 * critical section that continues beyond the start of "func()" must have
2758 * executed a memory barrier after the call_rcu() but before the beginning
2759 * of that RCU read-side critical section. Note that these guarantees
2760 * include CPUs that are offline, idle, or executing in user mode, as
2761 * well as CPUs that are executing in the kernel.
2763 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2764 * resulting RCU callback function "func()", then both CPU A and CPU B are
2765 * guaranteed to execute a full memory barrier during the time interval
2766 * between the call to call_rcu() and the invocation of "func()" -- even
2767 * if CPU A and CPU B are the same CPU (but again only if the system has
2768 * more than one CPU).
2770 * Implementation of these memory-ordering guarantees is described here:
2771 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
2773 void call_rcu(struct rcu_head *head, rcu_callback_t func)
2775 static atomic_t doublefrees;
2776 unsigned long flags;
2777 struct rcu_data *rdp;
2780 /* Misaligned rcu_head! */
2781 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2783 if (debug_rcu_head_queue(head)) {
2785 * Probable double call_rcu(), so leak the callback.
2786 * Use rcu:rcu_callback trace event to find the previous
2787 * time callback was passed to call_rcu().
2789 if (atomic_inc_return(&doublefrees) < 4) {
2790 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
2793 WRITE_ONCE(head->func, rcu_leak_callback);
2798 kasan_record_aux_stack_noalloc(head);
2799 local_irq_save(flags);
2800 rdp = this_cpu_ptr(&rcu_data);
2802 /* Add the callback to our list. */
2803 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2804 // This can trigger due to call_rcu() from offline CPU:
2805 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2806 WARN_ON_ONCE(!rcu_is_watching());
2807 // Very early boot, before rcu_init(). Initialize if needed
2808 // and then drop through to queue the callback.
2809 if (rcu_segcblist_empty(&rdp->cblist))
2810 rcu_segcblist_init(&rdp->cblist);
2814 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2815 return; // Enqueued onto ->nocb_bypass, so just leave.
2816 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2817 rcu_segcblist_enqueue(&rdp->cblist, head);
2818 if (__is_kvfree_rcu_offset((unsigned long)func))
2819 trace_rcu_kvfree_callback(rcu_state.name, head,
2820 (unsigned long)func,
2821 rcu_segcblist_n_cbs(&rdp->cblist));
2823 trace_rcu_callback(rcu_state.name, head,
2824 rcu_segcblist_n_cbs(&rdp->cblist));
2826 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
2828 /* Go handle any RCU core processing required. */
2829 if (unlikely(rcu_rdp_is_offloaded(rdp))) {
2830 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2832 __call_rcu_core(rdp, head, flags);
2833 local_irq_restore(flags);
2836 EXPORT_SYMBOL_GPL(call_rcu);
2839 /* Maximum number of jiffies to wait before draining a batch. */
2840 #define KFREE_DRAIN_JIFFIES (5 * HZ)
2841 #define KFREE_N_BATCHES 2
2842 #define FREE_N_CHANNELS 2
2845 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2846 * @nr_records: Number of active pointers in the array
2847 * @next: Next bulk object in the block chain
2848 * @records: Array of the kvfree_rcu() pointers
2850 struct kvfree_rcu_bulk_data {
2851 unsigned long nr_records;
2852 struct kvfree_rcu_bulk_data *next;
2857 * This macro defines how many entries the "records" array
2858 * will contain. It is based on the fact that the size of
2859 * kvfree_rcu_bulk_data structure becomes exactly one page.
2861 #define KVFREE_BULK_MAX_ENTR \
2862 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2865 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2866 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2867 * @head_free: List of kfree_rcu() objects waiting for a grace period
2868 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
2869 * @krcp: Pointer to @kfree_rcu_cpu structure
2872 struct kfree_rcu_cpu_work {
2873 struct rcu_work rcu_work;
2874 struct rcu_head *head_free;
2875 struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
2876 struct kfree_rcu_cpu *krcp;
2880 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2881 * @head: List of kfree_rcu() objects not yet waiting for a grace period
2882 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
2883 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2884 * @lock: Synchronize access to this structure
2885 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
2886 * @initialized: The @rcu_work fields have been initialized
2887 * @count: Number of objects for which GP not started
2889 * A simple cache list that contains objects for reuse purpose.
2890 * In order to save some per-cpu space the list is singular.
2891 * Even though it is lockless an access has to be protected by the
2893 * @page_cache_work: A work to refill the cache when it is empty
2894 * @backoff_page_cache_fill: Delay cache refills
2895 * @work_in_progress: Indicates that page_cache_work is running
2896 * @hrtimer: A hrtimer for scheduling a page_cache_work
2897 * @nr_bkv_objs: number of allocated objects at @bkvcache.
2899 * This is a per-CPU structure. The reason that it is not included in
2900 * the rcu_data structure is to permit this code to be extracted from
2901 * the RCU files. Such extraction could allow further optimization of
2902 * the interactions with the slab allocators.
2904 struct kfree_rcu_cpu {
2905 struct rcu_head *head;
2906 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
2907 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
2908 raw_spinlock_t lock;
2909 struct delayed_work monitor_work;
2913 struct delayed_work page_cache_work;
2914 atomic_t backoff_page_cache_fill;
2915 atomic_t work_in_progress;
2916 struct hrtimer hrtimer;
2918 struct llist_head bkvcache;
2922 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
2923 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
2926 static __always_inline void
2927 debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
2929 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
2932 for (i = 0; i < bhead->nr_records; i++)
2933 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
2937 static inline struct kfree_rcu_cpu *
2938 krc_this_cpu_lock(unsigned long *flags)
2940 struct kfree_rcu_cpu *krcp;
2942 local_irq_save(*flags); // For safely calling this_cpu_ptr().
2943 krcp = this_cpu_ptr(&krc);
2944 raw_spin_lock(&krcp->lock);
2950 krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
2952 raw_spin_unlock_irqrestore(&krcp->lock, flags);
2955 static inline struct kvfree_rcu_bulk_data *
2956 get_cached_bnode(struct kfree_rcu_cpu *krcp)
2958 if (!krcp->nr_bkv_objs)
2961 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
2962 return (struct kvfree_rcu_bulk_data *)
2963 llist_del_first(&krcp->bkvcache);
2967 put_cached_bnode(struct kfree_rcu_cpu *krcp,
2968 struct kvfree_rcu_bulk_data *bnode)
2971 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
2974 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
2975 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
2980 drain_page_cache(struct kfree_rcu_cpu *krcp)
2982 unsigned long flags;
2983 struct llist_node *page_list, *pos, *n;
2986 raw_spin_lock_irqsave(&krcp->lock, flags);
2987 page_list = llist_del_all(&krcp->bkvcache);
2988 WRITE_ONCE(krcp->nr_bkv_objs, 0);
2989 raw_spin_unlock_irqrestore(&krcp->lock, flags);
2991 llist_for_each_safe(pos, n, page_list) {
2992 free_page((unsigned long)pos);
3000 * This function is invoked in workqueue context after a grace period.
3001 * It frees all the objects queued on ->bkvhead_free or ->head_free.
3003 static void kfree_rcu_work(struct work_struct *work)
3005 unsigned long flags;
3006 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3007 struct rcu_head *head, *next;
3008 struct kfree_rcu_cpu *krcp;
3009 struct kfree_rcu_cpu_work *krwp;
3012 krwp = container_of(to_rcu_work(work),
3013 struct kfree_rcu_cpu_work, rcu_work);
3016 raw_spin_lock_irqsave(&krcp->lock, flags);
3017 // Channels 1 and 2.
3018 for (i = 0; i < FREE_N_CHANNELS; i++) {
3019 bkvhead[i] = krwp->bkvhead_free[i];
3020 krwp->bkvhead_free[i] = NULL;
3024 head = krwp->head_free;
3025 krwp->head_free = NULL;
3026 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3028 // Handle the first two channels.
3029 for (i = 0; i < FREE_N_CHANNELS; i++) {
3030 for (; bkvhead[i]; bkvhead[i] = bnext) {
3031 bnext = bkvhead[i]->next;
3032 debug_rcu_bhead_unqueue(bkvhead[i]);
3034 rcu_lock_acquire(&rcu_callback_map);
3035 if (i == 0) { // kmalloc() / kfree().
3036 trace_rcu_invoke_kfree_bulk_callback(
3037 rcu_state.name, bkvhead[i]->nr_records,
3038 bkvhead[i]->records);
3040 kfree_bulk(bkvhead[i]->nr_records,
3041 bkvhead[i]->records);
3042 } else { // vmalloc() / vfree().
3043 for (j = 0; j < bkvhead[i]->nr_records; j++) {
3044 trace_rcu_invoke_kvfree_callback(
3046 bkvhead[i]->records[j], 0);
3048 vfree(bkvhead[i]->records[j]);
3051 rcu_lock_release(&rcu_callback_map);
3053 raw_spin_lock_irqsave(&krcp->lock, flags);
3054 if (put_cached_bnode(krcp, bkvhead[i]))
3056 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3059 free_page((unsigned long) bkvhead[i]);
3061 cond_resched_tasks_rcu_qs();
3066 * This is used when the "bulk" path can not be used for the
3067 * double-argument of kvfree_rcu(). This happens when the
3068 * page-cache is empty, which means that objects are instead
3069 * queued on a linked list through their rcu_head structures.
3070 * This list is named "Channel 3".
3072 for (; head; head = next) {
3073 unsigned long offset = (unsigned long)head->func;
3074 void *ptr = (void *)head - offset;
3077 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3078 rcu_lock_acquire(&rcu_callback_map);
3079 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3081 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3084 rcu_lock_release(&rcu_callback_map);
3085 cond_resched_tasks_rcu_qs();
3090 need_offload_krc(struct kfree_rcu_cpu *krcp)
3094 for (i = 0; i < FREE_N_CHANNELS; i++)
3095 if (krcp->bkvhead[i])
3098 return !!krcp->head;
3102 schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3104 long delay, delay_left;
3106 delay = READ_ONCE(krcp->count) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
3107 if (delayed_work_pending(&krcp->monitor_work)) {
3108 delay_left = krcp->monitor_work.timer.expires - jiffies;
3109 if (delay < delay_left)
3110 mod_delayed_work(system_wq, &krcp->monitor_work, delay);
3113 queue_delayed_work(system_wq, &krcp->monitor_work, delay);
3117 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3119 static void kfree_rcu_monitor(struct work_struct *work)
3121 struct kfree_rcu_cpu *krcp = container_of(work,
3122 struct kfree_rcu_cpu, monitor_work.work);
3123 unsigned long flags;
3126 raw_spin_lock_irqsave(&krcp->lock, flags);
3128 // Attempt to start a new batch.
3129 for (i = 0; i < KFREE_N_BATCHES; i++) {
3130 struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3132 // Try to detach bkvhead or head and attach it over any
3133 // available corresponding free channel. It can be that
3134 // a previous RCU batch is in progress, it means that
3135 // immediately to queue another one is not possible so
3136 // in that case the monitor work is rearmed.
3137 if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3138 (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3139 (krcp->head && !krwp->head_free)) {
3140 // Channel 1 corresponds to the SLAB-pointer bulk path.
3141 // Channel 2 corresponds to vmalloc-pointer bulk path.
3142 for (j = 0; j < FREE_N_CHANNELS; j++) {
3143 if (!krwp->bkvhead_free[j]) {
3144 krwp->bkvhead_free[j] = krcp->bkvhead[j];
3145 krcp->bkvhead[j] = NULL;
3149 // Channel 3 corresponds to both SLAB and vmalloc
3150 // objects queued on the linked list.
3151 if (!krwp->head_free) {
3152 krwp->head_free = krcp->head;
3156 WRITE_ONCE(krcp->count, 0);
3158 // One work is per one batch, so there are three
3159 // "free channels", the batch can handle. It can
3160 // be that the work is in the pending state when
3161 // channels have been detached following by each
3163 queue_rcu_work(system_wq, &krwp->rcu_work);
3167 // If there is nothing to detach, it means that our job is
3168 // successfully done here. In case of having at least one
3169 // of the channels that is still busy we should rearm the
3170 // work to repeat an attempt. Because previous batches are
3171 // still in progress.
3172 if (need_offload_krc(krcp))
3173 schedule_delayed_monitor_work(krcp);
3175 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3178 static enum hrtimer_restart
3179 schedule_page_work_fn(struct hrtimer *t)
3181 struct kfree_rcu_cpu *krcp =
3182 container_of(t, struct kfree_rcu_cpu, hrtimer);
3184 queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3185 return HRTIMER_NORESTART;
3188 static void fill_page_cache_func(struct work_struct *work)
3190 struct kvfree_rcu_bulk_data *bnode;
3191 struct kfree_rcu_cpu *krcp =
3192 container_of(work, struct kfree_rcu_cpu,
3193 page_cache_work.work);
3194 unsigned long flags;
3199 nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3200 1 : rcu_min_cached_objs;
3202 for (i = 0; i < nr_pages; i++) {
3203 bnode = (struct kvfree_rcu_bulk_data *)
3204 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3209 raw_spin_lock_irqsave(&krcp->lock, flags);
3210 pushed = put_cached_bnode(krcp, bnode);
3211 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3214 free_page((unsigned long) bnode);
3219 atomic_set(&krcp->work_in_progress, 0);
3220 atomic_set(&krcp->backoff_page_cache_fill, 0);
3224 run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3226 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3227 !atomic_xchg(&krcp->work_in_progress, 1)) {
3228 if (atomic_read(&krcp->backoff_page_cache_fill)) {
3229 queue_delayed_work(system_wq,
3230 &krcp->page_cache_work,
3231 msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3233 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3234 krcp->hrtimer.function = schedule_page_work_fn;
3235 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3240 // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3241 // state specified by flags. If can_alloc is true, the caller must
3242 // be schedulable and not be holding any locks or mutexes that might be
3243 // acquired by the memory allocator or anything that it might invoke.
3244 // Returns true if ptr was successfully recorded, else the caller must
3247 add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3248 unsigned long *flags, void *ptr, bool can_alloc)
3250 struct kvfree_rcu_bulk_data *bnode;
3253 *krcp = krc_this_cpu_lock(flags);
3254 if (unlikely(!(*krcp)->initialized))
3257 idx = !!is_vmalloc_addr(ptr);
3259 /* Check if a new block is required. */
3260 if (!(*krcp)->bkvhead[idx] ||
3261 (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3262 bnode = get_cached_bnode(*krcp);
3263 if (!bnode && can_alloc) {
3264 krc_this_cpu_unlock(*krcp, *flags);
3266 // __GFP_NORETRY - allows a light-weight direct reclaim
3267 // what is OK from minimizing of fallback hitting point of
3268 // view. Apart of that it forbids any OOM invoking what is
3269 // also beneficial since we are about to release memory soon.
3271 // __GFP_NOMEMALLOC - prevents from consuming of all the
3272 // memory reserves. Please note we have a fallback path.
3274 // __GFP_NOWARN - it is supposed that an allocation can
3275 // be failed under low memory or high memory pressure
3277 bnode = (struct kvfree_rcu_bulk_data *)
3278 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3279 *krcp = krc_this_cpu_lock(flags);
3285 /* Initialize the new block. */
3286 bnode->nr_records = 0;
3287 bnode->next = (*krcp)->bkvhead[idx];
3289 /* Attach it to the head. */
3290 (*krcp)->bkvhead[idx] = bnode;
3293 /* Finally insert. */
3294 (*krcp)->bkvhead[idx]->records
3295 [(*krcp)->bkvhead[idx]->nr_records++] = ptr;
3301 * Queue a request for lazy invocation of the appropriate free routine
3302 * after a grace period. Please note that three paths are maintained,
3303 * two for the common case using arrays of pointers and a third one that
3304 * is used only when the main paths cannot be used, for example, due to
3307 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3308 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3309 * be free'd in workqueue context. This allows us to: batch requests together to
3310 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3312 void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3314 unsigned long flags;
3315 struct kfree_rcu_cpu *krcp;
3320 ptr = (void *) head - (unsigned long) func;
3323 * Please note there is a limitation for the head-less
3324 * variant, that is why there is a clear rule for such
3325 * objects: it can be used from might_sleep() context
3326 * only. For other places please embed an rcu_head to
3330 ptr = (unsigned long *) func;
3333 // Queue the object but don't yet schedule the batch.
3334 if (debug_rcu_head_queue(ptr)) {
3335 // Probable double kfree_rcu(), just leak.
3336 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3339 // Mark as success and leave.
3343 kasan_record_aux_stack_noalloc(ptr);
3344 success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3346 run_page_cache_worker(krcp);
3349 // Inline if kvfree_rcu(one_arg) call.
3353 head->next = krcp->head;
3358 WRITE_ONCE(krcp->count, krcp->count + 1);
3360 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3361 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
3362 schedule_delayed_monitor_work(krcp);
3365 krc_this_cpu_unlock(krcp, flags);
3368 * Inline kvfree() after synchronize_rcu(). We can do
3369 * it from might_sleep() context only, so the current
3370 * CPU can pass the QS state.
3373 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3378 EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3380 static unsigned long
3381 kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3384 unsigned long count = 0;
3386 /* Snapshot count of all CPUs */
3387 for_each_possible_cpu(cpu) {
3388 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3390 count += READ_ONCE(krcp->count);
3391 count += READ_ONCE(krcp->nr_bkv_objs);
3392 atomic_set(&krcp->backoff_page_cache_fill, 1);
3395 return count == 0 ? SHRINK_EMPTY : count;
3398 static unsigned long
3399 kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3403 for_each_possible_cpu(cpu) {
3405 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3407 count = krcp->count;
3408 count += drain_page_cache(krcp);
3409 kfree_rcu_monitor(&krcp->monitor_work.work);
3411 sc->nr_to_scan -= count;
3414 if (sc->nr_to_scan <= 0)
3418 return freed == 0 ? SHRINK_STOP : freed;
3421 static struct shrinker kfree_rcu_shrinker = {
3422 .count_objects = kfree_rcu_shrink_count,
3423 .scan_objects = kfree_rcu_shrink_scan,
3425 .seeks = DEFAULT_SEEKS,
3428 void __init kfree_rcu_scheduler_running(void)
3431 unsigned long flags;
3433 for_each_possible_cpu(cpu) {
3434 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3436 raw_spin_lock_irqsave(&krcp->lock, flags);
3437 if (need_offload_krc(krcp))
3438 schedule_delayed_monitor_work(krcp);
3439 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3444 * During early boot, any blocking grace-period wait automatically
3445 * implies a grace period.
3447 * Later on, this could in theory be the case for kernels built with
3448 * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
3449 * is not a common case. Furthermore, this optimization would cause
3450 * the rcu_gp_oldstate structure to expand by 50%, so this potential
3451 * grace-period optimization is ignored once the scheduler is running.
3453 static int rcu_blocking_is_gp(void)
3455 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
3457 might_sleep(); /* Check for RCU read-side critical section. */
3462 * synchronize_rcu - wait until a grace period has elapsed.
3464 * Control will return to the caller some time after a full grace
3465 * period has elapsed, in other words after all currently executing RCU
3466 * read-side critical sections have completed. Note, however, that
3467 * upon return from synchronize_rcu(), the caller might well be executing
3468 * concurrently with new RCU read-side critical sections that began while
3469 * synchronize_rcu() was waiting.
3471 * RCU read-side critical sections are delimited by rcu_read_lock()
3472 * and rcu_read_unlock(), and may be nested. In addition, but only in
3473 * v5.0 and later, regions of code across which interrupts, preemption,
3474 * or softirqs have been disabled also serve as RCU read-side critical
3475 * sections. This includes hardware interrupt handlers, softirq handlers,
3478 * Note that this guarantee implies further memory-ordering guarantees.
3479 * On systems with more than one CPU, when synchronize_rcu() returns,
3480 * each CPU is guaranteed to have executed a full memory barrier since
3481 * the end of its last RCU read-side critical section whose beginning
3482 * preceded the call to synchronize_rcu(). In addition, each CPU having
3483 * an RCU read-side critical section that extends beyond the return from
3484 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3485 * after the beginning of synchronize_rcu() and before the beginning of
3486 * that RCU read-side critical section. Note that these guarantees include
3487 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3488 * that are executing in the kernel.
3490 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3491 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3492 * to have executed a full memory barrier during the execution of
3493 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3494 * again only if the system has more than one CPU).
3496 * Implementation of these memory-ordering guarantees is described here:
3497 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3499 void synchronize_rcu(void)
3501 unsigned long flags;
3502 struct rcu_node *rnp;
3504 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3505 lock_is_held(&rcu_lock_map) ||
3506 lock_is_held(&rcu_sched_lock_map),
3507 "Illegal synchronize_rcu() in RCU read-side critical section");
3508 if (!rcu_blocking_is_gp()) {
3509 if (rcu_gp_is_expedited())
3510 synchronize_rcu_expedited();
3512 wait_rcu_gp(call_rcu);
3516 // Context allows vacuous grace periods.
3517 // Note well that this code runs with !PREEMPT && !SMP.
3518 // In addition, all code that advances grace periods runs at
3519 // process level. Therefore, this normal GP overlaps with other
3520 // normal GPs only by being fully nested within them, which allows
3521 // reuse of ->gp_seq_polled_snap.
3522 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
3523 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
3525 // Update the normal grace-period counters to record
3526 // this grace period, but only those used by the boot CPU.
3527 // The rcu_scheduler_starting() will take care of the rest of
3529 local_irq_save(flags);
3530 WARN_ON_ONCE(num_online_cpus() > 1);
3531 rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
3532 for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
3533 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
3534 local_irq_restore(flags);
3536 EXPORT_SYMBOL_GPL(synchronize_rcu);
3539 * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
3540 * @rgosp: Place to put state cookie
3542 * Stores into @rgosp a value that will always be treated by functions
3543 * like poll_state_synchronize_rcu_full() as a cookie whose grace period
3544 * has already completed.
3546 void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3548 rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
3549 rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
3551 EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
3554 * get_state_synchronize_rcu - Snapshot current RCU state
3556 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3557 * or poll_state_synchronize_rcu() to determine whether or not a full
3558 * grace period has elapsed in the meantime.
3560 unsigned long get_state_synchronize_rcu(void)
3563 * Any prior manipulation of RCU-protected data must happen
3564 * before the load from ->gp_seq.
3567 return rcu_seq_snap(&rcu_state.gp_seq_polled);
3569 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3572 * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
3573 * @rgosp: location to place combined normal/expedited grace-period state
3575 * Places the normal and expedited grace-period states in @rgosp. This
3576 * state value can be passed to a later call to cond_synchronize_rcu_full()
3577 * or poll_state_synchronize_rcu_full() to determine whether or not a
3578 * grace period (whether normal or expedited) has elapsed in the meantime.
3579 * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
3580 * long, but is guaranteed to see all grace periods. In contrast, the
3581 * combined state occupies less memory, but can sometimes fail to take
3582 * grace periods into account.
3584 * This does not guarantee that the needed grace period will actually
3587 void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3589 struct rcu_node *rnp = rcu_get_root();
3592 * Any prior manipulation of RCU-protected data must happen
3593 * before the loads from ->gp_seq and ->expedited_sequence.
3596 rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
3597 rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
3599 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
3602 * Helper function for start_poll_synchronize_rcu() and
3603 * start_poll_synchronize_rcu_full().
3605 static void start_poll_synchronize_rcu_common(void)
3607 unsigned long flags;
3609 struct rcu_data *rdp;
3610 struct rcu_node *rnp;
3612 lockdep_assert_irqs_enabled();
3613 local_irq_save(flags);
3614 rdp = this_cpu_ptr(&rcu_data);
3616 raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3617 // Note it is possible for a grace period to have elapsed between
3618 // the above call to get_state_synchronize_rcu() and the below call
3619 // to rcu_seq_snap. This is OK, the worst that happens is that we
3620 // get a grace period that no one needed. These accesses are ordered
3621 // by smp_mb(), and we are accessing them in the opposite order
3622 // from which they are updated at grace-period start, as required.
3623 needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
3624 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3626 rcu_gp_kthread_wake();
3630 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3632 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3633 * or poll_state_synchronize_rcu() to determine whether or not a full
3634 * grace period has elapsed in the meantime. If the needed grace period
3635 * is not already slated to start, notifies RCU core of the need for that
3638 * Interrupts must be enabled for the case where it is necessary to awaken
3639 * the grace-period kthread.
3641 unsigned long start_poll_synchronize_rcu(void)
3643 unsigned long gp_seq = get_state_synchronize_rcu();
3645 start_poll_synchronize_rcu_common();
3648 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3651 * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
3652 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3654 * Places the normal and expedited grace-period states in *@rgos. This
3655 * state value can be passed to a later call to cond_synchronize_rcu_full()
3656 * or poll_state_synchronize_rcu_full() to determine whether or not a
3657 * grace period (whether normal or expedited) has elapsed in the meantime.
3658 * If the needed grace period is not already slated to start, notifies
3659 * RCU core of the need for that grace period.
3661 * Interrupts must be enabled for the case where it is necessary to awaken
3662 * the grace-period kthread.
3664 void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3666 get_state_synchronize_rcu_full(rgosp);
3668 start_poll_synchronize_rcu_common();
3670 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
3673 * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
3674 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3676 * If a full RCU grace period has elapsed since the earlier call from
3677 * which @oldstate was obtained, return @true, otherwise return @false.
3678 * If @false is returned, it is the caller's responsibility to invoke this
3679 * function later on until it does return @true. Alternatively, the caller
3680 * can explicitly wait for a grace period, for example, by passing @oldstate
3681 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3683 * Yes, this function does not take counter wrap into account.
3684 * But counter wrap is harmless. If the counter wraps, we have waited for
3685 * more than a billion grace periods (and way more on a 64-bit system!).
3686 * Those needing to keep old state values for very long time periods
3687 * (many hours even on 32-bit systems) should check them occasionally and
3688 * either refresh them or set a flag indicating that the grace period has
3689 * completed. Alternatively, they can use get_completed_synchronize_rcu()
3690 * to get a guaranteed-completed grace-period state.
3692 * This function provides the same memory-ordering guarantees that
3693 * would be provided by a synchronize_rcu() that was invoked at the call
3694 * to the function that provided @oldstate, and that returned at the end
3697 bool poll_state_synchronize_rcu(unsigned long oldstate)
3699 if (oldstate == RCU_GET_STATE_COMPLETED ||
3700 rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) {
3701 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3706 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3709 * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
3710 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3712 * If a full RCU grace period has elapsed since the earlier call from
3713 * which *rgosp was obtained, return @true, otherwise return @false.
3714 * If @false is returned, it is the caller's responsibility to invoke this
3715 * function later on until it does return @true. Alternatively, the caller
3716 * can explicitly wait for a grace period, for example, by passing @rgosp
3717 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3719 * Yes, this function does not take counter wrap into account.
3720 * But counter wrap is harmless. If the counter wraps, we have waited
3721 * for more than a billion grace periods (and way more on a 64-bit
3722 * system!). Those needing to keep rcu_gp_oldstate values for very
3723 * long time periods (many hours even on 32-bit systems) should check
3724 * them occasionally and either refresh them or set a flag indicating
3725 * that the grace period has completed. Alternatively, they can use
3726 * get_completed_synchronize_rcu_full() to get a guaranteed-completed
3727 * grace-period state.
3729 * This function provides the same memory-ordering guarantees that would
3730 * be provided by a synchronize_rcu() that was invoked at the call to
3731 * the function that provided @rgosp, and that returned at the end of this
3732 * function. And this guarantee requires that the root rcu_node structure's
3733 * ->gp_seq field be checked instead of that of the rcu_state structure.
3734 * The problem is that the just-ending grace-period's callbacks can be
3735 * invoked between the time that the root rcu_node structure's ->gp_seq
3736 * field is updated and the time that the rcu_state structure's ->gp_seq
3737 * field is updated. Therefore, if a single synchronize_rcu() is to
3738 * cause a subsequent poll_state_synchronize_rcu_full() to return @true,
3739 * then the root rcu_node structure is the one that needs to be polled.
3741 bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3743 struct rcu_node *rnp = rcu_get_root();
3745 smp_mb(); // Order against root rcu_node structure grace-period cleanup.
3746 if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED ||
3747 rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) ||
3748 rgosp->rgos_exp == RCU_GET_STATE_COMPLETED ||
3749 rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) {
3750 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3755 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full);
3758 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3759 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
3761 * If a full RCU grace period has elapsed since the earlier call to
3762 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3763 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3765 * Yes, this function does not take counter wrap into account.
3766 * But counter wrap is harmless. If the counter wraps, we have waited for
3767 * more than 2 billion grace periods (and way more on a 64-bit system!),
3768 * so waiting for a couple of additional grace periods should be just fine.
3770 * This function provides the same memory-ordering guarantees that
3771 * would be provided by a synchronize_rcu() that was invoked at the call
3772 * to the function that provided @oldstate and that returned at the end
3775 void cond_synchronize_rcu(unsigned long oldstate)
3777 if (!poll_state_synchronize_rcu(oldstate))
3780 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3783 * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
3784 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
3786 * If a full RCU grace period has elapsed since the call to
3787 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
3788 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
3789 * obtained, just return. Otherwise, invoke synchronize_rcu() to wait
3790 * for a full grace period.
3792 * Yes, this function does not take counter wrap into account.
3793 * But counter wrap is harmless. If the counter wraps, we have waited for
3794 * more than 2 billion grace periods (and way more on a 64-bit system!),
3795 * so waiting for a couple of additional grace periods should be just fine.
3797 * This function provides the same memory-ordering guarantees that
3798 * would be provided by a synchronize_rcu() that was invoked at the call
3799 * to the function that provided @rgosp and that returned at the end of
3802 void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3804 if (!poll_state_synchronize_rcu_full(rgosp))
3807 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full);
3810 * Check to see if there is any immediate RCU-related work to be done by
3811 * the current CPU, returning 1 if so and zero otherwise. The checks are
3812 * in order of increasing expense: checks that can be carried out against
3813 * CPU-local state are performed first. However, we must check for CPU
3814 * stalls first, else we might not get a chance.
3816 static int rcu_pending(int user)
3818 bool gp_in_progress;
3819 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3820 struct rcu_node *rnp = rdp->mynode;
3822 lockdep_assert_irqs_disabled();
3824 /* Check for CPU stalls, if enabled. */
3825 check_cpu_stall(rdp);
3827 /* Does this CPU need a deferred NOCB wakeup? */
3828 if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3831 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3832 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3835 /* Is the RCU core waiting for a quiescent state from this CPU? */
3836 gp_in_progress = rcu_gp_in_progress();
3837 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3840 /* Does this CPU have callbacks ready to invoke? */
3841 if (!rcu_rdp_is_offloaded(rdp) &&
3842 rcu_segcblist_ready_cbs(&rdp->cblist))
3845 /* Has RCU gone idle with this CPU needing another grace period? */
3846 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3847 !rcu_rdp_is_offloaded(rdp) &&
3848 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3851 /* Have RCU grace period completed or started? */
3852 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3853 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3861 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3862 * the compiler is expected to optimize this away.
3864 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3866 trace_rcu_barrier(rcu_state.name, s, cpu,
3867 atomic_read(&rcu_state.barrier_cpu_count), done);
3871 * RCU callback function for rcu_barrier(). If we are last, wake
3872 * up the task executing rcu_barrier().
3874 * Note that the value of rcu_state.barrier_sequence must be captured
3875 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3876 * other CPUs might count the value down to zero before this CPU gets
3877 * around to invoking rcu_barrier_trace(), which might result in bogus
3878 * data from the next instance of rcu_barrier().
3880 static void rcu_barrier_callback(struct rcu_head *rhp)
3882 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3884 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3885 rcu_barrier_trace(TPS("LastCB"), -1, s);
3886 complete(&rcu_state.barrier_completion);
3888 rcu_barrier_trace(TPS("CB"), -1, s);
3893 * If needed, entrain an rcu_barrier() callback on rdp->cblist.
3895 static void rcu_barrier_entrain(struct rcu_data *rdp)
3897 unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence);
3898 unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap);
3900 lockdep_assert_held(&rcu_state.barrier_lock);
3901 if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq))
3903 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3904 rdp->barrier_head.func = rcu_barrier_callback;
3905 debug_rcu_head_queue(&rdp->barrier_head);
3907 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
3908 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
3909 atomic_inc(&rcu_state.barrier_cpu_count);
3911 debug_rcu_head_unqueue(&rdp->barrier_head);
3912 rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence);
3914 rcu_nocb_unlock(rdp);
3915 smp_store_release(&rdp->barrier_seq_snap, gseq);
3919 * Called with preemption disabled, and from cross-cpu IRQ context.
3921 static void rcu_barrier_handler(void *cpu_in)
3923 uintptr_t cpu = (uintptr_t)cpu_in;
3924 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3926 lockdep_assert_irqs_disabled();
3927 WARN_ON_ONCE(cpu != rdp->cpu);
3928 WARN_ON_ONCE(cpu != smp_processor_id());
3929 raw_spin_lock(&rcu_state.barrier_lock);
3930 rcu_barrier_entrain(rdp);
3931 raw_spin_unlock(&rcu_state.barrier_lock);
3935 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3937 * Note that this primitive does not necessarily wait for an RCU grace period
3938 * to complete. For example, if there are no RCU callbacks queued anywhere
3939 * in the system, then rcu_barrier() is within its rights to return
3940 * immediately, without waiting for anything, much less an RCU grace period.
3942 void rcu_barrier(void)
3945 unsigned long flags;
3947 struct rcu_data *rdp;
3948 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3950 rcu_barrier_trace(TPS("Begin"), -1, s);
3952 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3953 mutex_lock(&rcu_state.barrier_mutex);
3955 /* Did someone else do our work for us? */
3956 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
3957 rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence);
3958 smp_mb(); /* caller's subsequent code after above check. */
3959 mutex_unlock(&rcu_state.barrier_mutex);
3963 /* Mark the start of the barrier operation. */
3964 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
3965 rcu_seq_start(&rcu_state.barrier_sequence);
3966 gseq = rcu_state.barrier_sequence;
3967 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
3970 * Initialize the count to two rather than to zero in order
3971 * to avoid a too-soon return to zero in case of an immediate
3972 * invocation of the just-enqueued callback (or preemption of
3973 * this task). Exclude CPU-hotplug operations to ensure that no
3974 * offline non-offloaded CPU has callbacks queued.
3976 init_completion(&rcu_state.barrier_completion);
3977 atomic_set(&rcu_state.barrier_cpu_count, 2);
3978 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
3981 * Force each CPU with callbacks to register a new callback.
3982 * When that callback is invoked, we will know that all of the
3983 * corresponding CPU's preceding callbacks have been invoked.
3985 for_each_possible_cpu(cpu) {
3986 rdp = per_cpu_ptr(&rcu_data, cpu);
3988 if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq)
3990 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
3991 if (!rcu_segcblist_n_cbs(&rdp->cblist)) {
3992 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
3993 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
3994 rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence);
3997 if (!rcu_rdp_cpu_online(rdp)) {
3998 rcu_barrier_entrain(rdp);
3999 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4000 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4001 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence);
4004 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4005 if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) {
4006 schedule_timeout_uninterruptible(1);
4009 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4010 rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence);
4014 * Now that we have an rcu_barrier_callback() callback on each
4015 * CPU, and thus each counted, remove the initial count.
4017 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4018 complete(&rcu_state.barrier_completion);
4020 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4021 wait_for_completion(&rcu_state.barrier_completion);
4023 /* Mark the end of the barrier operation. */
4024 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4025 rcu_seq_end(&rcu_state.barrier_sequence);
4026 gseq = rcu_state.barrier_sequence;
4027 for_each_possible_cpu(cpu) {
4028 rdp = per_cpu_ptr(&rcu_data, cpu);
4030 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4033 /* Other rcu_barrier() invocations can now safely proceed. */
4034 mutex_unlock(&rcu_state.barrier_mutex);
4036 EXPORT_SYMBOL_GPL(rcu_barrier);
4039 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4040 * first CPU in a given leaf rcu_node structure coming online. The caller
4041 * must hold the corresponding leaf rcu_node ->lock with interrupts
4044 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4048 struct rcu_node *rnp = rnp_leaf;
4050 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4051 WARN_ON_ONCE(rnp->wait_blkd_tasks);
4053 mask = rnp->grpmask;
4057 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4058 oldmask = rnp->qsmaskinit;
4059 rnp->qsmaskinit |= mask;
4060 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4067 * Do boot-time initialization of a CPU's per-CPU RCU data.
4070 rcu_boot_init_percpu_data(int cpu)
4072 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
4073 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4075 /* Set up local state, ensuring consistent view of global state. */
4076 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4077 INIT_WORK(&rdp->strict_work, strict_work_handler);
4078 WARN_ON_ONCE(ct->dynticks_nesting != 1);
4079 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)));
4080 rdp->barrier_seq_snap = rcu_state.barrier_sequence;
4081 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4082 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4083 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4084 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4085 rdp->last_sched_clock = jiffies;
4087 rcu_boot_init_nocb_percpu_data(rdp);
4091 * Invoked early in the CPU-online process, when pretty much all services
4092 * are available. The incoming CPU is not present.
4094 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4095 * offline event can be happening at a given time. Note also that we can
4096 * accept some slop in the rsp->gp_seq access due to the fact that this
4097 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4098 * And any offloaded callbacks are being numbered elsewhere.
4100 int rcutree_prepare_cpu(unsigned int cpu)
4102 unsigned long flags;
4103 struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
4104 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4105 struct rcu_node *rnp = rcu_get_root();
4107 /* Set up local state, ensuring consistent view of global state. */
4108 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4109 rdp->qlen_last_fqs_check = 0;
4110 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4111 rdp->blimit = blimit;
4112 ct->dynticks_nesting = 1; /* CPU not up, no tearing. */
4113 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4116 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4119 if (!rcu_segcblist_is_enabled(&rdp->cblist))
4120 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4123 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4124 * propagation up the rcu_node tree will happen at the beginning
4125 * of the next grace period.
4128 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4129 rdp->beenonline = true; /* We have now been online. */
4130 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4131 rdp->gp_seq_needed = rdp->gp_seq;
4132 rdp->cpu_no_qs.b.norm = true;
4133 rdp->core_needs_qs = false;
4134 rdp->rcu_iw_pending = false;
4135 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4136 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4137 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4138 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4139 rcu_spawn_one_boost_kthread(rnp);
4140 rcu_spawn_cpu_nocb_kthread(cpu);
4141 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4147 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4149 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4151 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4153 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4157 * Near the end of the CPU-online process. Pretty much all services
4158 * enabled, and the CPU is now very much alive.
4160 int rcutree_online_cpu(unsigned int cpu)
4162 unsigned long flags;
4163 struct rcu_data *rdp;
4164 struct rcu_node *rnp;
4166 rdp = per_cpu_ptr(&rcu_data, cpu);
4168 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4169 rnp->ffmask |= rdp->grpmask;
4170 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4171 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4172 return 0; /* Too early in boot for scheduler work. */
4173 sync_sched_exp_online_cleanup(cpu);
4174 rcutree_affinity_setting(cpu, -1);
4176 // Stop-machine done, so allow nohz_full to disable tick.
4177 tick_dep_clear(TICK_DEP_BIT_RCU);
4182 * Near the beginning of the process. The CPU is still very much alive
4183 * with pretty much all services enabled.
4185 int rcutree_offline_cpu(unsigned int cpu)
4187 unsigned long flags;
4188 struct rcu_data *rdp;
4189 struct rcu_node *rnp;
4191 rdp = per_cpu_ptr(&rcu_data, cpu);
4193 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4194 rnp->ffmask &= ~rdp->grpmask;
4195 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4197 rcutree_affinity_setting(cpu, cpu);
4199 // nohz_full CPUs need the tick for stop-machine to work quickly
4200 tick_dep_set(TICK_DEP_BIT_RCU);
4205 * Mark the specified CPU as being online so that subsequent grace periods
4206 * (both expedited and normal) will wait on it. Note that this means that
4207 * incoming CPUs are not allowed to use RCU read-side critical sections
4208 * until this function is called. Failing to observe this restriction
4209 * will result in lockdep splats.
4211 * Note that this function is special in that it is invoked directly
4212 * from the incoming CPU rather than from the cpuhp_step mechanism.
4213 * This is because this function must be invoked at a precise location.
4215 void rcu_cpu_starting(unsigned int cpu)
4217 unsigned long flags;
4219 struct rcu_data *rdp;
4220 struct rcu_node *rnp;
4223 rdp = per_cpu_ptr(&rcu_data, cpu);
4224 if (rdp->cpu_started)
4226 rdp->cpu_started = true;
4229 mask = rdp->grpmask;
4230 local_irq_save(flags);
4231 arch_spin_lock(&rcu_state.ofl_lock);
4232 rcu_dynticks_eqs_online();
4233 raw_spin_lock(&rcu_state.barrier_lock);
4234 raw_spin_lock_rcu_node(rnp);
4235 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4236 raw_spin_unlock(&rcu_state.barrier_lock);
4237 newcpu = !(rnp->expmaskinitnext & mask);
4238 rnp->expmaskinitnext |= mask;
4239 /* Allow lockless access for expedited grace periods. */
4240 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4241 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4242 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4243 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4244 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4246 /* An incoming CPU should never be blocking a grace period. */
4247 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4248 /* rcu_report_qs_rnp() *really* wants some flags to restore */
4249 unsigned long flags2;
4251 local_irq_save(flags2);
4252 rcu_disable_urgency_upon_qs(rdp);
4253 /* Report QS -after- changing ->qsmaskinitnext! */
4254 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags2);
4256 raw_spin_unlock_rcu_node(rnp);
4258 arch_spin_unlock(&rcu_state.ofl_lock);
4259 local_irq_restore(flags);
4260 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4264 * The outgoing function has no further need of RCU, so remove it from
4265 * the rcu_node tree's ->qsmaskinitnext bit masks.
4267 * Note that this function is special in that it is invoked directly
4268 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4269 * This is because this function must be invoked at a precise location.
4271 void rcu_report_dead(unsigned int cpu)
4273 unsigned long flags, seq_flags;
4275 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4276 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4278 // Do any dangling deferred wakeups.
4279 do_nocb_deferred_wakeup(rdp);
4281 /* QS for any half-done expedited grace period. */
4282 rcu_report_exp_rdp(rdp);
4283 rcu_preempt_deferred_qs(current);
4285 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4286 mask = rdp->grpmask;
4287 local_irq_save(seq_flags);
4288 arch_spin_lock(&rcu_state.ofl_lock);
4289 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4290 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4291 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4292 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4293 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4294 rcu_disable_urgency_upon_qs(rdp);
4295 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4296 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4298 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4299 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4300 arch_spin_unlock(&rcu_state.ofl_lock);
4301 local_irq_restore(seq_flags);
4303 rdp->cpu_started = false;
4306 #ifdef CONFIG_HOTPLUG_CPU
4308 * The outgoing CPU has just passed through the dying-idle state, and we
4309 * are being invoked from the CPU that was IPIed to continue the offline
4310 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4312 void rcutree_migrate_callbacks(int cpu)
4314 unsigned long flags;
4315 struct rcu_data *my_rdp;
4316 struct rcu_node *my_rnp;
4317 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4320 if (rcu_rdp_is_offloaded(rdp) ||
4321 rcu_segcblist_empty(&rdp->cblist))
4322 return; /* No callbacks to migrate. */
4324 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4325 WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
4326 rcu_barrier_entrain(rdp);
4327 my_rdp = this_cpu_ptr(&rcu_data);
4328 my_rnp = my_rdp->mynode;
4329 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4330 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4331 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4332 /* Leverage recent GPs and set GP for new callbacks. */
4333 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4334 rcu_advance_cbs(my_rnp, my_rdp);
4335 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4336 raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
4337 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4338 rcu_segcblist_disable(&rdp->cblist);
4339 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
4340 check_cb_ovld_locked(my_rdp, my_rnp);
4341 if (rcu_rdp_is_offloaded(my_rdp)) {
4342 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4343 __call_rcu_nocb_wake(my_rdp, true, flags);
4345 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4346 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4349 rcu_gp_kthread_wake();
4350 lockdep_assert_irqs_enabled();
4351 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4352 !rcu_segcblist_empty(&rdp->cblist),
4353 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4354 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4355 rcu_segcblist_first_cb(&rdp->cblist));
4360 * On non-huge systems, use expedited RCU grace periods to make suspend
4361 * and hibernation run faster.
4363 static int rcu_pm_notify(struct notifier_block *self,
4364 unsigned long action, void *hcpu)
4367 case PM_HIBERNATION_PREPARE:
4368 case PM_SUSPEND_PREPARE:
4371 case PM_POST_HIBERNATION:
4372 case PM_POST_SUSPEND:
4373 rcu_unexpedite_gp();
4381 #ifdef CONFIG_RCU_EXP_KTHREAD
4382 struct kthread_worker *rcu_exp_gp_kworker;
4383 struct kthread_worker *rcu_exp_par_gp_kworker;
4385 static void __init rcu_start_exp_gp_kworkers(void)
4387 const char *par_gp_kworker_name = "rcu_exp_par_gp_kthread_worker";
4388 const char *gp_kworker_name = "rcu_exp_gp_kthread_worker";
4389 struct sched_param param = { .sched_priority = kthread_prio };
4391 rcu_exp_gp_kworker = kthread_create_worker(0, gp_kworker_name);
4392 if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
4393 pr_err("Failed to create %s!\n", gp_kworker_name);
4397 rcu_exp_par_gp_kworker = kthread_create_worker(0, par_gp_kworker_name);
4398 if (IS_ERR_OR_NULL(rcu_exp_par_gp_kworker)) {
4399 pr_err("Failed to create %s!\n", par_gp_kworker_name);
4400 kthread_destroy_worker(rcu_exp_gp_kworker);
4404 sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, ¶m);
4405 sched_setscheduler_nocheck(rcu_exp_par_gp_kworker->task, SCHED_FIFO,
4409 static inline void rcu_alloc_par_gp_wq(void)
4412 #else /* !CONFIG_RCU_EXP_KTHREAD */
4413 struct workqueue_struct *rcu_par_gp_wq;
4415 static void __init rcu_start_exp_gp_kworkers(void)
4419 static inline void rcu_alloc_par_gp_wq(void)
4421 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4422 WARN_ON(!rcu_par_gp_wq);
4424 #endif /* CONFIG_RCU_EXP_KTHREAD */
4427 * Spawn the kthreads that handle RCU's grace periods.
4429 static int __init rcu_spawn_gp_kthread(void)
4431 unsigned long flags;
4432 struct rcu_node *rnp;
4433 struct sched_param sp;
4434 struct task_struct *t;
4435 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4437 rcu_scheduler_fully_active = 1;
4438 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4439 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4442 sp.sched_priority = kthread_prio;
4443 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4445 rnp = rcu_get_root();
4446 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4447 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4448 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4449 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4450 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4451 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4453 /* This is a pre-SMP initcall, we expect a single CPU */
4454 WARN_ON(num_online_cpus() > 1);
4456 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
4457 * due to rcu_scheduler_fully_active.
4459 rcu_spawn_cpu_nocb_kthread(smp_processor_id());
4460 rcu_spawn_one_boost_kthread(rdp->mynode);
4461 rcu_spawn_core_kthreads();
4462 /* Create kthread worker for expedited GPs */
4463 rcu_start_exp_gp_kworkers();
4466 early_initcall(rcu_spawn_gp_kthread);
4469 * This function is invoked towards the end of the scheduler's
4470 * initialization process. Before this is called, the idle task might
4471 * contain synchronous grace-period primitives (during which time, this idle
4472 * task is booting the system, and such primitives are no-ops). After this
4473 * function is called, any synchronous grace-period primitives are run as
4474 * expedited, with the requesting task driving the grace period forward.
4475 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4476 * runtime RCU functionality.
4478 void rcu_scheduler_starting(void)
4480 unsigned long flags;
4481 struct rcu_node *rnp;
4483 WARN_ON(num_online_cpus() != 1);
4484 WARN_ON(nr_context_switches() > 0);
4485 rcu_test_sync_prims();
4487 // Fix up the ->gp_seq counters.
4488 local_irq_save(flags);
4489 rcu_for_each_node_breadth_first(rnp)
4490 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
4491 local_irq_restore(flags);
4493 // Switch out of early boot mode.
4494 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4495 rcu_test_sync_prims();
4499 * Helper function for rcu_init() that initializes the rcu_state structure.
4501 static void __init rcu_init_one(void)
4503 static const char * const buf[] = RCU_NODE_NAME_INIT;
4504 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4505 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4506 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4508 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4512 struct rcu_node *rnp;
4514 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4516 /* Silence gcc 4.8 false positive about array index out of range. */
4517 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4518 panic("rcu_init_one: rcu_num_lvls out of range");
4520 /* Initialize the level-tracking arrays. */
4522 for (i = 1; i < rcu_num_lvls; i++)
4523 rcu_state.level[i] =
4524 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4525 rcu_init_levelspread(levelspread, num_rcu_lvl);
4527 /* Initialize the elements themselves, starting from the leaves. */
4529 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4530 cpustride *= levelspread[i];
4531 rnp = rcu_state.level[i];
4532 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4533 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4534 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4535 &rcu_node_class[i], buf[i]);
4536 raw_spin_lock_init(&rnp->fqslock);
4537 lockdep_set_class_and_name(&rnp->fqslock,
4538 &rcu_fqs_class[i], fqs[i]);
4539 rnp->gp_seq = rcu_state.gp_seq;
4540 rnp->gp_seq_needed = rcu_state.gp_seq;
4541 rnp->completedqs = rcu_state.gp_seq;
4543 rnp->qsmaskinit = 0;
4544 rnp->grplo = j * cpustride;
4545 rnp->grphi = (j + 1) * cpustride - 1;
4546 if (rnp->grphi >= nr_cpu_ids)
4547 rnp->grphi = nr_cpu_ids - 1;
4553 rnp->grpnum = j % levelspread[i - 1];
4554 rnp->grpmask = BIT(rnp->grpnum);
4555 rnp->parent = rcu_state.level[i - 1] +
4556 j / levelspread[i - 1];
4559 INIT_LIST_HEAD(&rnp->blkd_tasks);
4560 rcu_init_one_nocb(rnp);
4561 init_waitqueue_head(&rnp->exp_wq[0]);
4562 init_waitqueue_head(&rnp->exp_wq[1]);
4563 init_waitqueue_head(&rnp->exp_wq[2]);
4564 init_waitqueue_head(&rnp->exp_wq[3]);
4565 spin_lock_init(&rnp->exp_lock);
4566 mutex_init(&rnp->boost_kthread_mutex);
4567 raw_spin_lock_init(&rnp->exp_poll_lock);
4568 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
4569 INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
4573 init_swait_queue_head(&rcu_state.gp_wq);
4574 init_swait_queue_head(&rcu_state.expedited_wq);
4575 rnp = rcu_first_leaf_node();
4576 for_each_possible_cpu(i) {
4577 while (i > rnp->grphi)
4579 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4580 rcu_boot_init_percpu_data(i);
4585 * Force priority from the kernel command-line into range.
4587 static void __init sanitize_kthread_prio(void)
4589 int kthread_prio_in = kthread_prio;
4591 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4592 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4594 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4596 else if (kthread_prio < 0)
4598 else if (kthread_prio > 99)
4601 if (kthread_prio != kthread_prio_in)
4602 pr_alert("%s: Limited prio to %d from %d\n",
4603 __func__, kthread_prio, kthread_prio_in);
4607 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4608 * replace the definitions in tree.h because those are needed to size
4609 * the ->node array in the rcu_state structure.
4611 void rcu_init_geometry(void)
4615 static unsigned long old_nr_cpu_ids;
4616 int rcu_capacity[RCU_NUM_LVLS];
4617 static bool initialized;
4621 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4622 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4624 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4628 old_nr_cpu_ids = nr_cpu_ids;
4632 * Initialize any unspecified boot parameters.
4633 * The default values of jiffies_till_first_fqs and
4634 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4635 * value, which is a function of HZ, then adding one for each
4636 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4638 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4639 if (jiffies_till_first_fqs == ULONG_MAX)
4640 jiffies_till_first_fqs = d;
4641 if (jiffies_till_next_fqs == ULONG_MAX)
4642 jiffies_till_next_fqs = d;
4643 adjust_jiffies_till_sched_qs();
4645 /* If the compile-time values are accurate, just leave. */
4646 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4647 nr_cpu_ids == NR_CPUS)
4649 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4650 rcu_fanout_leaf, nr_cpu_ids);
4653 * The boot-time rcu_fanout_leaf parameter must be at least two
4654 * and cannot exceed the number of bits in the rcu_node masks.
4655 * Complain and fall back to the compile-time values if this
4656 * limit is exceeded.
4658 if (rcu_fanout_leaf < 2 ||
4659 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4660 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4666 * Compute number of nodes that can be handled an rcu_node tree
4667 * with the given number of levels.
4669 rcu_capacity[0] = rcu_fanout_leaf;
4670 for (i = 1; i < RCU_NUM_LVLS; i++)
4671 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4674 * The tree must be able to accommodate the configured number of CPUs.
4675 * If this limit is exceeded, fall back to the compile-time values.
4677 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4678 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4683 /* Calculate the number of levels in the tree. */
4684 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4686 rcu_num_lvls = i + 1;
4688 /* Calculate the number of rcu_nodes at each level of the tree. */
4689 for (i = 0; i < rcu_num_lvls; i++) {
4690 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4691 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4694 /* Calculate the total number of rcu_node structures. */
4696 for (i = 0; i < rcu_num_lvls; i++)
4697 rcu_num_nodes += num_rcu_lvl[i];
4701 * Dump out the structure of the rcu_node combining tree associated
4702 * with the rcu_state structure.
4704 static void __init rcu_dump_rcu_node_tree(void)
4707 struct rcu_node *rnp;
4709 pr_info("rcu_node tree layout dump\n");
4711 rcu_for_each_node_breadth_first(rnp) {
4712 if (rnp->level != level) {
4717 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4722 struct workqueue_struct *rcu_gp_wq;
4724 static void __init kfree_rcu_batch_init(void)
4729 /* Clamp it to [0:100] seconds interval. */
4730 if (rcu_delay_page_cache_fill_msec < 0 ||
4731 rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
4733 rcu_delay_page_cache_fill_msec =
4734 clamp(rcu_delay_page_cache_fill_msec, 0,
4735 (int) (100 * MSEC_PER_SEC));
4737 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4738 rcu_delay_page_cache_fill_msec);
4741 for_each_possible_cpu(cpu) {
4742 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4744 for (i = 0; i < KFREE_N_BATCHES; i++) {
4745 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4746 krcp->krw_arr[i].krcp = krcp;
4749 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4750 INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
4751 krcp->initialized = true;
4753 if (register_shrinker(&kfree_rcu_shrinker, "rcu-kfree"))
4754 pr_err("Failed to register kfree_rcu() shrinker!\n");
4757 void __init rcu_init(void)
4759 int cpu = smp_processor_id();
4761 rcu_early_boot_tests();
4763 kfree_rcu_batch_init();
4764 rcu_bootup_announce();
4765 sanitize_kthread_prio();
4766 rcu_init_geometry();
4769 rcu_dump_rcu_node_tree();
4771 open_softirq(RCU_SOFTIRQ, rcu_core_si);
4774 * We don't need protection against CPU-hotplug here because
4775 * this is called early in boot, before either interrupts
4776 * or the scheduler are operational.
4778 pm_notifier(rcu_pm_notify, 0);
4779 WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
4780 rcutree_prepare_cpu(cpu);
4781 rcu_cpu_starting(cpu);
4782 rcutree_online_cpu(cpu);
4784 /* Create workqueue for Tree SRCU and for expedited GPs. */
4785 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4786 WARN_ON(!rcu_gp_wq);
4787 rcu_alloc_par_gp_wq();
4789 /* Fill in default value for rcutree.qovld boot parameter. */
4790 /* -After- the rcu_node ->lock fields are initialized! */
4792 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4796 // Kick-start any polled grace periods that started early.
4797 if (!(per_cpu_ptr(&rcu_data, cpu)->mynode->exp_seq_poll_rq & 0x1))
4798 (void)start_poll_synchronize_rcu_expedited();
4801 #include "tree_stall.h"
4802 #include "tree_exp.h"
4803 #include "tree_nocb.h"
4804 #include "tree_plugin.h"