1 // SPDX-License-Identifier: GPL-2.0+
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
5 * Copyright IBM Corporation, 2008
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
14 * For detailed explanation of Read-Copy Update mechanism see -
18 #define pr_fmt(fmt) "rcu: " fmt
20 #include <linux/types.h>
21 #include <linux/kernel.h>
22 #include <linux/init.h>
23 #include <linux/spinlock.h>
24 #include <linux/smp.h>
25 #include <linux/rcupdate_wait.h>
26 #include <linux/interrupt.h>
27 #include <linux/sched.h>
28 #include <linux/sched/debug.h>
29 #include <linux/nmi.h>
30 #include <linux/atomic.h>
31 #include <linux/bitops.h>
32 #include <linux/export.h>
33 #include <linux/completion.h>
34 #include <linux/moduleparam.h>
35 #include <linux/panic.h>
36 #include <linux/panic_notifier.h>
37 #include <linux/percpu.h>
38 #include <linux/notifier.h>
39 #include <linux/cpu.h>
40 #include <linux/mutex.h>
41 #include <linux/time.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/wait.h>
44 #include <linux/kthread.h>
45 #include <uapi/linux/sched/types.h>
46 #include <linux/prefetch.h>
47 #include <linux/delay.h>
48 #include <linux/random.h>
49 #include <linux/trace_events.h>
50 #include <linux/suspend.h>
51 #include <linux/ftrace.h>
52 #include <linux/tick.h>
53 #include <linux/sysrq.h>
54 #include <linux/kprobes.h>
55 #include <linux/gfp.h>
56 #include <linux/oom.h>
57 #include <linux/smpboot.h>
58 #include <linux/jiffies.h>
59 #include <linux/slab.h>
60 #include <linux/sched/isolation.h>
61 #include <linux/sched/clock.h>
62 #include <linux/vmalloc.h>
64 #include <linux/kasan.h>
65 #include <linux/context_tracking.h>
66 #include "../time/tick-internal.h"
71 #ifdef MODULE_PARAM_PREFIX
72 #undef MODULE_PARAM_PREFIX
74 #define MODULE_PARAM_PREFIX "rcutree."
76 /* Data structures. */
78 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data
, rcu_data
) = {
80 #ifdef CONFIG_RCU_NOCB_CPU
81 .cblist
.flags
= SEGCBLIST_RCU_CORE
,
84 static struct rcu_state rcu_state
= {
85 .level
= { &rcu_state
.node
[0] },
86 .gp_state
= RCU_GP_IDLE
,
87 .gp_seq
= (0UL - 300UL) << RCU_SEQ_CTR_SHIFT
,
88 .barrier_mutex
= __MUTEX_INITIALIZER(rcu_state
.barrier_mutex
),
89 .barrier_lock
= __RAW_SPIN_LOCK_UNLOCKED(rcu_state
.barrier_lock
),
92 .exp_mutex
= __MUTEX_INITIALIZER(rcu_state
.exp_mutex
),
93 .exp_wake_mutex
= __MUTEX_INITIALIZER(rcu_state
.exp_wake_mutex
),
94 .ofl_lock
= __ARCH_SPIN_LOCK_UNLOCKED
,
97 /* Dump rcu_node combining tree at boot to verify correct setup. */
98 static bool dump_tree
;
99 module_param(dump_tree
, bool, 0444);
100 /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
101 static bool use_softirq
= !IS_ENABLED(CONFIG_PREEMPT_RT
);
102 #ifndef CONFIG_PREEMPT_RT
103 module_param(use_softirq
, bool, 0444);
105 /* Control rcu_node-tree auto-balancing at boot time. */
106 static bool rcu_fanout_exact
;
107 module_param(rcu_fanout_exact
, bool, 0444);
108 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
109 static int rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
110 module_param(rcu_fanout_leaf
, int, 0444);
111 int rcu_num_lvls __read_mostly
= RCU_NUM_LVLS
;
112 /* Number of rcu_nodes at specified level. */
113 int num_rcu_lvl
[] = NUM_RCU_LVL_INIT
;
114 int rcu_num_nodes __read_mostly
= NUM_RCU_NODES
; /* Total # rcu_nodes in use. */
117 * The rcu_scheduler_active variable is initialized to the value
118 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
119 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
120 * RCU can assume that there is but one task, allowing RCU to (for example)
121 * optimize synchronize_rcu() to a simple barrier(). When this variable
122 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
123 * to detect real grace periods. This variable is also used to suppress
124 * boot-time false positives from lockdep-RCU error checking. Finally, it
125 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
126 * is fully initialized, including all of its kthreads having been spawned.
128 int rcu_scheduler_active __read_mostly
;
129 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
132 * The rcu_scheduler_fully_active variable transitions from zero to one
133 * during the early_initcall() processing, which is after the scheduler
134 * is capable of creating new tasks. So RCU processing (for example,
135 * creating tasks for RCU priority boosting) must be delayed until after
136 * rcu_scheduler_fully_active transitions from zero to one. We also
137 * currently delay invocation of any RCU callbacks until after this point.
139 * It might later prove better for people registering RCU callbacks during
140 * early boot to take responsibility for these callbacks, but one step at
143 static int rcu_scheduler_fully_active __read_mostly
;
145 static void rcu_report_qs_rnp(unsigned long mask
, struct rcu_node
*rnp
,
146 unsigned long gps
, unsigned long flags
);
147 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
);
148 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
);
149 static void rcu_boost_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
150 static void invoke_rcu_core(void);
151 static void rcu_report_exp_rdp(struct rcu_data
*rdp
);
152 static void sync_sched_exp_online_cleanup(int cpu
);
153 static void check_cb_ovld_locked(struct rcu_data
*rdp
, struct rcu_node
*rnp
);
154 static bool rcu_rdp_is_offloaded(struct rcu_data
*rdp
);
157 * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
158 * real-time priority(enabling/disabling) is controlled by
159 * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
161 static int kthread_prio
= IS_ENABLED(CONFIG_RCU_BOOST
) ? 1 : 0;
162 module_param(kthread_prio
, int, 0444);
164 /* Delay in jiffies for grace-period initialization delays, debug only. */
166 static int gp_preinit_delay
;
167 module_param(gp_preinit_delay
, int, 0444);
168 static int gp_init_delay
;
169 module_param(gp_init_delay
, int, 0444);
170 static int gp_cleanup_delay
;
171 module_param(gp_cleanup_delay
, int, 0444);
173 // Add delay to rcu_read_unlock() for strict grace periods.
174 static int rcu_unlock_delay
;
175 #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
176 module_param(rcu_unlock_delay
, int, 0444);
180 * This rcu parameter is runtime-read-only. It reflects
181 * a minimum allowed number of objects which can be cached
182 * per-CPU. Object size is equal to one page. This value
183 * can be changed at boot time.
185 static int rcu_min_cached_objs
= 5;
186 module_param(rcu_min_cached_objs
, int, 0444);
188 // A page shrinker can ask for pages to be freed to make them
189 // available for other parts of the system. This usually happens
190 // under low memory conditions, and in that case we should also
191 // defer page-cache filling for a short time period.
193 // The default value is 5 seconds, which is long enough to reduce
194 // interference with the shrinker while it asks other systems to
195 // drain their caches.
196 static int rcu_delay_page_cache_fill_msec
= 5000;
197 module_param(rcu_delay_page_cache_fill_msec
, int, 0444);
199 /* Retrieve RCU kthreads priority for rcutorture */
200 int rcu_get_gp_kthreads_prio(void)
204 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio
);
207 * Number of grace periods between delays, normalized by the duration of
208 * the delay. The longer the delay, the more the grace periods between
209 * each delay. The reason for this normalization is that it means that,
210 * for non-zero delays, the overall slowdown of grace periods is constant
211 * regardless of the duration of the delay. This arrangement balances
212 * the need for long delays to increase some race probabilities with the
213 * need for fast grace periods to increase other race probabilities.
215 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
218 * Compute the mask of online CPUs for the specified rcu_node structure.
219 * This will not be stable unless the rcu_node structure's ->lock is
220 * held, but the bit corresponding to the current CPU will be stable
223 static unsigned long rcu_rnp_online_cpus(struct rcu_node
*rnp
)
225 return READ_ONCE(rnp
->qsmaskinitnext
);
229 * Is the CPU corresponding to the specified rcu_data structure online
230 * from RCU's perspective? This perspective is given by that structure's
231 * ->qsmaskinitnext field rather than by the global cpu_online_mask.
233 static bool rcu_rdp_cpu_online(struct rcu_data
*rdp
)
235 return !!(rdp
->grpmask
& rcu_rnp_online_cpus(rdp
->mynode
));
239 * Return true if an RCU grace period is in progress. The READ_ONCE()s
240 * permit this function to be invoked without holding the root rcu_node
241 * structure's ->lock, but of course results can be subject to change.
243 static int rcu_gp_in_progress(void)
245 return rcu_seq_state(rcu_seq_current(&rcu_state
.gp_seq
));
249 * Return the number of callbacks queued on the specified CPU.
250 * Handles both the nocbs and normal cases.
252 static long rcu_get_n_cbs_cpu(int cpu
)
254 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
256 if (rcu_segcblist_is_enabled(&rdp
->cblist
))
257 return rcu_segcblist_n_cbs(&rdp
->cblist
);
261 void rcu_softirq_qs(void)
264 rcu_preempt_deferred_qs(current
);
265 rcu_tasks_qs(current
, false);
269 * Reset the current CPU's ->dynticks counter to indicate that the
270 * newly onlined CPU is no longer in an extended quiescent state.
271 * This will either leave the counter unchanged, or increment it
272 * to the next non-quiescent value.
274 * The non-atomic test/increment sequence works because the upper bits
275 * of the ->dynticks counter are manipulated only by the corresponding CPU,
276 * or when the corresponding CPU is offline.
278 static void rcu_dynticks_eqs_online(void)
280 if (ct_dynticks() & RCU_DYNTICKS_IDX
)
282 ct_state_inc(RCU_DYNTICKS_IDX
);
286 * Snapshot the ->dynticks counter with full ordering so as to allow
287 * stable comparison of this counter with past and future snapshots.
289 static int rcu_dynticks_snap(int cpu
)
291 smp_mb(); // Fundamental RCU ordering guarantee.
292 return ct_dynticks_cpu_acquire(cpu
);
296 * Return true if the snapshot returned from rcu_dynticks_snap()
297 * indicates that RCU is in an extended quiescent state.
299 static bool rcu_dynticks_in_eqs(int snap
)
301 return !(snap
& RCU_DYNTICKS_IDX
);
304 /* Return true if the specified CPU is currently idle from an RCU viewpoint. */
305 bool rcu_is_idle_cpu(int cpu
)
307 return rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu
));
311 * Return true if the CPU corresponding to the specified rcu_data
312 * structure has spent some time in an extended quiescent state since
313 * rcu_dynticks_snap() returned the specified snapshot.
315 static bool rcu_dynticks_in_eqs_since(struct rcu_data
*rdp
, int snap
)
317 return snap
!= rcu_dynticks_snap(rdp
->cpu
);
321 * Return true if the referenced integer is zero while the specified
322 * CPU remains within a single extended quiescent state.
324 bool rcu_dynticks_zero_in_eqs(int cpu
, int *vp
)
328 // If not quiescent, force back to earlier extended quiescent state.
329 snap
= ct_dynticks_cpu(cpu
) & ~RCU_DYNTICKS_IDX
;
330 smp_rmb(); // Order ->dynticks and *vp reads.
332 return false; // Non-zero, so report failure;
333 smp_rmb(); // Order *vp read and ->dynticks re-read.
335 // If still in the same extended quiescent state, we are good!
336 return snap
== ct_dynticks_cpu(cpu
);
340 * Let the RCU core know that this CPU has gone through the scheduler,
341 * which is a quiescent state. This is called when the need for a
342 * quiescent state is urgent, so we burn an atomic operation and full
343 * memory barriers to let the RCU core know about it, regardless of what
344 * this CPU might (or might not) do in the near future.
346 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
348 * The caller must have disabled interrupts and must not be idle.
350 notrace
void rcu_momentary_dyntick_idle(void)
354 raw_cpu_write(rcu_data
.rcu_need_heavy_qs
, false);
355 seq
= ct_state_inc(2 * RCU_DYNTICKS_IDX
);
356 /* It is illegal to call this from idle state. */
357 WARN_ON_ONCE(!(seq
& RCU_DYNTICKS_IDX
));
358 rcu_preempt_deferred_qs(current
);
360 EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle
);
363 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
365 * If the current CPU is idle and running at a first-level (not nested)
366 * interrupt, or directly, from idle, return true.
368 * The caller must have at least disabled IRQs.
370 static int rcu_is_cpu_rrupt_from_idle(void)
375 * Usually called from the tick; but also used from smp_function_call()
376 * for expedited grace periods. This latter can result in running from
377 * the idle task, instead of an actual IPI.
379 lockdep_assert_irqs_disabled();
381 /* Check for counter underflows */
382 RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
383 "RCU dynticks_nesting counter underflow!");
384 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
385 "RCU dynticks_nmi_nesting counter underflow/zero!");
387 /* Are we at first interrupt nesting level? */
388 nesting
= ct_dynticks_nmi_nesting();
393 * If we're not in an interrupt, we must be in the idle task!
395 WARN_ON_ONCE(!nesting
&& !is_idle_task(current
));
397 /* Does CPU appear to be idle from an RCU standpoint? */
398 return ct_dynticks_nesting() == 0;
401 #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
402 // Maximum callbacks per rcu_do_batch ...
403 #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
404 static long blimit
= DEFAULT_RCU_BLIMIT
;
405 #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
406 static long qhimark
= DEFAULT_RCU_QHIMARK
;
407 #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
408 static long qlowmark
= DEFAULT_RCU_QLOMARK
;
409 #define DEFAULT_RCU_QOVLD_MULT 2
410 #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
411 static long qovld
= DEFAULT_RCU_QOVLD
; // If this many pending, hammer QS.
412 static long qovld_calc
= -1; // No pre-initialization lock acquisitions!
414 module_param(blimit
, long, 0444);
415 module_param(qhimark
, long, 0444);
416 module_param(qlowmark
, long, 0444);
417 module_param(qovld
, long, 0444);
419 static ulong jiffies_till_first_fqs
= IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD
) ? 0 : ULONG_MAX
;
420 static ulong jiffies_till_next_fqs
= ULONG_MAX
;
421 static bool rcu_kick_kthreads
;
422 static int rcu_divisor
= 7;
423 module_param(rcu_divisor
, int, 0644);
425 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
426 static long rcu_resched_ns
= 3 * NSEC_PER_MSEC
;
427 module_param(rcu_resched_ns
, long, 0644);
430 * How long the grace period must be before we start recruiting
431 * quiescent-state help from rcu_note_context_switch().
433 static ulong jiffies_till_sched_qs
= ULONG_MAX
;
434 module_param(jiffies_till_sched_qs
, ulong
, 0444);
435 static ulong jiffies_to_sched_qs
; /* See adjust_jiffies_till_sched_qs(). */
436 module_param(jiffies_to_sched_qs
, ulong
, 0444); /* Display only! */
439 * Make sure that we give the grace-period kthread time to detect any
440 * idle CPUs before taking active measures to force quiescent states.
441 * However, don't go below 100 milliseconds, adjusted upwards for really
444 static void adjust_jiffies_till_sched_qs(void)
448 /* If jiffies_till_sched_qs was specified, respect the request. */
449 if (jiffies_till_sched_qs
!= ULONG_MAX
) {
450 WRITE_ONCE(jiffies_to_sched_qs
, jiffies_till_sched_qs
);
453 /* Otherwise, set to third fqs scan, but bound below on large system. */
454 j
= READ_ONCE(jiffies_till_first_fqs
) +
455 2 * READ_ONCE(jiffies_till_next_fqs
);
456 if (j
< HZ
/ 10 + nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
)
457 j
= HZ
/ 10 + nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
458 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j
);
459 WRITE_ONCE(jiffies_to_sched_qs
, j
);
462 static int param_set_first_fqs_jiffies(const char *val
, const struct kernel_param
*kp
)
465 int ret
= kstrtoul(val
, 0, &j
);
468 WRITE_ONCE(*(ulong
*)kp
->arg
, (j
> HZ
) ? HZ
: j
);
469 adjust_jiffies_till_sched_qs();
474 static int param_set_next_fqs_jiffies(const char *val
, const struct kernel_param
*kp
)
477 int ret
= kstrtoul(val
, 0, &j
);
480 WRITE_ONCE(*(ulong
*)kp
->arg
, (j
> HZ
) ? HZ
: (j
?: 1));
481 adjust_jiffies_till_sched_qs();
486 static const struct kernel_param_ops first_fqs_jiffies_ops
= {
487 .set
= param_set_first_fqs_jiffies
,
488 .get
= param_get_ulong
,
491 static const struct kernel_param_ops next_fqs_jiffies_ops
= {
492 .set
= param_set_next_fqs_jiffies
,
493 .get
= param_get_ulong
,
496 module_param_cb(jiffies_till_first_fqs
, &first_fqs_jiffies_ops
, &jiffies_till_first_fqs
, 0644);
497 module_param_cb(jiffies_till_next_fqs
, &next_fqs_jiffies_ops
, &jiffies_till_next_fqs
, 0644);
498 module_param(rcu_kick_kthreads
, bool, 0644);
500 static void force_qs_rnp(int (*f
)(struct rcu_data
*rdp
));
501 static int rcu_pending(int user
);
504 * Return the number of RCU GPs completed thus far for debug & stats.
506 unsigned long rcu_get_gp_seq(void)
508 return READ_ONCE(rcu_state
.gp_seq
);
510 EXPORT_SYMBOL_GPL(rcu_get_gp_seq
);
513 * Return the number of RCU expedited batches completed thus far for
514 * debug & stats. Odd numbers mean that a batch is in progress, even
515 * numbers mean idle. The value returned will thus be roughly double
516 * the cumulative batches since boot.
518 unsigned long rcu_exp_batches_completed(void)
520 return rcu_state
.expedited_sequence
;
522 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed
);
525 * Return the root node of the rcu_state structure.
527 static struct rcu_node
*rcu_get_root(void)
529 return &rcu_state
.node
[0];
533 * Send along grace-period-related data for rcutorture diagnostics.
535 void rcutorture_get_gp_data(enum rcutorture_type test_type
, int *flags
,
536 unsigned long *gp_seq
)
540 *flags
= READ_ONCE(rcu_state
.gp_flags
);
541 *gp_seq
= rcu_seq_current(&rcu_state
.gp_seq
);
547 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data
);
549 #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
551 * An empty function that will trigger a reschedule on
552 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
554 static void late_wakeup_func(struct irq_work
*work
)
558 static DEFINE_PER_CPU(struct irq_work
, late_wakeup_work
) =
559 IRQ_WORK_INIT(late_wakeup_func
);
564 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
565 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
567 * In these cases the late RCU wake ups aren't supported in the resched loops and our
568 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
569 * get re-enabled again.
571 noinstr
void rcu_irq_work_resched(void)
573 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
575 if (IS_ENABLED(CONFIG_GENERIC_ENTRY
) && !(current
->flags
& PF_VCPU
))
578 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK
) && (current
->flags
& PF_VCPU
))
581 instrumentation_begin();
582 if (do_nocb_deferred_wakeup(rdp
) && need_resched()) {
583 irq_work_queue(this_cpu_ptr(&late_wakeup_work
));
585 instrumentation_end();
587 #endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
589 #ifdef CONFIG_PROVE_RCU
591 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
593 void rcu_irq_exit_check_preempt(void)
595 lockdep_assert_irqs_disabled();
597 RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
598 "RCU dynticks_nesting counter underflow/zero!");
599 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
601 "Bad RCU dynticks_nmi_nesting counter\n");
602 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
603 "RCU in extended quiescent state!");
605 #endif /* #ifdef CONFIG_PROVE_RCU */
607 #ifdef CONFIG_NO_HZ_FULL
609 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
611 * The scheduler tick is not normally enabled when CPUs enter the kernel
612 * from nohz_full userspace execution. After all, nohz_full userspace
613 * execution is an RCU quiescent state and the time executing in the kernel
614 * is quite short. Except of course when it isn't. And it is not hard to
615 * cause a large system to spend tens of seconds or even minutes looping
616 * in the kernel, which can cause a number of problems, include RCU CPU
619 * Therefore, if a nohz_full CPU fails to report a quiescent state
620 * in a timely manner, the RCU grace-period kthread sets that CPU's
621 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
622 * exception will invoke this function, which will turn on the scheduler
623 * tick, which will enable RCU to detect that CPU's quiescent states,
624 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
625 * The tick will be disabled once a quiescent state is reported for
628 * Of course, in carefully tuned systems, there might never be an
629 * interrupt or exception. In that case, the RCU grace-period kthread
630 * will eventually cause one to happen. However, in less carefully
631 * controlled environments, this function allows RCU to get what it
632 * needs without creating otherwise useless interruptions.
634 void __rcu_irq_enter_check_tick(void)
636 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
638 // If we're here from NMI there's nothing to do.
642 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
643 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
645 if (!tick_nohz_full_cpu(rdp
->cpu
) ||
646 !READ_ONCE(rdp
->rcu_urgent_qs
) ||
647 READ_ONCE(rdp
->rcu_forced_tick
)) {
648 // RCU doesn't need nohz_full help from this CPU, or it is
649 // already getting that help.
653 // We get here only when not in an extended quiescent state and
654 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
655 // already watching and (2) The fact that we are in an interrupt
656 // handler and that the rcu_node lock is an irq-disabled lock
657 // prevents self-deadlock. So we can safely recheck under the lock.
658 // Note that the nohz_full state currently cannot change.
659 raw_spin_lock_rcu_node(rdp
->mynode
);
660 if (rdp
->rcu_urgent_qs
&& !rdp
->rcu_forced_tick
) {
661 // A nohz_full CPU is in the kernel and RCU needs a
662 // quiescent state. Turn on the tick!
663 WRITE_ONCE(rdp
->rcu_forced_tick
, true);
664 tick_dep_set_cpu(rdp
->cpu
, TICK_DEP_BIT_RCU
);
666 raw_spin_unlock_rcu_node(rdp
->mynode
);
668 #endif /* CONFIG_NO_HZ_FULL */
671 * Check to see if any future non-offloaded RCU-related work will need
672 * to be done by the current CPU, even if none need be done immediately,
673 * returning 1 if so. This function is part of the RCU implementation;
674 * it is -not- an exported member of the RCU API. This is used by
675 * the idle-entry code to figure out whether it is safe to disable the
676 * scheduler-clock interrupt.
678 * Just check whether or not this CPU has non-offloaded RCU callbacks
681 int rcu_needs_cpu(void)
683 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data
)->cblist
) &&
684 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data
));
688 * If any sort of urgency was applied to the current CPU (for example,
689 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
690 * to get to a quiescent state, disable it.
692 static void rcu_disable_urgency_upon_qs(struct rcu_data
*rdp
)
694 raw_lockdep_assert_held_rcu_node(rdp
->mynode
);
695 WRITE_ONCE(rdp
->rcu_urgent_qs
, false);
696 WRITE_ONCE(rdp
->rcu_need_heavy_qs
, false);
697 if (tick_nohz_full_cpu(rdp
->cpu
) && rdp
->rcu_forced_tick
) {
698 tick_dep_clear_cpu(rdp
->cpu
, TICK_DEP_BIT_RCU
);
699 WRITE_ONCE(rdp
->rcu_forced_tick
, false);
704 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
706 * Return true if RCU is watching the running CPU, which means that this
707 * CPU can safely enter RCU read-side critical sections. In other words,
708 * if the current CPU is not in its idle loop or is in an interrupt or
709 * NMI handler, return true.
711 * Make notrace because it can be called by the internal functions of
712 * ftrace, and making this notrace removes unnecessary recursion calls.
714 notrace
bool rcu_is_watching(void)
718 preempt_disable_notrace();
719 ret
= !rcu_dynticks_curr_cpu_in_eqs();
720 preempt_enable_notrace();
723 EXPORT_SYMBOL_GPL(rcu_is_watching
);
726 * If a holdout task is actually running, request an urgent quiescent
727 * state from its CPU. This is unsynchronized, so migrations can cause
728 * the request to go to the wrong CPU. Which is OK, all that will happen
729 * is that the CPU's next context switch will be a bit slower and next
730 * time around this task will generate another request.
732 void rcu_request_urgent_qs_task(struct task_struct
*t
)
739 return; /* This task is not running on that CPU. */
740 smp_store_release(per_cpu_ptr(&rcu_data
.rcu_urgent_qs
, cpu
), true);
743 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
746 * Is the current CPU online as far as RCU is concerned?
748 * Disable preemption to avoid false positives that could otherwise
749 * happen due to the current CPU number being sampled, this task being
750 * preempted, its old CPU being taken offline, resuming on some other CPU,
751 * then determining that its old CPU is now offline.
753 * Disable checking if in an NMI handler because we cannot safely
754 * report errors from NMI handlers anyway. In addition, it is OK to use
755 * RCU on an offline processor during initial boot, hence the check for
756 * rcu_scheduler_fully_active.
758 bool rcu_lockdep_current_cpu_online(void)
760 struct rcu_data
*rdp
;
763 if (in_nmi() || !rcu_scheduler_fully_active
)
765 preempt_disable_notrace();
766 rdp
= this_cpu_ptr(&rcu_data
);
768 * Strictly, we care here about the case where the current CPU is
769 * in rcu_cpu_starting() and thus has an excuse for rdp->grpmask
770 * not being up to date. So arch_spin_is_locked() might have a
771 * false positive if it's held by some *other* CPU, but that's
772 * OK because that just means a false *negative* on the warning.
774 if (rcu_rdp_cpu_online(rdp
) || arch_spin_is_locked(&rcu_state
.ofl_lock
))
776 preempt_enable_notrace();
779 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online
);
781 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
784 * When trying to report a quiescent state on behalf of some other CPU,
785 * it is our responsibility to check for and handle potential overflow
786 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
787 * After all, the CPU might be in deep idle state, and thus executing no
790 static void rcu_gpnum_ovf(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
792 raw_lockdep_assert_held_rcu_node(rnp
);
793 if (ULONG_CMP_LT(rcu_seq_current(&rdp
->gp_seq
) + ULONG_MAX
/ 4,
795 WRITE_ONCE(rdp
->gpwrap
, true);
796 if (ULONG_CMP_LT(rdp
->rcu_iw_gp_seq
+ ULONG_MAX
/ 4, rnp
->gp_seq
))
797 rdp
->rcu_iw_gp_seq
= rnp
->gp_seq
+ ULONG_MAX
/ 4;
801 * Snapshot the specified CPU's dynticks counter so that we can later
802 * credit them with an implicit quiescent state. Return 1 if this CPU
803 * is in dynticks idle mode, which is an extended quiescent state.
805 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
807 rdp
->dynticks_snap
= rcu_dynticks_snap(rdp
->cpu
);
808 if (rcu_dynticks_in_eqs(rdp
->dynticks_snap
)) {
809 trace_rcu_fqs(rcu_state
.name
, rdp
->gp_seq
, rdp
->cpu
, TPS("dti"));
810 rcu_gpnum_ovf(rdp
->mynode
, rdp
);
817 * Return true if the specified CPU has passed through a quiescent
818 * state by virtue of being in or having passed through an dynticks
819 * idle state since the last call to dyntick_save_progress_counter()
820 * for this same CPU, or by virtue of having been offline.
822 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
825 struct rcu_node
*rnp
= rdp
->mynode
;
828 * If the CPU passed through or entered a dynticks idle phase with
829 * no active irq/NMI handlers, then we can safely pretend that the CPU
830 * already acknowledged the request to pass through a quiescent
831 * state. Either way, that CPU cannot possibly be in an RCU
832 * read-side critical section that started before the beginning
833 * of the current RCU grace period.
835 if (rcu_dynticks_in_eqs_since(rdp
, rdp
->dynticks_snap
)) {
836 trace_rcu_fqs(rcu_state
.name
, rdp
->gp_seq
, rdp
->cpu
, TPS("dti"));
837 rcu_gpnum_ovf(rnp
, rdp
);
842 * Complain if a CPU that is considered to be offline from RCU's
843 * perspective has not yet reported a quiescent state. After all,
844 * the offline CPU should have reported a quiescent state during
845 * the CPU-offline process, or, failing that, by rcu_gp_init()
846 * if it ran concurrently with either the CPU going offline or the
847 * last task on a leaf rcu_node structure exiting its RCU read-side
848 * critical section while all CPUs corresponding to that structure
849 * are offline. This added warning detects bugs in any of these
852 * The rcu_node structure's ->lock is held here, which excludes
853 * the relevant portions the CPU-hotplug code, the grace-period
854 * initialization code, and the rcu_read_unlock() code paths.
856 * For more detail, please refer to the "Hotplug CPU" section
857 * of RCU's Requirements documentation.
859 if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp
))) {
860 struct rcu_node
*rnp1
;
862 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
863 __func__
, rnp
->grplo
, rnp
->grphi
, rnp
->level
,
864 (long)rnp
->gp_seq
, (long)rnp
->completedqs
);
865 for (rnp1
= rnp
; rnp1
; rnp1
= rnp1
->parent
)
866 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
867 __func__
, rnp1
->grplo
, rnp1
->grphi
, rnp1
->qsmask
, rnp1
->qsmaskinit
, rnp1
->qsmaskinitnext
, rnp1
->rcu_gp_init_mask
);
868 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
869 __func__
, rdp
->cpu
, ".o"[rcu_rdp_cpu_online(rdp
)],
870 (long)rdp
->rcu_onl_gp_seq
, rdp
->rcu_onl_gp_flags
,
871 (long)rdp
->rcu_ofl_gp_seq
, rdp
->rcu_ofl_gp_flags
);
872 return 1; /* Break things loose after complaining. */
876 * A CPU running for an extended time within the kernel can
877 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
878 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
879 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
880 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
881 * variable are safe because the assignments are repeated if this
882 * CPU failed to pass through a quiescent state. This code
883 * also checks .jiffies_resched in case jiffies_to_sched_qs
886 jtsq
= READ_ONCE(jiffies_to_sched_qs
);
887 if (!READ_ONCE(rdp
->rcu_need_heavy_qs
) &&
888 (time_after(jiffies
, rcu_state
.gp_start
+ jtsq
* 2) ||
889 time_after(jiffies
, rcu_state
.jiffies_resched
) ||
891 WRITE_ONCE(rdp
->rcu_need_heavy_qs
, true);
892 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
893 smp_store_release(&rdp
->rcu_urgent_qs
, true);
894 } else if (time_after(jiffies
, rcu_state
.gp_start
+ jtsq
)) {
895 WRITE_ONCE(rdp
->rcu_urgent_qs
, true);
899 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
900 * The above code handles this, but only for straight cond_resched().
901 * And some in-kernel loops check need_resched() before calling
902 * cond_resched(), which defeats the above code for CPUs that are
903 * running in-kernel with scheduling-clock interrupts disabled.
904 * So hit them over the head with the resched_cpu() hammer!
906 if (tick_nohz_full_cpu(rdp
->cpu
) &&
907 (time_after(jiffies
, READ_ONCE(rdp
->last_fqs_resched
) + jtsq
* 3) ||
909 WRITE_ONCE(rdp
->rcu_urgent_qs
, true);
910 resched_cpu(rdp
->cpu
);
911 WRITE_ONCE(rdp
->last_fqs_resched
, jiffies
);
915 * If more than halfway to RCU CPU stall-warning time, invoke
916 * resched_cpu() more frequently to try to loosen things up a bit.
917 * Also check to see if the CPU is getting hammered with interrupts,
918 * but only once per grace period, just to keep the IPIs down to
921 if (time_after(jiffies
, rcu_state
.jiffies_resched
)) {
922 if (time_after(jiffies
,
923 READ_ONCE(rdp
->last_fqs_resched
) + jtsq
)) {
924 resched_cpu(rdp
->cpu
);
925 WRITE_ONCE(rdp
->last_fqs_resched
, jiffies
);
927 if (IS_ENABLED(CONFIG_IRQ_WORK
) &&
928 !rdp
->rcu_iw_pending
&& rdp
->rcu_iw_gp_seq
!= rnp
->gp_seq
&&
929 (rnp
->ffmask
& rdp
->grpmask
)) {
930 rdp
->rcu_iw_pending
= true;
931 rdp
->rcu_iw_gp_seq
= rnp
->gp_seq
;
932 irq_work_queue_on(&rdp
->rcu_iw
, rdp
->cpu
);
939 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
940 static void trace_rcu_this_gp(struct rcu_node
*rnp
, struct rcu_data
*rdp
,
941 unsigned long gp_seq_req
, const char *s
)
943 trace_rcu_future_grace_period(rcu_state
.name
, READ_ONCE(rnp
->gp_seq
),
944 gp_seq_req
, rnp
->level
,
945 rnp
->grplo
, rnp
->grphi
, s
);
949 * rcu_start_this_gp - Request the start of a particular grace period
950 * @rnp_start: The leaf node of the CPU from which to start.
951 * @rdp: The rcu_data corresponding to the CPU from which to start.
952 * @gp_seq_req: The gp_seq of the grace period to start.
954 * Start the specified grace period, as needed to handle newly arrived
955 * callbacks. The required future grace periods are recorded in each
956 * rcu_node structure's ->gp_seq_needed field. Returns true if there
957 * is reason to awaken the grace-period kthread.
959 * The caller must hold the specified rcu_node structure's ->lock, which
960 * is why the caller is responsible for waking the grace-period kthread.
962 * Returns true if the GP thread needs to be awakened else false.
964 static bool rcu_start_this_gp(struct rcu_node
*rnp_start
, struct rcu_data
*rdp
,
965 unsigned long gp_seq_req
)
968 struct rcu_node
*rnp
;
971 * Use funnel locking to either acquire the root rcu_node
972 * structure's lock or bail out if the need for this grace period
973 * has already been recorded -- or if that grace period has in
974 * fact already started. If there is already a grace period in
975 * progress in a non-leaf node, no recording is needed because the
976 * end of the grace period will scan the leaf rcu_node structures.
977 * Note that rnp_start->lock must not be released.
979 raw_lockdep_assert_held_rcu_node(rnp_start
);
980 trace_rcu_this_gp(rnp_start
, rdp
, gp_seq_req
, TPS("Startleaf"));
981 for (rnp
= rnp_start
; 1; rnp
= rnp
->parent
) {
982 if (rnp
!= rnp_start
)
983 raw_spin_lock_rcu_node(rnp
);
984 if (ULONG_CMP_GE(rnp
->gp_seq_needed
, gp_seq_req
) ||
985 rcu_seq_started(&rnp
->gp_seq
, gp_seq_req
) ||
987 rcu_seq_state(rcu_seq_current(&rnp
->gp_seq
)))) {
988 trace_rcu_this_gp(rnp
, rdp
, gp_seq_req
,
992 WRITE_ONCE(rnp
->gp_seq_needed
, gp_seq_req
);
993 if (rcu_seq_state(rcu_seq_current(&rnp
->gp_seq
))) {
995 * We just marked the leaf or internal node, and a
996 * grace period is in progress, which means that
997 * rcu_gp_cleanup() will see the marking. Bail to
1000 trace_rcu_this_gp(rnp_start
, rdp
, gp_seq_req
,
1001 TPS("Startedleaf"));
1004 if (rnp
!= rnp_start
&& rnp
->parent
!= NULL
)
1005 raw_spin_unlock_rcu_node(rnp
);
1007 break; /* At root, and perhaps also leaf. */
1010 /* If GP already in progress, just leave, otherwise start one. */
1011 if (rcu_gp_in_progress()) {
1012 trace_rcu_this_gp(rnp
, rdp
, gp_seq_req
, TPS("Startedleafroot"));
1015 trace_rcu_this_gp(rnp
, rdp
, gp_seq_req
, TPS("Startedroot"));
1016 WRITE_ONCE(rcu_state
.gp_flags
, rcu_state
.gp_flags
| RCU_GP_FLAG_INIT
);
1017 WRITE_ONCE(rcu_state
.gp_req_activity
, jiffies
);
1018 if (!READ_ONCE(rcu_state
.gp_kthread
)) {
1019 trace_rcu_this_gp(rnp
, rdp
, gp_seq_req
, TPS("NoGPkthread"));
1022 trace_rcu_grace_period(rcu_state
.name
, data_race(rcu_state
.gp_seq
), TPS("newreq"));
1023 ret
= true; /* Caller must wake GP kthread. */
1025 /* Push furthest requested GP to leaf node and rcu_data structure. */
1026 if (ULONG_CMP_LT(gp_seq_req
, rnp
->gp_seq_needed
)) {
1027 WRITE_ONCE(rnp_start
->gp_seq_needed
, rnp
->gp_seq_needed
);
1028 WRITE_ONCE(rdp
->gp_seq_needed
, rnp
->gp_seq_needed
);
1030 if (rnp
!= rnp_start
)
1031 raw_spin_unlock_rcu_node(rnp
);
1036 * Clean up any old requests for the just-ended grace period. Also return
1037 * whether any additional grace periods have been requested.
1039 static bool rcu_future_gp_cleanup(struct rcu_node
*rnp
)
1042 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
1044 needmore
= ULONG_CMP_LT(rnp
->gp_seq
, rnp
->gp_seq_needed
);
1046 rnp
->gp_seq_needed
= rnp
->gp_seq
; /* Avoid counter wrap. */
1047 trace_rcu_this_gp(rnp
, rdp
, rnp
->gp_seq
,
1048 needmore
? TPS("CleanupMore") : TPS("Cleanup"));
1053 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1054 * interrupt or softirq handler, in which case we just might immediately
1055 * sleep upon return, resulting in a grace-period hang), and don't bother
1056 * awakening when there is nothing for the grace-period kthread to do
1057 * (as in several CPUs raced to awaken, we lost), and finally don't try
1058 * to awaken a kthread that has not yet been created. If all those checks
1059 * are passed, track some debug information and awaken.
1061 * So why do the self-wakeup when in an interrupt or softirq handler
1062 * in the grace-period kthread's context? Because the kthread might have
1063 * been interrupted just as it was going to sleep, and just after the final
1064 * pre-sleep check of the awaken condition. In this case, a wakeup really
1065 * is required, and is therefore supplied.
1067 static void rcu_gp_kthread_wake(void)
1069 struct task_struct
*t
= READ_ONCE(rcu_state
.gp_kthread
);
1071 if ((current
== t
&& !in_hardirq() && !in_serving_softirq()) ||
1072 !READ_ONCE(rcu_state
.gp_flags
) || !t
)
1074 WRITE_ONCE(rcu_state
.gp_wake_time
, jiffies
);
1075 WRITE_ONCE(rcu_state
.gp_wake_seq
, READ_ONCE(rcu_state
.gp_seq
));
1076 swake_up_one(&rcu_state
.gp_wq
);
1080 * If there is room, assign a ->gp_seq number to any callbacks on this
1081 * CPU that have not already been assigned. Also accelerate any callbacks
1082 * that were previously assigned a ->gp_seq number that has since proven
1083 * to be too conservative, which can happen if callbacks get assigned a
1084 * ->gp_seq number while RCU is idle, but with reference to a non-root
1085 * rcu_node structure. This function is idempotent, so it does not hurt
1086 * to call it repeatedly. Returns an flag saying that we should awaken
1087 * the RCU grace-period kthread.
1089 * The caller must hold rnp->lock with interrupts disabled.
1091 static bool rcu_accelerate_cbs(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1093 unsigned long gp_seq_req
;
1096 rcu_lockdep_assert_cblist_protected(rdp
);
1097 raw_lockdep_assert_held_rcu_node(rnp
);
1099 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1100 if (!rcu_segcblist_pend_cbs(&rdp
->cblist
))
1103 trace_rcu_segcb_stats(&rdp
->cblist
, TPS("SegCbPreAcc"));
1106 * Callbacks are often registered with incomplete grace-period
1107 * information. Something about the fact that getting exact
1108 * information requires acquiring a global lock... RCU therefore
1109 * makes a conservative estimate of the grace period number at which
1110 * a given callback will become ready to invoke. The following
1111 * code checks this estimate and improves it when possible, thus
1112 * accelerating callback invocation to an earlier grace-period
1115 gp_seq_req
= rcu_seq_snap(&rcu_state
.gp_seq
);
1116 if (rcu_segcblist_accelerate(&rdp
->cblist
, gp_seq_req
))
1117 ret
= rcu_start_this_gp(rnp
, rdp
, gp_seq_req
);
1119 /* Trace depending on how much we were able to accelerate. */
1120 if (rcu_segcblist_restempty(&rdp
->cblist
, RCU_WAIT_TAIL
))
1121 trace_rcu_grace_period(rcu_state
.name
, gp_seq_req
, TPS("AccWaitCB"));
1123 trace_rcu_grace_period(rcu_state
.name
, gp_seq_req
, TPS("AccReadyCB"));
1125 trace_rcu_segcb_stats(&rdp
->cblist
, TPS("SegCbPostAcc"));
1131 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1132 * rcu_node structure's ->lock be held. It consults the cached value
1133 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1134 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1135 * while holding the leaf rcu_node structure's ->lock.
1137 static void rcu_accelerate_cbs_unlocked(struct rcu_node
*rnp
,
1138 struct rcu_data
*rdp
)
1143 rcu_lockdep_assert_cblist_protected(rdp
);
1144 c
= rcu_seq_snap(&rcu_state
.gp_seq
);
1145 if (!READ_ONCE(rdp
->gpwrap
) && ULONG_CMP_GE(rdp
->gp_seq_needed
, c
)) {
1146 /* Old request still live, so mark recent callbacks. */
1147 (void)rcu_segcblist_accelerate(&rdp
->cblist
, c
);
1150 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
1151 needwake
= rcu_accelerate_cbs(rnp
, rdp
);
1152 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
1154 rcu_gp_kthread_wake();
1158 * Move any callbacks whose grace period has completed to the
1159 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1160 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1161 * sublist. This function is idempotent, so it does not hurt to
1162 * invoke it repeatedly. As long as it is not invoked -too- often...
1163 * Returns true if the RCU grace-period kthread needs to be awakened.
1165 * The caller must hold rnp->lock with interrupts disabled.
1167 static bool rcu_advance_cbs(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1169 rcu_lockdep_assert_cblist_protected(rdp
);
1170 raw_lockdep_assert_held_rcu_node(rnp
);
1172 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1173 if (!rcu_segcblist_pend_cbs(&rdp
->cblist
))
1177 * Find all callbacks whose ->gp_seq numbers indicate that they
1178 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1180 rcu_segcblist_advance(&rdp
->cblist
, rnp
->gp_seq
);
1182 /* Classify any remaining callbacks. */
1183 return rcu_accelerate_cbs(rnp
, rdp
);
1187 * Move and classify callbacks, but only if doing so won't require
1188 * that the RCU grace-period kthread be awakened.
1190 static void __maybe_unused
rcu_advance_cbs_nowake(struct rcu_node
*rnp
,
1191 struct rcu_data
*rdp
)
1193 rcu_lockdep_assert_cblist_protected(rdp
);
1194 if (!rcu_seq_state(rcu_seq_current(&rnp
->gp_seq
)) || !raw_spin_trylock_rcu_node(rnp
))
1196 // The grace period cannot end while we hold the rcu_node lock.
1197 if (rcu_seq_state(rcu_seq_current(&rnp
->gp_seq
)))
1198 WARN_ON_ONCE(rcu_advance_cbs(rnp
, rdp
));
1199 raw_spin_unlock_rcu_node(rnp
);
1203 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1204 * quiescent state. This is intended to be invoked when the CPU notices
1205 * a new grace period.
1207 static void rcu_strict_gp_check_qs(void)
1209 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD
)) {
1216 * Update CPU-local rcu_data state to record the beginnings and ends of
1217 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1218 * structure corresponding to the current CPU, and must have irqs disabled.
1219 * Returns true if the grace-period kthread needs to be awakened.
1221 static bool __note_gp_changes(struct rcu_node
*rnp
, struct rcu_data
*rdp
)
1225 const bool offloaded
= rcu_rdp_is_offloaded(rdp
);
1227 raw_lockdep_assert_held_rcu_node(rnp
);
1229 if (rdp
->gp_seq
== rnp
->gp_seq
)
1230 return false; /* Nothing to do. */
1232 /* Handle the ends of any preceding grace periods first. */
1233 if (rcu_seq_completed_gp(rdp
->gp_seq
, rnp
->gp_seq
) ||
1234 unlikely(READ_ONCE(rdp
->gpwrap
))) {
1236 ret
= rcu_advance_cbs(rnp
, rdp
); /* Advance CBs. */
1237 rdp
->core_needs_qs
= false;
1238 trace_rcu_grace_period(rcu_state
.name
, rdp
->gp_seq
, TPS("cpuend"));
1241 ret
= rcu_accelerate_cbs(rnp
, rdp
); /* Recent CBs. */
1242 if (rdp
->core_needs_qs
)
1243 rdp
->core_needs_qs
= !!(rnp
->qsmask
& rdp
->grpmask
);
1246 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1247 if (rcu_seq_new_gp(rdp
->gp_seq
, rnp
->gp_seq
) ||
1248 unlikely(READ_ONCE(rdp
->gpwrap
))) {
1250 * If the current grace period is waiting for this CPU,
1251 * set up to detect a quiescent state, otherwise don't
1252 * go looking for one.
1254 trace_rcu_grace_period(rcu_state
.name
, rnp
->gp_seq
, TPS("cpustart"));
1255 need_qs
= !!(rnp
->qsmask
& rdp
->grpmask
);
1256 rdp
->cpu_no_qs
.b
.norm
= need_qs
;
1257 rdp
->core_needs_qs
= need_qs
;
1258 zero_cpu_stall_ticks(rdp
);
1260 rdp
->gp_seq
= rnp
->gp_seq
; /* Remember new grace-period state. */
1261 if (ULONG_CMP_LT(rdp
->gp_seq_needed
, rnp
->gp_seq_needed
) || rdp
->gpwrap
)
1262 WRITE_ONCE(rdp
->gp_seq_needed
, rnp
->gp_seq_needed
);
1263 if (IS_ENABLED(CONFIG_PROVE_RCU
) && READ_ONCE(rdp
->gpwrap
))
1264 WRITE_ONCE(rdp
->last_sched_clock
, jiffies
);
1265 WRITE_ONCE(rdp
->gpwrap
, false);
1266 rcu_gpnum_ovf(rnp
, rdp
);
1270 static void note_gp_changes(struct rcu_data
*rdp
)
1272 unsigned long flags
;
1274 struct rcu_node
*rnp
;
1276 local_irq_save(flags
);
1278 if ((rdp
->gp_seq
== rcu_seq_current(&rnp
->gp_seq
) &&
1279 !unlikely(READ_ONCE(rdp
->gpwrap
))) || /* w/out lock. */
1280 !raw_spin_trylock_rcu_node(rnp
)) { /* irqs already off, so later. */
1281 local_irq_restore(flags
);
1284 needwake
= __note_gp_changes(rnp
, rdp
);
1285 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1286 rcu_strict_gp_check_qs();
1288 rcu_gp_kthread_wake();
1291 static atomic_t
*rcu_gp_slow_suppress
;
1293 /* Register a counter to suppress debugging grace-period delays. */
1294 void rcu_gp_slow_register(atomic_t
*rgssp
)
1296 WARN_ON_ONCE(rcu_gp_slow_suppress
);
1298 WRITE_ONCE(rcu_gp_slow_suppress
, rgssp
);
1300 EXPORT_SYMBOL_GPL(rcu_gp_slow_register
);
1302 /* Unregister a counter, with NULL for not caring which. */
1303 void rcu_gp_slow_unregister(atomic_t
*rgssp
)
1305 WARN_ON_ONCE(rgssp
&& rgssp
!= rcu_gp_slow_suppress
);
1307 WRITE_ONCE(rcu_gp_slow_suppress
, NULL
);
1309 EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister
);
1311 static bool rcu_gp_slow_is_suppressed(void)
1313 atomic_t
*rgssp
= READ_ONCE(rcu_gp_slow_suppress
);
1315 return rgssp
&& atomic_read(rgssp
);
1318 static void rcu_gp_slow(int delay
)
1320 if (!rcu_gp_slow_is_suppressed() && delay
> 0 &&
1321 !(rcu_seq_ctr(rcu_state
.gp_seq
) % (rcu_num_nodes
* PER_RCU_NODE_PERIOD
* delay
)))
1322 schedule_timeout_idle(delay
);
1325 static unsigned long sleep_duration
;
1327 /* Allow rcutorture to stall the grace-period kthread. */
1328 void rcu_gp_set_torture_wait(int duration
)
1330 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST
) && duration
> 0)
1331 WRITE_ONCE(sleep_duration
, duration
);
1333 EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait
);
1335 /* Actually implement the aforementioned wait. */
1336 static void rcu_gp_torture_wait(void)
1338 unsigned long duration
;
1340 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST
))
1342 duration
= xchg(&sleep_duration
, 0UL);
1344 pr_alert("%s: Waiting %lu jiffies\n", __func__
, duration
);
1345 schedule_timeout_idle(duration
);
1346 pr_alert("%s: Wait complete\n", __func__
);
1351 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1354 static void rcu_strict_gp_boundary(void *unused
)
1359 // Has rcu_init() been invoked? This is used (for example) to determine
1360 // whether spinlocks may be acquired safely.
1361 static bool rcu_init_invoked(void)
1363 return !!rcu_state
.n_online_cpus
;
1366 // Make the polled API aware of the beginning of a grace period.
1367 static void rcu_poll_gp_seq_start(unsigned long *snap
)
1369 struct rcu_node
*rnp
= rcu_get_root();
1371 if (rcu_init_invoked())
1372 raw_lockdep_assert_held_rcu_node(rnp
);
1374 // If RCU was idle, note beginning of GP.
1375 if (!rcu_seq_state(rcu_state
.gp_seq_polled
))
1376 rcu_seq_start(&rcu_state
.gp_seq_polled
);
1378 // Either way, record current state.
1379 *snap
= rcu_state
.gp_seq_polled
;
1382 // Make the polled API aware of the end of a grace period.
1383 static void rcu_poll_gp_seq_end(unsigned long *snap
)
1385 struct rcu_node
*rnp
= rcu_get_root();
1387 if (rcu_init_invoked())
1388 raw_lockdep_assert_held_rcu_node(rnp
);
1390 // If the previously noted GP is still in effect, record the
1391 // end of that GP. Either way, zero counter to avoid counter-wrap
1393 if (*snap
&& *snap
== rcu_state
.gp_seq_polled
) {
1394 rcu_seq_end(&rcu_state
.gp_seq_polled
);
1395 rcu_state
.gp_seq_polled_snap
= 0;
1396 rcu_state
.gp_seq_polled_exp_snap
= 0;
1402 // Make the polled API aware of the beginning of a grace period, but
1403 // where caller does not hold the root rcu_node structure's lock.
1404 static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap
)
1406 struct rcu_node
*rnp
= rcu_get_root();
1408 if (rcu_init_invoked()) {
1409 lockdep_assert_irqs_enabled();
1410 raw_spin_lock_irq_rcu_node(rnp
);
1412 rcu_poll_gp_seq_start(snap
);
1413 if (rcu_init_invoked())
1414 raw_spin_unlock_irq_rcu_node(rnp
);
1417 // Make the polled API aware of the end of a grace period, but where
1418 // caller does not hold the root rcu_node structure's lock.
1419 static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap
)
1421 struct rcu_node
*rnp
= rcu_get_root();
1423 if (rcu_init_invoked()) {
1424 lockdep_assert_irqs_enabled();
1425 raw_spin_lock_irq_rcu_node(rnp
);
1427 rcu_poll_gp_seq_end(snap
);
1428 if (rcu_init_invoked())
1429 raw_spin_unlock_irq_rcu_node(rnp
);
1433 * Initialize a new grace period. Return false if no grace period required.
1435 static noinline_for_stack
bool rcu_gp_init(void)
1437 unsigned long flags
;
1438 unsigned long oldmask
;
1440 struct rcu_data
*rdp
;
1441 struct rcu_node
*rnp
= rcu_get_root();
1443 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1444 raw_spin_lock_irq_rcu_node(rnp
);
1445 if (!READ_ONCE(rcu_state
.gp_flags
)) {
1446 /* Spurious wakeup, tell caller to go back to sleep. */
1447 raw_spin_unlock_irq_rcu_node(rnp
);
1450 WRITE_ONCE(rcu_state
.gp_flags
, 0); /* Clear all flags: New GP. */
1452 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1454 * Grace period already in progress, don't start another.
1455 * Not supposed to be able to happen.
1457 raw_spin_unlock_irq_rcu_node(rnp
);
1461 /* Advance to a new grace period and initialize state. */
1462 record_gp_stall_check_time();
1463 /* Record GP times before starting GP, hence rcu_seq_start(). */
1464 rcu_seq_start(&rcu_state
.gp_seq
);
1465 ASSERT_EXCLUSIVE_WRITER(rcu_state
.gp_seq
);
1466 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
, TPS("start"));
1467 rcu_poll_gp_seq_start(&rcu_state
.gp_seq_polled_snap
);
1468 raw_spin_unlock_irq_rcu_node(rnp
);
1471 * Apply per-leaf buffered online and offline operations to
1472 * the rcu_node tree. Note that this new grace period need not
1473 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1474 * offlining path, when combined with checks in this function,
1475 * will handle CPUs that are currently going offline or that will
1476 * go offline later. Please also refer to "Hotplug CPU" section
1477 * of RCU's Requirements documentation.
1479 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_ONOFF
);
1480 /* Exclude CPU hotplug operations. */
1481 rcu_for_each_leaf_node(rnp
) {
1482 local_irq_save(flags
);
1483 arch_spin_lock(&rcu_state
.ofl_lock
);
1484 raw_spin_lock_rcu_node(rnp
);
1485 if (rnp
->qsmaskinit
== rnp
->qsmaskinitnext
&&
1486 !rnp
->wait_blkd_tasks
) {
1487 /* Nothing to do on this leaf rcu_node structure. */
1488 raw_spin_unlock_rcu_node(rnp
);
1489 arch_spin_unlock(&rcu_state
.ofl_lock
);
1490 local_irq_restore(flags
);
1494 /* Record old state, apply changes to ->qsmaskinit field. */
1495 oldmask
= rnp
->qsmaskinit
;
1496 rnp
->qsmaskinit
= rnp
->qsmaskinitnext
;
1498 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1499 if (!oldmask
!= !rnp
->qsmaskinit
) {
1500 if (!oldmask
) { /* First online CPU for rcu_node. */
1501 if (!rnp
->wait_blkd_tasks
) /* Ever offline? */
1502 rcu_init_new_rnp(rnp
);
1503 } else if (rcu_preempt_has_tasks(rnp
)) {
1504 rnp
->wait_blkd_tasks
= true; /* blocked tasks */
1505 } else { /* Last offline CPU and can propagate. */
1506 rcu_cleanup_dead_rnp(rnp
);
1511 * If all waited-on tasks from prior grace period are
1512 * done, and if all this rcu_node structure's CPUs are
1513 * still offline, propagate up the rcu_node tree and
1514 * clear ->wait_blkd_tasks. Otherwise, if one of this
1515 * rcu_node structure's CPUs has since come back online,
1516 * simply clear ->wait_blkd_tasks.
1518 if (rnp
->wait_blkd_tasks
&&
1519 (!rcu_preempt_has_tasks(rnp
) || rnp
->qsmaskinit
)) {
1520 rnp
->wait_blkd_tasks
= false;
1521 if (!rnp
->qsmaskinit
)
1522 rcu_cleanup_dead_rnp(rnp
);
1525 raw_spin_unlock_rcu_node(rnp
);
1526 arch_spin_unlock(&rcu_state
.ofl_lock
);
1527 local_irq_restore(flags
);
1529 rcu_gp_slow(gp_preinit_delay
); /* Races with CPU hotplug. */
1532 * Set the quiescent-state-needed bits in all the rcu_node
1533 * structures for all currently online CPUs in breadth-first
1534 * order, starting from the root rcu_node structure, relying on the
1535 * layout of the tree within the rcu_state.node[] array. Note that
1536 * other CPUs will access only the leaves of the hierarchy, thus
1537 * seeing that no grace period is in progress, at least until the
1538 * corresponding leaf node has been initialized.
1540 * The grace period cannot complete until the initialization
1541 * process finishes, because this kthread handles both.
1543 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_INIT
);
1544 rcu_for_each_node_breadth_first(rnp
) {
1545 rcu_gp_slow(gp_init_delay
);
1546 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1547 rdp
= this_cpu_ptr(&rcu_data
);
1548 rcu_preempt_check_blocked_tasks(rnp
);
1549 rnp
->qsmask
= rnp
->qsmaskinit
;
1550 WRITE_ONCE(rnp
->gp_seq
, rcu_state
.gp_seq
);
1551 if (rnp
== rdp
->mynode
)
1552 (void)__note_gp_changes(rnp
, rdp
);
1553 rcu_preempt_boost_start_gp(rnp
);
1554 trace_rcu_grace_period_init(rcu_state
.name
, rnp
->gp_seq
,
1555 rnp
->level
, rnp
->grplo
,
1556 rnp
->grphi
, rnp
->qsmask
);
1557 /* Quiescent states for tasks on any now-offline CPUs. */
1558 mask
= rnp
->qsmask
& ~rnp
->qsmaskinitnext
;
1559 rnp
->rcu_gp_init_mask
= mask
;
1560 if ((mask
|| rnp
->wait_blkd_tasks
) && rcu_is_leaf_node(rnp
))
1561 rcu_report_qs_rnp(mask
, rnp
, rnp
->gp_seq
, flags
);
1563 raw_spin_unlock_irq_rcu_node(rnp
);
1564 cond_resched_tasks_rcu_qs();
1565 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1568 // If strict, make all CPUs aware of new grace period.
1569 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD
))
1570 on_each_cpu(rcu_strict_gp_boundary
, NULL
, 0);
1576 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1579 static bool rcu_gp_fqs_check_wake(int *gfp
)
1581 struct rcu_node
*rnp
= rcu_get_root();
1583 // If under overload conditions, force an immediate FQS scan.
1584 if (*gfp
& RCU_GP_FLAG_OVLD
)
1587 // Someone like call_rcu() requested a force-quiescent-state scan.
1588 *gfp
= READ_ONCE(rcu_state
.gp_flags
);
1589 if (*gfp
& RCU_GP_FLAG_FQS
)
1592 // The current grace period has completed.
1593 if (!READ_ONCE(rnp
->qsmask
) && !rcu_preempt_blocked_readers_cgp(rnp
))
1600 * Do one round of quiescent-state forcing.
1602 static void rcu_gp_fqs(bool first_time
)
1604 struct rcu_node
*rnp
= rcu_get_root();
1606 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1607 WRITE_ONCE(rcu_state
.n_force_qs
, rcu_state
.n_force_qs
+ 1);
1609 /* Collect dyntick-idle snapshots. */
1610 force_qs_rnp(dyntick_save_progress_counter
);
1612 /* Handle dyntick-idle and offline CPUs. */
1613 force_qs_rnp(rcu_implicit_dynticks_qs
);
1615 /* Clear flag to prevent immediate re-entry. */
1616 if (READ_ONCE(rcu_state
.gp_flags
) & RCU_GP_FLAG_FQS
) {
1617 raw_spin_lock_irq_rcu_node(rnp
);
1618 WRITE_ONCE(rcu_state
.gp_flags
,
1619 READ_ONCE(rcu_state
.gp_flags
) & ~RCU_GP_FLAG_FQS
);
1620 raw_spin_unlock_irq_rcu_node(rnp
);
1625 * Loop doing repeated quiescent-state forcing until the grace period ends.
1627 static noinline_for_stack
void rcu_gp_fqs_loop(void)
1629 bool first_gp_fqs
= true;
1633 struct rcu_node
*rnp
= rcu_get_root();
1635 j
= READ_ONCE(jiffies_till_first_fqs
);
1636 if (rcu_state
.cbovld
)
1637 gf
= RCU_GP_FLAG_OVLD
;
1640 if (rcu_state
.cbovld
) {
1645 if (!ret
|| time_before(jiffies
+ j
, rcu_state
.jiffies_force_qs
)) {
1646 WRITE_ONCE(rcu_state
.jiffies_force_qs
, jiffies
+ j
);
1648 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1649 * update; required for stall checks.
1652 WRITE_ONCE(rcu_state
.jiffies_kick_kthreads
,
1653 jiffies
+ (j
? 3 * j
: 2));
1655 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
,
1657 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_WAIT_FQS
);
1658 (void)swait_event_idle_timeout_exclusive(rcu_state
.gp_wq
,
1659 rcu_gp_fqs_check_wake(&gf
), j
);
1660 rcu_gp_torture_wait();
1661 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_DOING_FQS
);
1662 /* Locking provides needed memory barriers. */
1664 * Exit the loop if the root rcu_node structure indicates that the grace period
1665 * has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check
1666 * is required only for single-node rcu_node trees because readers blocking
1667 * the current grace period are queued only on leaf rcu_node structures.
1668 * For multi-node trees, checking the root node's ->qsmask suffices, because a
1669 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
1670 * the corresponding leaf nodes have passed through their quiescent state.
1672 if (!READ_ONCE(rnp
->qsmask
) &&
1673 !rcu_preempt_blocked_readers_cgp(rnp
))
1675 /* If time for quiescent-state forcing, do it. */
1676 if (!time_after(rcu_state
.jiffies_force_qs
, jiffies
) ||
1677 (gf
& (RCU_GP_FLAG_FQS
| RCU_GP_FLAG_OVLD
))) {
1678 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
,
1680 rcu_gp_fqs(first_gp_fqs
);
1683 first_gp_fqs
= false;
1684 gf
= rcu_state
.cbovld
? RCU_GP_FLAG_OVLD
: 0;
1686 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
,
1688 cond_resched_tasks_rcu_qs();
1689 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1690 ret
= 0; /* Force full wait till next FQS. */
1691 j
= READ_ONCE(jiffies_till_next_fqs
);
1693 /* Deal with stray signal. */
1694 cond_resched_tasks_rcu_qs();
1695 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1696 WARN_ON(signal_pending(current
));
1697 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
,
1699 ret
= 1; /* Keep old FQS timing. */
1701 if (time_after(jiffies
, rcu_state
.jiffies_force_qs
))
1704 j
= rcu_state
.jiffies_force_qs
- j
;
1711 * Clean up after the old grace period.
1713 static noinline
void rcu_gp_cleanup(void)
1716 bool needgp
= false;
1717 unsigned long gp_duration
;
1718 unsigned long new_gp_seq
;
1720 struct rcu_data
*rdp
;
1721 struct rcu_node
*rnp
= rcu_get_root();
1722 struct swait_queue_head
*sq
;
1724 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1725 raw_spin_lock_irq_rcu_node(rnp
);
1726 rcu_state
.gp_end
= jiffies
;
1727 gp_duration
= rcu_state
.gp_end
- rcu_state
.gp_start
;
1728 if (gp_duration
> rcu_state
.gp_max
)
1729 rcu_state
.gp_max
= gp_duration
;
1732 * We know the grace period is complete, but to everyone else
1733 * it appears to still be ongoing. But it is also the case
1734 * that to everyone else it looks like there is nothing that
1735 * they can do to advance the grace period. It is therefore
1736 * safe for us to drop the lock in order to mark the grace
1737 * period as completed in all of the rcu_node structures.
1739 rcu_poll_gp_seq_end(&rcu_state
.gp_seq_polled_snap
);
1740 raw_spin_unlock_irq_rcu_node(rnp
);
1743 * Propagate new ->gp_seq value to rcu_node structures so that
1744 * other CPUs don't have to wait until the start of the next grace
1745 * period to process their callbacks. This also avoids some nasty
1746 * RCU grace-period initialization races by forcing the end of
1747 * the current grace period to be completely recorded in all of
1748 * the rcu_node structures before the beginning of the next grace
1749 * period is recorded in any of the rcu_node structures.
1751 new_gp_seq
= rcu_state
.gp_seq
;
1752 rcu_seq_end(&new_gp_seq
);
1753 rcu_for_each_node_breadth_first(rnp
) {
1754 raw_spin_lock_irq_rcu_node(rnp
);
1755 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
)))
1756 dump_blkd_tasks(rnp
, 10);
1757 WARN_ON_ONCE(rnp
->qsmask
);
1758 WRITE_ONCE(rnp
->gp_seq
, new_gp_seq
);
1760 smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
1761 rdp
= this_cpu_ptr(&rcu_data
);
1762 if (rnp
== rdp
->mynode
)
1763 needgp
= __note_gp_changes(rnp
, rdp
) || needgp
;
1764 /* smp_mb() provided by prior unlock-lock pair. */
1765 needgp
= rcu_future_gp_cleanup(rnp
) || needgp
;
1766 // Reset overload indication for CPUs no longer overloaded
1767 if (rcu_is_leaf_node(rnp
))
1768 for_each_leaf_node_cpu_mask(rnp
, cpu
, rnp
->cbovldmask
) {
1769 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
1770 check_cb_ovld_locked(rdp
, rnp
);
1772 sq
= rcu_nocb_gp_get(rnp
);
1773 raw_spin_unlock_irq_rcu_node(rnp
);
1774 rcu_nocb_gp_cleanup(sq
);
1775 cond_resched_tasks_rcu_qs();
1776 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1777 rcu_gp_slow(gp_cleanup_delay
);
1779 rnp
= rcu_get_root();
1780 raw_spin_lock_irq_rcu_node(rnp
); /* GP before ->gp_seq update. */
1782 /* Declare grace period done, trace first to use old GP number. */
1783 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
, TPS("end"));
1784 rcu_seq_end(&rcu_state
.gp_seq
);
1785 ASSERT_EXCLUSIVE_WRITER(rcu_state
.gp_seq
);
1786 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_IDLE
);
1787 /* Check for GP requests since above loop. */
1788 rdp
= this_cpu_ptr(&rcu_data
);
1789 if (!needgp
&& ULONG_CMP_LT(rnp
->gp_seq
, rnp
->gp_seq_needed
)) {
1790 trace_rcu_this_gp(rnp
, rdp
, rnp
->gp_seq_needed
,
1791 TPS("CleanupMore"));
1794 /* Advance CBs to reduce false positives below. */
1795 offloaded
= rcu_rdp_is_offloaded(rdp
);
1796 if ((offloaded
|| !rcu_accelerate_cbs(rnp
, rdp
)) && needgp
) {
1798 // We get here if a grace period was needed (“needgp”)
1799 // and the above call to rcu_accelerate_cbs() did not set
1800 // the RCU_GP_FLAG_INIT bit in ->gp_state (which records
1801 // the need for another grace period). The purpose
1802 // of the “offloaded” check is to avoid invoking
1803 // rcu_accelerate_cbs() on an offloaded CPU because we do not
1804 // hold the ->nocb_lock needed to safely access an offloaded
1805 // ->cblist. We do not want to acquire that lock because
1806 // it can be heavily contended during callback floods.
1808 WRITE_ONCE(rcu_state
.gp_flags
, RCU_GP_FLAG_INIT
);
1809 WRITE_ONCE(rcu_state
.gp_req_activity
, jiffies
);
1810 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
, TPS("newreq"));
1813 // We get here either if there is no need for an
1814 // additional grace period or if rcu_accelerate_cbs() has
1815 // already set the RCU_GP_FLAG_INIT bit in ->gp_flags.
1816 // So all we need to do is to clear all of the other
1819 WRITE_ONCE(rcu_state
.gp_flags
, rcu_state
.gp_flags
& RCU_GP_FLAG_INIT
);
1821 raw_spin_unlock_irq_rcu_node(rnp
);
1823 // If strict, make all CPUs aware of the end of the old grace period.
1824 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD
))
1825 on_each_cpu(rcu_strict_gp_boundary
, NULL
, 0);
1829 * Body of kthread that handles grace periods.
1831 static int __noreturn
rcu_gp_kthread(void *unused
)
1833 rcu_bind_gp_kthread();
1836 /* Handle grace-period start. */
1838 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
,
1840 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_WAIT_GPS
);
1841 swait_event_idle_exclusive(rcu_state
.gp_wq
,
1842 READ_ONCE(rcu_state
.gp_flags
) &
1844 rcu_gp_torture_wait();
1845 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_DONE_GPS
);
1846 /* Locking provides needed memory barrier. */
1849 cond_resched_tasks_rcu_qs();
1850 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
1851 WARN_ON(signal_pending(current
));
1852 trace_rcu_grace_period(rcu_state
.name
, rcu_state
.gp_seq
,
1856 /* Handle quiescent-state forcing. */
1859 /* Handle grace-period end. */
1860 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_CLEANUP
);
1862 WRITE_ONCE(rcu_state
.gp_state
, RCU_GP_CLEANED
);
1867 * Report a full set of quiescent states to the rcu_state data structure.
1868 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1869 * another grace period is required. Whether we wake the grace-period
1870 * kthread or it awakens itself for the next round of quiescent-state
1871 * forcing, that kthread will clean up after the just-completed grace
1872 * period. Note that the caller must hold rnp->lock, which is released
1875 static void rcu_report_qs_rsp(unsigned long flags
)
1876 __releases(rcu_get_root()->lock
)
1878 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1879 WARN_ON_ONCE(!rcu_gp_in_progress());
1880 WRITE_ONCE(rcu_state
.gp_flags
,
1881 READ_ONCE(rcu_state
.gp_flags
) | RCU_GP_FLAG_FQS
);
1882 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags
);
1883 rcu_gp_kthread_wake();
1887 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1888 * Allows quiescent states for a group of CPUs to be reported at one go
1889 * to the specified rcu_node structure, though all the CPUs in the group
1890 * must be represented by the same rcu_node structure (which need not be a
1891 * leaf rcu_node structure, though it often will be). The gps parameter
1892 * is the grace-period snapshot, which means that the quiescent states
1893 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1894 * must be held upon entry, and it is released before return.
1896 * As a special case, if mask is zero, the bit-already-cleared check is
1897 * disabled. This allows propagating quiescent state due to resumed tasks
1898 * during grace-period initialization.
1900 static void rcu_report_qs_rnp(unsigned long mask
, struct rcu_node
*rnp
,
1901 unsigned long gps
, unsigned long flags
)
1902 __releases(rnp
->lock
)
1904 unsigned long oldmask
= 0;
1905 struct rcu_node
*rnp_c
;
1907 raw_lockdep_assert_held_rcu_node(rnp
);
1909 /* Walk up the rcu_node hierarchy. */
1911 if ((!(rnp
->qsmask
& mask
) && mask
) || rnp
->gp_seq
!= gps
) {
1914 * Our bit has already been cleared, or the
1915 * relevant grace period is already over, so done.
1917 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1920 WARN_ON_ONCE(oldmask
); /* Any child must be all zeroed! */
1921 WARN_ON_ONCE(!rcu_is_leaf_node(rnp
) &&
1922 rcu_preempt_blocked_readers_cgp(rnp
));
1923 WRITE_ONCE(rnp
->qsmask
, rnp
->qsmask
& ~mask
);
1924 trace_rcu_quiescent_state_report(rcu_state
.name
, rnp
->gp_seq
,
1925 mask
, rnp
->qsmask
, rnp
->level
,
1926 rnp
->grplo
, rnp
->grphi
,
1928 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
1930 /* Other bits still set at this level, so done. */
1931 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1934 rnp
->completedqs
= rnp
->gp_seq
;
1935 mask
= rnp
->grpmask
;
1936 if (rnp
->parent
== NULL
) {
1938 /* No more levels. Exit loop holding root lock. */
1942 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1945 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
1946 oldmask
= READ_ONCE(rnp_c
->qsmask
);
1950 * Get here if we are the last CPU to pass through a quiescent
1951 * state for this grace period. Invoke rcu_report_qs_rsp()
1952 * to clean up and start the next grace period if one is needed.
1954 rcu_report_qs_rsp(flags
); /* releases rnp->lock. */
1958 * Record a quiescent state for all tasks that were previously queued
1959 * on the specified rcu_node structure and that were blocking the current
1960 * RCU grace period. The caller must hold the corresponding rnp->lock with
1961 * irqs disabled, and this lock is released upon return, but irqs remain
1964 static void __maybe_unused
1965 rcu_report_unblock_qs_rnp(struct rcu_node
*rnp
, unsigned long flags
)
1966 __releases(rnp
->lock
)
1970 struct rcu_node
*rnp_p
;
1972 raw_lockdep_assert_held_rcu_node(rnp
);
1973 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU
)) ||
1974 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp
)) ||
1976 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
1977 return; /* Still need more quiescent states! */
1980 rnp
->completedqs
= rnp
->gp_seq
;
1981 rnp_p
= rnp
->parent
;
1982 if (rnp_p
== NULL
) {
1984 * Only one rcu_node structure in the tree, so don't
1985 * try to report up to its nonexistent parent!
1987 rcu_report_qs_rsp(flags
);
1991 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1993 mask
= rnp
->grpmask
;
1994 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
1995 raw_spin_lock_rcu_node(rnp_p
); /* irqs already disabled. */
1996 rcu_report_qs_rnp(mask
, rnp_p
, gps
, flags
);
2000 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2001 * structure. This must be called from the specified CPU.
2004 rcu_report_qs_rdp(struct rcu_data
*rdp
)
2006 unsigned long flags
;
2008 bool needwake
= false;
2009 bool needacc
= false;
2010 struct rcu_node
*rnp
;
2012 WARN_ON_ONCE(rdp
->cpu
!= smp_processor_id());
2014 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2015 if (rdp
->cpu_no_qs
.b
.norm
|| rdp
->gp_seq
!= rnp
->gp_seq
||
2019 * The grace period in which this quiescent state was
2020 * recorded has ended, so don't report it upwards.
2021 * We will instead need a new quiescent state that lies
2022 * within the current grace period.
2024 rdp
->cpu_no_qs
.b
.norm
= true; /* need qs for new gp. */
2025 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2028 mask
= rdp
->grpmask
;
2029 rdp
->core_needs_qs
= false;
2030 if ((rnp
->qsmask
& mask
) == 0) {
2031 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2034 * This GP can't end until cpu checks in, so all of our
2035 * callbacks can be processed during the next GP.
2037 * NOCB kthreads have their own way to deal with that...
2039 if (!rcu_rdp_is_offloaded(rdp
)) {
2040 needwake
= rcu_accelerate_cbs(rnp
, rdp
);
2041 } else if (!rcu_segcblist_completely_offloaded(&rdp
->cblist
)) {
2043 * ...but NOCB kthreads may miss or delay callbacks acceleration
2044 * if in the middle of a (de-)offloading process.
2049 rcu_disable_urgency_upon_qs(rdp
);
2050 rcu_report_qs_rnp(mask
, rnp
, rnp
->gp_seq
, flags
);
2051 /* ^^^ Released rnp->lock */
2053 rcu_gp_kthread_wake();
2056 rcu_nocb_lock_irqsave(rdp
, flags
);
2057 rcu_accelerate_cbs_unlocked(rnp
, rdp
);
2058 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2064 * Check to see if there is a new grace period of which this CPU
2065 * is not yet aware, and if so, set up local rcu_data state for it.
2066 * Otherwise, see if this CPU has just passed through its first
2067 * quiescent state for this grace period, and record that fact if so.
2070 rcu_check_quiescent_state(struct rcu_data
*rdp
)
2072 /* Check for grace-period ends and beginnings. */
2073 note_gp_changes(rdp
);
2076 * Does this CPU still need to do its part for current grace period?
2077 * If no, return and let the other CPUs do their part as well.
2079 if (!rdp
->core_needs_qs
)
2083 * Was there a quiescent state since the beginning of the grace
2084 * period? If no, then exit and wait for the next call.
2086 if (rdp
->cpu_no_qs
.b
.norm
)
2090 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2093 rcu_report_qs_rdp(rdp
);
2097 * Near the end of the offline process. Trace the fact that this CPU
2100 int rcutree_dying_cpu(unsigned int cpu
)
2103 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2104 struct rcu_node
*rnp
= rdp
->mynode
;
2106 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2109 blkd
= !!(rnp
->qsmask
& rdp
->grpmask
);
2110 trace_rcu_grace_period(rcu_state
.name
, READ_ONCE(rnp
->gp_seq
),
2111 blkd
? TPS("cpuofl-bgp") : TPS("cpuofl"));
2116 * All CPUs for the specified rcu_node structure have gone offline,
2117 * and all tasks that were preempted within an RCU read-side critical
2118 * section while running on one of those CPUs have since exited their RCU
2119 * read-side critical section. Some other CPU is reporting this fact with
2120 * the specified rcu_node structure's ->lock held and interrupts disabled.
2121 * This function therefore goes up the tree of rcu_node structures,
2122 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2123 * the leaf rcu_node structure's ->qsmaskinit field has already been
2126 * This function does check that the specified rcu_node structure has
2127 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2128 * prematurely. That said, invoking it after the fact will cost you
2129 * a needless lock acquisition. So once it has done its work, don't
2132 static void rcu_cleanup_dead_rnp(struct rcu_node
*rnp_leaf
)
2135 struct rcu_node
*rnp
= rnp_leaf
;
2137 raw_lockdep_assert_held_rcu_node(rnp_leaf
);
2138 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
) ||
2139 WARN_ON_ONCE(rnp_leaf
->qsmaskinit
) ||
2140 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf
)))
2143 mask
= rnp
->grpmask
;
2147 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
2148 rnp
->qsmaskinit
&= ~mask
;
2149 /* Between grace periods, so better already be zero! */
2150 WARN_ON_ONCE(rnp
->qsmask
);
2151 if (rnp
->qsmaskinit
) {
2152 raw_spin_unlock_rcu_node(rnp
);
2153 /* irqs remain disabled. */
2156 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
2161 * The CPU has been completely removed, and some other CPU is reporting
2162 * this fact from process context. Do the remainder of the cleanup.
2163 * There can only be one CPU hotplug operation at a time, so no need for
2166 int rcutree_dead_cpu(unsigned int cpu
)
2168 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2169 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
2171 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
))
2174 WRITE_ONCE(rcu_state
.n_online_cpus
, rcu_state
.n_online_cpus
- 1);
2175 /* Adjust any no-longer-needed kthreads. */
2176 rcu_boost_kthread_setaffinity(rnp
, -1);
2177 // Stop-machine done, so allow nohz_full to disable tick.
2178 tick_dep_clear(TICK_DEP_BIT_RCU
);
2183 * Invoke any RCU callbacks that have made it to the end of their grace
2184 * period. Throttle as specified by rdp->blimit.
2186 static void rcu_do_batch(struct rcu_data
*rdp
)
2189 bool __maybe_unused empty
;
2190 unsigned long flags
;
2191 struct rcu_head
*rhp
;
2192 struct rcu_cblist rcl
= RCU_CBLIST_INITIALIZER(rcl
);
2194 long pending
, tlimit
= 0;
2196 /* If no callbacks are ready, just return. */
2197 if (!rcu_segcblist_ready_cbs(&rdp
->cblist
)) {
2198 trace_rcu_batch_start(rcu_state
.name
,
2199 rcu_segcblist_n_cbs(&rdp
->cblist
), 0);
2200 trace_rcu_batch_end(rcu_state
.name
, 0,
2201 !rcu_segcblist_empty(&rdp
->cblist
),
2202 need_resched(), is_idle_task(current
),
2203 rcu_is_callbacks_kthread(rdp
));
2208 * Extract the list of ready callbacks, disabling IRQs to prevent
2209 * races with call_rcu() from interrupt handlers. Leave the
2210 * callback counts, as rcu_barrier() needs to be conservative.
2212 rcu_nocb_lock_irqsave(rdp
, flags
);
2213 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2214 pending
= rcu_segcblist_n_cbs(&rdp
->cblist
);
2215 div
= READ_ONCE(rcu_divisor
);
2216 div
= div
< 0 ? 7 : div
> sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div
;
2217 bl
= max(rdp
->blimit
, pending
>> div
);
2218 if (in_serving_softirq() && unlikely(bl
> 100)) {
2219 long rrn
= READ_ONCE(rcu_resched_ns
);
2221 rrn
= rrn
< NSEC_PER_MSEC
? NSEC_PER_MSEC
: rrn
> NSEC_PER_SEC
? NSEC_PER_SEC
: rrn
;
2222 tlimit
= local_clock() + rrn
;
2224 trace_rcu_batch_start(rcu_state
.name
,
2225 rcu_segcblist_n_cbs(&rdp
->cblist
), bl
);
2226 rcu_segcblist_extract_done_cbs(&rdp
->cblist
, &rcl
);
2227 if (rcu_rdp_is_offloaded(rdp
))
2228 rdp
->qlen_last_fqs_check
= rcu_segcblist_n_cbs(&rdp
->cblist
);
2230 trace_rcu_segcb_stats(&rdp
->cblist
, TPS("SegCbDequeued"));
2231 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2233 /* Invoke callbacks. */
2234 tick_dep_set_task(current
, TICK_DEP_BIT_RCU
);
2235 rhp
= rcu_cblist_dequeue(&rcl
);
2237 for (; rhp
; rhp
= rcu_cblist_dequeue(&rcl
)) {
2241 debug_rcu_head_unqueue(rhp
);
2243 rcu_lock_acquire(&rcu_callback_map
);
2244 trace_rcu_invoke_callback(rcu_state
.name
, rhp
);
2247 WRITE_ONCE(rhp
->func
, (rcu_callback_t
)0L);
2250 rcu_lock_release(&rcu_callback_map
);
2253 * Stop only if limit reached and CPU has something to do.
2255 if (in_serving_softirq()) {
2256 if (count
>= bl
&& (need_resched() || !is_idle_task(current
)))
2259 * Make sure we don't spend too much time here and deprive other
2260 * softirq vectors of CPU cycles.
2262 if (unlikely(tlimit
)) {
2263 /* only call local_clock() every 32 callbacks */
2264 if (likely((count
& 31) || local_clock() < tlimit
))
2266 /* Exceeded the time limit, so leave. */
2271 lockdep_assert_irqs_enabled();
2272 cond_resched_tasks_rcu_qs();
2273 lockdep_assert_irqs_enabled();
2278 rcu_nocb_lock_irqsave(rdp
, flags
);
2279 rdp
->n_cbs_invoked
+= count
;
2280 trace_rcu_batch_end(rcu_state
.name
, count
, !!rcl
.head
, need_resched(),
2281 is_idle_task(current
), rcu_is_callbacks_kthread(rdp
));
2283 /* Update counts and requeue any remaining callbacks. */
2284 rcu_segcblist_insert_done_cbs(&rdp
->cblist
, &rcl
);
2285 rcu_segcblist_add_len(&rdp
->cblist
, -count
);
2287 /* Reinstate batch limit if we have worked down the excess. */
2288 count
= rcu_segcblist_n_cbs(&rdp
->cblist
);
2289 if (rdp
->blimit
>= DEFAULT_MAX_RCU_BLIMIT
&& count
<= qlowmark
)
2290 rdp
->blimit
= blimit
;
2292 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2293 if (count
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
2294 rdp
->qlen_last_fqs_check
= 0;
2295 rdp
->n_force_qs_snap
= READ_ONCE(rcu_state
.n_force_qs
);
2296 } else if (count
< rdp
->qlen_last_fqs_check
- qhimark
)
2297 rdp
->qlen_last_fqs_check
= count
;
2300 * The following usually indicates a double call_rcu(). To track
2301 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2303 empty
= rcu_segcblist_empty(&rdp
->cblist
);
2304 WARN_ON_ONCE(count
== 0 && !empty
);
2305 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU
) &&
2306 count
!= 0 && empty
);
2307 WARN_ON_ONCE(count
== 0 && rcu_segcblist_n_segment_cbs(&rdp
->cblist
) != 0);
2308 WARN_ON_ONCE(!empty
&& rcu_segcblist_n_segment_cbs(&rdp
->cblist
) == 0);
2310 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2312 tick_dep_clear_task(current
, TICK_DEP_BIT_RCU
);
2316 * This function is invoked from each scheduling-clock interrupt,
2317 * and checks to see if this CPU is in a non-context-switch quiescent
2318 * state, for example, user mode or idle loop. It also schedules RCU
2319 * core processing. If the current grace period has gone on too long,
2320 * it will ask the scheduler to manufacture a context switch for the sole
2321 * purpose of providing the needed quiescent state.
2323 void rcu_sched_clock_irq(int user
)
2327 if (IS_ENABLED(CONFIG_PROVE_RCU
)) {
2329 WARN_ON_ONCE(time_before(j
, __this_cpu_read(rcu_data
.last_sched_clock
)));
2330 __this_cpu_write(rcu_data
.last_sched_clock
, j
);
2332 trace_rcu_utilization(TPS("Start scheduler-tick"));
2333 lockdep_assert_irqs_disabled();
2334 raw_cpu_inc(rcu_data
.ticks_this_gp
);
2335 /* The load-acquire pairs with the store-release setting to true. */
2336 if (smp_load_acquire(this_cpu_ptr(&rcu_data
.rcu_urgent_qs
))) {
2337 /* Idle and userspace execution already are quiescent states. */
2338 if (!rcu_is_cpu_rrupt_from_idle() && !user
) {
2339 set_tsk_need_resched(current
);
2340 set_preempt_need_resched();
2342 __this_cpu_write(rcu_data
.rcu_urgent_qs
, false);
2344 rcu_flavor_sched_clock_irq(user
);
2345 if (rcu_pending(user
))
2347 if (user
|| rcu_is_cpu_rrupt_from_idle())
2348 rcu_note_voluntary_context_switch(current
);
2349 lockdep_assert_irqs_disabled();
2351 trace_rcu_utilization(TPS("End scheduler-tick"));
2355 * Scan the leaf rcu_node structures. For each structure on which all
2356 * CPUs have reported a quiescent state and on which there are tasks
2357 * blocking the current grace period, initiate RCU priority boosting.
2358 * Otherwise, invoke the specified function to check dyntick state for
2359 * each CPU that has not yet reported a quiescent state.
2361 static void force_qs_rnp(int (*f
)(struct rcu_data
*rdp
))
2364 unsigned long flags
;
2366 struct rcu_data
*rdp
;
2367 struct rcu_node
*rnp
;
2369 rcu_state
.cbovld
= rcu_state
.cbovldnext
;
2370 rcu_state
.cbovldnext
= false;
2371 rcu_for_each_leaf_node(rnp
) {
2372 cond_resched_tasks_rcu_qs();
2374 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
2375 rcu_state
.cbovldnext
|= !!rnp
->cbovldmask
;
2376 if (rnp
->qsmask
== 0) {
2377 if (rcu_preempt_blocked_readers_cgp(rnp
)) {
2379 * No point in scanning bits because they
2380 * are all zero. But we might need to
2381 * priority-boost blocked readers.
2383 rcu_initiate_boost(rnp
, flags
);
2384 /* rcu_initiate_boost() releases rnp->lock */
2387 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2390 for_each_leaf_node_cpu_mask(rnp
, cpu
, rnp
->qsmask
) {
2391 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
2393 mask
|= rdp
->grpmask
;
2394 rcu_disable_urgency_upon_qs(rdp
);
2398 /* Idle/offline CPUs, report (releases rnp->lock). */
2399 rcu_report_qs_rnp(mask
, rnp
, rnp
->gp_seq
, flags
);
2401 /* Nothing to do here, so just drop the lock. */
2402 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
2408 * Force quiescent states on reluctant CPUs, and also detect which
2409 * CPUs are in dyntick-idle mode.
2411 void rcu_force_quiescent_state(void)
2413 unsigned long flags
;
2415 struct rcu_node
*rnp
;
2416 struct rcu_node
*rnp_old
= NULL
;
2418 /* Funnel through hierarchy to reduce memory contention. */
2419 rnp
= __this_cpu_read(rcu_data
.mynode
);
2420 for (; rnp
!= NULL
; rnp
= rnp
->parent
) {
2421 ret
= (READ_ONCE(rcu_state
.gp_flags
) & RCU_GP_FLAG_FQS
) ||
2422 !raw_spin_trylock(&rnp
->fqslock
);
2423 if (rnp_old
!= NULL
)
2424 raw_spin_unlock(&rnp_old
->fqslock
);
2429 /* rnp_old == rcu_get_root(), rnp == NULL. */
2431 /* Reached the root of the rcu_node tree, acquire lock. */
2432 raw_spin_lock_irqsave_rcu_node(rnp_old
, flags
);
2433 raw_spin_unlock(&rnp_old
->fqslock
);
2434 if (READ_ONCE(rcu_state
.gp_flags
) & RCU_GP_FLAG_FQS
) {
2435 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2436 return; /* Someone beat us to it. */
2438 WRITE_ONCE(rcu_state
.gp_flags
,
2439 READ_ONCE(rcu_state
.gp_flags
) | RCU_GP_FLAG_FQS
);
2440 raw_spin_unlock_irqrestore_rcu_node(rnp_old
, flags
);
2441 rcu_gp_kthread_wake();
2443 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state
);
2445 // Workqueue handler for an RCU reader for kernels enforcing struct RCU
2447 static void strict_work_handler(struct work_struct
*work
)
2453 /* Perform RCU core processing work for the current CPU. */
2454 static __latent_entropy
void rcu_core(void)
2456 unsigned long flags
;
2457 struct rcu_data
*rdp
= raw_cpu_ptr(&rcu_data
);
2458 struct rcu_node
*rnp
= rdp
->mynode
;
2460 * On RT rcu_core() can be preempted when IRQs aren't disabled.
2461 * Therefore this function can race with concurrent NOCB (de-)offloading
2462 * on this CPU and the below condition must be considered volatile.
2463 * However if we race with:
2465 * _ Offloading: In the worst case we accelerate or process callbacks
2466 * concurrently with NOCB kthreads. We are guaranteed to
2467 * call rcu_nocb_lock() if that happens.
2469 * _ Deoffloading: In the worst case we miss callbacks acceleration or
2470 * processing. This is fine because the early stage
2471 * of deoffloading invokes rcu_core() after setting
2472 * SEGCBLIST_RCU_CORE. So we guarantee that we'll process
2473 * what could have been dismissed without the need to wait
2474 * for the next rcu_pending() check in the next jiffy.
2476 const bool do_batch
= !rcu_segcblist_completely_offloaded(&rdp
->cblist
);
2478 if (cpu_is_offline(smp_processor_id()))
2480 trace_rcu_utilization(TPS("Start RCU core"));
2481 WARN_ON_ONCE(!rdp
->beenonline
);
2483 /* Report any deferred quiescent states if preemption enabled. */
2484 if (IS_ENABLED(CONFIG_PREEMPT_COUNT
) && (!(preempt_count() & PREEMPT_MASK
))) {
2485 rcu_preempt_deferred_qs(current
);
2486 } else if (rcu_preempt_need_deferred_qs(current
)) {
2487 set_tsk_need_resched(current
);
2488 set_preempt_need_resched();
2491 /* Update RCU state based on any recent quiescent states. */
2492 rcu_check_quiescent_state(rdp
);
2494 /* No grace period and unregistered callbacks? */
2495 if (!rcu_gp_in_progress() &&
2496 rcu_segcblist_is_enabled(&rdp
->cblist
) && do_batch
) {
2497 rcu_nocb_lock_irqsave(rdp
, flags
);
2498 if (!rcu_segcblist_restempty(&rdp
->cblist
, RCU_NEXT_READY_TAIL
))
2499 rcu_accelerate_cbs_unlocked(rnp
, rdp
);
2500 rcu_nocb_unlock_irqrestore(rdp
, flags
);
2503 rcu_check_gp_start_stall(rnp
, rdp
, rcu_jiffies_till_stall_check());
2505 /* If there are callbacks ready, invoke them. */
2506 if (do_batch
&& rcu_segcblist_ready_cbs(&rdp
->cblist
) &&
2507 likely(READ_ONCE(rcu_scheduler_fully_active
))) {
2509 /* Re-invoke RCU core processing if there are callbacks remaining. */
2510 if (rcu_segcblist_ready_cbs(&rdp
->cblist
))
2514 /* Do any needed deferred wakeups of rcuo kthreads. */
2515 do_nocb_deferred_wakeup(rdp
);
2516 trace_rcu_utilization(TPS("End RCU core"));
2518 // If strict GPs, schedule an RCU reader in a clean environment.
2519 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD
))
2520 queue_work_on(rdp
->cpu
, rcu_gp_wq
, &rdp
->strict_work
);
2523 static void rcu_core_si(struct softirq_action
*h
)
2528 static void rcu_wake_cond(struct task_struct
*t
, int status
)
2531 * If the thread is yielding, only wake it when this
2532 * is invoked from idle
2534 if (t
&& (status
!= RCU_KTHREAD_YIELDING
|| is_idle_task(current
)))
2538 static void invoke_rcu_core_kthread(void)
2540 struct task_struct
*t
;
2541 unsigned long flags
;
2543 local_irq_save(flags
);
2544 __this_cpu_write(rcu_data
.rcu_cpu_has_work
, 1);
2545 t
= __this_cpu_read(rcu_data
.rcu_cpu_kthread_task
);
2546 if (t
!= NULL
&& t
!= current
)
2547 rcu_wake_cond(t
, __this_cpu_read(rcu_data
.rcu_cpu_kthread_status
));
2548 local_irq_restore(flags
);
2552 * Wake up this CPU's rcuc kthread to do RCU core processing.
2554 static void invoke_rcu_core(void)
2556 if (!cpu_online(smp_processor_id()))
2559 raise_softirq(RCU_SOFTIRQ
);
2561 invoke_rcu_core_kthread();
2564 static void rcu_cpu_kthread_park(unsigned int cpu
)
2566 per_cpu(rcu_data
.rcu_cpu_kthread_status
, cpu
) = RCU_KTHREAD_OFFCPU
;
2569 static int rcu_cpu_kthread_should_run(unsigned int cpu
)
2571 return __this_cpu_read(rcu_data
.rcu_cpu_has_work
);
2575 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2576 * the RCU softirq used in configurations of RCU that do not support RCU
2577 * priority boosting.
2579 static void rcu_cpu_kthread(unsigned int cpu
)
2581 unsigned int *statusp
= this_cpu_ptr(&rcu_data
.rcu_cpu_kthread_status
);
2582 char work
, *workp
= this_cpu_ptr(&rcu_data
.rcu_cpu_has_work
);
2583 unsigned long *j
= this_cpu_ptr(&rcu_data
.rcuc_activity
);
2586 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2587 for (spincnt
= 0; spincnt
< 10; spincnt
++) {
2588 WRITE_ONCE(*j
, jiffies
);
2590 *statusp
= RCU_KTHREAD_RUNNING
;
2591 local_irq_disable();
2599 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2600 *statusp
= RCU_KTHREAD_WAITING
;
2604 *statusp
= RCU_KTHREAD_YIELDING
;
2605 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2606 schedule_timeout_idle(2);
2607 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2608 *statusp
= RCU_KTHREAD_WAITING
;
2609 WRITE_ONCE(*j
, jiffies
);
2612 static struct smp_hotplug_thread rcu_cpu_thread_spec
= {
2613 .store
= &rcu_data
.rcu_cpu_kthread_task
,
2614 .thread_should_run
= rcu_cpu_kthread_should_run
,
2615 .thread_fn
= rcu_cpu_kthread
,
2616 .thread_comm
= "rcuc/%u",
2617 .setup
= rcu_cpu_kthread_setup
,
2618 .park
= rcu_cpu_kthread_park
,
2622 * Spawn per-CPU RCU core processing kthreads.
2624 static int __init
rcu_spawn_core_kthreads(void)
2628 for_each_possible_cpu(cpu
)
2629 per_cpu(rcu_data
.rcu_cpu_has_work
, cpu
) = 0;
2632 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec
),
2633 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__
);
2638 * Handle any core-RCU processing required by a call_rcu() invocation.
2640 static void __call_rcu_core(struct rcu_data
*rdp
, struct rcu_head
*head
,
2641 unsigned long flags
)
2644 * If called from an extended quiescent state, invoke the RCU
2645 * core in order to force a re-evaluation of RCU's idleness.
2647 if (!rcu_is_watching())
2650 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2651 if (irqs_disabled_flags(flags
) || cpu_is_offline(smp_processor_id()))
2655 * Force the grace period if too many callbacks or too long waiting.
2656 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2657 * if some other CPU has recently done so. Also, don't bother
2658 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2659 * is the only one waiting for a grace period to complete.
2661 if (unlikely(rcu_segcblist_n_cbs(&rdp
->cblist
) >
2662 rdp
->qlen_last_fqs_check
+ qhimark
)) {
2664 /* Are we ignoring a completed grace period? */
2665 note_gp_changes(rdp
);
2667 /* Start a new grace period if one not already started. */
2668 if (!rcu_gp_in_progress()) {
2669 rcu_accelerate_cbs_unlocked(rdp
->mynode
, rdp
);
2671 /* Give the grace period a kick. */
2672 rdp
->blimit
= DEFAULT_MAX_RCU_BLIMIT
;
2673 if (READ_ONCE(rcu_state
.n_force_qs
) == rdp
->n_force_qs_snap
&&
2674 rcu_segcblist_first_pend_cb(&rdp
->cblist
) != head
)
2675 rcu_force_quiescent_state();
2676 rdp
->n_force_qs_snap
= READ_ONCE(rcu_state
.n_force_qs
);
2677 rdp
->qlen_last_fqs_check
= rcu_segcblist_n_cbs(&rdp
->cblist
);
2683 * RCU callback function to leak a callback.
2685 static void rcu_leak_callback(struct rcu_head
*rhp
)
2690 * Check and if necessary update the leaf rcu_node structure's
2691 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2692 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2693 * structure's ->lock.
2695 static void check_cb_ovld_locked(struct rcu_data
*rdp
, struct rcu_node
*rnp
)
2697 raw_lockdep_assert_held_rcu_node(rnp
);
2698 if (qovld_calc
<= 0)
2699 return; // Early boot and wildcard value set.
2700 if (rcu_segcblist_n_cbs(&rdp
->cblist
) >= qovld_calc
)
2701 WRITE_ONCE(rnp
->cbovldmask
, rnp
->cbovldmask
| rdp
->grpmask
);
2703 WRITE_ONCE(rnp
->cbovldmask
, rnp
->cbovldmask
& ~rdp
->grpmask
);
2707 * Check and if necessary update the leaf rcu_node structure's
2708 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2709 * number of queued RCU callbacks. No locks need be held, but the
2710 * caller must have disabled interrupts.
2712 * Note that this function ignores the possibility that there are a lot
2713 * of callbacks all of which have already seen the end of their respective
2714 * grace periods. This omission is due to the need for no-CBs CPUs to
2715 * be holding ->nocb_lock to do this check, which is too heavy for a
2716 * common-case operation.
2718 static void check_cb_ovld(struct rcu_data
*rdp
)
2720 struct rcu_node
*const rnp
= rdp
->mynode
;
2722 if (qovld_calc
<= 0 ||
2723 ((rcu_segcblist_n_cbs(&rdp
->cblist
) >= qovld_calc
) ==
2724 !!(READ_ONCE(rnp
->cbovldmask
) & rdp
->grpmask
)))
2725 return; // Early boot wildcard value or already set correctly.
2726 raw_spin_lock_rcu_node(rnp
);
2727 check_cb_ovld_locked(rdp
, rnp
);
2728 raw_spin_unlock_rcu_node(rnp
);
2732 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2733 * @head: structure to be used for queueing the RCU updates.
2734 * @func: actual callback function to be invoked after the grace period
2736 * The callback function will be invoked some time after a full grace
2737 * period elapses, in other words after all pre-existing RCU read-side
2738 * critical sections have completed. However, the callback function
2739 * might well execute concurrently with RCU read-side critical sections
2740 * that started after call_rcu() was invoked.
2742 * RCU read-side critical sections are delimited by rcu_read_lock()
2743 * and rcu_read_unlock(), and may be nested. In addition, but only in
2744 * v5.0 and later, regions of code across which interrupts, preemption,
2745 * or softirqs have been disabled also serve as RCU read-side critical
2746 * sections. This includes hardware interrupt handlers, softirq handlers,
2749 * Note that all CPUs must agree that the grace period extended beyond
2750 * all pre-existing RCU read-side critical section. On systems with more
2751 * than one CPU, this means that when "func()" is invoked, each CPU is
2752 * guaranteed to have executed a full memory barrier since the end of its
2753 * last RCU read-side critical section whose beginning preceded the call
2754 * to call_rcu(). It also means that each CPU executing an RCU read-side
2755 * critical section that continues beyond the start of "func()" must have
2756 * executed a memory barrier after the call_rcu() but before the beginning
2757 * of that RCU read-side critical section. Note that these guarantees
2758 * include CPUs that are offline, idle, or executing in user mode, as
2759 * well as CPUs that are executing in the kernel.
2761 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2762 * resulting RCU callback function "func()", then both CPU A and CPU B are
2763 * guaranteed to execute a full memory barrier during the time interval
2764 * between the call to call_rcu() and the invocation of "func()" -- even
2765 * if CPU A and CPU B are the same CPU (but again only if the system has
2766 * more than one CPU).
2768 * Implementation of these memory-ordering guarantees is described here:
2769 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
2771 void call_rcu(struct rcu_head
*head
, rcu_callback_t func
)
2773 static atomic_t doublefrees
;
2774 unsigned long flags
;
2775 struct rcu_data
*rdp
;
2778 /* Misaligned rcu_head! */
2779 WARN_ON_ONCE((unsigned long)head
& (sizeof(void *) - 1));
2781 if (debug_rcu_head_queue(head
)) {
2783 * Probable double call_rcu(), so leak the callback.
2784 * Use rcu:rcu_callback trace event to find the previous
2785 * time callback was passed to call_rcu().
2787 if (atomic_inc_return(&doublefrees
) < 4) {
2788 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__
, head
, head
->func
);
2791 WRITE_ONCE(head
->func
, rcu_leak_callback
);
2796 kasan_record_aux_stack_noalloc(head
);
2797 local_irq_save(flags
);
2798 rdp
= this_cpu_ptr(&rcu_data
);
2800 /* Add the callback to our list. */
2801 if (unlikely(!rcu_segcblist_is_enabled(&rdp
->cblist
))) {
2802 // This can trigger due to call_rcu() from offline CPU:
2803 WARN_ON_ONCE(rcu_scheduler_active
!= RCU_SCHEDULER_INACTIVE
);
2804 WARN_ON_ONCE(!rcu_is_watching());
2805 // Very early boot, before rcu_init(). Initialize if needed
2806 // and then drop through to queue the callback.
2807 if (rcu_segcblist_empty(&rdp
->cblist
))
2808 rcu_segcblist_init(&rdp
->cblist
);
2812 if (rcu_nocb_try_bypass(rdp
, head
, &was_alldone
, flags
))
2813 return; // Enqueued onto ->nocb_bypass, so just leave.
2814 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2815 rcu_segcblist_enqueue(&rdp
->cblist
, head
);
2816 if (__is_kvfree_rcu_offset((unsigned long)func
))
2817 trace_rcu_kvfree_callback(rcu_state
.name
, head
,
2818 (unsigned long)func
,
2819 rcu_segcblist_n_cbs(&rdp
->cblist
));
2821 trace_rcu_callback(rcu_state
.name
, head
,
2822 rcu_segcblist_n_cbs(&rdp
->cblist
));
2824 trace_rcu_segcb_stats(&rdp
->cblist
, TPS("SegCBQueued"));
2826 /* Go handle any RCU core processing required. */
2827 if (unlikely(rcu_rdp_is_offloaded(rdp
))) {
2828 __call_rcu_nocb_wake(rdp
, was_alldone
, flags
); /* unlocks */
2830 __call_rcu_core(rdp
, head
, flags
);
2831 local_irq_restore(flags
);
2834 EXPORT_SYMBOL_GPL(call_rcu
);
2837 /* Maximum number of jiffies to wait before draining a batch. */
2838 #define KFREE_DRAIN_JIFFIES (5 * HZ)
2839 #define KFREE_N_BATCHES 2
2840 #define FREE_N_CHANNELS 2
2843 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2844 * @nr_records: Number of active pointers in the array
2845 * @next: Next bulk object in the block chain
2846 * @records: Array of the kvfree_rcu() pointers
2848 struct kvfree_rcu_bulk_data
{
2849 unsigned long nr_records
;
2850 struct kvfree_rcu_bulk_data
*next
;
2855 * This macro defines how many entries the "records" array
2856 * will contain. It is based on the fact that the size of
2857 * kvfree_rcu_bulk_data structure becomes exactly one page.
2859 #define KVFREE_BULK_MAX_ENTR \
2860 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2863 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2864 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2865 * @head_free: List of kfree_rcu() objects waiting for a grace period
2866 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
2867 * @krcp: Pointer to @kfree_rcu_cpu structure
2870 struct kfree_rcu_cpu_work
{
2871 struct rcu_work rcu_work
;
2872 struct rcu_head
*head_free
;
2873 struct kvfree_rcu_bulk_data
*bkvhead_free
[FREE_N_CHANNELS
];
2874 struct kfree_rcu_cpu
*krcp
;
2878 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2879 * @head: List of kfree_rcu() objects not yet waiting for a grace period
2880 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
2881 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2882 * @lock: Synchronize access to this structure
2883 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
2884 * @initialized: The @rcu_work fields have been initialized
2885 * @count: Number of objects for which GP not started
2887 * A simple cache list that contains objects for reuse purpose.
2888 * In order to save some per-cpu space the list is singular.
2889 * Even though it is lockless an access has to be protected by the
2891 * @page_cache_work: A work to refill the cache when it is empty
2892 * @backoff_page_cache_fill: Delay cache refills
2893 * @work_in_progress: Indicates that page_cache_work is running
2894 * @hrtimer: A hrtimer for scheduling a page_cache_work
2895 * @nr_bkv_objs: number of allocated objects at @bkvcache.
2897 * This is a per-CPU structure. The reason that it is not included in
2898 * the rcu_data structure is to permit this code to be extracted from
2899 * the RCU files. Such extraction could allow further optimization of
2900 * the interactions with the slab allocators.
2902 struct kfree_rcu_cpu
{
2903 struct rcu_head
*head
;
2904 struct kvfree_rcu_bulk_data
*bkvhead
[FREE_N_CHANNELS
];
2905 struct kfree_rcu_cpu_work krw_arr
[KFREE_N_BATCHES
];
2906 raw_spinlock_t lock
;
2907 struct delayed_work monitor_work
;
2911 struct delayed_work page_cache_work
;
2912 atomic_t backoff_page_cache_fill
;
2913 atomic_t work_in_progress
;
2914 struct hrtimer hrtimer
;
2916 struct llist_head bkvcache
;
2920 static DEFINE_PER_CPU(struct kfree_rcu_cpu
, krc
) = {
2921 .lock
= __RAW_SPIN_LOCK_UNLOCKED(krc
.lock
),
2924 static __always_inline
void
2925 debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data
*bhead
)
2927 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
2930 for (i
= 0; i
< bhead
->nr_records
; i
++)
2931 debug_rcu_head_unqueue((struct rcu_head
*)(bhead
->records
[i
]));
2935 static inline struct kfree_rcu_cpu
*
2936 krc_this_cpu_lock(unsigned long *flags
)
2938 struct kfree_rcu_cpu
*krcp
;
2940 local_irq_save(*flags
); // For safely calling this_cpu_ptr().
2941 krcp
= this_cpu_ptr(&krc
);
2942 raw_spin_lock(&krcp
->lock
);
2948 krc_this_cpu_unlock(struct kfree_rcu_cpu
*krcp
, unsigned long flags
)
2950 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
2953 static inline struct kvfree_rcu_bulk_data
*
2954 get_cached_bnode(struct kfree_rcu_cpu
*krcp
)
2956 if (!krcp
->nr_bkv_objs
)
2959 WRITE_ONCE(krcp
->nr_bkv_objs
, krcp
->nr_bkv_objs
- 1);
2960 return (struct kvfree_rcu_bulk_data
*)
2961 llist_del_first(&krcp
->bkvcache
);
2965 put_cached_bnode(struct kfree_rcu_cpu
*krcp
,
2966 struct kvfree_rcu_bulk_data
*bnode
)
2969 if (krcp
->nr_bkv_objs
>= rcu_min_cached_objs
)
2972 llist_add((struct llist_node
*) bnode
, &krcp
->bkvcache
);
2973 WRITE_ONCE(krcp
->nr_bkv_objs
, krcp
->nr_bkv_objs
+ 1);
2978 drain_page_cache(struct kfree_rcu_cpu
*krcp
)
2980 unsigned long flags
;
2981 struct llist_node
*page_list
, *pos
, *n
;
2984 raw_spin_lock_irqsave(&krcp
->lock
, flags
);
2985 page_list
= llist_del_all(&krcp
->bkvcache
);
2986 WRITE_ONCE(krcp
->nr_bkv_objs
, 0);
2987 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
2989 llist_for_each_safe(pos
, n
, page_list
) {
2990 free_page((unsigned long)pos
);
2998 * This function is invoked in workqueue context after a grace period.
2999 * It frees all the objects queued on ->bkvhead_free or ->head_free.
3001 static void kfree_rcu_work(struct work_struct
*work
)
3003 unsigned long flags
;
3004 struct kvfree_rcu_bulk_data
*bkvhead
[FREE_N_CHANNELS
], *bnext
;
3005 struct rcu_head
*head
, *next
;
3006 struct kfree_rcu_cpu
*krcp
;
3007 struct kfree_rcu_cpu_work
*krwp
;
3010 krwp
= container_of(to_rcu_work(work
),
3011 struct kfree_rcu_cpu_work
, rcu_work
);
3014 raw_spin_lock_irqsave(&krcp
->lock
, flags
);
3015 // Channels 1 and 2.
3016 for (i
= 0; i
< FREE_N_CHANNELS
; i
++) {
3017 bkvhead
[i
] = krwp
->bkvhead_free
[i
];
3018 krwp
->bkvhead_free
[i
] = NULL
;
3022 head
= krwp
->head_free
;
3023 krwp
->head_free
= NULL
;
3024 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
3026 // Handle the first two channels.
3027 for (i
= 0; i
< FREE_N_CHANNELS
; i
++) {
3028 for (; bkvhead
[i
]; bkvhead
[i
] = bnext
) {
3029 bnext
= bkvhead
[i
]->next
;
3030 debug_rcu_bhead_unqueue(bkvhead
[i
]);
3032 rcu_lock_acquire(&rcu_callback_map
);
3033 if (i
== 0) { // kmalloc() / kfree().
3034 trace_rcu_invoke_kfree_bulk_callback(
3035 rcu_state
.name
, bkvhead
[i
]->nr_records
,
3036 bkvhead
[i
]->records
);
3038 kfree_bulk(bkvhead
[i
]->nr_records
,
3039 bkvhead
[i
]->records
);
3040 } else { // vmalloc() / vfree().
3041 for (j
= 0; j
< bkvhead
[i
]->nr_records
; j
++) {
3042 trace_rcu_invoke_kvfree_callback(
3044 bkvhead
[i
]->records
[j
], 0);
3046 vfree(bkvhead
[i
]->records
[j
]);
3049 rcu_lock_release(&rcu_callback_map
);
3051 raw_spin_lock_irqsave(&krcp
->lock
, flags
);
3052 if (put_cached_bnode(krcp
, bkvhead
[i
]))
3054 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
3057 free_page((unsigned long) bkvhead
[i
]);
3059 cond_resched_tasks_rcu_qs();
3064 * This is used when the "bulk" path can not be used for the
3065 * double-argument of kvfree_rcu(). This happens when the
3066 * page-cache is empty, which means that objects are instead
3067 * queued on a linked list through their rcu_head structures.
3068 * This list is named "Channel 3".
3070 for (; head
; head
= next
) {
3071 unsigned long offset
= (unsigned long)head
->func
;
3072 void *ptr
= (void *)head
- offset
;
3075 debug_rcu_head_unqueue((struct rcu_head
*)ptr
);
3076 rcu_lock_acquire(&rcu_callback_map
);
3077 trace_rcu_invoke_kvfree_callback(rcu_state
.name
, head
, offset
);
3079 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset
)))
3082 rcu_lock_release(&rcu_callback_map
);
3083 cond_resched_tasks_rcu_qs();
3088 need_offload_krc(struct kfree_rcu_cpu
*krcp
)
3092 for (i
= 0; i
< FREE_N_CHANNELS
; i
++)
3093 if (krcp
->bkvhead
[i
])
3096 return !!krcp
->head
;
3100 schedule_delayed_monitor_work(struct kfree_rcu_cpu
*krcp
)
3102 long delay
, delay_left
;
3104 delay
= READ_ONCE(krcp
->count
) >= KVFREE_BULK_MAX_ENTR
? 1:KFREE_DRAIN_JIFFIES
;
3105 if (delayed_work_pending(&krcp
->monitor_work
)) {
3106 delay_left
= krcp
->monitor_work
.timer
.expires
- jiffies
;
3107 if (delay
< delay_left
)
3108 mod_delayed_work(system_wq
, &krcp
->monitor_work
, delay
);
3111 queue_delayed_work(system_wq
, &krcp
->monitor_work
, delay
);
3115 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3117 static void kfree_rcu_monitor(struct work_struct
*work
)
3119 struct kfree_rcu_cpu
*krcp
= container_of(work
,
3120 struct kfree_rcu_cpu
, monitor_work
.work
);
3121 unsigned long flags
;
3124 raw_spin_lock_irqsave(&krcp
->lock
, flags
);
3126 // Attempt to start a new batch.
3127 for (i
= 0; i
< KFREE_N_BATCHES
; i
++) {
3128 struct kfree_rcu_cpu_work
*krwp
= &(krcp
->krw_arr
[i
]);
3130 // Try to detach bkvhead or head and attach it over any
3131 // available corresponding free channel. It can be that
3132 // a previous RCU batch is in progress, it means that
3133 // immediately to queue another one is not possible so
3134 // in that case the monitor work is rearmed.
3135 if ((krcp
->bkvhead
[0] && !krwp
->bkvhead_free
[0]) ||
3136 (krcp
->bkvhead
[1] && !krwp
->bkvhead_free
[1]) ||
3137 (krcp
->head
&& !krwp
->head_free
)) {
3138 // Channel 1 corresponds to the SLAB-pointer bulk path.
3139 // Channel 2 corresponds to vmalloc-pointer bulk path.
3140 for (j
= 0; j
< FREE_N_CHANNELS
; j
++) {
3141 if (!krwp
->bkvhead_free
[j
]) {
3142 krwp
->bkvhead_free
[j
] = krcp
->bkvhead
[j
];
3143 krcp
->bkvhead
[j
] = NULL
;
3147 // Channel 3 corresponds to both SLAB and vmalloc
3148 // objects queued on the linked list.
3149 if (!krwp
->head_free
) {
3150 krwp
->head_free
= krcp
->head
;
3154 WRITE_ONCE(krcp
->count
, 0);
3156 // One work is per one batch, so there are three
3157 // "free channels", the batch can handle. It can
3158 // be that the work is in the pending state when
3159 // channels have been detached following by each
3161 queue_rcu_work(system_wq
, &krwp
->rcu_work
);
3165 // If there is nothing to detach, it means that our job is
3166 // successfully done here. In case of having at least one
3167 // of the channels that is still busy we should rearm the
3168 // work to repeat an attempt. Because previous batches are
3169 // still in progress.
3170 if (need_offload_krc(krcp
))
3171 schedule_delayed_monitor_work(krcp
);
3173 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
3176 static enum hrtimer_restart
3177 schedule_page_work_fn(struct hrtimer
*t
)
3179 struct kfree_rcu_cpu
*krcp
=
3180 container_of(t
, struct kfree_rcu_cpu
, hrtimer
);
3182 queue_delayed_work(system_highpri_wq
, &krcp
->page_cache_work
, 0);
3183 return HRTIMER_NORESTART
;
3186 static void fill_page_cache_func(struct work_struct
*work
)
3188 struct kvfree_rcu_bulk_data
*bnode
;
3189 struct kfree_rcu_cpu
*krcp
=
3190 container_of(work
, struct kfree_rcu_cpu
,
3191 page_cache_work
.work
);
3192 unsigned long flags
;
3197 nr_pages
= atomic_read(&krcp
->backoff_page_cache_fill
) ?
3198 1 : rcu_min_cached_objs
;
3200 for (i
= 0; i
< nr_pages
; i
++) {
3201 bnode
= (struct kvfree_rcu_bulk_data
*)
3202 __get_free_page(GFP_KERNEL
| __GFP_NORETRY
| __GFP_NOMEMALLOC
| __GFP_NOWARN
);
3207 raw_spin_lock_irqsave(&krcp
->lock
, flags
);
3208 pushed
= put_cached_bnode(krcp
, bnode
);
3209 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
3212 free_page((unsigned long) bnode
);
3217 atomic_set(&krcp
->work_in_progress
, 0);
3218 atomic_set(&krcp
->backoff_page_cache_fill
, 0);
3222 run_page_cache_worker(struct kfree_rcu_cpu
*krcp
)
3224 if (rcu_scheduler_active
== RCU_SCHEDULER_RUNNING
&&
3225 !atomic_xchg(&krcp
->work_in_progress
, 1)) {
3226 if (atomic_read(&krcp
->backoff_page_cache_fill
)) {
3227 queue_delayed_work(system_wq
,
3228 &krcp
->page_cache_work
,
3229 msecs_to_jiffies(rcu_delay_page_cache_fill_msec
));
3231 hrtimer_init(&krcp
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3232 krcp
->hrtimer
.function
= schedule_page_work_fn
;
3233 hrtimer_start(&krcp
->hrtimer
, 0, HRTIMER_MODE_REL
);
3238 // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3239 // state specified by flags. If can_alloc is true, the caller must
3240 // be schedulable and not be holding any locks or mutexes that might be
3241 // acquired by the memory allocator or anything that it might invoke.
3242 // Returns true if ptr was successfully recorded, else the caller must
3245 add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu
**krcp
,
3246 unsigned long *flags
, void *ptr
, bool can_alloc
)
3248 struct kvfree_rcu_bulk_data
*bnode
;
3251 *krcp
= krc_this_cpu_lock(flags
);
3252 if (unlikely(!(*krcp
)->initialized
))
3255 idx
= !!is_vmalloc_addr(ptr
);
3257 /* Check if a new block is required. */
3258 if (!(*krcp
)->bkvhead
[idx
] ||
3259 (*krcp
)->bkvhead
[idx
]->nr_records
== KVFREE_BULK_MAX_ENTR
) {
3260 bnode
= get_cached_bnode(*krcp
);
3261 if (!bnode
&& can_alloc
) {
3262 krc_this_cpu_unlock(*krcp
, *flags
);
3264 // __GFP_NORETRY - allows a light-weight direct reclaim
3265 // what is OK from minimizing of fallback hitting point of
3266 // view. Apart of that it forbids any OOM invoking what is
3267 // also beneficial since we are about to release memory soon.
3269 // __GFP_NOMEMALLOC - prevents from consuming of all the
3270 // memory reserves. Please note we have a fallback path.
3272 // __GFP_NOWARN - it is supposed that an allocation can
3273 // be failed under low memory or high memory pressure
3275 bnode
= (struct kvfree_rcu_bulk_data
*)
3276 __get_free_page(GFP_KERNEL
| __GFP_NORETRY
| __GFP_NOMEMALLOC
| __GFP_NOWARN
);
3277 *krcp
= krc_this_cpu_lock(flags
);
3283 /* Initialize the new block. */
3284 bnode
->nr_records
= 0;
3285 bnode
->next
= (*krcp
)->bkvhead
[idx
];
3287 /* Attach it to the head. */
3288 (*krcp
)->bkvhead
[idx
] = bnode
;
3291 /* Finally insert. */
3292 (*krcp
)->bkvhead
[idx
]->records
3293 [(*krcp
)->bkvhead
[idx
]->nr_records
++] = ptr
;
3299 * Queue a request for lazy invocation of the appropriate free routine
3300 * after a grace period. Please note that three paths are maintained,
3301 * two for the common case using arrays of pointers and a third one that
3302 * is used only when the main paths cannot be used, for example, due to
3305 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3306 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3307 * be free'd in workqueue context. This allows us to: batch requests together to
3308 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3310 void kvfree_call_rcu(struct rcu_head
*head
, rcu_callback_t func
)
3312 unsigned long flags
;
3313 struct kfree_rcu_cpu
*krcp
;
3318 ptr
= (void *) head
- (unsigned long) func
;
3321 * Please note there is a limitation for the head-less
3322 * variant, that is why there is a clear rule for such
3323 * objects: it can be used from might_sleep() context
3324 * only. For other places please embed an rcu_head to
3328 ptr
= (unsigned long *) func
;
3331 // Queue the object but don't yet schedule the batch.
3332 if (debug_rcu_head_queue(ptr
)) {
3333 // Probable double kfree_rcu(), just leak.
3334 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3337 // Mark as success and leave.
3341 kasan_record_aux_stack_noalloc(ptr
);
3342 success
= add_ptr_to_bulk_krc_lock(&krcp
, &flags
, ptr
, !head
);
3344 run_page_cache_worker(krcp
);
3347 // Inline if kvfree_rcu(one_arg) call.
3351 head
->next
= krcp
->head
;
3356 WRITE_ONCE(krcp
->count
, krcp
->count
+ 1);
3358 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3359 if (rcu_scheduler_active
== RCU_SCHEDULER_RUNNING
)
3360 schedule_delayed_monitor_work(krcp
);
3363 krc_this_cpu_unlock(krcp
, flags
);
3366 * Inline kvfree() after synchronize_rcu(). We can do
3367 * it from might_sleep() context only, so the current
3368 * CPU can pass the QS state.
3371 debug_rcu_head_unqueue((struct rcu_head
*) ptr
);
3376 EXPORT_SYMBOL_GPL(kvfree_call_rcu
);
3378 static unsigned long
3379 kfree_rcu_shrink_count(struct shrinker
*shrink
, struct shrink_control
*sc
)
3382 unsigned long count
= 0;
3384 /* Snapshot count of all CPUs */
3385 for_each_possible_cpu(cpu
) {
3386 struct kfree_rcu_cpu
*krcp
= per_cpu_ptr(&krc
, cpu
);
3388 count
+= READ_ONCE(krcp
->count
);
3389 count
+= READ_ONCE(krcp
->nr_bkv_objs
);
3390 atomic_set(&krcp
->backoff_page_cache_fill
, 1);
3393 return count
== 0 ? SHRINK_EMPTY
: count
;
3396 static unsigned long
3397 kfree_rcu_shrink_scan(struct shrinker
*shrink
, struct shrink_control
*sc
)
3401 for_each_possible_cpu(cpu
) {
3403 struct kfree_rcu_cpu
*krcp
= per_cpu_ptr(&krc
, cpu
);
3405 count
= krcp
->count
;
3406 count
+= drain_page_cache(krcp
);
3407 kfree_rcu_monitor(&krcp
->monitor_work
.work
);
3409 sc
->nr_to_scan
-= count
;
3412 if (sc
->nr_to_scan
<= 0)
3416 return freed
== 0 ? SHRINK_STOP
: freed
;
3419 static struct shrinker kfree_rcu_shrinker
= {
3420 .count_objects
= kfree_rcu_shrink_count
,
3421 .scan_objects
= kfree_rcu_shrink_scan
,
3423 .seeks
= DEFAULT_SEEKS
,
3426 void __init
kfree_rcu_scheduler_running(void)
3429 unsigned long flags
;
3431 for_each_possible_cpu(cpu
) {
3432 struct kfree_rcu_cpu
*krcp
= per_cpu_ptr(&krc
, cpu
);
3434 raw_spin_lock_irqsave(&krcp
->lock
, flags
);
3435 if (need_offload_krc(krcp
))
3436 schedule_delayed_monitor_work(krcp
);
3437 raw_spin_unlock_irqrestore(&krcp
->lock
, flags
);
3442 * During early boot, any blocking grace-period wait automatically
3443 * implies a grace period.
3445 * Later on, this could in theory be the case for kernels built with
3446 * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
3447 * is not a common case. Furthermore, this optimization would cause
3448 * the rcu_gp_oldstate structure to expand by 50%, so this potential
3449 * grace-period optimization is ignored once the scheduler is running.
3451 static int rcu_blocking_is_gp(void)
3453 if (rcu_scheduler_active
!= RCU_SCHEDULER_INACTIVE
)
3455 might_sleep(); /* Check for RCU read-side critical section. */
3460 * synchronize_rcu - wait until a grace period has elapsed.
3462 * Control will return to the caller some time after a full grace
3463 * period has elapsed, in other words after all currently executing RCU
3464 * read-side critical sections have completed. Note, however, that
3465 * upon return from synchronize_rcu(), the caller might well be executing
3466 * concurrently with new RCU read-side critical sections that began while
3467 * synchronize_rcu() was waiting.
3469 * RCU read-side critical sections are delimited by rcu_read_lock()
3470 * and rcu_read_unlock(), and may be nested. In addition, but only in
3471 * v5.0 and later, regions of code across which interrupts, preemption,
3472 * or softirqs have been disabled also serve as RCU read-side critical
3473 * sections. This includes hardware interrupt handlers, softirq handlers,
3476 * Note that this guarantee implies further memory-ordering guarantees.
3477 * On systems with more than one CPU, when synchronize_rcu() returns,
3478 * each CPU is guaranteed to have executed a full memory barrier since
3479 * the end of its last RCU read-side critical section whose beginning
3480 * preceded the call to synchronize_rcu(). In addition, each CPU having
3481 * an RCU read-side critical section that extends beyond the return from
3482 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3483 * after the beginning of synchronize_rcu() and before the beginning of
3484 * that RCU read-side critical section. Note that these guarantees include
3485 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3486 * that are executing in the kernel.
3488 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3489 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3490 * to have executed a full memory barrier during the execution of
3491 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3492 * again only if the system has more than one CPU).
3494 * Implementation of these memory-ordering guarantees is described here:
3495 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3497 void synchronize_rcu(void)
3499 unsigned long flags
;
3500 struct rcu_node
*rnp
;
3502 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map
) ||
3503 lock_is_held(&rcu_lock_map
) ||
3504 lock_is_held(&rcu_sched_lock_map
),
3505 "Illegal synchronize_rcu() in RCU read-side critical section");
3506 if (!rcu_blocking_is_gp()) {
3507 if (rcu_gp_is_expedited())
3508 synchronize_rcu_expedited();
3510 wait_rcu_gp(call_rcu
);
3514 // Context allows vacuous grace periods.
3515 // Note well that this code runs with !PREEMPT && !SMP.
3516 // In addition, all code that advances grace periods runs at
3517 // process level. Therefore, this normal GP overlaps with other
3518 // normal GPs only by being fully nested within them, which allows
3519 // reuse of ->gp_seq_polled_snap.
3520 rcu_poll_gp_seq_start_unlocked(&rcu_state
.gp_seq_polled_snap
);
3521 rcu_poll_gp_seq_end_unlocked(&rcu_state
.gp_seq_polled_snap
);
3523 // Update the normal grace-period counters to record
3524 // this grace period, but only those used by the boot CPU.
3525 // The rcu_scheduler_starting() will take care of the rest of
3527 local_irq_save(flags
);
3528 WARN_ON_ONCE(num_online_cpus() > 1);
3529 rcu_state
.gp_seq
+= (1 << RCU_SEQ_CTR_SHIFT
);
3530 for (rnp
= this_cpu_ptr(&rcu_data
)->mynode
; rnp
; rnp
= rnp
->parent
)
3531 rnp
->gp_seq_needed
= rnp
->gp_seq
= rcu_state
.gp_seq
;
3532 local_irq_restore(flags
);
3534 EXPORT_SYMBOL_GPL(synchronize_rcu
);
3537 * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
3538 * @rgosp: Place to put state cookie
3540 * Stores into @rgosp a value that will always be treated by functions
3541 * like poll_state_synchronize_rcu_full() as a cookie whose grace period
3542 * has already completed.
3544 void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate
*rgosp
)
3546 rgosp
->rgos_norm
= RCU_GET_STATE_COMPLETED
;
3547 rgosp
->rgos_exp
= RCU_GET_STATE_COMPLETED
;
3549 EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full
);
3552 * get_state_synchronize_rcu - Snapshot current RCU state
3554 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3555 * or poll_state_synchronize_rcu() to determine whether or not a full
3556 * grace period has elapsed in the meantime.
3558 unsigned long get_state_synchronize_rcu(void)
3561 * Any prior manipulation of RCU-protected data must happen
3562 * before the load from ->gp_seq.
3565 return rcu_seq_snap(&rcu_state
.gp_seq_polled
);
3567 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu
);
3570 * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
3571 * @rgosp: location to place combined normal/expedited grace-period state
3573 * Places the normal and expedited grace-period states in @rgosp. This
3574 * state value can be passed to a later call to cond_synchronize_rcu_full()
3575 * or poll_state_synchronize_rcu_full() to determine whether or not a
3576 * grace period (whether normal or expedited) has elapsed in the meantime.
3577 * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
3578 * long, but is guaranteed to see all grace periods. In contrast, the
3579 * combined state occupies less memory, but can sometimes fail to take
3580 * grace periods into account.
3582 * This does not guarantee that the needed grace period will actually
3585 void get_state_synchronize_rcu_full(struct rcu_gp_oldstate
*rgosp
)
3587 struct rcu_node
*rnp
= rcu_get_root();
3590 * Any prior manipulation of RCU-protected data must happen
3591 * before the loads from ->gp_seq and ->expedited_sequence.
3594 rgosp
->rgos_norm
= rcu_seq_snap(&rnp
->gp_seq
);
3595 rgosp
->rgos_exp
= rcu_seq_snap(&rcu_state
.expedited_sequence
);
3597 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full
);
3600 * Helper function for start_poll_synchronize_rcu() and
3601 * start_poll_synchronize_rcu_full().
3603 static void start_poll_synchronize_rcu_common(void)
3605 unsigned long flags
;
3607 struct rcu_data
*rdp
;
3608 struct rcu_node
*rnp
;
3610 lockdep_assert_irqs_enabled();
3611 local_irq_save(flags
);
3612 rdp
= this_cpu_ptr(&rcu_data
);
3614 raw_spin_lock_rcu_node(rnp
); // irqs already disabled.
3615 // Note it is possible for a grace period to have elapsed between
3616 // the above call to get_state_synchronize_rcu() and the below call
3617 // to rcu_seq_snap. This is OK, the worst that happens is that we
3618 // get a grace period that no one needed. These accesses are ordered
3619 // by smp_mb(), and we are accessing them in the opposite order
3620 // from which they are updated at grace-period start, as required.
3621 needwake
= rcu_start_this_gp(rnp
, rdp
, rcu_seq_snap(&rcu_state
.gp_seq
));
3622 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
3624 rcu_gp_kthread_wake();
3628 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3630 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3631 * or poll_state_synchronize_rcu() to determine whether or not a full
3632 * grace period has elapsed in the meantime. If the needed grace period
3633 * is not already slated to start, notifies RCU core of the need for that
3636 * Interrupts must be enabled for the case where it is necessary to awaken
3637 * the grace-period kthread.
3639 unsigned long start_poll_synchronize_rcu(void)
3641 unsigned long gp_seq
= get_state_synchronize_rcu();
3643 start_poll_synchronize_rcu_common();
3646 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu
);
3649 * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
3650 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3652 * Places the normal and expedited grace-period states in *@rgos. This
3653 * state value can be passed to a later call to cond_synchronize_rcu_full()
3654 * or poll_state_synchronize_rcu_full() to determine whether or not a
3655 * grace period (whether normal or expedited) has elapsed in the meantime.
3656 * If the needed grace period is not already slated to start, notifies
3657 * RCU core of the need for that grace period.
3659 * Interrupts must be enabled for the case where it is necessary to awaken
3660 * the grace-period kthread.
3662 void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate
*rgosp
)
3664 get_state_synchronize_rcu_full(rgosp
);
3666 start_poll_synchronize_rcu_common();
3668 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full
);
3671 * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
3672 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3674 * If a full RCU grace period has elapsed since the earlier call from
3675 * which @oldstate was obtained, return @true, otherwise return @false.
3676 * If @false is returned, it is the caller's responsibility to invoke this
3677 * function later on until it does return @true. Alternatively, the caller
3678 * can explicitly wait for a grace period, for example, by passing @oldstate
3679 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3681 * Yes, this function does not take counter wrap into account.
3682 * But counter wrap is harmless. If the counter wraps, we have waited for
3683 * more than a billion grace periods (and way more on a 64-bit system!).
3684 * Those needing to keep old state values for very long time periods
3685 * (many hours even on 32-bit systems) should check them occasionally and
3686 * either refresh them or set a flag indicating that the grace period has
3687 * completed. Alternatively, they can use get_completed_synchronize_rcu()
3688 * to get a guaranteed-completed grace-period state.
3690 * This function provides the same memory-ordering guarantees that
3691 * would be provided by a synchronize_rcu() that was invoked at the call
3692 * to the function that provided @oldstate, and that returned at the end
3695 bool poll_state_synchronize_rcu(unsigned long oldstate
)
3697 if (oldstate
== RCU_GET_STATE_COMPLETED
||
3698 rcu_seq_done_exact(&rcu_state
.gp_seq_polled
, oldstate
)) {
3699 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3704 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu
);
3707 * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
3708 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3710 * If a full RCU grace period has elapsed since the earlier call from
3711 * which *rgosp was obtained, return @true, otherwise return @false.
3712 * If @false is returned, it is the caller's responsibility to invoke this
3713 * function later on until it does return @true. Alternatively, the caller
3714 * can explicitly wait for a grace period, for example, by passing @rgosp
3715 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3717 * Yes, this function does not take counter wrap into account.
3718 * But counter wrap is harmless. If the counter wraps, we have waited
3719 * for more than a billion grace periods (and way more on a 64-bit
3720 * system!). Those needing to keep rcu_gp_oldstate values for very
3721 * long time periods (many hours even on 32-bit systems) should check
3722 * them occasionally and either refresh them or set a flag indicating
3723 * that the grace period has completed. Alternatively, they can use
3724 * get_completed_synchronize_rcu_full() to get a guaranteed-completed
3725 * grace-period state.
3727 * This function provides the same memory-ordering guarantees that would
3728 * be provided by a synchronize_rcu() that was invoked at the call to
3729 * the function that provided @rgosp, and that returned at the end of this
3730 * function. And this guarantee requires that the root rcu_node structure's
3731 * ->gp_seq field be checked instead of that of the rcu_state structure.
3732 * The problem is that the just-ending grace-period's callbacks can be
3733 * invoked between the time that the root rcu_node structure's ->gp_seq
3734 * field is updated and the time that the rcu_state structure's ->gp_seq
3735 * field is updated. Therefore, if a single synchronize_rcu() is to
3736 * cause a subsequent poll_state_synchronize_rcu_full() to return @true,
3737 * then the root rcu_node structure is the one that needs to be polled.
3739 bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate
*rgosp
)
3741 struct rcu_node
*rnp
= rcu_get_root();
3743 smp_mb(); // Order against root rcu_node structure grace-period cleanup.
3744 if (rgosp
->rgos_norm
== RCU_GET_STATE_COMPLETED
||
3745 rcu_seq_done_exact(&rnp
->gp_seq
, rgosp
->rgos_norm
) ||
3746 rgosp
->rgos_exp
== RCU_GET_STATE_COMPLETED
||
3747 rcu_seq_done_exact(&rcu_state
.expedited_sequence
, rgosp
->rgos_exp
)) {
3748 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3753 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full
);
3756 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3757 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
3759 * If a full RCU grace period has elapsed since the earlier call to
3760 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3761 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3763 * Yes, this function does not take counter wrap into account.
3764 * But counter wrap is harmless. If the counter wraps, we have waited for
3765 * more than 2 billion grace periods (and way more on a 64-bit system!),
3766 * so waiting for a couple of additional grace periods should be just fine.
3768 * This function provides the same memory-ordering guarantees that
3769 * would be provided by a synchronize_rcu() that was invoked at the call
3770 * to the function that provided @oldstate and that returned at the end
3773 void cond_synchronize_rcu(unsigned long oldstate
)
3775 if (!poll_state_synchronize_rcu(oldstate
))
3778 EXPORT_SYMBOL_GPL(cond_synchronize_rcu
);
3781 * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
3782 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
3784 * If a full RCU grace period has elapsed since the call to
3785 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
3786 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
3787 * obtained, just return. Otherwise, invoke synchronize_rcu() to wait
3788 * for a full grace period.
3790 * Yes, this function does not take counter wrap into account.
3791 * But counter wrap is harmless. If the counter wraps, we have waited for
3792 * more than 2 billion grace periods (and way more on a 64-bit system!),
3793 * so waiting for a couple of additional grace periods should be just fine.
3795 * This function provides the same memory-ordering guarantees that
3796 * would be provided by a synchronize_rcu() that was invoked at the call
3797 * to the function that provided @rgosp and that returned at the end of
3800 void cond_synchronize_rcu_full(struct rcu_gp_oldstate
*rgosp
)
3802 if (!poll_state_synchronize_rcu_full(rgosp
))
3805 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full
);
3808 * Check to see if there is any immediate RCU-related work to be done by
3809 * the current CPU, returning 1 if so and zero otherwise. The checks are
3810 * in order of increasing expense: checks that can be carried out against
3811 * CPU-local state are performed first. However, we must check for CPU
3812 * stalls first, else we might not get a chance.
3814 static int rcu_pending(int user
)
3816 bool gp_in_progress
;
3817 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
3818 struct rcu_node
*rnp
= rdp
->mynode
;
3820 lockdep_assert_irqs_disabled();
3822 /* Check for CPU stalls, if enabled. */
3823 check_cpu_stall(rdp
);
3825 /* Does this CPU need a deferred NOCB wakeup? */
3826 if (rcu_nocb_need_deferred_wakeup(rdp
, RCU_NOCB_WAKE
))
3829 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3830 if ((user
|| rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3833 /* Is the RCU core waiting for a quiescent state from this CPU? */
3834 gp_in_progress
= rcu_gp_in_progress();
3835 if (rdp
->core_needs_qs
&& !rdp
->cpu_no_qs
.b
.norm
&& gp_in_progress
)
3838 /* Does this CPU have callbacks ready to invoke? */
3839 if (!rcu_rdp_is_offloaded(rdp
) &&
3840 rcu_segcblist_ready_cbs(&rdp
->cblist
))
3843 /* Has RCU gone idle with this CPU needing another grace period? */
3844 if (!gp_in_progress
&& rcu_segcblist_is_enabled(&rdp
->cblist
) &&
3845 !rcu_rdp_is_offloaded(rdp
) &&
3846 !rcu_segcblist_restempty(&rdp
->cblist
, RCU_NEXT_READY_TAIL
))
3849 /* Have RCU grace period completed or started? */
3850 if (rcu_seq_current(&rnp
->gp_seq
) != rdp
->gp_seq
||
3851 unlikely(READ_ONCE(rdp
->gpwrap
))) /* outside lock */
3859 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3860 * the compiler is expected to optimize this away.
3862 static void rcu_barrier_trace(const char *s
, int cpu
, unsigned long done
)
3864 trace_rcu_barrier(rcu_state
.name
, s
, cpu
,
3865 atomic_read(&rcu_state
.barrier_cpu_count
), done
);
3869 * RCU callback function for rcu_barrier(). If we are last, wake
3870 * up the task executing rcu_barrier().
3872 * Note that the value of rcu_state.barrier_sequence must be captured
3873 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3874 * other CPUs might count the value down to zero before this CPU gets
3875 * around to invoking rcu_barrier_trace(), which might result in bogus
3876 * data from the next instance of rcu_barrier().
3878 static void rcu_barrier_callback(struct rcu_head
*rhp
)
3880 unsigned long __maybe_unused s
= rcu_state
.barrier_sequence
;
3882 if (atomic_dec_and_test(&rcu_state
.barrier_cpu_count
)) {
3883 rcu_barrier_trace(TPS("LastCB"), -1, s
);
3884 complete(&rcu_state
.barrier_completion
);
3886 rcu_barrier_trace(TPS("CB"), -1, s
);
3891 * If needed, entrain an rcu_barrier() callback on rdp->cblist.
3893 static void rcu_barrier_entrain(struct rcu_data
*rdp
)
3895 unsigned long gseq
= READ_ONCE(rcu_state
.barrier_sequence
);
3896 unsigned long lseq
= READ_ONCE(rdp
->barrier_seq_snap
);
3898 lockdep_assert_held(&rcu_state
.barrier_lock
);
3899 if (rcu_seq_state(lseq
) || !rcu_seq_state(gseq
) || rcu_seq_ctr(lseq
) != rcu_seq_ctr(gseq
))
3901 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state
.barrier_sequence
);
3902 rdp
->barrier_head
.func
= rcu_barrier_callback
;
3903 debug_rcu_head_queue(&rdp
->barrier_head
);
3905 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp
, NULL
, jiffies
));
3906 if (rcu_segcblist_entrain(&rdp
->cblist
, &rdp
->barrier_head
)) {
3907 atomic_inc(&rcu_state
.barrier_cpu_count
);
3909 debug_rcu_head_unqueue(&rdp
->barrier_head
);
3910 rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state
.barrier_sequence
);
3912 rcu_nocb_unlock(rdp
);
3913 smp_store_release(&rdp
->barrier_seq_snap
, gseq
);
3917 * Called with preemption disabled, and from cross-cpu IRQ context.
3919 static void rcu_barrier_handler(void *cpu_in
)
3921 uintptr_t cpu
= (uintptr_t)cpu_in
;
3922 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
3924 lockdep_assert_irqs_disabled();
3925 WARN_ON_ONCE(cpu
!= rdp
->cpu
);
3926 WARN_ON_ONCE(cpu
!= smp_processor_id());
3927 raw_spin_lock(&rcu_state
.barrier_lock
);
3928 rcu_barrier_entrain(rdp
);
3929 raw_spin_unlock(&rcu_state
.barrier_lock
);
3933 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3935 * Note that this primitive does not necessarily wait for an RCU grace period
3936 * to complete. For example, if there are no RCU callbacks queued anywhere
3937 * in the system, then rcu_barrier() is within its rights to return
3938 * immediately, without waiting for anything, much less an RCU grace period.
3940 void rcu_barrier(void)
3943 unsigned long flags
;
3945 struct rcu_data
*rdp
;
3946 unsigned long s
= rcu_seq_snap(&rcu_state
.barrier_sequence
);
3948 rcu_barrier_trace(TPS("Begin"), -1, s
);
3950 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3951 mutex_lock(&rcu_state
.barrier_mutex
);
3953 /* Did someone else do our work for us? */
3954 if (rcu_seq_done(&rcu_state
.barrier_sequence
, s
)) {
3955 rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state
.barrier_sequence
);
3956 smp_mb(); /* caller's subsequent code after above check. */
3957 mutex_unlock(&rcu_state
.barrier_mutex
);
3961 /* Mark the start of the barrier operation. */
3962 raw_spin_lock_irqsave(&rcu_state
.barrier_lock
, flags
);
3963 rcu_seq_start(&rcu_state
.barrier_sequence
);
3964 gseq
= rcu_state
.barrier_sequence
;
3965 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state
.barrier_sequence
);
3968 * Initialize the count to two rather than to zero in order
3969 * to avoid a too-soon return to zero in case of an immediate
3970 * invocation of the just-enqueued callback (or preemption of
3971 * this task). Exclude CPU-hotplug operations to ensure that no
3972 * offline non-offloaded CPU has callbacks queued.
3974 init_completion(&rcu_state
.barrier_completion
);
3975 atomic_set(&rcu_state
.barrier_cpu_count
, 2);
3976 raw_spin_unlock_irqrestore(&rcu_state
.barrier_lock
, flags
);
3979 * Force each CPU with callbacks to register a new callback.
3980 * When that callback is invoked, we will know that all of the
3981 * corresponding CPU's preceding callbacks have been invoked.
3983 for_each_possible_cpu(cpu
) {
3984 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
3986 if (smp_load_acquire(&rdp
->barrier_seq_snap
) == gseq
)
3988 raw_spin_lock_irqsave(&rcu_state
.barrier_lock
, flags
);
3989 if (!rcu_segcblist_n_cbs(&rdp
->cblist
)) {
3990 WRITE_ONCE(rdp
->barrier_seq_snap
, gseq
);
3991 raw_spin_unlock_irqrestore(&rcu_state
.barrier_lock
, flags
);
3992 rcu_barrier_trace(TPS("NQ"), cpu
, rcu_state
.barrier_sequence
);
3995 if (!rcu_rdp_cpu_online(rdp
)) {
3996 rcu_barrier_entrain(rdp
);
3997 WARN_ON_ONCE(READ_ONCE(rdp
->barrier_seq_snap
) != gseq
);
3998 raw_spin_unlock_irqrestore(&rcu_state
.barrier_lock
, flags
);
3999 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu
, rcu_state
.barrier_sequence
);
4002 raw_spin_unlock_irqrestore(&rcu_state
.barrier_lock
, flags
);
4003 if (smp_call_function_single(cpu
, rcu_barrier_handler
, (void *)cpu
, 1)) {
4004 schedule_timeout_uninterruptible(1);
4007 WARN_ON_ONCE(READ_ONCE(rdp
->barrier_seq_snap
) != gseq
);
4008 rcu_barrier_trace(TPS("OnlineQ"), cpu
, rcu_state
.barrier_sequence
);
4012 * Now that we have an rcu_barrier_callback() callback on each
4013 * CPU, and thus each counted, remove the initial count.
4015 if (atomic_sub_and_test(2, &rcu_state
.barrier_cpu_count
))
4016 complete(&rcu_state
.barrier_completion
);
4018 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4019 wait_for_completion(&rcu_state
.barrier_completion
);
4021 /* Mark the end of the barrier operation. */
4022 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state
.barrier_sequence
);
4023 rcu_seq_end(&rcu_state
.barrier_sequence
);
4024 gseq
= rcu_state
.barrier_sequence
;
4025 for_each_possible_cpu(cpu
) {
4026 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4028 WRITE_ONCE(rdp
->barrier_seq_snap
, gseq
);
4031 /* Other rcu_barrier() invocations can now safely proceed. */
4032 mutex_unlock(&rcu_state
.barrier_mutex
);
4034 EXPORT_SYMBOL_GPL(rcu_barrier
);
4037 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4038 * first CPU in a given leaf rcu_node structure coming online. The caller
4039 * must hold the corresponding leaf rcu_node ->lock with interrupts
4042 static void rcu_init_new_rnp(struct rcu_node
*rnp_leaf
)
4046 struct rcu_node
*rnp
= rnp_leaf
;
4048 raw_lockdep_assert_held_rcu_node(rnp_leaf
);
4049 WARN_ON_ONCE(rnp
->wait_blkd_tasks
);
4051 mask
= rnp
->grpmask
;
4055 raw_spin_lock_rcu_node(rnp
); /* Interrupts already disabled. */
4056 oldmask
= rnp
->qsmaskinit
;
4057 rnp
->qsmaskinit
|= mask
;
4058 raw_spin_unlock_rcu_node(rnp
); /* Interrupts remain disabled. */
4065 * Do boot-time initialization of a CPU's per-CPU RCU data.
4068 rcu_boot_init_percpu_data(int cpu
)
4070 struct context_tracking
*ct
= this_cpu_ptr(&context_tracking
);
4071 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4073 /* Set up local state, ensuring consistent view of global state. */
4074 rdp
->grpmask
= leaf_node_cpu_bit(rdp
->mynode
, cpu
);
4075 INIT_WORK(&rdp
->strict_work
, strict_work_handler
);
4076 WARN_ON_ONCE(ct
->dynticks_nesting
!= 1);
4077 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu
)));
4078 rdp
->barrier_seq_snap
= rcu_state
.barrier_sequence
;
4079 rdp
->rcu_ofl_gp_seq
= rcu_state
.gp_seq
;
4080 rdp
->rcu_ofl_gp_flags
= RCU_GP_CLEANED
;
4081 rdp
->rcu_onl_gp_seq
= rcu_state
.gp_seq
;
4082 rdp
->rcu_onl_gp_flags
= RCU_GP_CLEANED
;
4083 rdp
->last_sched_clock
= jiffies
;
4085 rcu_boot_init_nocb_percpu_data(rdp
);
4089 * Invoked early in the CPU-online process, when pretty much all services
4090 * are available. The incoming CPU is not present.
4092 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4093 * offline event can be happening at a given time. Note also that we can
4094 * accept some slop in the rsp->gp_seq access due to the fact that this
4095 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4096 * And any offloaded callbacks are being numbered elsewhere.
4098 int rcutree_prepare_cpu(unsigned int cpu
)
4100 unsigned long flags
;
4101 struct context_tracking
*ct
= per_cpu_ptr(&context_tracking
, cpu
);
4102 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4103 struct rcu_node
*rnp
= rcu_get_root();
4105 /* Set up local state, ensuring consistent view of global state. */
4106 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4107 rdp
->qlen_last_fqs_check
= 0;
4108 rdp
->n_force_qs_snap
= READ_ONCE(rcu_state
.n_force_qs
);
4109 rdp
->blimit
= blimit
;
4110 ct
->dynticks_nesting
= 1; /* CPU not up, no tearing. */
4111 raw_spin_unlock_rcu_node(rnp
); /* irqs remain disabled. */
4114 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4117 if (!rcu_segcblist_is_enabled(&rdp
->cblist
))
4118 rcu_segcblist_init(&rdp
->cblist
); /* Re-enable callbacks. */
4121 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4122 * propagation up the rcu_node tree will happen at the beginning
4123 * of the next grace period.
4126 raw_spin_lock_rcu_node(rnp
); /* irqs already disabled. */
4127 rdp
->beenonline
= true; /* We have now been online. */
4128 rdp
->gp_seq
= READ_ONCE(rnp
->gp_seq
);
4129 rdp
->gp_seq_needed
= rdp
->gp_seq
;
4130 rdp
->cpu_no_qs
.b
.norm
= true;
4131 rdp
->core_needs_qs
= false;
4132 rdp
->rcu_iw_pending
= false;
4133 rdp
->rcu_iw
= IRQ_WORK_INIT_HARD(rcu_iw_handler
);
4134 rdp
->rcu_iw_gp_seq
= rdp
->gp_seq
- 1;
4135 trace_rcu_grace_period(rcu_state
.name
, rdp
->gp_seq
, TPS("cpuonl"));
4136 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4137 rcu_spawn_one_boost_kthread(rnp
);
4138 rcu_spawn_cpu_nocb_kthread(cpu
);
4139 WRITE_ONCE(rcu_state
.n_online_cpus
, rcu_state
.n_online_cpus
+ 1);
4145 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4147 static void rcutree_affinity_setting(unsigned int cpu
, int outgoing
)
4149 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4151 rcu_boost_kthread_setaffinity(rdp
->mynode
, outgoing
);
4155 * Near the end of the CPU-online process. Pretty much all services
4156 * enabled, and the CPU is now very much alive.
4158 int rcutree_online_cpu(unsigned int cpu
)
4160 unsigned long flags
;
4161 struct rcu_data
*rdp
;
4162 struct rcu_node
*rnp
;
4164 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4166 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4167 rnp
->ffmask
|= rdp
->grpmask
;
4168 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4169 if (rcu_scheduler_active
== RCU_SCHEDULER_INACTIVE
)
4170 return 0; /* Too early in boot for scheduler work. */
4171 sync_sched_exp_online_cleanup(cpu
);
4172 rcutree_affinity_setting(cpu
, -1);
4174 // Stop-machine done, so allow nohz_full to disable tick.
4175 tick_dep_clear(TICK_DEP_BIT_RCU
);
4180 * Near the beginning of the process. The CPU is still very much alive
4181 * with pretty much all services enabled.
4183 int rcutree_offline_cpu(unsigned int cpu
)
4185 unsigned long flags
;
4186 struct rcu_data
*rdp
;
4187 struct rcu_node
*rnp
;
4189 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4191 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4192 rnp
->ffmask
&= ~rdp
->grpmask
;
4193 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4195 rcutree_affinity_setting(cpu
, cpu
);
4197 // nohz_full CPUs need the tick for stop-machine to work quickly
4198 tick_dep_set(TICK_DEP_BIT_RCU
);
4203 * Mark the specified CPU as being online so that subsequent grace periods
4204 * (both expedited and normal) will wait on it. Note that this means that
4205 * incoming CPUs are not allowed to use RCU read-side critical sections
4206 * until this function is called. Failing to observe this restriction
4207 * will result in lockdep splats.
4209 * Note that this function is special in that it is invoked directly
4210 * from the incoming CPU rather than from the cpuhp_step mechanism.
4211 * This is because this function must be invoked at a precise location.
4213 void rcu_cpu_starting(unsigned int cpu
)
4215 unsigned long flags
;
4217 struct rcu_data
*rdp
;
4218 struct rcu_node
*rnp
;
4221 rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4222 if (rdp
->cpu_started
)
4224 rdp
->cpu_started
= true;
4227 mask
= rdp
->grpmask
;
4228 local_irq_save(flags
);
4229 arch_spin_lock(&rcu_state
.ofl_lock
);
4230 rcu_dynticks_eqs_online();
4231 raw_spin_lock(&rcu_state
.barrier_lock
);
4232 raw_spin_lock_rcu_node(rnp
);
4233 WRITE_ONCE(rnp
->qsmaskinitnext
, rnp
->qsmaskinitnext
| mask
);
4234 raw_spin_unlock(&rcu_state
.barrier_lock
);
4235 newcpu
= !(rnp
->expmaskinitnext
& mask
);
4236 rnp
->expmaskinitnext
|= mask
;
4237 /* Allow lockless access for expedited grace periods. */
4238 smp_store_release(&rcu_state
.ncpus
, rcu_state
.ncpus
+ newcpu
); /* ^^^ */
4239 ASSERT_EXCLUSIVE_WRITER(rcu_state
.ncpus
);
4240 rcu_gpnum_ovf(rnp
, rdp
); /* Offline-induced counter wrap? */
4241 rdp
->rcu_onl_gp_seq
= READ_ONCE(rcu_state
.gp_seq
);
4242 rdp
->rcu_onl_gp_flags
= READ_ONCE(rcu_state
.gp_flags
);
4244 /* An incoming CPU should never be blocking a grace period. */
4245 if (WARN_ON_ONCE(rnp
->qsmask
& mask
)) { /* RCU waiting on incoming CPU? */
4246 /* rcu_report_qs_rnp() *really* wants some flags to restore */
4247 unsigned long flags2
;
4249 local_irq_save(flags2
);
4250 rcu_disable_urgency_upon_qs(rdp
);
4251 /* Report QS -after- changing ->qsmaskinitnext! */
4252 rcu_report_qs_rnp(mask
, rnp
, rnp
->gp_seq
, flags2
);
4254 raw_spin_unlock_rcu_node(rnp
);
4256 arch_spin_unlock(&rcu_state
.ofl_lock
);
4257 local_irq_restore(flags
);
4258 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4262 * The outgoing function has no further need of RCU, so remove it from
4263 * the rcu_node tree's ->qsmaskinitnext bit masks.
4265 * Note that this function is special in that it is invoked directly
4266 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4267 * This is because this function must be invoked at a precise location.
4269 void rcu_report_dead(unsigned int cpu
)
4271 unsigned long flags
, seq_flags
;
4273 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4274 struct rcu_node
*rnp
= rdp
->mynode
; /* Outgoing CPU's rdp & rnp. */
4276 // Do any dangling deferred wakeups.
4277 do_nocb_deferred_wakeup(rdp
);
4279 /* QS for any half-done expedited grace period. */
4280 rcu_report_exp_rdp(rdp
);
4281 rcu_preempt_deferred_qs(current
);
4283 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4284 mask
= rdp
->grpmask
;
4285 local_irq_save(seq_flags
);
4286 arch_spin_lock(&rcu_state
.ofl_lock
);
4287 raw_spin_lock_irqsave_rcu_node(rnp
, flags
); /* Enforce GP memory-order guarantee. */
4288 rdp
->rcu_ofl_gp_seq
= READ_ONCE(rcu_state
.gp_seq
);
4289 rdp
->rcu_ofl_gp_flags
= READ_ONCE(rcu_state
.gp_flags
);
4290 if (rnp
->qsmask
& mask
) { /* RCU waiting on outgoing CPU? */
4291 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4292 rcu_disable_urgency_upon_qs(rdp
);
4293 rcu_report_qs_rnp(mask
, rnp
, rnp
->gp_seq
, flags
);
4294 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4296 WRITE_ONCE(rnp
->qsmaskinitnext
, rnp
->qsmaskinitnext
& ~mask
);
4297 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4298 arch_spin_unlock(&rcu_state
.ofl_lock
);
4299 local_irq_restore(seq_flags
);
4301 rdp
->cpu_started
= false;
4304 #ifdef CONFIG_HOTPLUG_CPU
4306 * The outgoing CPU has just passed through the dying-idle state, and we
4307 * are being invoked from the CPU that was IPIed to continue the offline
4308 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4310 void rcutree_migrate_callbacks(int cpu
)
4312 unsigned long flags
;
4313 struct rcu_data
*my_rdp
;
4314 struct rcu_node
*my_rnp
;
4315 struct rcu_data
*rdp
= per_cpu_ptr(&rcu_data
, cpu
);
4318 if (rcu_rdp_is_offloaded(rdp
) ||
4319 rcu_segcblist_empty(&rdp
->cblist
))
4320 return; /* No callbacks to migrate. */
4322 raw_spin_lock_irqsave(&rcu_state
.barrier_lock
, flags
);
4323 WARN_ON_ONCE(rcu_rdp_cpu_online(rdp
));
4324 rcu_barrier_entrain(rdp
);
4325 my_rdp
= this_cpu_ptr(&rcu_data
);
4326 my_rnp
= my_rdp
->mynode
;
4327 rcu_nocb_lock(my_rdp
); /* irqs already disabled. */
4328 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp
, NULL
, jiffies
));
4329 raw_spin_lock_rcu_node(my_rnp
); /* irqs already disabled. */
4330 /* Leverage recent GPs and set GP for new callbacks. */
4331 needwake
= rcu_advance_cbs(my_rnp
, rdp
) ||
4332 rcu_advance_cbs(my_rnp
, my_rdp
);
4333 rcu_segcblist_merge(&my_rdp
->cblist
, &rdp
->cblist
);
4334 raw_spin_unlock(&rcu_state
.barrier_lock
); /* irqs remain disabled. */
4335 needwake
= needwake
|| rcu_advance_cbs(my_rnp
, my_rdp
);
4336 rcu_segcblist_disable(&rdp
->cblist
);
4337 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp
->cblist
) != !rcu_segcblist_n_cbs(&my_rdp
->cblist
));
4338 check_cb_ovld_locked(my_rdp
, my_rnp
);
4339 if (rcu_rdp_is_offloaded(my_rdp
)) {
4340 raw_spin_unlock_rcu_node(my_rnp
); /* irqs remain disabled. */
4341 __call_rcu_nocb_wake(my_rdp
, true, flags
);
4343 rcu_nocb_unlock(my_rdp
); /* irqs remain disabled. */
4344 raw_spin_unlock_irqrestore_rcu_node(my_rnp
, flags
);
4347 rcu_gp_kthread_wake();
4348 lockdep_assert_irqs_enabled();
4349 WARN_ONCE(rcu_segcblist_n_cbs(&rdp
->cblist
) != 0 ||
4350 !rcu_segcblist_empty(&rdp
->cblist
),
4351 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4352 cpu
, rcu_segcblist_n_cbs(&rdp
->cblist
),
4353 rcu_segcblist_first_cb(&rdp
->cblist
));
4358 * On non-huge systems, use expedited RCU grace periods to make suspend
4359 * and hibernation run faster.
4361 static int rcu_pm_notify(struct notifier_block
*self
,
4362 unsigned long action
, void *hcpu
)
4365 case PM_HIBERNATION_PREPARE
:
4366 case PM_SUSPEND_PREPARE
:
4369 case PM_POST_HIBERNATION
:
4370 case PM_POST_SUSPEND
:
4371 rcu_unexpedite_gp();
4379 #ifdef CONFIG_RCU_EXP_KTHREAD
4380 struct kthread_worker
*rcu_exp_gp_kworker
;
4381 struct kthread_worker
*rcu_exp_par_gp_kworker
;
4383 static void __init
rcu_start_exp_gp_kworkers(void)
4385 const char *par_gp_kworker_name
= "rcu_exp_par_gp_kthread_worker";
4386 const char *gp_kworker_name
= "rcu_exp_gp_kthread_worker";
4387 struct sched_param param
= { .sched_priority
= kthread_prio
};
4389 rcu_exp_gp_kworker
= kthread_create_worker(0, gp_kworker_name
);
4390 if (IS_ERR_OR_NULL(rcu_exp_gp_kworker
)) {
4391 pr_err("Failed to create %s!\n", gp_kworker_name
);
4395 rcu_exp_par_gp_kworker
= kthread_create_worker(0, par_gp_kworker_name
);
4396 if (IS_ERR_OR_NULL(rcu_exp_par_gp_kworker
)) {
4397 pr_err("Failed to create %s!\n", par_gp_kworker_name
);
4398 kthread_destroy_worker(rcu_exp_gp_kworker
);
4402 sched_setscheduler_nocheck(rcu_exp_gp_kworker
->task
, SCHED_FIFO
, ¶m
);
4403 sched_setscheduler_nocheck(rcu_exp_par_gp_kworker
->task
, SCHED_FIFO
,
4407 static inline void rcu_alloc_par_gp_wq(void)
4410 #else /* !CONFIG_RCU_EXP_KTHREAD */
4411 struct workqueue_struct
*rcu_par_gp_wq
;
4413 static void __init
rcu_start_exp_gp_kworkers(void)
4417 static inline void rcu_alloc_par_gp_wq(void)
4419 rcu_par_gp_wq
= alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM
, 0);
4420 WARN_ON(!rcu_par_gp_wq
);
4422 #endif /* CONFIG_RCU_EXP_KTHREAD */
4425 * Spawn the kthreads that handle RCU's grace periods.
4427 static int __init
rcu_spawn_gp_kthread(void)
4429 unsigned long flags
;
4430 struct rcu_node
*rnp
;
4431 struct sched_param sp
;
4432 struct task_struct
*t
;
4433 struct rcu_data
*rdp
= this_cpu_ptr(&rcu_data
);
4435 rcu_scheduler_fully_active
= 1;
4436 t
= kthread_create(rcu_gp_kthread
, NULL
, "%s", rcu_state
.name
);
4437 if (WARN_ONCE(IS_ERR(t
), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__
))
4440 sp
.sched_priority
= kthread_prio
;
4441 sched_setscheduler_nocheck(t
, SCHED_FIFO
, &sp
);
4443 rnp
= rcu_get_root();
4444 raw_spin_lock_irqsave_rcu_node(rnp
, flags
);
4445 WRITE_ONCE(rcu_state
.gp_activity
, jiffies
);
4446 WRITE_ONCE(rcu_state
.gp_req_activity
, jiffies
);
4447 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4448 smp_store_release(&rcu_state
.gp_kthread
, t
); /* ^^^ */
4449 raw_spin_unlock_irqrestore_rcu_node(rnp
, flags
);
4451 /* This is a pre-SMP initcall, we expect a single CPU */
4452 WARN_ON(num_online_cpus() > 1);
4454 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
4455 * due to rcu_scheduler_fully_active.
4457 rcu_spawn_cpu_nocb_kthread(smp_processor_id());
4458 rcu_spawn_one_boost_kthread(rdp
->mynode
);
4459 rcu_spawn_core_kthreads();
4460 /* Create kthread worker for expedited GPs */
4461 rcu_start_exp_gp_kworkers();
4464 early_initcall(rcu_spawn_gp_kthread
);
4467 * This function is invoked towards the end of the scheduler's
4468 * initialization process. Before this is called, the idle task might
4469 * contain synchronous grace-period primitives (during which time, this idle
4470 * task is booting the system, and such primitives are no-ops). After this
4471 * function is called, any synchronous grace-period primitives are run as
4472 * expedited, with the requesting task driving the grace period forward.
4473 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4474 * runtime RCU functionality.
4476 void rcu_scheduler_starting(void)
4478 unsigned long flags
;
4479 struct rcu_node
*rnp
;
4481 WARN_ON(num_online_cpus() != 1);
4482 WARN_ON(nr_context_switches() > 0);
4483 rcu_test_sync_prims();
4485 // Fix up the ->gp_seq counters.
4486 local_irq_save(flags
);
4487 rcu_for_each_node_breadth_first(rnp
)
4488 rnp
->gp_seq_needed
= rnp
->gp_seq
= rcu_state
.gp_seq
;
4489 local_irq_restore(flags
);
4491 // Switch out of early boot mode.
4492 rcu_scheduler_active
= RCU_SCHEDULER_INIT
;
4493 rcu_test_sync_prims();
4497 * Helper function for rcu_init() that initializes the rcu_state structure.
4499 static void __init
rcu_init_one(void)
4501 static const char * const buf
[] = RCU_NODE_NAME_INIT
;
4502 static const char * const fqs
[] = RCU_FQS_NAME_INIT
;
4503 static struct lock_class_key rcu_node_class
[RCU_NUM_LVLS
];
4504 static struct lock_class_key rcu_fqs_class
[RCU_NUM_LVLS
];
4506 int levelspread
[RCU_NUM_LVLS
]; /* kids/node in each level. */
4510 struct rcu_node
*rnp
;
4512 BUILD_BUG_ON(RCU_NUM_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
4514 /* Silence gcc 4.8 false positive about array index out of range. */
4515 if (rcu_num_lvls
<= 0 || rcu_num_lvls
> RCU_NUM_LVLS
)
4516 panic("rcu_init_one: rcu_num_lvls out of range");
4518 /* Initialize the level-tracking arrays. */
4520 for (i
= 1; i
< rcu_num_lvls
; i
++)
4521 rcu_state
.level
[i
] =
4522 rcu_state
.level
[i
- 1] + num_rcu_lvl
[i
- 1];
4523 rcu_init_levelspread(levelspread
, num_rcu_lvl
);
4525 /* Initialize the elements themselves, starting from the leaves. */
4527 for (i
= rcu_num_lvls
- 1; i
>= 0; i
--) {
4528 cpustride
*= levelspread
[i
];
4529 rnp
= rcu_state
.level
[i
];
4530 for (j
= 0; j
< num_rcu_lvl
[i
]; j
++, rnp
++) {
4531 raw_spin_lock_init(&ACCESS_PRIVATE(rnp
, lock
));
4532 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp
, lock
),
4533 &rcu_node_class
[i
], buf
[i
]);
4534 raw_spin_lock_init(&rnp
->fqslock
);
4535 lockdep_set_class_and_name(&rnp
->fqslock
,
4536 &rcu_fqs_class
[i
], fqs
[i
]);
4537 rnp
->gp_seq
= rcu_state
.gp_seq
;
4538 rnp
->gp_seq_needed
= rcu_state
.gp_seq
;
4539 rnp
->completedqs
= rcu_state
.gp_seq
;
4541 rnp
->qsmaskinit
= 0;
4542 rnp
->grplo
= j
* cpustride
;
4543 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
4544 if (rnp
->grphi
>= nr_cpu_ids
)
4545 rnp
->grphi
= nr_cpu_ids
- 1;
4551 rnp
->grpnum
= j
% levelspread
[i
- 1];
4552 rnp
->grpmask
= BIT(rnp
->grpnum
);
4553 rnp
->parent
= rcu_state
.level
[i
- 1] +
4554 j
/ levelspread
[i
- 1];
4557 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
4558 rcu_init_one_nocb(rnp
);
4559 init_waitqueue_head(&rnp
->exp_wq
[0]);
4560 init_waitqueue_head(&rnp
->exp_wq
[1]);
4561 init_waitqueue_head(&rnp
->exp_wq
[2]);
4562 init_waitqueue_head(&rnp
->exp_wq
[3]);
4563 spin_lock_init(&rnp
->exp_lock
);
4564 mutex_init(&rnp
->boost_kthread_mutex
);
4565 raw_spin_lock_init(&rnp
->exp_poll_lock
);
4566 rnp
->exp_seq_poll_rq
= RCU_GET_STATE_COMPLETED
;
4567 INIT_WORK(&rnp
->exp_poll_wq
, sync_rcu_do_polled_gp
);
4571 init_swait_queue_head(&rcu_state
.gp_wq
);
4572 init_swait_queue_head(&rcu_state
.expedited_wq
);
4573 rnp
= rcu_first_leaf_node();
4574 for_each_possible_cpu(i
) {
4575 while (i
> rnp
->grphi
)
4577 per_cpu_ptr(&rcu_data
, i
)->mynode
= rnp
;
4578 rcu_boot_init_percpu_data(i
);
4583 * Force priority from the kernel command-line into range.
4585 static void __init
sanitize_kthread_prio(void)
4587 int kthread_prio_in
= kthread_prio
;
4589 if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 2
4590 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST
))
4592 else if (IS_ENABLED(CONFIG_RCU_BOOST
) && kthread_prio
< 1)
4594 else if (kthread_prio
< 0)
4596 else if (kthread_prio
> 99)
4599 if (kthread_prio
!= kthread_prio_in
)
4600 pr_alert("%s: Limited prio to %d from %d\n",
4601 __func__
, kthread_prio
, kthread_prio_in
);
4605 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4606 * replace the definitions in tree.h because those are needed to size
4607 * the ->node array in the rcu_state structure.
4609 void rcu_init_geometry(void)
4613 static unsigned long old_nr_cpu_ids
;
4614 int rcu_capacity
[RCU_NUM_LVLS
];
4615 static bool initialized
;
4619 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4620 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4622 WARN_ON_ONCE(old_nr_cpu_ids
!= nr_cpu_ids
);
4626 old_nr_cpu_ids
= nr_cpu_ids
;
4630 * Initialize any unspecified boot parameters.
4631 * The default values of jiffies_till_first_fqs and
4632 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4633 * value, which is a function of HZ, then adding one for each
4634 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4636 d
= RCU_JIFFIES_TILL_FORCE_QS
+ nr_cpu_ids
/ RCU_JIFFIES_FQS_DIV
;
4637 if (jiffies_till_first_fqs
== ULONG_MAX
)
4638 jiffies_till_first_fqs
= d
;
4639 if (jiffies_till_next_fqs
== ULONG_MAX
)
4640 jiffies_till_next_fqs
= d
;
4641 adjust_jiffies_till_sched_qs();
4643 /* If the compile-time values are accurate, just leave. */
4644 if (rcu_fanout_leaf
== RCU_FANOUT_LEAF
&&
4645 nr_cpu_ids
== NR_CPUS
)
4647 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4648 rcu_fanout_leaf
, nr_cpu_ids
);
4651 * The boot-time rcu_fanout_leaf parameter must be at least two
4652 * and cannot exceed the number of bits in the rcu_node masks.
4653 * Complain and fall back to the compile-time values if this
4654 * limit is exceeded.
4656 if (rcu_fanout_leaf
< 2 ||
4657 rcu_fanout_leaf
> sizeof(unsigned long) * 8) {
4658 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4664 * Compute number of nodes that can be handled an rcu_node tree
4665 * with the given number of levels.
4667 rcu_capacity
[0] = rcu_fanout_leaf
;
4668 for (i
= 1; i
< RCU_NUM_LVLS
; i
++)
4669 rcu_capacity
[i
] = rcu_capacity
[i
- 1] * RCU_FANOUT
;
4672 * The tree must be able to accommodate the configured number of CPUs.
4673 * If this limit is exceeded, fall back to the compile-time values.
4675 if (nr_cpu_ids
> rcu_capacity
[RCU_NUM_LVLS
- 1]) {
4676 rcu_fanout_leaf
= RCU_FANOUT_LEAF
;
4681 /* Calculate the number of levels in the tree. */
4682 for (i
= 0; nr_cpu_ids
> rcu_capacity
[i
]; i
++) {
4684 rcu_num_lvls
= i
+ 1;
4686 /* Calculate the number of rcu_nodes at each level of the tree. */
4687 for (i
= 0; i
< rcu_num_lvls
; i
++) {
4688 int cap
= rcu_capacity
[(rcu_num_lvls
- 1) - i
];
4689 num_rcu_lvl
[i
] = DIV_ROUND_UP(nr_cpu_ids
, cap
);
4692 /* Calculate the total number of rcu_node structures. */
4694 for (i
= 0; i
< rcu_num_lvls
; i
++)
4695 rcu_num_nodes
+= num_rcu_lvl
[i
];
4699 * Dump out the structure of the rcu_node combining tree associated
4700 * with the rcu_state structure.
4702 static void __init
rcu_dump_rcu_node_tree(void)
4705 struct rcu_node
*rnp
;
4707 pr_info("rcu_node tree layout dump\n");
4709 rcu_for_each_node_breadth_first(rnp
) {
4710 if (rnp
->level
!= level
) {
4715 pr_cont("%d:%d ^%d ", rnp
->grplo
, rnp
->grphi
, rnp
->grpnum
);
4720 struct workqueue_struct
*rcu_gp_wq
;
4722 static void __init
kfree_rcu_batch_init(void)
4727 /* Clamp it to [0:100] seconds interval. */
4728 if (rcu_delay_page_cache_fill_msec
< 0 ||
4729 rcu_delay_page_cache_fill_msec
> 100 * MSEC_PER_SEC
) {
4731 rcu_delay_page_cache_fill_msec
=
4732 clamp(rcu_delay_page_cache_fill_msec
, 0,
4733 (int) (100 * MSEC_PER_SEC
));
4735 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4736 rcu_delay_page_cache_fill_msec
);
4739 for_each_possible_cpu(cpu
) {
4740 struct kfree_rcu_cpu
*krcp
= per_cpu_ptr(&krc
, cpu
);
4742 for (i
= 0; i
< KFREE_N_BATCHES
; i
++) {
4743 INIT_RCU_WORK(&krcp
->krw_arr
[i
].rcu_work
, kfree_rcu_work
);
4744 krcp
->krw_arr
[i
].krcp
= krcp
;
4747 INIT_DELAYED_WORK(&krcp
->monitor_work
, kfree_rcu_monitor
);
4748 INIT_DELAYED_WORK(&krcp
->page_cache_work
, fill_page_cache_func
);
4749 krcp
->initialized
= true;
4751 if (register_shrinker(&kfree_rcu_shrinker
, "rcu-kfree"))
4752 pr_err("Failed to register kfree_rcu() shrinker!\n");
4755 void __init
rcu_init(void)
4757 int cpu
= smp_processor_id();
4759 rcu_early_boot_tests();
4761 kfree_rcu_batch_init();
4762 rcu_bootup_announce();
4763 sanitize_kthread_prio();
4764 rcu_init_geometry();
4767 rcu_dump_rcu_node_tree();
4769 open_softirq(RCU_SOFTIRQ
, rcu_core_si
);
4772 * We don't need protection against CPU-hotplug here because
4773 * this is called early in boot, before either interrupts
4774 * or the scheduler are operational.
4776 pm_notifier(rcu_pm_notify
, 0);
4777 WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
4778 rcutree_prepare_cpu(cpu
);
4779 rcu_cpu_starting(cpu
);
4780 rcutree_online_cpu(cpu
);
4782 /* Create workqueue for Tree SRCU and for expedited GPs. */
4783 rcu_gp_wq
= alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM
, 0);
4784 WARN_ON(!rcu_gp_wq
);
4785 rcu_alloc_par_gp_wq();
4787 /* Fill in default value for rcutree.qovld boot parameter. */
4788 /* -After- the rcu_node ->lock fields are initialized! */
4790 qovld_calc
= DEFAULT_RCU_QOVLD_MULT
* qhimark
;
4794 // Kick-start any polled grace periods that started early.
4795 if (!(per_cpu_ptr(&rcu_data
, cpu
)->mynode
->exp_seq_poll_rq
& 0x1))
4796 (void)start_poll_synchronize_rcu_expedited();
4799 #include "tree_stall.h"
4800 #include "tree_exp.h"
4801 #include "tree_nocb.h"
4802 #include "tree_plugin.h"