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/percpu.h>
36 #include <linux/notifier.h>
37 #include <linux/cpu.h>
38 #include <linux/mutex.h>
39 #include <linux/time.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/wait.h>
42 #include <linux/kthread.h>
43 #include <uapi/linux/sched/types.h>
44 #include <linux/prefetch.h>
45 #include <linux/delay.h>
46 #include <linux/random.h>
47 #include <linux/trace_events.h>
48 #include <linux/suspend.h>
49 #include <linux/ftrace.h>
50 #include <linux/tick.h>
51 #include <linux/sysrq.h>
52 #include <linux/kprobes.h>
53 #include <linux/gfp.h>
54 #include <linux/oom.h>
55 #include <linux/smpboot.h>
56 #include <linux/jiffies.h>
57 #include <linux/slab.h>
58 #include <linux/sched/isolation.h>
59 #include <linux/sched/clock.h>
60 #include <linux/vmalloc.h>
62 #include <linux/kasan.h>
63 #include "../time/tick-internal.h"
68 #ifdef MODULE_PARAM_PREFIX
69 #undef MODULE_PARAM_PREFIX
71 #define MODULE_PARAM_PREFIX "rcutree."
73 /* Data structures. */
76 * Steal a bit from the bottom of ->dynticks for idle entry/exit
77 * control. Initially this is for TLB flushing.
79 #define RCU_DYNTICK_CTRL_MASK 0x1
80 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
82 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
83 .dynticks_nesting = 1,
84 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
85 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
87 static struct rcu_state rcu_state = {
88 .level = { &rcu_state.node[0] },
89 .gp_state = RCU_GP_IDLE,
90 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
91 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
94 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
95 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
96 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
99 /* Dump rcu_node combining tree at boot to verify correct setup. */
100 static bool dump_tree;
101 module_param(dump_tree, bool, 0444);
102 /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
103 static bool use_softirq = true;
104 module_param(use_softirq, bool, 0444);
105 /* Control rcu_node-tree auto-balancing at boot time. */
106 static bool rcu_fanout_exact;
107 module_param(rcu_fanout_exact, bool, 0444);
108 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
109 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
110 module_param(rcu_fanout_leaf, int, 0444);
111 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
112 /* Number of rcu_nodes at specified level. */
113 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
114 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
117 * The rcu_scheduler_active variable is initialized to the value
118 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
119 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
120 * RCU can assume that there is but one task, allowing RCU to (for example)
121 * optimize synchronize_rcu() to a simple barrier(). When this variable
122 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
123 * to detect real grace periods. This variable is also used to suppress
124 * boot-time false positives from lockdep-RCU error checking. Finally, it
125 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
126 * is fully initialized, including all of its kthreads having been spawned.
128 int rcu_scheduler_active __read_mostly;
129 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
132 * The rcu_scheduler_fully_active variable transitions from zero to one
133 * during the early_initcall() processing, which is after the scheduler
134 * is capable of creating new tasks. So RCU processing (for example,
135 * creating tasks for RCU priority boosting) must be delayed until after
136 * rcu_scheduler_fully_active transitions from zero to one. We also
137 * currently delay invocation of any RCU callbacks until after this point.
139 * It might later prove better for people registering RCU callbacks during
140 * early boot to take responsibility for these callbacks, but one step at
143 static int rcu_scheduler_fully_active __read_mostly;
145 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
146 unsigned long gps, unsigned long flags);
147 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
148 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
149 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
150 static void invoke_rcu_core(void);
151 static void rcu_report_exp_rdp(struct rcu_data *rdp);
152 static void sync_sched_exp_online_cleanup(int cpu);
153 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
155 /* rcuc/rcub kthread realtime priority */
156 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
157 module_param(kthread_prio, int, 0444);
159 /* Delay in jiffies for grace-period initialization delays, debug only. */
161 static int gp_preinit_delay;
162 module_param(gp_preinit_delay, int, 0444);
163 static int gp_init_delay;
164 module_param(gp_init_delay, int, 0444);
165 static int gp_cleanup_delay;
166 module_param(gp_cleanup_delay, int, 0444);
168 // Add delay to rcu_read_unlock() for strict grace periods.
169 static int rcu_unlock_delay;
170 #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
171 module_param(rcu_unlock_delay, int, 0444);
175 * This rcu parameter is runtime-read-only. It reflects
176 * a minimum allowed number of objects which can be cached
177 * per-CPU. Object size is equal to one page. This value
178 * can be changed at boot time.
180 static int rcu_min_cached_objs = 5;
181 module_param(rcu_min_cached_objs, int, 0444);
183 /* Retrieve RCU kthreads priority for rcutorture */
184 int rcu_get_gp_kthreads_prio(void)
188 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
191 * Number of grace periods between delays, normalized by the duration of
192 * the delay. The longer the delay, the more the grace periods between
193 * each delay. The reason for this normalization is that it means that,
194 * for non-zero delays, the overall slowdown of grace periods is constant
195 * regardless of the duration of the delay. This arrangement balances
196 * the need for long delays to increase some race probabilities with the
197 * need for fast grace periods to increase other race probabilities.
199 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
202 * Compute the mask of online CPUs for the specified rcu_node structure.
203 * This will not be stable unless the rcu_node structure's ->lock is
204 * held, but the bit corresponding to the current CPU will be stable
207 static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
209 return READ_ONCE(rnp->qsmaskinitnext);
213 * Return true if an RCU grace period is in progress. The READ_ONCE()s
214 * permit this function to be invoked without holding the root rcu_node
215 * structure's ->lock, but of course results can be subject to change.
217 static int rcu_gp_in_progress(void)
219 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
223 * Return the number of callbacks queued on the specified CPU.
224 * Handles both the nocbs and normal cases.
226 static long rcu_get_n_cbs_cpu(int cpu)
228 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
230 if (rcu_segcblist_is_enabled(&rdp->cblist))
231 return rcu_segcblist_n_cbs(&rdp->cblist);
235 void rcu_softirq_qs(void)
238 rcu_preempt_deferred_qs(current);
242 * Record entry into an extended quiescent state. This is only to be
243 * called when not already in an extended quiescent state, that is,
244 * RCU is watching prior to the call to this function and is no longer
245 * watching upon return.
247 static noinstr void rcu_dynticks_eqs_enter(void)
249 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
253 * CPUs seeing atomic_add_return() must see prior RCU read-side
254 * critical sections, and we also must force ordering with the
257 rcu_dynticks_task_trace_enter(); // Before ->dynticks update!
258 seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
259 // RCU is no longer watching. Better be in extended quiescent state!
260 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
261 (seq & RCU_DYNTICK_CTRL_CTR));
262 /* Better not have special action (TLB flush) pending! */
263 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
264 (seq & RCU_DYNTICK_CTRL_MASK));
268 * Record exit from an extended quiescent state. This is only to be
269 * called from an extended quiescent state, that is, RCU is not watching
270 * prior to the call to this function and is watching upon return.
272 static noinstr void rcu_dynticks_eqs_exit(void)
274 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
278 * CPUs seeing atomic_add_return() must see prior idle sojourns,
279 * and we also must force ordering with the next RCU read-side
282 seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
283 // RCU is now watching. Better not be in an extended quiescent state!
284 rcu_dynticks_task_trace_exit(); // After ->dynticks update!
285 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
286 !(seq & RCU_DYNTICK_CTRL_CTR));
287 if (seq & RCU_DYNTICK_CTRL_MASK) {
288 arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
289 smp_mb__after_atomic(); /* _exit after clearing mask. */
294 * Reset the current CPU's ->dynticks counter to indicate that the
295 * newly onlined CPU is no longer in an extended quiescent state.
296 * This will either leave the counter unchanged, or increment it
297 * to the next non-quiescent value.
299 * The non-atomic test/increment sequence works because the upper bits
300 * of the ->dynticks counter are manipulated only by the corresponding CPU,
301 * or when the corresponding CPU is offline.
303 static void rcu_dynticks_eqs_online(void)
305 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
307 if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
309 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
313 * Is the current CPU in an extended quiescent state?
315 * No ordering, as we are sampling CPU-local information.
317 static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void)
319 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
321 return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
325 * Snapshot the ->dynticks counter with full ordering so as to allow
326 * stable comparison of this counter with past and future snapshots.
328 static int rcu_dynticks_snap(struct rcu_data *rdp)
330 int snap = atomic_add_return(0, &rdp->dynticks);
332 return snap & ~RCU_DYNTICK_CTRL_MASK;
336 * Return true if the snapshot returned from rcu_dynticks_snap()
337 * indicates that RCU is in an extended quiescent state.
339 static bool rcu_dynticks_in_eqs(int snap)
341 return !(snap & RCU_DYNTICK_CTRL_CTR);
344 /* Return true if the specified CPU is currently idle from an RCU viewpoint. */
345 bool rcu_is_idle_cpu(int cpu)
347 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
349 return rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
353 * Return true if the CPU corresponding to the specified rcu_data
354 * structure has spent some time in an extended quiescent state since
355 * rcu_dynticks_snap() returned the specified snapshot.
357 static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
359 return snap != rcu_dynticks_snap(rdp);
363 * Return true if the referenced integer is zero while the specified
364 * CPU remains within a single extended quiescent state.
366 bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
368 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
371 // If not quiescent, force back to earlier extended quiescent state.
372 snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK |
373 RCU_DYNTICK_CTRL_CTR);
375 smp_rmb(); // Order ->dynticks and *vp reads.
377 return false; // Non-zero, so report failure;
378 smp_rmb(); // Order *vp read and ->dynticks re-read.
380 // If still in the same extended quiescent state, we are good!
381 return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK);
385 * Set the special (bottom) bit of the specified CPU so that it
386 * will take special action (such as flushing its TLB) on the
387 * next exit from an extended quiescent state. Returns true if
388 * the bit was successfully set, or false if the CPU was not in
389 * an extended quiescent state.
391 bool rcu_eqs_special_set(int cpu)
396 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
398 new_old = atomic_read(&rdp->dynticks);
401 if (old & RCU_DYNTICK_CTRL_CTR)
403 new = old | RCU_DYNTICK_CTRL_MASK;
404 new_old = atomic_cmpxchg(&rdp->dynticks, old, new);
405 } while (new_old != old);
410 * Let the RCU core know that this CPU has gone through the scheduler,
411 * which is a quiescent state. This is called when the need for a
412 * quiescent state is urgent, so we burn an atomic operation and full
413 * memory barriers to let the RCU core know about it, regardless of what
414 * this CPU might (or might not) do in the near future.
416 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
418 * The caller must have disabled interrupts and must not be idle.
420 notrace void rcu_momentary_dyntick_idle(void)
424 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
425 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
426 &this_cpu_ptr(&rcu_data)->dynticks);
427 /* It is illegal to call this from idle state. */
428 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
429 rcu_preempt_deferred_qs(current);
431 EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
434 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
436 * If the current CPU is idle and running at a first-level (not nested)
437 * interrupt, or directly, from idle, return true.
439 * The caller must have at least disabled IRQs.
441 static int rcu_is_cpu_rrupt_from_idle(void)
446 * Usually called from the tick; but also used from smp_function_call()
447 * for expedited grace periods. This latter can result in running from
448 * the idle task, instead of an actual IPI.
450 lockdep_assert_irqs_disabled();
452 /* Check for counter underflows */
453 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
454 "RCU dynticks_nesting counter underflow!");
455 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
456 "RCU dynticks_nmi_nesting counter underflow/zero!");
458 /* Are we at first interrupt nesting level? */
459 nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting);
464 * If we're not in an interrupt, we must be in the idle task!
466 WARN_ON_ONCE(!nesting && !is_idle_task(current));
468 /* Does CPU appear to be idle from an RCU standpoint? */
469 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
472 #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
473 // Maximum callbacks per rcu_do_batch ...
474 #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
475 static long blimit = DEFAULT_RCU_BLIMIT;
476 #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
477 static long qhimark = DEFAULT_RCU_QHIMARK;
478 #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
479 static long qlowmark = DEFAULT_RCU_QLOMARK;
480 #define DEFAULT_RCU_QOVLD_MULT 2
481 #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
482 static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
483 static long qovld_calc = -1; // No pre-initialization lock acquisitions!
485 module_param(blimit, long, 0444);
486 module_param(qhimark, long, 0444);
487 module_param(qlowmark, long, 0444);
488 module_param(qovld, long, 0444);
490 static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
491 static ulong jiffies_till_next_fqs = ULONG_MAX;
492 static bool rcu_kick_kthreads;
493 static int rcu_divisor = 7;
494 module_param(rcu_divisor, int, 0644);
496 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
497 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
498 module_param(rcu_resched_ns, long, 0644);
501 * How long the grace period must be before we start recruiting
502 * quiescent-state help from rcu_note_context_switch().
504 static ulong jiffies_till_sched_qs = ULONG_MAX;
505 module_param(jiffies_till_sched_qs, ulong, 0444);
506 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
507 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
510 * Make sure that we give the grace-period kthread time to detect any
511 * idle CPUs before taking active measures to force quiescent states.
512 * However, don't go below 100 milliseconds, adjusted upwards for really
515 static void adjust_jiffies_till_sched_qs(void)
519 /* If jiffies_till_sched_qs was specified, respect the request. */
520 if (jiffies_till_sched_qs != ULONG_MAX) {
521 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
524 /* Otherwise, set to third fqs scan, but bound below on large system. */
525 j = READ_ONCE(jiffies_till_first_fqs) +
526 2 * READ_ONCE(jiffies_till_next_fqs);
527 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
528 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
529 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
530 WRITE_ONCE(jiffies_to_sched_qs, j);
533 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
536 int ret = kstrtoul(val, 0, &j);
539 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
540 adjust_jiffies_till_sched_qs();
545 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
548 int ret = kstrtoul(val, 0, &j);
551 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
552 adjust_jiffies_till_sched_qs();
557 static const struct kernel_param_ops first_fqs_jiffies_ops = {
558 .set = param_set_first_fqs_jiffies,
559 .get = param_get_ulong,
562 static const struct kernel_param_ops next_fqs_jiffies_ops = {
563 .set = param_set_next_fqs_jiffies,
564 .get = param_get_ulong,
567 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
568 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
569 module_param(rcu_kick_kthreads, bool, 0644);
571 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
572 static int rcu_pending(int user);
575 * Return the number of RCU GPs completed thus far for debug & stats.
577 unsigned long rcu_get_gp_seq(void)
579 return READ_ONCE(rcu_state.gp_seq);
581 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
584 * Return the number of RCU expedited batches completed thus far for
585 * debug & stats. Odd numbers mean that a batch is in progress, even
586 * numbers mean idle. The value returned will thus be roughly double
587 * the cumulative batches since boot.
589 unsigned long rcu_exp_batches_completed(void)
591 return rcu_state.expedited_sequence;
593 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
596 * Return the root node of the rcu_state structure.
598 static struct rcu_node *rcu_get_root(void)
600 return &rcu_state.node[0];
604 * Send along grace-period-related data for rcutorture diagnostics.
606 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
607 unsigned long *gp_seq)
611 *flags = READ_ONCE(rcu_state.gp_flags);
612 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
618 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
621 * Enter an RCU extended quiescent state, which can be either the
622 * idle loop or adaptive-tickless usermode execution.
624 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
625 * the possibility of usermode upcalls having messed up our count
626 * of interrupt nesting level during the prior busy period.
628 static noinstr void rcu_eqs_enter(bool user)
630 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
632 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
633 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
634 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
635 rdp->dynticks_nesting == 0);
636 if (rdp->dynticks_nesting != 1) {
637 // RCU will still be watching, so just do accounting and leave.
638 rdp->dynticks_nesting--;
642 lockdep_assert_irqs_disabled();
643 instrumentation_begin();
644 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
645 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
646 rdp = this_cpu_ptr(&rcu_data);
647 rcu_prepare_for_idle();
648 rcu_preempt_deferred_qs(current);
650 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
651 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
653 instrumentation_end();
654 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
655 // RCU is watching here ...
656 rcu_dynticks_eqs_enter();
657 // ... but is no longer watching here.
658 rcu_dynticks_task_enter();
662 * rcu_idle_enter - inform RCU that current CPU is entering idle
664 * Enter idle mode, in other words, -leave- the mode in which RCU
665 * read-side critical sections can occur. (Though RCU read-side
666 * critical sections can occur in irq handlers in idle, a possibility
667 * handled by irq_enter() and irq_exit().)
669 * If you add or remove a call to rcu_idle_enter(), be sure to test with
670 * CONFIG_RCU_EQS_DEBUG=y.
672 void rcu_idle_enter(void)
674 lockdep_assert_irqs_disabled();
675 rcu_eqs_enter(false);
677 EXPORT_SYMBOL_GPL(rcu_idle_enter);
679 #ifdef CONFIG_NO_HZ_FULL
682 * An empty function that will trigger a reschedule on
683 * IRQ tail once IRQs get re-enabled on userspace resume.
685 static void late_wakeup_func(struct irq_work *work)
689 static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
690 IRQ_WORK_INIT(late_wakeup_func);
693 * rcu_user_enter - inform RCU that we are resuming userspace.
695 * Enter RCU idle mode right before resuming userspace. No use of RCU
696 * is permitted between this call and rcu_user_exit(). This way the
697 * CPU doesn't need to maintain the tick for RCU maintenance purposes
698 * when the CPU runs in userspace.
700 * If you add or remove a call to rcu_user_enter(), be sure to test with
701 * CONFIG_RCU_EQS_DEBUG=y.
703 noinstr void rcu_user_enter(void)
705 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
707 lockdep_assert_irqs_disabled();
710 * Other than generic entry implementation, we may be past the last
711 * rescheduling opportunity in the entry code. Trigger a self IPI
712 * that will fire and reschedule once we resume in user/guest mode.
714 instrumentation_begin();
715 if (!IS_ENABLED(CONFIG_GENERIC_ENTRY) || (current->flags & PF_VCPU)) {
716 if (do_nocb_deferred_wakeup(rdp) && need_resched())
717 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
719 instrumentation_end();
724 #endif /* CONFIG_NO_HZ_FULL */
727 * rcu_nmi_exit - inform RCU of exit from NMI context
729 * If we are returning from the outermost NMI handler that interrupted an
730 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
731 * to let the RCU grace-period handling know that the CPU is back to
734 * If you add or remove a call to rcu_nmi_exit(), be sure to test
735 * with CONFIG_RCU_EQS_DEBUG=y.
737 noinstr void rcu_nmi_exit(void)
739 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
741 instrumentation_begin();
743 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
744 * (We are exiting an NMI handler, so RCU better be paying attention
747 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
748 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
751 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
752 * leave it in non-RCU-idle state.
754 if (rdp->dynticks_nmi_nesting != 1) {
755 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
756 atomic_read(&rdp->dynticks));
757 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
758 rdp->dynticks_nmi_nesting - 2);
759 instrumentation_end();
763 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
764 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
765 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
768 rcu_prepare_for_idle();
770 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
771 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
772 instrumentation_end();
774 // RCU is watching here ...
775 rcu_dynticks_eqs_enter();
776 // ... but is no longer watching here.
779 rcu_dynticks_task_enter();
783 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
785 * Exit from an interrupt handler, which might possibly result in entering
786 * idle mode, in other words, leaving the mode in which read-side critical
787 * sections can occur. The caller must have disabled interrupts.
789 * This code assumes that the idle loop never does anything that might
790 * result in unbalanced calls to irq_enter() and irq_exit(). If your
791 * architecture's idle loop violates this assumption, RCU will give you what
792 * you deserve, good and hard. But very infrequently and irreproducibly.
794 * Use things like work queues to work around this limitation.
796 * You have been warned.
798 * If you add or remove a call to rcu_irq_exit(), be sure to test with
799 * CONFIG_RCU_EQS_DEBUG=y.
801 void noinstr rcu_irq_exit(void)
803 lockdep_assert_irqs_disabled();
808 * rcu_irq_exit_preempt - Inform RCU that current CPU is exiting irq
809 * towards in kernel preemption
811 * Same as rcu_irq_exit() but has a sanity check that scheduling is safe
812 * from RCU point of view. Invoked from return from interrupt before kernel
815 void rcu_irq_exit_preempt(void)
817 lockdep_assert_irqs_disabled();
820 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
821 "RCU dynticks_nesting counter underflow/zero!");
822 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
824 "Bad RCU dynticks_nmi_nesting counter\n");
825 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
826 "RCU in extended quiescent state!");
829 #ifdef CONFIG_PROVE_RCU
831 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
833 void rcu_irq_exit_check_preempt(void)
835 lockdep_assert_irqs_disabled();
837 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
838 "RCU dynticks_nesting counter underflow/zero!");
839 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
841 "Bad RCU dynticks_nmi_nesting counter\n");
842 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
843 "RCU in extended quiescent state!");
845 #endif /* #ifdef CONFIG_PROVE_RCU */
848 * Wrapper for rcu_irq_exit() where interrupts are enabled.
850 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
851 * with CONFIG_RCU_EQS_DEBUG=y.
853 void rcu_irq_exit_irqson(void)
857 local_irq_save(flags);
859 local_irq_restore(flags);
863 * Exit an RCU extended quiescent state, which can be either the
864 * idle loop or adaptive-tickless usermode execution.
866 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
867 * allow for the possibility of usermode upcalls messing up our count of
868 * interrupt nesting level during the busy period that is just now starting.
870 static void noinstr rcu_eqs_exit(bool user)
872 struct rcu_data *rdp;
875 lockdep_assert_irqs_disabled();
876 rdp = this_cpu_ptr(&rcu_data);
877 oldval = rdp->dynticks_nesting;
878 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
880 // RCU was already watching, so just do accounting and leave.
881 rdp->dynticks_nesting++;
884 rcu_dynticks_task_exit();
885 // RCU is not watching here ...
886 rcu_dynticks_eqs_exit();
887 // ... but is watching here.
888 instrumentation_begin();
890 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
891 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
893 rcu_cleanup_after_idle();
894 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
895 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
896 WRITE_ONCE(rdp->dynticks_nesting, 1);
897 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
898 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
899 instrumentation_end();
903 * rcu_idle_exit - inform RCU that current CPU is leaving idle
905 * Exit idle mode, in other words, -enter- the mode in which RCU
906 * read-side critical sections can occur.
908 * If you add or remove a call to rcu_idle_exit(), be sure to test with
909 * CONFIG_RCU_EQS_DEBUG=y.
911 void rcu_idle_exit(void)
915 local_irq_save(flags);
917 local_irq_restore(flags);
919 EXPORT_SYMBOL_GPL(rcu_idle_exit);
921 #ifdef CONFIG_NO_HZ_FULL
923 * rcu_user_exit - inform RCU that we are exiting userspace.
925 * Exit RCU idle mode while entering the kernel because it can
926 * run a RCU read side critical section anytime.
928 * If you add or remove a call to rcu_user_exit(), be sure to test with
929 * CONFIG_RCU_EQS_DEBUG=y.
931 void noinstr rcu_user_exit(void)
937 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
939 * The scheduler tick is not normally enabled when CPUs enter the kernel
940 * from nohz_full userspace execution. After all, nohz_full userspace
941 * execution is an RCU quiescent state and the time executing in the kernel
942 * is quite short. Except of course when it isn't. And it is not hard to
943 * cause a large system to spend tens of seconds or even minutes looping
944 * in the kernel, which can cause a number of problems, include RCU CPU
947 * Therefore, if a nohz_full CPU fails to report a quiescent state
948 * in a timely manner, the RCU grace-period kthread sets that CPU's
949 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
950 * exception will invoke this function, which will turn on the scheduler
951 * tick, which will enable RCU to detect that CPU's quiescent states,
952 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
953 * The tick will be disabled once a quiescent state is reported for
956 * Of course, in carefully tuned systems, there might never be an
957 * interrupt or exception. In that case, the RCU grace-period kthread
958 * will eventually cause one to happen. However, in less carefully
959 * controlled environments, this function allows RCU to get what it
960 * needs without creating otherwise useless interruptions.
962 void __rcu_irq_enter_check_tick(void)
964 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
966 // If we're here from NMI there's nothing to do.
970 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
971 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
973 if (!tick_nohz_full_cpu(rdp->cpu) ||
974 !READ_ONCE(rdp->rcu_urgent_qs) ||
975 READ_ONCE(rdp->rcu_forced_tick)) {
976 // RCU doesn't need nohz_full help from this CPU, or it is
977 // already getting that help.
981 // We get here only when not in an extended quiescent state and
982 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
983 // already watching and (2) The fact that we are in an interrupt
984 // handler and that the rcu_node lock is an irq-disabled lock
985 // prevents self-deadlock. So we can safely recheck under the lock.
986 // Note that the nohz_full state currently cannot change.
987 raw_spin_lock_rcu_node(rdp->mynode);
988 if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
989 // A nohz_full CPU is in the kernel and RCU needs a
990 // quiescent state. Turn on the tick!
991 WRITE_ONCE(rdp->rcu_forced_tick, true);
992 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
994 raw_spin_unlock_rcu_node(rdp->mynode);
996 #endif /* CONFIG_NO_HZ_FULL */
999 * rcu_nmi_enter - inform RCU of entry to NMI context
1001 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
1002 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
1003 * that the CPU is active. This implementation permits nested NMIs, as
1004 * long as the nesting level does not overflow an int. (You will probably
1005 * run out of stack space first.)
1007 * If you add or remove a call to rcu_nmi_enter(), be sure to test
1008 * with CONFIG_RCU_EQS_DEBUG=y.
1010 noinstr void rcu_nmi_enter(void)
1013 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1015 /* Complain about underflow. */
1016 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
1019 * If idle from RCU viewpoint, atomically increment ->dynticks
1020 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1021 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1022 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1023 * to be in the outermost NMI handler that interrupted an RCU-idle
1024 * period (observation due to Andy Lutomirski).
1026 if (rcu_dynticks_curr_cpu_in_eqs()) {
1029 rcu_dynticks_task_exit();
1031 // RCU is not watching here ...
1032 rcu_dynticks_eqs_exit();
1033 // ... but is watching here.
1036 instrumentation_begin();
1037 rcu_cleanup_after_idle();
1038 instrumentation_end();
1041 instrumentation_begin();
1042 // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
1043 instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
1044 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
1045 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
1048 } else if (!in_nmi()) {
1049 instrumentation_begin();
1050 rcu_irq_enter_check_tick();
1051 instrumentation_end();
1053 instrumentation_begin();
1056 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1057 rdp->dynticks_nmi_nesting,
1058 rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
1059 instrumentation_end();
1060 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
1061 rdp->dynticks_nmi_nesting + incby);
1066 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1068 * Enter an interrupt handler, which might possibly result in exiting
1069 * idle mode, in other words, entering the mode in which read-side critical
1070 * sections can occur. The caller must have disabled interrupts.
1072 * Note that the Linux kernel is fully capable of entering an interrupt
1073 * handler that it never exits, for example when doing upcalls to user mode!
1074 * This code assumes that the idle loop never does upcalls to user mode.
1075 * If your architecture's idle loop does do upcalls to user mode (or does
1076 * anything else that results in unbalanced calls to the irq_enter() and
1077 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1078 * But very infrequently and irreproducibly.
1080 * Use things like work queues to work around this limitation.
1082 * You have been warned.
1084 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1085 * CONFIG_RCU_EQS_DEBUG=y.
1087 noinstr void rcu_irq_enter(void)
1089 lockdep_assert_irqs_disabled();
1094 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1096 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1097 * with CONFIG_RCU_EQS_DEBUG=y.
1099 void rcu_irq_enter_irqson(void)
1101 unsigned long flags;
1103 local_irq_save(flags);
1105 local_irq_restore(flags);
1109 * If any sort of urgency was applied to the current CPU (for example,
1110 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
1111 * to get to a quiescent state, disable it.
1113 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
1115 raw_lockdep_assert_held_rcu_node(rdp->mynode);
1116 WRITE_ONCE(rdp->rcu_urgent_qs, false);
1117 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
1118 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
1119 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1120 WRITE_ONCE(rdp->rcu_forced_tick, false);
1125 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
1127 * Return true if RCU is watching the running CPU, which means that this
1128 * CPU can safely enter RCU read-side critical sections. In other words,
1129 * if the current CPU is not in its idle loop or is in an interrupt or
1130 * NMI handler, return true.
1132 * Make notrace because it can be called by the internal functions of
1133 * ftrace, and making this notrace removes unnecessary recursion calls.
1135 notrace bool rcu_is_watching(void)
1139 preempt_disable_notrace();
1140 ret = !rcu_dynticks_curr_cpu_in_eqs();
1141 preempt_enable_notrace();
1144 EXPORT_SYMBOL_GPL(rcu_is_watching);
1147 * If a holdout task is actually running, request an urgent quiescent
1148 * state from its CPU. This is unsynchronized, so migrations can cause
1149 * the request to go to the wrong CPU. Which is OK, all that will happen
1150 * is that the CPU's next context switch will be a bit slower and next
1151 * time around this task will generate another request.
1153 void rcu_request_urgent_qs_task(struct task_struct *t)
1160 return; /* This task is not running on that CPU. */
1161 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
1164 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1167 * Is the current CPU online as far as RCU is concerned?
1169 * Disable preemption to avoid false positives that could otherwise
1170 * happen due to the current CPU number being sampled, this task being
1171 * preempted, its old CPU being taken offline, resuming on some other CPU,
1172 * then determining that its old CPU is now offline.
1174 * Disable checking if in an NMI handler because we cannot safely
1175 * report errors from NMI handlers anyway. In addition, it is OK to use
1176 * RCU on an offline processor during initial boot, hence the check for
1177 * rcu_scheduler_fully_active.
1179 bool rcu_lockdep_current_cpu_online(void)
1181 struct rcu_data *rdp;
1182 struct rcu_node *rnp;
1185 if (in_nmi() || !rcu_scheduler_fully_active)
1187 preempt_disable_notrace();
1188 rdp = this_cpu_ptr(&rcu_data);
1190 if (rdp->grpmask & rcu_rnp_online_cpus(rnp) || READ_ONCE(rnp->ofl_seq) & 0x1)
1192 preempt_enable_notrace();
1195 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1197 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1200 * We are reporting a quiescent state on behalf of some other CPU, so
1201 * it is our responsibility to check for and handle potential overflow
1202 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1203 * After all, the CPU might be in deep idle state, and thus executing no
1206 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1208 raw_lockdep_assert_held_rcu_node(rnp);
1209 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1211 WRITE_ONCE(rdp->gpwrap, true);
1212 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1213 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1217 * Snapshot the specified CPU's dynticks counter so that we can later
1218 * credit them with an implicit quiescent state. Return 1 if this CPU
1219 * is in dynticks idle mode, which is an extended quiescent state.
1221 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1223 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1224 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1225 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1226 rcu_gpnum_ovf(rdp->mynode, rdp);
1233 * Return true if the specified CPU has passed through a quiescent
1234 * state by virtue of being in or having passed through an dynticks
1235 * idle state since the last call to dyntick_save_progress_counter()
1236 * for this same CPU, or by virtue of having been offline.
1238 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1243 struct rcu_node *rnp = rdp->mynode;
1246 * If the CPU passed through or entered a dynticks idle phase with
1247 * no active irq/NMI handlers, then we can safely pretend that the CPU
1248 * already acknowledged the request to pass through a quiescent
1249 * state. Either way, that CPU cannot possibly be in an RCU
1250 * read-side critical section that started before the beginning
1251 * of the current RCU grace period.
1253 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1254 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1255 rcu_gpnum_ovf(rnp, rdp);
1260 * Complain if a CPU that is considered to be offline from RCU's
1261 * perspective has not yet reported a quiescent state. After all,
1262 * the offline CPU should have reported a quiescent state during
1263 * the CPU-offline process, or, failing that, by rcu_gp_init()
1264 * if it ran concurrently with either the CPU going offline or the
1265 * last task on a leaf rcu_node structure exiting its RCU read-side
1266 * critical section while all CPUs corresponding to that structure
1267 * are offline. This added warning detects bugs in any of these
1270 * The rcu_node structure's ->lock is held here, which excludes
1271 * the relevant portions the CPU-hotplug code, the grace-period
1272 * initialization code, and the rcu_read_unlock() code paths.
1274 * For more detail, please refer to the "Hotplug CPU" section
1275 * of RCU's Requirements documentation.
1277 if (WARN_ON_ONCE(!(rdp->grpmask & rcu_rnp_online_cpus(rnp)))) {
1279 struct rcu_node *rnp1;
1281 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1282 __func__, rnp->grplo, rnp->grphi, rnp->level,
1283 (long)rnp->gp_seq, (long)rnp->completedqs);
1284 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1285 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1286 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1287 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1288 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1289 __func__, rdp->cpu, ".o"[onl],
1290 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1291 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1292 return 1; /* Break things loose after complaining. */
1296 * A CPU running for an extended time within the kernel can
1297 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1298 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1299 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1300 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1301 * variable are safe because the assignments are repeated if this
1302 * CPU failed to pass through a quiescent state. This code
1303 * also checks .jiffies_resched in case jiffies_to_sched_qs
1306 jtsq = READ_ONCE(jiffies_to_sched_qs);
1307 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1308 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1309 if (!READ_ONCE(*rnhqp) &&
1310 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1311 time_after(jiffies, rcu_state.jiffies_resched) ||
1312 rcu_state.cbovld)) {
1313 WRITE_ONCE(*rnhqp, true);
1314 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1315 smp_store_release(ruqp, true);
1316 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1317 WRITE_ONCE(*ruqp, true);
1321 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1322 * The above code handles this, but only for straight cond_resched().
1323 * And some in-kernel loops check need_resched() before calling
1324 * cond_resched(), which defeats the above code for CPUs that are
1325 * running in-kernel with scheduling-clock interrupts disabled.
1326 * So hit them over the head with the resched_cpu() hammer!
1328 if (tick_nohz_full_cpu(rdp->cpu) &&
1329 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
1330 rcu_state.cbovld)) {
1331 WRITE_ONCE(*ruqp, true);
1332 resched_cpu(rdp->cpu);
1333 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1337 * If more than halfway to RCU CPU stall-warning time, invoke
1338 * resched_cpu() more frequently to try to loosen things up a bit.
1339 * Also check to see if the CPU is getting hammered with interrupts,
1340 * but only once per grace period, just to keep the IPIs down to
1343 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1344 if (time_after(jiffies,
1345 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1346 resched_cpu(rdp->cpu);
1347 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1349 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1350 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1351 (rnp->ffmask & rdp->grpmask)) {
1352 rdp->rcu_iw_pending = true;
1353 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1354 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1361 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1362 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1363 unsigned long gp_seq_req, const char *s)
1365 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
1366 gp_seq_req, rnp->level,
1367 rnp->grplo, rnp->grphi, s);
1371 * rcu_start_this_gp - Request the start of a particular grace period
1372 * @rnp_start: The leaf node of the CPU from which to start.
1373 * @rdp: The rcu_data corresponding to the CPU from which to start.
1374 * @gp_seq_req: The gp_seq of the grace period to start.
1376 * Start the specified grace period, as needed to handle newly arrived
1377 * callbacks. The required future grace periods are recorded in each
1378 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1379 * is reason to awaken the grace-period kthread.
1381 * The caller must hold the specified rcu_node structure's ->lock, which
1382 * is why the caller is responsible for waking the grace-period kthread.
1384 * Returns true if the GP thread needs to be awakened else false.
1386 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1387 unsigned long gp_seq_req)
1390 struct rcu_node *rnp;
1393 * Use funnel locking to either acquire the root rcu_node
1394 * structure's lock or bail out if the need for this grace period
1395 * has already been recorded -- or if that grace period has in
1396 * fact already started. If there is already a grace period in
1397 * progress in a non-leaf node, no recording is needed because the
1398 * end of the grace period will scan the leaf rcu_node structures.
1399 * Note that rnp_start->lock must not be released.
1401 raw_lockdep_assert_held_rcu_node(rnp_start);
1402 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1403 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1404 if (rnp != rnp_start)
1405 raw_spin_lock_rcu_node(rnp);
1406 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1407 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1408 (rnp != rnp_start &&
1409 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1410 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1414 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1415 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1417 * We just marked the leaf or internal node, and a
1418 * grace period is in progress, which means that
1419 * rcu_gp_cleanup() will see the marking. Bail to
1420 * reduce contention.
1422 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1423 TPS("Startedleaf"));
1426 if (rnp != rnp_start && rnp->parent != NULL)
1427 raw_spin_unlock_rcu_node(rnp);
1429 break; /* At root, and perhaps also leaf. */
1432 /* If GP already in progress, just leave, otherwise start one. */
1433 if (rcu_gp_in_progress()) {
1434 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1437 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1438 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1439 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1440 if (!READ_ONCE(rcu_state.gp_kthread)) {
1441 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1444 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1445 ret = true; /* Caller must wake GP kthread. */
1447 /* Push furthest requested GP to leaf node and rcu_data structure. */
1448 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1449 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1450 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1452 if (rnp != rnp_start)
1453 raw_spin_unlock_rcu_node(rnp);
1458 * Clean up any old requests for the just-ended grace period. Also return
1459 * whether any additional grace periods have been requested.
1461 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1464 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1466 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1468 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1469 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1470 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1475 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1476 * interrupt or softirq handler, in which case we just might immediately
1477 * sleep upon return, resulting in a grace-period hang), and don't bother
1478 * awakening when there is nothing for the grace-period kthread to do
1479 * (as in several CPUs raced to awaken, we lost), and finally don't try
1480 * to awaken a kthread that has not yet been created. If all those checks
1481 * are passed, track some debug information and awaken.
1483 * So why do the self-wakeup when in an interrupt or softirq handler
1484 * in the grace-period kthread's context? Because the kthread might have
1485 * been interrupted just as it was going to sleep, and just after the final
1486 * pre-sleep check of the awaken condition. In this case, a wakeup really
1487 * is required, and is therefore supplied.
1489 static void rcu_gp_kthread_wake(void)
1491 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1493 if ((current == t && !in_irq() && !in_serving_softirq()) ||
1494 !READ_ONCE(rcu_state.gp_flags) || !t)
1496 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1497 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1498 swake_up_one(&rcu_state.gp_wq);
1502 * If there is room, assign a ->gp_seq number to any callbacks on this
1503 * CPU that have not already been assigned. Also accelerate any callbacks
1504 * that were previously assigned a ->gp_seq number that has since proven
1505 * to be too conservative, which can happen if callbacks get assigned a
1506 * ->gp_seq number while RCU is idle, but with reference to a non-root
1507 * rcu_node structure. This function is idempotent, so it does not hurt
1508 * to call it repeatedly. Returns an flag saying that we should awaken
1509 * the RCU grace-period kthread.
1511 * The caller must hold rnp->lock with interrupts disabled.
1513 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1515 unsigned long gp_seq_req;
1518 rcu_lockdep_assert_cblist_protected(rdp);
1519 raw_lockdep_assert_held_rcu_node(rnp);
1521 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1522 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1526 * Callbacks are often registered with incomplete grace-period
1527 * information. Something about the fact that getting exact
1528 * information requires acquiring a global lock... RCU therefore
1529 * makes a conservative estimate of the grace period number at which
1530 * a given callback will become ready to invoke. The following
1531 * code checks this estimate and improves it when possible, thus
1532 * accelerating callback invocation to an earlier grace-period
1535 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1536 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1537 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1539 /* Trace depending on how much we were able to accelerate. */
1540 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1541 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1543 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1549 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1550 * rcu_node structure's ->lock be held. It consults the cached value
1551 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1552 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1553 * while holding the leaf rcu_node structure's ->lock.
1555 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1556 struct rcu_data *rdp)
1561 rcu_lockdep_assert_cblist_protected(rdp);
1562 c = rcu_seq_snap(&rcu_state.gp_seq);
1563 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1564 /* Old request still live, so mark recent callbacks. */
1565 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1568 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1569 needwake = rcu_accelerate_cbs(rnp, rdp);
1570 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1572 rcu_gp_kthread_wake();
1576 * Move any callbacks whose grace period has completed to the
1577 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1578 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1579 * sublist. This function is idempotent, so it does not hurt to
1580 * invoke it repeatedly. As long as it is not invoked -too- often...
1581 * Returns true if the RCU grace-period kthread needs to be awakened.
1583 * The caller must hold rnp->lock with interrupts disabled.
1585 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1587 rcu_lockdep_assert_cblist_protected(rdp);
1588 raw_lockdep_assert_held_rcu_node(rnp);
1590 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1591 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1595 * Find all callbacks whose ->gp_seq numbers indicate that they
1596 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1598 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1600 /* Classify any remaining callbacks. */
1601 return rcu_accelerate_cbs(rnp, rdp);
1605 * Move and classify callbacks, but only if doing so won't require
1606 * that the RCU grace-period kthread be awakened.
1608 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1609 struct rcu_data *rdp)
1611 rcu_lockdep_assert_cblist_protected(rdp);
1612 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1613 !raw_spin_trylock_rcu_node(rnp))
1615 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1616 raw_spin_unlock_rcu_node(rnp);
1620 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1621 * quiescent state. This is intended to be invoked when the CPU notices
1622 * a new grace period.
1624 static void rcu_strict_gp_check_qs(void)
1626 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1633 * Update CPU-local rcu_data state to record the beginnings and ends of
1634 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1635 * structure corresponding to the current CPU, and must have irqs disabled.
1636 * Returns true if the grace-period kthread needs to be awakened.
1638 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1642 const bool offloaded = rcu_segcblist_is_offloaded(&rdp->cblist);
1644 raw_lockdep_assert_held_rcu_node(rnp);
1646 if (rdp->gp_seq == rnp->gp_seq)
1647 return false; /* Nothing to do. */
1649 /* Handle the ends of any preceding grace periods first. */
1650 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1651 unlikely(READ_ONCE(rdp->gpwrap))) {
1653 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1654 rdp->core_needs_qs = false;
1655 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1658 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1659 if (rdp->core_needs_qs)
1660 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1663 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1664 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1665 unlikely(READ_ONCE(rdp->gpwrap))) {
1667 * If the current grace period is waiting for this CPU,
1668 * set up to detect a quiescent state, otherwise don't
1669 * go looking for one.
1671 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1672 need_qs = !!(rnp->qsmask & rdp->grpmask);
1673 rdp->cpu_no_qs.b.norm = need_qs;
1674 rdp->core_needs_qs = need_qs;
1675 zero_cpu_stall_ticks(rdp);
1677 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1678 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1679 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1680 WRITE_ONCE(rdp->gpwrap, false);
1681 rcu_gpnum_ovf(rnp, rdp);
1685 static void note_gp_changes(struct rcu_data *rdp)
1687 unsigned long flags;
1689 struct rcu_node *rnp;
1691 local_irq_save(flags);
1693 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1694 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1695 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1696 local_irq_restore(flags);
1699 needwake = __note_gp_changes(rnp, rdp);
1700 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1701 rcu_strict_gp_check_qs();
1703 rcu_gp_kthread_wake();
1706 static void rcu_gp_slow(int delay)
1709 !(rcu_seq_ctr(rcu_state.gp_seq) %
1710 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1711 schedule_timeout_idle(delay);
1714 static unsigned long sleep_duration;
1716 /* Allow rcutorture to stall the grace-period kthread. */
1717 void rcu_gp_set_torture_wait(int duration)
1719 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1720 WRITE_ONCE(sleep_duration, duration);
1722 EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1724 /* Actually implement the aforementioned wait. */
1725 static void rcu_gp_torture_wait(void)
1727 unsigned long duration;
1729 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1731 duration = xchg(&sleep_duration, 0UL);
1733 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1734 schedule_timeout_idle(duration);
1735 pr_alert("%s: Wait complete\n", __func__);
1740 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1743 static void rcu_strict_gp_boundary(void *unused)
1749 * Initialize a new grace period. Return false if no grace period required.
1751 static bool rcu_gp_init(void)
1753 unsigned long firstseq;
1754 unsigned long flags;
1755 unsigned long oldmask;
1757 struct rcu_data *rdp;
1758 struct rcu_node *rnp = rcu_get_root();
1760 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1761 raw_spin_lock_irq_rcu_node(rnp);
1762 if (!READ_ONCE(rcu_state.gp_flags)) {
1763 /* Spurious wakeup, tell caller to go back to sleep. */
1764 raw_spin_unlock_irq_rcu_node(rnp);
1767 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1769 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1771 * Grace period already in progress, don't start another.
1772 * Not supposed to be able to happen.
1774 raw_spin_unlock_irq_rcu_node(rnp);
1778 /* Advance to a new grace period and initialize state. */
1779 record_gp_stall_check_time();
1780 /* Record GP times before starting GP, hence rcu_seq_start(). */
1781 rcu_seq_start(&rcu_state.gp_seq);
1782 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1783 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1784 raw_spin_unlock_irq_rcu_node(rnp);
1787 * Apply per-leaf buffered online and offline operations to
1788 * the rcu_node tree. Note that this new grace period need not
1789 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1790 * offlining path, when combined with checks in this function,
1791 * will handle CPUs that are currently going offline or that will
1792 * go offline later. Please also refer to "Hotplug CPU" section
1793 * of RCU's Requirements documentation.
1795 rcu_state.gp_state = RCU_GP_ONOFF;
1796 rcu_for_each_leaf_node(rnp) {
1797 smp_mb(); // Pair with barriers used when updating ->ofl_seq to odd values.
1798 firstseq = READ_ONCE(rnp->ofl_seq);
1800 while (firstseq == READ_ONCE(rnp->ofl_seq))
1801 schedule_timeout_idle(1); // Can't wake unless RCU is watching.
1802 smp_mb(); // Pair with barriers used when updating ->ofl_seq to even values.
1803 raw_spin_lock(&rcu_state.ofl_lock);
1804 raw_spin_lock_irq_rcu_node(rnp);
1805 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1806 !rnp->wait_blkd_tasks) {
1807 /* Nothing to do on this leaf rcu_node structure. */
1808 raw_spin_unlock_irq_rcu_node(rnp);
1809 raw_spin_unlock(&rcu_state.ofl_lock);
1813 /* Record old state, apply changes to ->qsmaskinit field. */
1814 oldmask = rnp->qsmaskinit;
1815 rnp->qsmaskinit = rnp->qsmaskinitnext;
1817 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1818 if (!oldmask != !rnp->qsmaskinit) {
1819 if (!oldmask) { /* First online CPU for rcu_node. */
1820 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1821 rcu_init_new_rnp(rnp);
1822 } else if (rcu_preempt_has_tasks(rnp)) {
1823 rnp->wait_blkd_tasks = true; /* blocked tasks */
1824 } else { /* Last offline CPU and can propagate. */
1825 rcu_cleanup_dead_rnp(rnp);
1830 * If all waited-on tasks from prior grace period are
1831 * done, and if all this rcu_node structure's CPUs are
1832 * still offline, propagate up the rcu_node tree and
1833 * clear ->wait_blkd_tasks. Otherwise, if one of this
1834 * rcu_node structure's CPUs has since come back online,
1835 * simply clear ->wait_blkd_tasks.
1837 if (rnp->wait_blkd_tasks &&
1838 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1839 rnp->wait_blkd_tasks = false;
1840 if (!rnp->qsmaskinit)
1841 rcu_cleanup_dead_rnp(rnp);
1844 raw_spin_unlock_irq_rcu_node(rnp);
1845 raw_spin_unlock(&rcu_state.ofl_lock);
1847 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1850 * Set the quiescent-state-needed bits in all the rcu_node
1851 * structures for all currently online CPUs in breadth-first
1852 * order, starting from the root rcu_node structure, relying on the
1853 * layout of the tree within the rcu_state.node[] array. Note that
1854 * other CPUs will access only the leaves of the hierarchy, thus
1855 * seeing that no grace period is in progress, at least until the
1856 * corresponding leaf node has been initialized.
1858 * The grace period cannot complete until the initialization
1859 * process finishes, because this kthread handles both.
1861 rcu_state.gp_state = RCU_GP_INIT;
1862 rcu_for_each_node_breadth_first(rnp) {
1863 rcu_gp_slow(gp_init_delay);
1864 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1865 rdp = this_cpu_ptr(&rcu_data);
1866 rcu_preempt_check_blocked_tasks(rnp);
1867 rnp->qsmask = rnp->qsmaskinit;
1868 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1869 if (rnp == rdp->mynode)
1870 (void)__note_gp_changes(rnp, rdp);
1871 rcu_preempt_boost_start_gp(rnp);
1872 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1873 rnp->level, rnp->grplo,
1874 rnp->grphi, rnp->qsmask);
1875 /* Quiescent states for tasks on any now-offline CPUs. */
1876 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1877 rnp->rcu_gp_init_mask = mask;
1878 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1879 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1881 raw_spin_unlock_irq_rcu_node(rnp);
1882 cond_resched_tasks_rcu_qs();
1883 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1886 // If strict, make all CPUs aware of new grace period.
1887 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1888 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1894 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1897 static bool rcu_gp_fqs_check_wake(int *gfp)
1899 struct rcu_node *rnp = rcu_get_root();
1901 // If under overload conditions, force an immediate FQS scan.
1902 if (*gfp & RCU_GP_FLAG_OVLD)
1905 // Someone like call_rcu() requested a force-quiescent-state scan.
1906 *gfp = READ_ONCE(rcu_state.gp_flags);
1907 if (*gfp & RCU_GP_FLAG_FQS)
1910 // The current grace period has completed.
1911 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1918 * Do one round of quiescent-state forcing.
1920 static void rcu_gp_fqs(bool first_time)
1922 struct rcu_node *rnp = rcu_get_root();
1924 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1925 rcu_state.n_force_qs++;
1927 /* Collect dyntick-idle snapshots. */
1928 force_qs_rnp(dyntick_save_progress_counter);
1930 /* Handle dyntick-idle and offline CPUs. */
1931 force_qs_rnp(rcu_implicit_dynticks_qs);
1933 /* Clear flag to prevent immediate re-entry. */
1934 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1935 raw_spin_lock_irq_rcu_node(rnp);
1936 WRITE_ONCE(rcu_state.gp_flags,
1937 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1938 raw_spin_unlock_irq_rcu_node(rnp);
1943 * Loop doing repeated quiescent-state forcing until the grace period ends.
1945 static void rcu_gp_fqs_loop(void)
1951 struct rcu_node *rnp = rcu_get_root();
1953 first_gp_fqs = true;
1954 j = READ_ONCE(jiffies_till_first_fqs);
1955 if (rcu_state.cbovld)
1956 gf = RCU_GP_FLAG_OVLD;
1960 rcu_state.jiffies_force_qs = jiffies + j;
1961 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1962 jiffies + (j ? 3 * j : 2));
1964 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1966 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1967 ret = swait_event_idle_timeout_exclusive(
1968 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1969 rcu_gp_torture_wait();
1970 rcu_state.gp_state = RCU_GP_DOING_FQS;
1971 /* Locking provides needed memory barriers. */
1972 /* If grace period done, leave loop. */
1973 if (!READ_ONCE(rnp->qsmask) &&
1974 !rcu_preempt_blocked_readers_cgp(rnp))
1976 /* If time for quiescent-state forcing, do it. */
1977 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1978 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1979 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1981 rcu_gp_fqs(first_gp_fqs);
1984 first_gp_fqs = false;
1985 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1987 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1989 cond_resched_tasks_rcu_qs();
1990 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1991 ret = 0; /* Force full wait till next FQS. */
1992 j = READ_ONCE(jiffies_till_next_fqs);
1994 /* Deal with stray signal. */
1995 cond_resched_tasks_rcu_qs();
1996 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1997 WARN_ON(signal_pending(current));
1998 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2000 ret = 1; /* Keep old FQS timing. */
2002 if (time_after(jiffies, rcu_state.jiffies_force_qs))
2005 j = rcu_state.jiffies_force_qs - j;
2012 * Clean up after the old grace period.
2014 static void rcu_gp_cleanup(void)
2017 bool needgp = false;
2018 unsigned long gp_duration;
2019 unsigned long new_gp_seq;
2021 struct rcu_data *rdp;
2022 struct rcu_node *rnp = rcu_get_root();
2023 struct swait_queue_head *sq;
2025 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2026 raw_spin_lock_irq_rcu_node(rnp);
2027 rcu_state.gp_end = jiffies;
2028 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
2029 if (gp_duration > rcu_state.gp_max)
2030 rcu_state.gp_max = gp_duration;
2033 * We know the grace period is complete, but to everyone else
2034 * it appears to still be ongoing. But it is also the case
2035 * that to everyone else it looks like there is nothing that
2036 * they can do to advance the grace period. It is therefore
2037 * safe for us to drop the lock in order to mark the grace
2038 * period as completed in all of the rcu_node structures.
2040 raw_spin_unlock_irq_rcu_node(rnp);
2043 * Propagate new ->gp_seq value to rcu_node structures so that
2044 * other CPUs don't have to wait until the start of the next grace
2045 * period to process their callbacks. This also avoids some nasty
2046 * RCU grace-period initialization races by forcing the end of
2047 * the current grace period to be completely recorded in all of
2048 * the rcu_node structures before the beginning of the next grace
2049 * period is recorded in any of the rcu_node structures.
2051 new_gp_seq = rcu_state.gp_seq;
2052 rcu_seq_end(&new_gp_seq);
2053 rcu_for_each_node_breadth_first(rnp) {
2054 raw_spin_lock_irq_rcu_node(rnp);
2055 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2056 dump_blkd_tasks(rnp, 10);
2057 WARN_ON_ONCE(rnp->qsmask);
2058 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2059 rdp = this_cpu_ptr(&rcu_data);
2060 if (rnp == rdp->mynode)
2061 needgp = __note_gp_changes(rnp, rdp) || needgp;
2062 /* smp_mb() provided by prior unlock-lock pair. */
2063 needgp = rcu_future_gp_cleanup(rnp) || needgp;
2064 // Reset overload indication for CPUs no longer overloaded
2065 if (rcu_is_leaf_node(rnp))
2066 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
2067 rdp = per_cpu_ptr(&rcu_data, cpu);
2068 check_cb_ovld_locked(rdp, rnp);
2070 sq = rcu_nocb_gp_get(rnp);
2071 raw_spin_unlock_irq_rcu_node(rnp);
2072 rcu_nocb_gp_cleanup(sq);
2073 cond_resched_tasks_rcu_qs();
2074 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2075 rcu_gp_slow(gp_cleanup_delay);
2077 rnp = rcu_get_root();
2078 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2080 /* Declare grace period done, trace first to use old GP number. */
2081 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2082 rcu_seq_end(&rcu_state.gp_seq);
2083 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2084 rcu_state.gp_state = RCU_GP_IDLE;
2085 /* Check for GP requests since above loop. */
2086 rdp = this_cpu_ptr(&rcu_data);
2087 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2088 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2089 TPS("CleanupMore"));
2092 /* Advance CBs to reduce false positives below. */
2093 offloaded = rcu_segcblist_is_offloaded(&rdp->cblist);
2094 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
2095 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2096 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2097 trace_rcu_grace_period(rcu_state.name,
2101 WRITE_ONCE(rcu_state.gp_flags,
2102 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2104 raw_spin_unlock_irq_rcu_node(rnp);
2106 // If strict, make all CPUs aware of the end of the old grace period.
2107 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2108 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
2112 * Body of kthread that handles grace periods.
2114 static int __noreturn rcu_gp_kthread(void *unused)
2116 rcu_bind_gp_kthread();
2119 /* Handle grace-period start. */
2121 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2123 rcu_state.gp_state = RCU_GP_WAIT_GPS;
2124 swait_event_idle_exclusive(rcu_state.gp_wq,
2125 READ_ONCE(rcu_state.gp_flags) &
2127 rcu_gp_torture_wait();
2128 rcu_state.gp_state = RCU_GP_DONE_GPS;
2129 /* Locking provides needed memory barrier. */
2132 cond_resched_tasks_rcu_qs();
2133 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2134 WARN_ON(signal_pending(current));
2135 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2139 /* Handle quiescent-state forcing. */
2142 /* Handle grace-period end. */
2143 rcu_state.gp_state = RCU_GP_CLEANUP;
2145 rcu_state.gp_state = RCU_GP_CLEANED;
2150 * Report a full set of quiescent states to the rcu_state data structure.
2151 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2152 * another grace period is required. Whether we wake the grace-period
2153 * kthread or it awakens itself for the next round of quiescent-state
2154 * forcing, that kthread will clean up after the just-completed grace
2155 * period. Note that the caller must hold rnp->lock, which is released
2158 static void rcu_report_qs_rsp(unsigned long flags)
2159 __releases(rcu_get_root()->lock)
2161 raw_lockdep_assert_held_rcu_node(rcu_get_root());
2162 WARN_ON_ONCE(!rcu_gp_in_progress());
2163 WRITE_ONCE(rcu_state.gp_flags,
2164 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2165 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2166 rcu_gp_kthread_wake();
2170 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2171 * Allows quiescent states for a group of CPUs to be reported at one go
2172 * to the specified rcu_node structure, though all the CPUs in the group
2173 * must be represented by the same rcu_node structure (which need not be a
2174 * leaf rcu_node structure, though it often will be). The gps parameter
2175 * is the grace-period snapshot, which means that the quiescent states
2176 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2177 * must be held upon entry, and it is released before return.
2179 * As a special case, if mask is zero, the bit-already-cleared check is
2180 * disabled. This allows propagating quiescent state due to resumed tasks
2181 * during grace-period initialization.
2183 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2184 unsigned long gps, unsigned long flags)
2185 __releases(rnp->lock)
2187 unsigned long oldmask = 0;
2188 struct rcu_node *rnp_c;
2190 raw_lockdep_assert_held_rcu_node(rnp);
2192 /* Walk up the rcu_node hierarchy. */
2194 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2197 * Our bit has already been cleared, or the
2198 * relevant grace period is already over, so done.
2200 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2203 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2204 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2205 rcu_preempt_blocked_readers_cgp(rnp));
2206 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2207 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2208 mask, rnp->qsmask, rnp->level,
2209 rnp->grplo, rnp->grphi,
2211 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2213 /* Other bits still set at this level, so done. */
2214 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2217 rnp->completedqs = rnp->gp_seq;
2218 mask = rnp->grpmask;
2219 if (rnp->parent == NULL) {
2221 /* No more levels. Exit loop holding root lock. */
2225 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2228 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2229 oldmask = READ_ONCE(rnp_c->qsmask);
2233 * Get here if we are the last CPU to pass through a quiescent
2234 * state for this grace period. Invoke rcu_report_qs_rsp()
2235 * to clean up and start the next grace period if one is needed.
2237 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2241 * Record a quiescent state for all tasks that were previously queued
2242 * on the specified rcu_node structure and that were blocking the current
2243 * RCU grace period. The caller must hold the corresponding rnp->lock with
2244 * irqs disabled, and this lock is released upon return, but irqs remain
2247 static void __maybe_unused
2248 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2249 __releases(rnp->lock)
2253 struct rcu_node *rnp_p;
2255 raw_lockdep_assert_held_rcu_node(rnp);
2256 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2257 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2259 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2260 return; /* Still need more quiescent states! */
2263 rnp->completedqs = rnp->gp_seq;
2264 rnp_p = rnp->parent;
2265 if (rnp_p == NULL) {
2267 * Only one rcu_node structure in the tree, so don't
2268 * try to report up to its nonexistent parent!
2270 rcu_report_qs_rsp(flags);
2274 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2276 mask = rnp->grpmask;
2277 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2278 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2279 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2283 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2284 * structure. This must be called from the specified CPU.
2287 rcu_report_qs_rdp(struct rcu_data *rdp)
2289 unsigned long flags;
2291 bool needwake = false;
2292 const bool offloaded = rcu_segcblist_is_offloaded(&rdp->cblist);
2293 struct rcu_node *rnp;
2295 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2297 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2298 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2302 * The grace period in which this quiescent state was
2303 * recorded has ended, so don't report it upwards.
2304 * We will instead need a new quiescent state that lies
2305 * within the current grace period.
2307 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2308 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2311 mask = rdp->grpmask;
2312 rdp->core_needs_qs = false;
2313 if ((rnp->qsmask & mask) == 0) {
2314 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2317 * This GP can't end until cpu checks in, so all of our
2318 * callbacks can be processed during the next GP.
2321 needwake = rcu_accelerate_cbs(rnp, rdp);
2323 rcu_disable_urgency_upon_qs(rdp);
2324 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2325 /* ^^^ Released rnp->lock */
2327 rcu_gp_kthread_wake();
2332 * Check to see if there is a new grace period of which this CPU
2333 * is not yet aware, and if so, set up local rcu_data state for it.
2334 * Otherwise, see if this CPU has just passed through its first
2335 * quiescent state for this grace period, and record that fact if so.
2338 rcu_check_quiescent_state(struct rcu_data *rdp)
2340 /* Check for grace-period ends and beginnings. */
2341 note_gp_changes(rdp);
2344 * Does this CPU still need to do its part for current grace period?
2345 * If no, return and let the other CPUs do their part as well.
2347 if (!rdp->core_needs_qs)
2351 * Was there a quiescent state since the beginning of the grace
2352 * period? If no, then exit and wait for the next call.
2354 if (rdp->cpu_no_qs.b.norm)
2358 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2361 rcu_report_qs_rdp(rdp);
2365 * Near the end of the offline process. Trace the fact that this CPU
2368 int rcutree_dying_cpu(unsigned int cpu)
2371 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2372 struct rcu_node *rnp = rdp->mynode;
2374 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2377 blkd = !!(rnp->qsmask & rdp->grpmask);
2378 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2379 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2384 * All CPUs for the specified rcu_node structure have gone offline,
2385 * and all tasks that were preempted within an RCU read-side critical
2386 * section while running on one of those CPUs have since exited their RCU
2387 * read-side critical section. Some other CPU is reporting this fact with
2388 * the specified rcu_node structure's ->lock held and interrupts disabled.
2389 * This function therefore goes up the tree of rcu_node structures,
2390 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2391 * the leaf rcu_node structure's ->qsmaskinit field has already been
2394 * This function does check that the specified rcu_node structure has
2395 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2396 * prematurely. That said, invoking it after the fact will cost you
2397 * a needless lock acquisition. So once it has done its work, don't
2400 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2403 struct rcu_node *rnp = rnp_leaf;
2405 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2406 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2407 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2408 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2411 mask = rnp->grpmask;
2415 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2416 rnp->qsmaskinit &= ~mask;
2417 /* Between grace periods, so better already be zero! */
2418 WARN_ON_ONCE(rnp->qsmask);
2419 if (rnp->qsmaskinit) {
2420 raw_spin_unlock_rcu_node(rnp);
2421 /* irqs remain disabled. */
2424 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2429 * The CPU has been completely removed, and some other CPU is reporting
2430 * this fact from process context. Do the remainder of the cleanup.
2431 * There can only be one CPU hotplug operation at a time, so no need for
2434 int rcutree_dead_cpu(unsigned int cpu)
2436 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2437 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2439 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2442 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
2443 /* Adjust any no-longer-needed kthreads. */
2444 rcu_boost_kthread_setaffinity(rnp, -1);
2445 /* Do any needed no-CB deferred wakeups from this CPU. */
2446 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2448 // Stop-machine done, so allow nohz_full to disable tick.
2449 tick_dep_clear(TICK_DEP_BIT_RCU);
2454 * Invoke any RCU callbacks that have made it to the end of their grace
2455 * period. Thottle as specified by rdp->blimit.
2457 static void rcu_do_batch(struct rcu_data *rdp)
2460 unsigned long flags;
2461 const bool offloaded = rcu_segcblist_is_offloaded(&rdp->cblist);
2462 struct rcu_head *rhp;
2463 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2465 long pending, tlimit = 0;
2467 /* If no callbacks are ready, just return. */
2468 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2469 trace_rcu_batch_start(rcu_state.name,
2470 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2471 trace_rcu_batch_end(rcu_state.name, 0,
2472 !rcu_segcblist_empty(&rdp->cblist),
2473 need_resched(), is_idle_task(current),
2474 rcu_is_callbacks_kthread());
2479 * Extract the list of ready callbacks, disabling to prevent
2480 * races with call_rcu() from interrupt handlers. Leave the
2481 * callback counts, as rcu_barrier() needs to be conservative.
2483 local_irq_save(flags);
2485 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2486 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2487 div = READ_ONCE(rcu_divisor);
2488 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2489 bl = max(rdp->blimit, pending >> div);
2490 if (unlikely(bl > 100)) {
2491 long rrn = READ_ONCE(rcu_resched_ns);
2493 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2494 tlimit = local_clock() + rrn;
2496 trace_rcu_batch_start(rcu_state.name,
2497 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2498 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2500 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2501 rcu_nocb_unlock_irqrestore(rdp, flags);
2503 /* Invoke callbacks. */
2504 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2505 rhp = rcu_cblist_dequeue(&rcl);
2506 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2509 debug_rcu_head_unqueue(rhp);
2511 rcu_lock_acquire(&rcu_callback_map);
2512 trace_rcu_invoke_callback(rcu_state.name, rhp);
2515 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2518 rcu_lock_release(&rcu_callback_map);
2521 * Stop only if limit reached and CPU has something to do.
2522 * Note: The rcl structure counts down from zero.
2524 if (-rcl.len >= bl && !offloaded &&
2526 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2528 if (unlikely(tlimit)) {
2529 /* only call local_clock() every 32 callbacks */
2530 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2532 /* Exceeded the time limit, so leave. */
2536 WARN_ON_ONCE(in_serving_softirq());
2538 lockdep_assert_irqs_enabled();
2539 cond_resched_tasks_rcu_qs();
2540 lockdep_assert_irqs_enabled();
2545 local_irq_save(flags);
2548 rdp->n_cbs_invoked += count;
2549 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2550 is_idle_task(current), rcu_is_callbacks_kthread());
2552 /* Update counts and requeue any remaining callbacks. */
2553 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2554 smp_mb(); /* List handling before counting for rcu_barrier(). */
2555 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2557 /* Reinstate batch limit if we have worked down the excess. */
2558 count = rcu_segcblist_n_cbs(&rdp->cblist);
2559 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2560 rdp->blimit = blimit;
2562 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2563 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2564 rdp->qlen_last_fqs_check = 0;
2565 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2566 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2567 rdp->qlen_last_fqs_check = count;
2570 * The following usually indicates a double call_rcu(). To track
2571 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2573 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2574 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2575 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2577 rcu_nocb_unlock_irqrestore(rdp, flags);
2579 /* Re-invoke RCU core processing if there are callbacks remaining. */
2580 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2582 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2586 * This function is invoked from each scheduling-clock interrupt,
2587 * and checks to see if this CPU is in a non-context-switch quiescent
2588 * state, for example, user mode or idle loop. It also schedules RCU
2589 * core processing. If the current grace period has gone on too long,
2590 * it will ask the scheduler to manufacture a context switch for the sole
2591 * purpose of providing a providing the needed quiescent state.
2593 void rcu_sched_clock_irq(int user)
2595 trace_rcu_utilization(TPS("Start scheduler-tick"));
2596 raw_cpu_inc(rcu_data.ticks_this_gp);
2597 /* The load-acquire pairs with the store-release setting to true. */
2598 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2599 /* Idle and userspace execution already are quiescent states. */
2600 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2601 set_tsk_need_resched(current);
2602 set_preempt_need_resched();
2604 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2606 rcu_flavor_sched_clock_irq(user);
2607 if (rcu_pending(user))
2610 trace_rcu_utilization(TPS("End scheduler-tick"));
2614 * Scan the leaf rcu_node structures. For each structure on which all
2615 * CPUs have reported a quiescent state and on which there are tasks
2616 * blocking the current grace period, initiate RCU priority boosting.
2617 * Otherwise, invoke the specified function to check dyntick state for
2618 * each CPU that has not yet reported a quiescent state.
2620 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2623 unsigned long flags;
2625 struct rcu_data *rdp;
2626 struct rcu_node *rnp;
2628 rcu_state.cbovld = rcu_state.cbovldnext;
2629 rcu_state.cbovldnext = false;
2630 rcu_for_each_leaf_node(rnp) {
2631 cond_resched_tasks_rcu_qs();
2633 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2634 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2635 if (rnp->qsmask == 0) {
2636 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2638 * No point in scanning bits because they
2639 * are all zero. But we might need to
2640 * priority-boost blocked readers.
2642 rcu_initiate_boost(rnp, flags);
2643 /* rcu_initiate_boost() releases rnp->lock */
2646 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2649 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2650 rdp = per_cpu_ptr(&rcu_data, cpu);
2652 mask |= rdp->grpmask;
2653 rcu_disable_urgency_upon_qs(rdp);
2657 /* Idle/offline CPUs, report (releases rnp->lock). */
2658 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2660 /* Nothing to do here, so just drop the lock. */
2661 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2667 * Force quiescent states on reluctant CPUs, and also detect which
2668 * CPUs are in dyntick-idle mode.
2670 void rcu_force_quiescent_state(void)
2672 unsigned long flags;
2674 struct rcu_node *rnp;
2675 struct rcu_node *rnp_old = NULL;
2677 /* Funnel through hierarchy to reduce memory contention. */
2678 rnp = __this_cpu_read(rcu_data.mynode);
2679 for (; rnp != NULL; rnp = rnp->parent) {
2680 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2681 !raw_spin_trylock(&rnp->fqslock);
2682 if (rnp_old != NULL)
2683 raw_spin_unlock(&rnp_old->fqslock);
2688 /* rnp_old == rcu_get_root(), rnp == NULL. */
2690 /* Reached the root of the rcu_node tree, acquire lock. */
2691 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2692 raw_spin_unlock(&rnp_old->fqslock);
2693 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2694 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2695 return; /* Someone beat us to it. */
2697 WRITE_ONCE(rcu_state.gp_flags,
2698 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2699 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2700 rcu_gp_kthread_wake();
2702 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2704 // Workqueue handler for an RCU reader for kernels enforcing struct RCU
2706 static void strict_work_handler(struct work_struct *work)
2712 /* Perform RCU core processing work for the current CPU. */
2713 static __latent_entropy void rcu_core(void)
2715 unsigned long flags;
2716 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2717 struct rcu_node *rnp = rdp->mynode;
2718 const bool offloaded = rcu_segcblist_is_offloaded(&rdp->cblist);
2720 if (cpu_is_offline(smp_processor_id()))
2722 trace_rcu_utilization(TPS("Start RCU core"));
2723 WARN_ON_ONCE(!rdp->beenonline);
2725 /* Report any deferred quiescent states if preemption enabled. */
2726 if (!(preempt_count() & PREEMPT_MASK)) {
2727 rcu_preempt_deferred_qs(current);
2728 } else if (rcu_preempt_need_deferred_qs(current)) {
2729 set_tsk_need_resched(current);
2730 set_preempt_need_resched();
2733 /* Update RCU state based on any recent quiescent states. */
2734 rcu_check_quiescent_state(rdp);
2736 /* No grace period and unregistered callbacks? */
2737 if (!rcu_gp_in_progress() &&
2738 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2739 local_irq_save(flags);
2740 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2741 rcu_accelerate_cbs_unlocked(rnp, rdp);
2742 local_irq_restore(flags);
2745 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2747 /* If there are callbacks ready, invoke them. */
2748 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2749 likely(READ_ONCE(rcu_scheduler_fully_active)))
2752 /* Do any needed deferred wakeups of rcuo kthreads. */
2753 do_nocb_deferred_wakeup(rdp);
2754 trace_rcu_utilization(TPS("End RCU core"));
2756 // If strict GPs, schedule an RCU reader in a clean environment.
2757 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2758 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2761 static void rcu_core_si(struct softirq_action *h)
2766 static void rcu_wake_cond(struct task_struct *t, int status)
2769 * If the thread is yielding, only wake it when this
2770 * is invoked from idle
2772 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2776 static void invoke_rcu_core_kthread(void)
2778 struct task_struct *t;
2779 unsigned long flags;
2781 local_irq_save(flags);
2782 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2783 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2784 if (t != NULL && t != current)
2785 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2786 local_irq_restore(flags);
2790 * Wake up this CPU's rcuc kthread to do RCU core processing.
2792 static void invoke_rcu_core(void)
2794 if (!cpu_online(smp_processor_id()))
2797 raise_softirq(RCU_SOFTIRQ);
2799 invoke_rcu_core_kthread();
2802 static void rcu_cpu_kthread_park(unsigned int cpu)
2804 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2807 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2809 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2813 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2814 * the RCU softirq used in configurations of RCU that do not support RCU
2815 * priority boosting.
2817 static void rcu_cpu_kthread(unsigned int cpu)
2819 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2820 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2823 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2824 for (spincnt = 0; spincnt < 10; spincnt++) {
2826 *statusp = RCU_KTHREAD_RUNNING;
2827 local_irq_disable();
2835 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2836 *statusp = RCU_KTHREAD_WAITING;
2840 *statusp = RCU_KTHREAD_YIELDING;
2841 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2842 schedule_timeout_idle(2);
2843 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2844 *statusp = RCU_KTHREAD_WAITING;
2847 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2848 .store = &rcu_data.rcu_cpu_kthread_task,
2849 .thread_should_run = rcu_cpu_kthread_should_run,
2850 .thread_fn = rcu_cpu_kthread,
2851 .thread_comm = "rcuc/%u",
2852 .setup = rcu_cpu_kthread_setup,
2853 .park = rcu_cpu_kthread_park,
2857 * Spawn per-CPU RCU core processing kthreads.
2859 static int __init rcu_spawn_core_kthreads(void)
2863 for_each_possible_cpu(cpu)
2864 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2865 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2867 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2868 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2871 early_initcall(rcu_spawn_core_kthreads);
2874 * Handle any core-RCU processing required by a call_rcu() invocation.
2876 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2877 unsigned long flags)
2880 * If called from an extended quiescent state, invoke the RCU
2881 * core in order to force a re-evaluation of RCU's idleness.
2883 if (!rcu_is_watching())
2886 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2887 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2891 * Force the grace period if too many callbacks or too long waiting.
2892 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2893 * if some other CPU has recently done so. Also, don't bother
2894 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2895 * is the only one waiting for a grace period to complete.
2897 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2898 rdp->qlen_last_fqs_check + qhimark)) {
2900 /* Are we ignoring a completed grace period? */
2901 note_gp_changes(rdp);
2903 /* Start a new grace period if one not already started. */
2904 if (!rcu_gp_in_progress()) {
2905 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2907 /* Give the grace period a kick. */
2908 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2909 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2910 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2911 rcu_force_quiescent_state();
2912 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2913 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2919 * RCU callback function to leak a callback.
2921 static void rcu_leak_callback(struct rcu_head *rhp)
2926 * Check and if necessary update the leaf rcu_node structure's
2927 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2928 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2929 * structure's ->lock.
2931 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2933 raw_lockdep_assert_held_rcu_node(rnp);
2934 if (qovld_calc <= 0)
2935 return; // Early boot and wildcard value set.
2936 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2937 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2939 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2943 * Check and if necessary update the leaf rcu_node structure's
2944 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2945 * number of queued RCU callbacks. No locks need be held, but the
2946 * caller must have disabled interrupts.
2948 * Note that this function ignores the possibility that there are a lot
2949 * of callbacks all of which have already seen the end of their respective
2950 * grace periods. This omission is due to the need for no-CBs CPUs to
2951 * be holding ->nocb_lock to do this check, which is too heavy for a
2952 * common-case operation.
2954 static void check_cb_ovld(struct rcu_data *rdp)
2956 struct rcu_node *const rnp = rdp->mynode;
2958 if (qovld_calc <= 0 ||
2959 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2960 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2961 return; // Early boot wildcard value or already set correctly.
2962 raw_spin_lock_rcu_node(rnp);
2963 check_cb_ovld_locked(rdp, rnp);
2964 raw_spin_unlock_rcu_node(rnp);
2967 /* Helper function for call_rcu() and friends. */
2969 __call_rcu(struct rcu_head *head, rcu_callback_t func)
2971 unsigned long flags;
2972 struct rcu_data *rdp;
2975 /* Misaligned rcu_head! */
2976 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2978 if (debug_rcu_head_queue(head)) {
2980 * Probable double call_rcu(), so leak the callback.
2981 * Use rcu:rcu_callback trace event to find the previous
2982 * time callback was passed to __call_rcu().
2984 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2986 WRITE_ONCE(head->func, rcu_leak_callback);
2991 local_irq_save(flags);
2992 kasan_record_aux_stack(head);
2993 rdp = this_cpu_ptr(&rcu_data);
2995 /* Add the callback to our list. */
2996 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2997 // This can trigger due to call_rcu() from offline CPU:
2998 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2999 WARN_ON_ONCE(!rcu_is_watching());
3000 // Very early boot, before rcu_init(). Initialize if needed
3001 // and then drop through to queue the callback.
3002 if (rcu_segcblist_empty(&rdp->cblist))
3003 rcu_segcblist_init(&rdp->cblist);
3007 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
3008 return; // Enqueued onto ->nocb_bypass, so just leave.
3009 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
3010 rcu_segcblist_enqueue(&rdp->cblist, head);
3011 if (__is_kvfree_rcu_offset((unsigned long)func))
3012 trace_rcu_kvfree_callback(rcu_state.name, head,
3013 (unsigned long)func,
3014 rcu_segcblist_n_cbs(&rdp->cblist));
3016 trace_rcu_callback(rcu_state.name, head,
3017 rcu_segcblist_n_cbs(&rdp->cblist));
3019 /* Go handle any RCU core processing required. */
3020 if (unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
3021 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
3023 __call_rcu_core(rdp, head, flags);
3024 local_irq_restore(flags);
3029 * call_rcu() - Queue an RCU callback for invocation after a grace period.
3030 * @head: structure to be used for queueing the RCU updates.
3031 * @func: actual callback function to be invoked after the grace period
3033 * The callback function will be invoked some time after a full grace
3034 * period elapses, in other words after all pre-existing RCU read-side
3035 * critical sections have completed. However, the callback function
3036 * might well execute concurrently with RCU read-side critical sections
3037 * that started after call_rcu() was invoked. RCU read-side critical
3038 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
3039 * may be nested. In addition, regions of code across which interrupts,
3040 * preemption, or softirqs have been disabled also serve as RCU read-side
3041 * critical sections. This includes hardware interrupt handlers, softirq
3042 * handlers, and NMI handlers.
3044 * Note that all CPUs must agree that the grace period extended beyond
3045 * all pre-existing RCU read-side critical section. On systems with more
3046 * than one CPU, this means that when "func()" is invoked, each CPU is
3047 * guaranteed to have executed a full memory barrier since the end of its
3048 * last RCU read-side critical section whose beginning preceded the call
3049 * to call_rcu(). It also means that each CPU executing an RCU read-side
3050 * critical section that continues beyond the start of "func()" must have
3051 * executed a memory barrier after the call_rcu() but before the beginning
3052 * of that RCU read-side critical section. Note that these guarantees
3053 * include CPUs that are offline, idle, or executing in user mode, as
3054 * well as CPUs that are executing in the kernel.
3056 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
3057 * resulting RCU callback function "func()", then both CPU A and CPU B are
3058 * guaranteed to execute a full memory barrier during the time interval
3059 * between the call to call_rcu() and the invocation of "func()" -- even
3060 * if CPU A and CPU B are the same CPU (but again only if the system has
3061 * more than one CPU).
3063 void call_rcu(struct rcu_head *head, rcu_callback_t func)
3065 __call_rcu(head, func);
3067 EXPORT_SYMBOL_GPL(call_rcu);
3070 /* Maximum number of jiffies to wait before draining a batch. */
3071 #define KFREE_DRAIN_JIFFIES (HZ / 50)
3072 #define KFREE_N_BATCHES 2
3073 #define FREE_N_CHANNELS 2
3076 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
3077 * @nr_records: Number of active pointers in the array
3078 * @next: Next bulk object in the block chain
3079 * @records: Array of the kvfree_rcu() pointers
3081 struct kvfree_rcu_bulk_data {
3082 unsigned long nr_records;
3083 struct kvfree_rcu_bulk_data *next;
3088 * This macro defines how many entries the "records" array
3089 * will contain. It is based on the fact that the size of
3090 * kvfree_rcu_bulk_data structure becomes exactly one page.
3092 #define KVFREE_BULK_MAX_ENTR \
3093 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
3096 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3097 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3098 * @head_free: List of kfree_rcu() objects waiting for a grace period
3099 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
3100 * @krcp: Pointer to @kfree_rcu_cpu structure
3103 struct kfree_rcu_cpu_work {
3104 struct rcu_work rcu_work;
3105 struct rcu_head *head_free;
3106 struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
3107 struct kfree_rcu_cpu *krcp;
3111 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3112 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3113 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
3114 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3115 * @lock: Synchronize access to this structure
3116 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3117 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
3118 * @initialized: The @rcu_work fields have been initialized
3119 * @count: Number of objects for which GP not started
3121 * A simple cache list that contains objects for reuse purpose.
3122 * In order to save some per-cpu space the list is singular.
3123 * Even though it is lockless an access has to be protected by the
3125 * @page_cache_work: A work to refill the cache when it is empty
3126 * @work_in_progress: Indicates that page_cache_work is running
3127 * @hrtimer: A hrtimer for scheduling a page_cache_work
3128 * @nr_bkv_objs: number of allocated objects at @bkvcache.
3130 * This is a per-CPU structure. The reason that it is not included in
3131 * the rcu_data structure is to permit this code to be extracted from
3132 * the RCU files. Such extraction could allow further optimization of
3133 * the interactions with the slab allocators.
3135 struct kfree_rcu_cpu {
3136 struct rcu_head *head;
3137 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
3138 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3139 raw_spinlock_t lock;
3140 struct delayed_work monitor_work;
3145 struct work_struct page_cache_work;
3146 atomic_t work_in_progress;
3147 struct hrtimer hrtimer;
3149 struct llist_head bkvcache;
3153 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3154 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3157 static __always_inline void
3158 debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
3160 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3163 for (i = 0; i < bhead->nr_records; i++)
3164 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3168 static inline struct kfree_rcu_cpu *
3169 krc_this_cpu_lock(unsigned long *flags)
3171 struct kfree_rcu_cpu *krcp;
3173 local_irq_save(*flags); // For safely calling this_cpu_ptr().
3174 krcp = this_cpu_ptr(&krc);
3175 raw_spin_lock(&krcp->lock);
3181 krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3183 raw_spin_unlock(&krcp->lock);
3184 local_irq_restore(flags);
3187 static inline struct kvfree_rcu_bulk_data *
3188 get_cached_bnode(struct kfree_rcu_cpu *krcp)
3190 if (!krcp->nr_bkv_objs)
3193 krcp->nr_bkv_objs--;
3194 return (struct kvfree_rcu_bulk_data *)
3195 llist_del_first(&krcp->bkvcache);
3199 put_cached_bnode(struct kfree_rcu_cpu *krcp,
3200 struct kvfree_rcu_bulk_data *bnode)
3203 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3206 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3207 krcp->nr_bkv_objs++;
3213 * This function is invoked in workqueue context after a grace period.
3214 * It frees all the objects queued on ->bhead_free or ->head_free.
3216 static void kfree_rcu_work(struct work_struct *work)
3218 unsigned long flags;
3219 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3220 struct rcu_head *head, *next;
3221 struct kfree_rcu_cpu *krcp;
3222 struct kfree_rcu_cpu_work *krwp;
3225 krwp = container_of(to_rcu_work(work),
3226 struct kfree_rcu_cpu_work, rcu_work);
3229 raw_spin_lock_irqsave(&krcp->lock, flags);
3230 // Channels 1 and 2.
3231 for (i = 0; i < FREE_N_CHANNELS; i++) {
3232 bkvhead[i] = krwp->bkvhead_free[i];
3233 krwp->bkvhead_free[i] = NULL;
3237 head = krwp->head_free;
3238 krwp->head_free = NULL;
3239 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3241 // Handle two first channels.
3242 for (i = 0; i < FREE_N_CHANNELS; i++) {
3243 for (; bkvhead[i]; bkvhead[i] = bnext) {
3244 bnext = bkvhead[i]->next;
3245 debug_rcu_bhead_unqueue(bkvhead[i]);
3247 rcu_lock_acquire(&rcu_callback_map);
3248 if (i == 0) { // kmalloc() / kfree().
3249 trace_rcu_invoke_kfree_bulk_callback(
3250 rcu_state.name, bkvhead[i]->nr_records,
3251 bkvhead[i]->records);
3253 kfree_bulk(bkvhead[i]->nr_records,
3254 bkvhead[i]->records);
3255 } else { // vmalloc() / vfree().
3256 for (j = 0; j < bkvhead[i]->nr_records; j++) {
3257 trace_rcu_invoke_kvfree_callback(
3259 bkvhead[i]->records[j], 0);
3261 vfree(bkvhead[i]->records[j]);
3264 rcu_lock_release(&rcu_callback_map);
3266 raw_spin_lock_irqsave(&krcp->lock, flags);
3267 if (put_cached_bnode(krcp, bkvhead[i]))
3269 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3272 free_page((unsigned long) bkvhead[i]);
3274 cond_resched_tasks_rcu_qs();
3279 * Emergency case only. It can happen under low memory
3280 * condition when an allocation gets failed, so the "bulk"
3281 * path can not be temporary maintained.
3283 for (; head; head = next) {
3284 unsigned long offset = (unsigned long)head->func;
3285 void *ptr = (void *)head - offset;
3288 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3289 rcu_lock_acquire(&rcu_callback_map);
3290 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3292 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3295 rcu_lock_release(&rcu_callback_map);
3296 cond_resched_tasks_rcu_qs();
3301 * Schedule the kfree batch RCU work to run in workqueue context after a GP.
3303 * This function is invoked by kfree_rcu_monitor() when the KFREE_DRAIN_JIFFIES
3304 * timeout has been reached.
3306 static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
3308 struct kfree_rcu_cpu_work *krwp;
3309 bool repeat = false;
3312 lockdep_assert_held(&krcp->lock);
3314 for (i = 0; i < KFREE_N_BATCHES; i++) {
3315 krwp = &(krcp->krw_arr[i]);
3318 * Try to detach bkvhead or head and attach it over any
3319 * available corresponding free channel. It can be that
3320 * a previous RCU batch is in progress, it means that
3321 * immediately to queue another one is not possible so
3322 * return false to tell caller to retry.
3324 if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3325 (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3326 (krcp->head && !krwp->head_free)) {
3327 // Channel 1 corresponds to SLAB ptrs.
3328 // Channel 2 corresponds to vmalloc ptrs.
3329 for (j = 0; j < FREE_N_CHANNELS; j++) {
3330 if (!krwp->bkvhead_free[j]) {
3331 krwp->bkvhead_free[j] = krcp->bkvhead[j];
3332 krcp->bkvhead[j] = NULL;
3336 // Channel 3 corresponds to emergency path.
3337 if (!krwp->head_free) {
3338 krwp->head_free = krcp->head;
3342 WRITE_ONCE(krcp->count, 0);
3345 * One work is per one batch, so there are three
3346 * "free channels", the batch can handle. It can
3347 * be that the work is in the pending state when
3348 * channels have been detached following by each
3351 queue_rcu_work(system_wq, &krwp->rcu_work);
3354 // Repeat if any "free" corresponding channel is still busy.
3355 if (krcp->bkvhead[0] || krcp->bkvhead[1] || krcp->head)
3362 static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp,
3363 unsigned long flags)
3365 // Attempt to start a new batch.
3366 krcp->monitor_todo = false;
3367 if (queue_kfree_rcu_work(krcp)) {
3368 // Success! Our job is done here.
3369 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3373 // Previous RCU batch still in progress, try again later.
3374 krcp->monitor_todo = true;
3375 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3376 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3380 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3381 * It invokes kfree_rcu_drain_unlock() to attempt to start another batch.
3383 static void kfree_rcu_monitor(struct work_struct *work)
3385 unsigned long flags;
3386 struct kfree_rcu_cpu *krcp = container_of(work, struct kfree_rcu_cpu,
3389 raw_spin_lock_irqsave(&krcp->lock, flags);
3390 if (krcp->monitor_todo)
3391 kfree_rcu_drain_unlock(krcp, flags);
3393 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3396 static enum hrtimer_restart
3397 schedule_page_work_fn(struct hrtimer *t)
3399 struct kfree_rcu_cpu *krcp =
3400 container_of(t, struct kfree_rcu_cpu, hrtimer);
3402 queue_work(system_highpri_wq, &krcp->page_cache_work);
3403 return HRTIMER_NORESTART;
3406 static void fill_page_cache_func(struct work_struct *work)
3408 struct kvfree_rcu_bulk_data *bnode;
3409 struct kfree_rcu_cpu *krcp =
3410 container_of(work, struct kfree_rcu_cpu,
3412 unsigned long flags;
3416 for (i = 0; i < rcu_min_cached_objs; i++) {
3417 bnode = (struct kvfree_rcu_bulk_data *)
3418 __get_free_page(GFP_KERNEL | __GFP_NOWARN);
3421 raw_spin_lock_irqsave(&krcp->lock, flags);
3422 pushed = put_cached_bnode(krcp, bnode);
3423 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3426 free_page((unsigned long) bnode);
3432 atomic_set(&krcp->work_in_progress, 0);
3436 run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3438 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3439 !atomic_xchg(&krcp->work_in_progress, 1)) {
3440 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC,
3442 krcp->hrtimer.function = schedule_page_work_fn;
3443 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3448 kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
3450 struct kvfree_rcu_bulk_data *bnode;
3453 if (unlikely(!krcp->initialized))
3456 lockdep_assert_held(&krcp->lock);
3457 idx = !!is_vmalloc_addr(ptr);
3459 /* Check if a new block is required. */
3460 if (!krcp->bkvhead[idx] ||
3461 krcp->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3462 bnode = get_cached_bnode(krcp);
3463 /* Switch to emergency path. */
3467 /* Initialize the new block. */
3468 bnode->nr_records = 0;
3469 bnode->next = krcp->bkvhead[idx];
3471 /* Attach it to the head. */
3472 krcp->bkvhead[idx] = bnode;
3475 /* Finally insert. */
3476 krcp->bkvhead[idx]->records
3477 [krcp->bkvhead[idx]->nr_records++] = ptr;
3483 * Queue a request for lazy invocation of appropriate free routine after a
3484 * grace period. Please note there are three paths are maintained, two are the
3485 * main ones that use array of pointers interface and third one is emergency
3486 * one, that is used only when the main path can not be maintained temporary,
3487 * due to memory pressure.
3489 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3490 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3491 * be free'd in workqueue context. This allows us to: batch requests together to
3492 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3494 void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3496 unsigned long flags;
3497 struct kfree_rcu_cpu *krcp;
3502 ptr = (void *) head - (unsigned long) func;
3505 * Please note there is a limitation for the head-less
3506 * variant, that is why there is a clear rule for such
3507 * objects: it can be used from might_sleep() context
3508 * only. For other places please embed an rcu_head to
3512 ptr = (unsigned long *) func;
3515 krcp = krc_this_cpu_lock(&flags);
3517 // Queue the object but don't yet schedule the batch.
3518 if (debug_rcu_head_queue(ptr)) {
3519 // Probable double kfree_rcu(), just leak.
3520 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3523 // Mark as success and leave.
3528 success = kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr);
3530 run_page_cache_worker(krcp);
3533 // Inline if kvfree_rcu(one_arg) call.
3537 head->next = krcp->head;
3542 WRITE_ONCE(krcp->count, krcp->count + 1);
3544 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3545 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3546 !krcp->monitor_todo) {
3547 krcp->monitor_todo = true;
3548 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3552 krc_this_cpu_unlock(krcp, flags);
3555 * Inline kvfree() after synchronize_rcu(). We can do
3556 * it from might_sleep() context only, so the current
3557 * CPU can pass the QS state.
3560 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3565 EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3567 static unsigned long
3568 kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3571 unsigned long count = 0;
3573 /* Snapshot count of all CPUs */
3574 for_each_possible_cpu(cpu) {
3575 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3577 count += READ_ONCE(krcp->count);
3583 static unsigned long
3584 kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3587 unsigned long flags;
3589 for_each_possible_cpu(cpu) {
3591 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3593 count = krcp->count;
3594 raw_spin_lock_irqsave(&krcp->lock, flags);
3595 if (krcp->monitor_todo)
3596 kfree_rcu_drain_unlock(krcp, flags);
3598 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3600 sc->nr_to_scan -= count;
3603 if (sc->nr_to_scan <= 0)
3607 return freed == 0 ? SHRINK_STOP : freed;
3610 static struct shrinker kfree_rcu_shrinker = {
3611 .count_objects = kfree_rcu_shrink_count,
3612 .scan_objects = kfree_rcu_shrink_scan,
3614 .seeks = DEFAULT_SEEKS,
3617 void __init kfree_rcu_scheduler_running(void)
3620 unsigned long flags;
3622 for_each_possible_cpu(cpu) {
3623 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3625 raw_spin_lock_irqsave(&krcp->lock, flags);
3626 if (!krcp->head || krcp->monitor_todo) {
3627 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3630 krcp->monitor_todo = true;
3631 schedule_delayed_work_on(cpu, &krcp->monitor_work,
3632 KFREE_DRAIN_JIFFIES);
3633 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3638 * During early boot, any blocking grace-period wait automatically
3639 * implies a grace period. Later on, this is never the case for PREEMPTION.
3641 * However, because a context switch is a grace period for !PREEMPTION, any
3642 * blocking grace-period wait automatically implies a grace period if
3643 * there is only one CPU online at any point time during execution of
3644 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
3645 * occasionally incorrectly indicate that there are multiple CPUs online
3646 * when there was in fact only one the whole time, as this just adds some
3647 * overhead: RCU still operates correctly.
3649 static int rcu_blocking_is_gp(void)
3653 if (IS_ENABLED(CONFIG_PREEMPTION))
3654 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
3655 might_sleep(); /* Check for RCU read-side critical section. */
3658 * If the rcu_state.n_online_cpus counter is equal to one,
3659 * there is only one CPU, and that CPU sees all prior accesses
3660 * made by any CPU that was online at the time of its access.
3661 * Furthermore, if this counter is equal to one, its value cannot
3662 * change until after the preempt_enable() below.
3664 * Furthermore, if rcu_state.n_online_cpus is equal to one here,
3665 * all later CPUs (both this one and any that come online later
3666 * on) are guaranteed to see all accesses prior to this point
3667 * in the code, without the need for additional memory barriers.
3668 * Those memory barriers are provided by CPU-hotplug code.
3670 ret = READ_ONCE(rcu_state.n_online_cpus) <= 1;
3676 * synchronize_rcu - wait until a grace period has elapsed.
3678 * Control will return to the caller some time after a full grace
3679 * period has elapsed, in other words after all currently executing RCU
3680 * read-side critical sections have completed. Note, however, that
3681 * upon return from synchronize_rcu(), the caller might well be executing
3682 * concurrently with new RCU read-side critical sections that began while
3683 * synchronize_rcu() was waiting. RCU read-side critical sections are
3684 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
3685 * In addition, regions of code across which interrupts, preemption, or
3686 * softirqs have been disabled also serve as RCU read-side critical
3687 * sections. This includes hardware interrupt handlers, softirq handlers,
3690 * Note that this guarantee implies further memory-ordering guarantees.
3691 * On systems with more than one CPU, when synchronize_rcu() returns,
3692 * each CPU is guaranteed to have executed a full memory barrier since
3693 * the end of its last RCU read-side critical section whose beginning
3694 * preceded the call to synchronize_rcu(). In addition, each CPU having
3695 * an RCU read-side critical section that extends beyond the return from
3696 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3697 * after the beginning of synchronize_rcu() and before the beginning of
3698 * that RCU read-side critical section. Note that these guarantees include
3699 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3700 * that are executing in the kernel.
3702 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3703 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3704 * to have executed a full memory barrier during the execution of
3705 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3706 * again only if the system has more than one CPU).
3708 void synchronize_rcu(void)
3710 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3711 lock_is_held(&rcu_lock_map) ||
3712 lock_is_held(&rcu_sched_lock_map),
3713 "Illegal synchronize_rcu() in RCU read-side critical section");
3714 if (rcu_blocking_is_gp())
3715 return; // Context allows vacuous grace periods.
3716 if (rcu_gp_is_expedited())
3717 synchronize_rcu_expedited();
3719 wait_rcu_gp(call_rcu);
3721 EXPORT_SYMBOL_GPL(synchronize_rcu);
3724 * get_state_synchronize_rcu - Snapshot current RCU state
3726 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3727 * to determine whether or not a full grace period has elapsed in the
3730 unsigned long get_state_synchronize_rcu(void)
3733 * Any prior manipulation of RCU-protected data must happen
3734 * before the load from ->gp_seq.
3737 return rcu_seq_snap(&rcu_state.gp_seq);
3739 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3742 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3744 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3746 * If a full RCU grace period has elapsed since the earlier call to
3747 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3748 * synchronize_rcu() to wait for a full grace period.
3750 * Yes, this function does not take counter wrap into account. But
3751 * counter wrap is harmless. If the counter wraps, we have waited for
3752 * more than 2 billion grace periods (and way more on a 64-bit system!),
3753 * so waiting for one additional grace period should be just fine.
3755 void cond_synchronize_rcu(unsigned long oldstate)
3757 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
3760 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3762 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3765 * Check to see if there is any immediate RCU-related work to be done by
3766 * the current CPU, returning 1 if so and zero otherwise. The checks are
3767 * in order of increasing expense: checks that can be carried out against
3768 * CPU-local state are performed first. However, we must check for CPU
3769 * stalls first, else we might not get a chance.
3771 static int rcu_pending(int user)
3773 bool gp_in_progress;
3774 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3775 struct rcu_node *rnp = rdp->mynode;
3777 /* Check for CPU stalls, if enabled. */
3778 check_cpu_stall(rdp);
3780 /* Does this CPU need a deferred NOCB wakeup? */
3781 if (rcu_nocb_need_deferred_wakeup(rdp))
3784 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3785 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3788 /* Is the RCU core waiting for a quiescent state from this CPU? */
3789 gp_in_progress = rcu_gp_in_progress();
3790 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3793 /* Does this CPU have callbacks ready to invoke? */
3794 if (!rcu_segcblist_is_offloaded(&rdp->cblist) &&
3795 rcu_segcblist_ready_cbs(&rdp->cblist))
3798 /* Has RCU gone idle with this CPU needing another grace period? */
3799 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3800 !rcu_segcblist_is_offloaded(&rdp->cblist) &&
3801 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3804 /* Have RCU grace period completed or started? */
3805 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3806 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3814 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3815 * the compiler is expected to optimize this away.
3817 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3819 trace_rcu_barrier(rcu_state.name, s, cpu,
3820 atomic_read(&rcu_state.barrier_cpu_count), done);
3824 * RCU callback function for rcu_barrier(). If we are last, wake
3825 * up the task executing rcu_barrier().
3827 * Note that the value of rcu_state.barrier_sequence must be captured
3828 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3829 * other CPUs might count the value down to zero before this CPU gets
3830 * around to invoking rcu_barrier_trace(), which might result in bogus
3831 * data from the next instance of rcu_barrier().
3833 static void rcu_barrier_callback(struct rcu_head *rhp)
3835 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3837 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3838 rcu_barrier_trace(TPS("LastCB"), -1, s);
3839 complete(&rcu_state.barrier_completion);
3841 rcu_barrier_trace(TPS("CB"), -1, s);
3846 * Called with preemption disabled, and from cross-cpu IRQ context.
3848 static void rcu_barrier_func(void *cpu_in)
3850 uintptr_t cpu = (uintptr_t)cpu_in;
3851 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3853 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3854 rdp->barrier_head.func = rcu_barrier_callback;
3855 debug_rcu_head_queue(&rdp->barrier_head);
3857 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
3858 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
3859 atomic_inc(&rcu_state.barrier_cpu_count);
3861 debug_rcu_head_unqueue(&rdp->barrier_head);
3862 rcu_barrier_trace(TPS("IRQNQ"), -1,
3863 rcu_state.barrier_sequence);
3865 rcu_nocb_unlock(rdp);
3869 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3871 * Note that this primitive does not necessarily wait for an RCU grace period
3872 * to complete. For example, if there are no RCU callbacks queued anywhere
3873 * in the system, then rcu_barrier() is within its rights to return
3874 * immediately, without waiting for anything, much less an RCU grace period.
3876 void rcu_barrier(void)
3879 struct rcu_data *rdp;
3880 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3882 rcu_barrier_trace(TPS("Begin"), -1, s);
3884 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3885 mutex_lock(&rcu_state.barrier_mutex);
3887 /* Did someone else do our work for us? */
3888 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
3889 rcu_barrier_trace(TPS("EarlyExit"), -1,
3890 rcu_state.barrier_sequence);
3891 smp_mb(); /* caller's subsequent code after above check. */
3892 mutex_unlock(&rcu_state.barrier_mutex);
3896 /* Mark the start of the barrier operation. */
3897 rcu_seq_start(&rcu_state.barrier_sequence);
3898 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
3901 * Initialize the count to two rather than to zero in order
3902 * to avoid a too-soon return to zero in case of an immediate
3903 * invocation of the just-enqueued callback (or preemption of
3904 * this task). Exclude CPU-hotplug operations to ensure that no
3905 * offline non-offloaded CPU has callbacks queued.
3907 init_completion(&rcu_state.barrier_completion);
3908 atomic_set(&rcu_state.barrier_cpu_count, 2);
3912 * Force each CPU with callbacks to register a new callback.
3913 * When that callback is invoked, we will know that all of the
3914 * corresponding CPU's preceding callbacks have been invoked.
3916 for_each_possible_cpu(cpu) {
3917 rdp = per_cpu_ptr(&rcu_data, cpu);
3918 if (cpu_is_offline(cpu) &&
3919 !rcu_segcblist_is_offloaded(&rdp->cblist))
3921 if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
3922 rcu_barrier_trace(TPS("OnlineQ"), cpu,
3923 rcu_state.barrier_sequence);
3924 smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
3925 } else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
3926 cpu_is_offline(cpu)) {
3927 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
3928 rcu_state.barrier_sequence);
3929 local_irq_disable();
3930 rcu_barrier_func((void *)cpu);
3932 } else if (cpu_is_offline(cpu)) {
3933 rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
3934 rcu_state.barrier_sequence);
3936 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
3937 rcu_state.barrier_sequence);
3943 * Now that we have an rcu_barrier_callback() callback on each
3944 * CPU, and thus each counted, remove the initial count.
3946 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
3947 complete(&rcu_state.barrier_completion);
3949 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3950 wait_for_completion(&rcu_state.barrier_completion);
3952 /* Mark the end of the barrier operation. */
3953 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
3954 rcu_seq_end(&rcu_state.barrier_sequence);
3956 /* Other rcu_barrier() invocations can now safely proceed. */
3957 mutex_unlock(&rcu_state.barrier_mutex);
3959 EXPORT_SYMBOL_GPL(rcu_barrier);
3962 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3963 * first CPU in a given leaf rcu_node structure coming online. The caller
3964 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3967 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3971 struct rcu_node *rnp = rnp_leaf;
3973 raw_lockdep_assert_held_rcu_node(rnp_leaf);
3974 WARN_ON_ONCE(rnp->wait_blkd_tasks);
3976 mask = rnp->grpmask;
3980 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3981 oldmask = rnp->qsmaskinit;
3982 rnp->qsmaskinit |= mask;
3983 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3990 * Do boot-time initialization of a CPU's per-CPU RCU data.
3993 rcu_boot_init_percpu_data(int cpu)
3995 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3997 /* Set up local state, ensuring consistent view of global state. */
3998 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3999 INIT_WORK(&rdp->strict_work, strict_work_handler);
4000 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
4001 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
4002 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4003 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4004 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4005 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4007 rcu_boot_init_nocb_percpu_data(rdp);
4011 * Invoked early in the CPU-online process, when pretty much all services
4012 * are available. The incoming CPU is not present.
4014 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4015 * offline event can be happening at a given time. Note also that we can
4016 * accept some slop in the rsp->gp_seq access due to the fact that this
4017 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4018 * And any offloaded callbacks are being numbered elsewhere.
4020 int rcutree_prepare_cpu(unsigned int cpu)
4022 unsigned long flags;
4023 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4024 struct rcu_node *rnp = rcu_get_root();
4026 /* Set up local state, ensuring consistent view of global state. */
4027 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4028 rdp->qlen_last_fqs_check = 0;
4029 rdp->n_force_qs_snap = rcu_state.n_force_qs;
4030 rdp->blimit = blimit;
4031 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
4032 !rcu_segcblist_is_offloaded(&rdp->cblist))
4033 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4034 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
4035 rcu_dynticks_eqs_online();
4036 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4039 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4040 * propagation up the rcu_node tree will happen at the beginning
4041 * of the next grace period.
4044 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4045 rdp->beenonline = true; /* We have now been online. */
4046 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4047 rdp->gp_seq_needed = rdp->gp_seq;
4048 rdp->cpu_no_qs.b.norm = true;
4049 rdp->core_needs_qs = false;
4050 rdp->rcu_iw_pending = false;
4051 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4052 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4053 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4054 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4055 rcu_prepare_kthreads(cpu);
4056 rcu_spawn_cpu_nocb_kthread(cpu);
4057 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4063 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4065 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4067 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4069 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4073 * Near the end of the CPU-online process. Pretty much all services
4074 * enabled, and the CPU is now very much alive.
4076 int rcutree_online_cpu(unsigned int cpu)
4078 unsigned long flags;
4079 struct rcu_data *rdp;
4080 struct rcu_node *rnp;
4082 rdp = per_cpu_ptr(&rcu_data, cpu);
4084 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4085 rnp->ffmask |= rdp->grpmask;
4086 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4087 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4088 return 0; /* Too early in boot for scheduler work. */
4089 sync_sched_exp_online_cleanup(cpu);
4090 rcutree_affinity_setting(cpu, -1);
4092 // Stop-machine done, so allow nohz_full to disable tick.
4093 tick_dep_clear(TICK_DEP_BIT_RCU);
4098 * Near the beginning of the process. The CPU is still very much alive
4099 * with pretty much all services enabled.
4101 int rcutree_offline_cpu(unsigned int cpu)
4103 unsigned long flags;
4104 struct rcu_data *rdp;
4105 struct rcu_node *rnp;
4107 rdp = per_cpu_ptr(&rcu_data, cpu);
4109 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4110 rnp->ffmask &= ~rdp->grpmask;
4111 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4113 rcutree_affinity_setting(cpu, cpu);
4115 // nohz_full CPUs need the tick for stop-machine to work quickly
4116 tick_dep_set(TICK_DEP_BIT_RCU);
4121 * Mark the specified CPU as being online so that subsequent grace periods
4122 * (both expedited and normal) will wait on it. Note that this means that
4123 * incoming CPUs are not allowed to use RCU read-side critical sections
4124 * until this function is called. Failing to observe this restriction
4125 * will result in lockdep splats.
4127 * Note that this function is special in that it is invoked directly
4128 * from the incoming CPU rather than from the cpuhp_step mechanism.
4129 * This is because this function must be invoked at a precise location.
4131 void rcu_cpu_starting(unsigned int cpu)
4133 unsigned long flags;
4135 struct rcu_data *rdp;
4136 struct rcu_node *rnp;
4139 rdp = per_cpu_ptr(&rcu_data, cpu);
4140 if (rdp->cpu_started)
4142 rdp->cpu_started = true;
4145 mask = rdp->grpmask;
4146 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4147 WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4148 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4149 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4150 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4151 newcpu = !(rnp->expmaskinitnext & mask);
4152 rnp->expmaskinitnext |= mask;
4153 /* Allow lockless access for expedited grace periods. */
4154 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4155 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4156 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4157 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4158 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4160 /* An incoming CPU should never be blocking a grace period. */
4161 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4162 rcu_disable_urgency_upon_qs(rdp);
4163 /* Report QS -after- changing ->qsmaskinitnext! */
4164 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4166 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4168 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4169 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4170 WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4171 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4175 * The outgoing function has no further need of RCU, so remove it from
4176 * the rcu_node tree's ->qsmaskinitnext bit masks.
4178 * Note that this function is special in that it is invoked directly
4179 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4180 * This is because this function must be invoked at a precise location.
4182 void rcu_report_dead(unsigned int cpu)
4184 unsigned long flags;
4186 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4187 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4189 /* QS for any half-done expedited grace period. */
4191 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
4193 rcu_preempt_deferred_qs(current);
4195 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4196 mask = rdp->grpmask;
4197 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4198 WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4199 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4200 raw_spin_lock(&rcu_state.ofl_lock);
4201 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4202 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4203 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4204 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4205 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4206 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4207 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4209 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4210 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4211 raw_spin_unlock(&rcu_state.ofl_lock);
4212 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4213 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4214 WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4216 rdp->cpu_started = false;
4219 #ifdef CONFIG_HOTPLUG_CPU
4221 * The outgoing CPU has just passed through the dying-idle state, and we
4222 * are being invoked from the CPU that was IPIed to continue the offline
4223 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4225 void rcutree_migrate_callbacks(int cpu)
4227 unsigned long flags;
4228 struct rcu_data *my_rdp;
4229 struct rcu_node *my_rnp;
4230 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4233 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
4234 rcu_segcblist_empty(&rdp->cblist))
4235 return; /* No callbacks to migrate. */
4237 local_irq_save(flags);
4238 my_rdp = this_cpu_ptr(&rcu_data);
4239 my_rnp = my_rdp->mynode;
4240 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4241 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4242 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4243 /* Leverage recent GPs and set GP for new callbacks. */
4244 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4245 rcu_advance_cbs(my_rnp, my_rdp);
4246 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4247 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4248 rcu_segcblist_disable(&rdp->cblist);
4249 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
4250 !rcu_segcblist_n_cbs(&my_rdp->cblist));
4251 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
4252 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4253 __call_rcu_nocb_wake(my_rdp, true, flags);
4255 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4256 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4259 rcu_gp_kthread_wake();
4260 lockdep_assert_irqs_enabled();
4261 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4262 !rcu_segcblist_empty(&rdp->cblist),
4263 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4264 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4265 rcu_segcblist_first_cb(&rdp->cblist));
4270 * On non-huge systems, use expedited RCU grace periods to make suspend
4271 * and hibernation run faster.
4273 static int rcu_pm_notify(struct notifier_block *self,
4274 unsigned long action, void *hcpu)
4277 case PM_HIBERNATION_PREPARE:
4278 case PM_SUSPEND_PREPARE:
4281 case PM_POST_HIBERNATION:
4282 case PM_POST_SUSPEND:
4283 rcu_unexpedite_gp();
4292 * Spawn the kthreads that handle RCU's grace periods.
4294 static int __init rcu_spawn_gp_kthread(void)
4296 unsigned long flags;
4297 int kthread_prio_in = kthread_prio;
4298 struct rcu_node *rnp;
4299 struct sched_param sp;
4300 struct task_struct *t;
4302 /* Force priority into range. */
4303 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4304 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4306 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4308 else if (kthread_prio < 0)
4310 else if (kthread_prio > 99)
4313 if (kthread_prio != kthread_prio_in)
4314 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4315 kthread_prio, kthread_prio_in);
4317 rcu_scheduler_fully_active = 1;
4318 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4319 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4322 sp.sched_priority = kthread_prio;
4323 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4325 rnp = rcu_get_root();
4326 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4327 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4328 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4329 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4330 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4331 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4333 rcu_spawn_nocb_kthreads();
4334 rcu_spawn_boost_kthreads();
4337 early_initcall(rcu_spawn_gp_kthread);
4340 * This function is invoked towards the end of the scheduler's
4341 * initialization process. Before this is called, the idle task might
4342 * contain synchronous grace-period primitives (during which time, this idle
4343 * task is booting the system, and such primitives are no-ops). After this
4344 * function is called, any synchronous grace-period primitives are run as
4345 * expedited, with the requesting task driving the grace period forward.
4346 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4347 * runtime RCU functionality.
4349 void rcu_scheduler_starting(void)
4351 WARN_ON(num_online_cpus() != 1);
4352 WARN_ON(nr_context_switches() > 0);
4353 rcu_test_sync_prims();
4354 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4355 rcu_test_sync_prims();
4359 * Helper function for rcu_init() that initializes the rcu_state structure.
4361 static void __init rcu_init_one(void)
4363 static const char * const buf[] = RCU_NODE_NAME_INIT;
4364 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4365 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4366 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4368 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4372 struct rcu_node *rnp;
4374 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4376 /* Silence gcc 4.8 false positive about array index out of range. */
4377 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4378 panic("rcu_init_one: rcu_num_lvls out of range");
4380 /* Initialize the level-tracking arrays. */
4382 for (i = 1; i < rcu_num_lvls; i++)
4383 rcu_state.level[i] =
4384 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4385 rcu_init_levelspread(levelspread, num_rcu_lvl);
4387 /* Initialize the elements themselves, starting from the leaves. */
4389 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4390 cpustride *= levelspread[i];
4391 rnp = rcu_state.level[i];
4392 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4393 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4394 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4395 &rcu_node_class[i], buf[i]);
4396 raw_spin_lock_init(&rnp->fqslock);
4397 lockdep_set_class_and_name(&rnp->fqslock,
4398 &rcu_fqs_class[i], fqs[i]);
4399 rnp->gp_seq = rcu_state.gp_seq;
4400 rnp->gp_seq_needed = rcu_state.gp_seq;
4401 rnp->completedqs = rcu_state.gp_seq;
4403 rnp->qsmaskinit = 0;
4404 rnp->grplo = j * cpustride;
4405 rnp->grphi = (j + 1) * cpustride - 1;
4406 if (rnp->grphi >= nr_cpu_ids)
4407 rnp->grphi = nr_cpu_ids - 1;
4413 rnp->grpnum = j % levelspread[i - 1];
4414 rnp->grpmask = BIT(rnp->grpnum);
4415 rnp->parent = rcu_state.level[i - 1] +
4416 j / levelspread[i - 1];
4419 INIT_LIST_HEAD(&rnp->blkd_tasks);
4420 rcu_init_one_nocb(rnp);
4421 init_waitqueue_head(&rnp->exp_wq[0]);
4422 init_waitqueue_head(&rnp->exp_wq[1]);
4423 init_waitqueue_head(&rnp->exp_wq[2]);
4424 init_waitqueue_head(&rnp->exp_wq[3]);
4425 spin_lock_init(&rnp->exp_lock);
4429 init_swait_queue_head(&rcu_state.gp_wq);
4430 init_swait_queue_head(&rcu_state.expedited_wq);
4431 rnp = rcu_first_leaf_node();
4432 for_each_possible_cpu(i) {
4433 while (i > rnp->grphi)
4435 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4436 rcu_boot_init_percpu_data(i);
4441 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4442 * replace the definitions in tree.h because those are needed to size
4443 * the ->node array in the rcu_state structure.
4445 static void __init rcu_init_geometry(void)
4449 int rcu_capacity[RCU_NUM_LVLS];
4452 * Initialize any unspecified boot parameters.
4453 * The default values of jiffies_till_first_fqs and
4454 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4455 * value, which is a function of HZ, then adding one for each
4456 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4458 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4459 if (jiffies_till_first_fqs == ULONG_MAX)
4460 jiffies_till_first_fqs = d;
4461 if (jiffies_till_next_fqs == ULONG_MAX)
4462 jiffies_till_next_fqs = d;
4463 adjust_jiffies_till_sched_qs();
4465 /* If the compile-time values are accurate, just leave. */
4466 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4467 nr_cpu_ids == NR_CPUS)
4469 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4470 rcu_fanout_leaf, nr_cpu_ids);
4473 * The boot-time rcu_fanout_leaf parameter must be at least two
4474 * and cannot exceed the number of bits in the rcu_node masks.
4475 * Complain and fall back to the compile-time values if this
4476 * limit is exceeded.
4478 if (rcu_fanout_leaf < 2 ||
4479 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4480 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4486 * Compute number of nodes that can be handled an rcu_node tree
4487 * with the given number of levels.
4489 rcu_capacity[0] = rcu_fanout_leaf;
4490 for (i = 1; i < RCU_NUM_LVLS; i++)
4491 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4494 * The tree must be able to accommodate the configured number of CPUs.
4495 * If this limit is exceeded, fall back to the compile-time values.
4497 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4498 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4503 /* Calculate the number of levels in the tree. */
4504 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4506 rcu_num_lvls = i + 1;
4508 /* Calculate the number of rcu_nodes at each level of the tree. */
4509 for (i = 0; i < rcu_num_lvls; i++) {
4510 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4511 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4514 /* Calculate the total number of rcu_node structures. */
4516 for (i = 0; i < rcu_num_lvls; i++)
4517 rcu_num_nodes += num_rcu_lvl[i];
4521 * Dump out the structure of the rcu_node combining tree associated
4522 * with the rcu_state structure.
4524 static void __init rcu_dump_rcu_node_tree(void)
4527 struct rcu_node *rnp;
4529 pr_info("rcu_node tree layout dump\n");
4531 rcu_for_each_node_breadth_first(rnp) {
4532 if (rnp->level != level) {
4537 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4542 struct workqueue_struct *rcu_gp_wq;
4543 struct workqueue_struct *rcu_par_gp_wq;
4545 static void __init kfree_rcu_batch_init(void)
4550 for_each_possible_cpu(cpu) {
4551 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4553 for (i = 0; i < KFREE_N_BATCHES; i++) {
4554 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4555 krcp->krw_arr[i].krcp = krcp;
4558 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4559 INIT_WORK(&krcp->page_cache_work, fill_page_cache_func);
4560 krcp->initialized = true;
4562 if (register_shrinker(&kfree_rcu_shrinker))
4563 pr_err("Failed to register kfree_rcu() shrinker!\n");
4566 void __init rcu_init(void)
4570 rcu_early_boot_tests();
4572 kfree_rcu_batch_init();
4573 rcu_bootup_announce();
4574 rcu_init_geometry();
4577 rcu_dump_rcu_node_tree();
4579 open_softirq(RCU_SOFTIRQ, rcu_core_si);
4582 * We don't need protection against CPU-hotplug here because
4583 * this is called early in boot, before either interrupts
4584 * or the scheduler are operational.
4586 pm_notifier(rcu_pm_notify, 0);
4587 for_each_online_cpu(cpu) {
4588 rcutree_prepare_cpu(cpu);
4589 rcu_cpu_starting(cpu);
4590 rcutree_online_cpu(cpu);
4593 /* Create workqueue for expedited GPs and for Tree SRCU. */
4594 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4595 WARN_ON(!rcu_gp_wq);
4596 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4597 WARN_ON(!rcu_par_gp_wq);
4600 /* Fill in default value for rcutree.qovld boot parameter. */
4601 /* -After- the rcu_node ->lock fields are initialized! */
4603 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4608 #include "tree_stall.h"
4609 #include "tree_exp.h"
4610 #include "tree_plugin.h"